From f46c0c1b0ea0bc67e24e4bf026a7e898c1af22a9 Mon Sep 17 00:00:00 2001 From: David Renshaw Date: Wed, 17 Jan 2024 02:17:50 -0500 Subject: [PATCH 001/101] llama : fix copy/paste error in llama_sampling_params comment (#4994) --- common/sampling.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/common/sampling.h b/common/sampling.h index f16ef97e3..2ee180376 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -17,7 +17,7 @@ typedef struct llama_sampling_params { float min_p = 0.05f; // 0.0 = disabled float tfs_z = 1.00f; // 1.0 = disabled float typical_p = 1.00f; // 1.0 = disabled - float temp = 0.80f; // 1.0 = disabled + float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) float penalty_repeat = 1.10f; // 1.0 = disabled float penalty_freq = 0.00f; // 0.0 = disabled From 75632936659772d5b2ce54b0b65319fecbaac2e6 Mon Sep 17 00:00:00 2001 From: Paul Tsochantaris Date: Wed, 17 Jan 2024 08:07:24 +0000 Subject: [PATCH 002/101] metal : remove unnecessary nil check (#4986) --- ggml-metal.m | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index a549e6713..8bb4edd64 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -2236,10 +2236,7 @@ static bool ggml_metal_graph_compute( #endif } - if (encoder != nil) { - [encoder endEncoding]; - encoder = nil; - } + [encoder endEncoding]; [command_buffer commit]; }); From 2b3a665d3917edf393761a24c4835447894df74a Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Wed, 17 Jan 2024 12:36:37 +0200 Subject: [PATCH 003/101] llama : use Q4_K for attn_v for Q2_K_S when n_gqa >= 4 (#4996) Co-authored-by: Iwan Kawrakow --- llama.cpp | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 765d20ddb..2c5983c67 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8477,7 +8477,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty } else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K; } else if (name.find("attn_v.weight") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) { + new_type = qs.model.hparams.n_gqa() >= 4 ? GGML_TYPE_Q4_K : GGML_TYPE_Q3_K; + } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && qs.model.hparams.n_gqa() >= 4) { + new_type = GGML_TYPE_Q4_K; + } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { new_type = qs.i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K; } From 4f4bf35f46600441dec2f941e667291eeb9a18d8 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 15:45:03 +0200 Subject: [PATCH 004/101] py : fix missing added_tokens_dict for SPM and BPE vocabs (#4971) * py : fix missing added_tokens_dict for SPM vocab * py : pad with unknown tokens when data is missing ggml-ci * py : fix BPE vocab conversion ggml-ci * py : fix padded dummy tokens (I hope) --- convert.py | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-) diff --git a/convert.py b/convert.py index 3b613eefc..b47bb6185 100755 --- a/convert.py +++ b/convert.py @@ -387,6 +387,7 @@ class BpeVocab: # GPT self.bpe_tokenizer = json.loads( open(str(fname_tokenizer), encoding="utf-8").read() ) + self.vocab = self.bpe_tokenizer["model"]["vocab"] added_tokens: dict[str, int] if fname_added_tokens is not None: # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab. @@ -405,7 +406,7 @@ class BpeVocab: # GPT if item["content"] not in self.bpe_tokenizer ) - vocab_size: int = len(self.bpe_tokenizer) + vocab_size: int = len(self.vocab) expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) actual_ids = sorted(added_tokens.values()) if expected_ids != actual_ids: @@ -415,6 +416,7 @@ class BpeVocab: # GPT ) items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) + self.added_tokens_dict = added_tokens self.added_tokens_list = [text for (text, idx) in items] self.vocab_size_base: int = vocab_size self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) @@ -422,10 +424,9 @@ class BpeVocab: # GPT self.fname_added_tokens = fname_added_tokens def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: - tokenizer = self.bpe_tokenizer - reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()} + reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} - for i, _ in enumerate(tokenizer): + for i, _ in enumerate(self.vocab): yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: @@ -466,6 +467,7 @@ class SentencePieceVocab: # LlaMa ) # Token pieces that were added to the base vocabulary. + self.added_tokens_dict = added_tokens self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] self.vocab_size_base = vocab_size self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) @@ -1006,6 +1008,7 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N ) for i in range(1, pad_count + 1): vocab.added_tokens_dict[f""] = -1 + vocab.added_tokens_list.append(f"") vocab.vocab_size = params.n_vocab return @@ -1097,6 +1100,8 @@ class OutputFile: scores.append(score) toktypes.append(toktype) + assert(len(tokens) == vocab.vocab_size) + return tokens, scores, toktypes def add_meta_vocab(self, vocab: Vocab) -> None: @@ -1373,15 +1378,14 @@ class VocabFactory: self.files[file] = file_path elif parent_file_path.exists(): self.files[file] = parent_file_path + print(f"Found vocab files: {self.files}") def _select_file(self, vocabtype: Optional[str]) -> Path: if vocabtype in ["spm", "bpe"]: - # For SentencePiece and BPE, return specific files as before - file_key = "tokenizer.model" if vocabtype == "spm" else "vocab.json" - if self.files[file_key]: - return self.files[file_key] - else: - raise FileNotFoundError(f"{vocabtype} {file_key} not found.") + for file_key in self.files.keys(): + if self.files[file_key]: + return self.files[file_key] + raise FileNotFoundError(f"{vocabtype} vocab not found.") elif vocabtype == "hfft": # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file return self.path From 0f83e727af0a7cadf90b7ecc1f8e35de1d0880bc Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 18:37:36 +0200 Subject: [PATCH 005/101] py : fix whitespace --- convert.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert.py b/convert.py index b47bb6185..e38ee5315 100755 --- a/convert.py +++ b/convert.py @@ -1100,7 +1100,7 @@ class OutputFile: scores.append(score) toktypes.append(toktype) - assert(len(tokens) == vocab.vocab_size) + assert len(tokens) == vocab.vocab_size return tokens, scores, toktypes From c918fe8dca8fa1c4602427e0a4b88e20046f6c34 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 18:38:39 +0200 Subject: [PATCH 006/101] metal : create autorelease pool during library build (#4970) * metal : create autorelease pool during library build ggml-ci * test : simplify ggml-ci --- .gitignore | 1 + Makefile | 5 ++++- ci/run.sh | 2 ++ ggml-metal.m | 19 +++++++++---------- tests/CMakeLists.txt | 1 + tests/test-autorelease.cpp | 28 ++++++++++++++++++++++++++++ 6 files changed, 45 insertions(+), 11 deletions(-) create mode 100644 tests/test-autorelease.cpp diff --git a/.gitignore b/.gitignore index fba207045..5ab81445d 100644 --- a/.gitignore +++ b/.gitignore @@ -105,3 +105,4 @@ poetry.toml /tests/test-tokenizer-1-bpe /tests/test-rope /tests/test-backend-ops +/tests/test-autorelease diff --git a/Makefile b/Makefile index 995b89f7a..a8658a596 100644 --- a/Makefile +++ b/Makefile @@ -9,7 +9,7 @@ TEST_TARGETS = \ tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \ tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \ tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope \ - tests/test-backend-ops + tests/test-backend-ops tests/test-autorelease # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -747,3 +747,6 @@ tests/test-c.o: tests/test-c.c llama.h tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/ci/run.sh b/ci/run.sh index 47a254f4c..86293f0db 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -179,6 +179,8 @@ function gg_run_open_llama_3b_v2 { wiki_test_60="${path_wiki}/wiki.test-60.raw" + ./bin/test-autorelease ${model_f16} + ./bin/quantize ${model_f16} ${model_q8_0} q8_0 ./bin/quantize ${model_f16} ${model_q4_0} q4_0 ./bin/quantize ${model_f16} ${model_q4_1} q4_1 diff --git a/ggml-metal.m b/ggml-metal.m index 8bb4edd64..66d4d675e 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -303,22 +303,21 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { return NULL; } - // dictionary of preprocessor macros - NSMutableDictionary * prep = [NSMutableDictionary dictionary]; + @autoreleasepool { + // dictionary of preprocessor macros + NSMutableDictionary * prep = [NSMutableDictionary dictionary]; #ifdef GGML_QKK_64 - prep[@"QK_K"] = @(64); + prep[@"QK_K"] = @(64); #endif - MTLCompileOptions* options = [MTLCompileOptions new]; - options.preprocessorMacros = prep; + MTLCompileOptions* options = [MTLCompileOptions new]; + options.preprocessorMacros = prep; - //[options setFastMathEnabled:false]; + //[options setFastMathEnabled:false]; - ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error]; - - [options release]; - [prep release]; + ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error]; + } } if (error) { diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 7c932240d..d7aaab843 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -49,6 +49,7 @@ llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # llama_build_and_test_executable(test-opt.cpp) # SLOW llama_build_and_test_executable(test-backend-ops.cpp) +llama_build_and_test_executable(test-autorelease.cpp) llama_build_and_test_executable(test-rope.cpp) diff --git a/tests/test-autorelease.cpp b/tests/test-autorelease.cpp new file mode 100644 index 000000000..289c6ba6c --- /dev/null +++ b/tests/test-autorelease.cpp @@ -0,0 +1,28 @@ +// ref: https://github.com/ggerganov/llama.cpp/issues/4952#issuecomment-1892864763 + +#include +#include +#include + +#include "llama.h" + +// This creates a new context inside a pthread and then tries to exit cleanly. +int main(int argc, char ** argv) { + if (argc < 2) { + printf("Usage: %s model.gguf\n", argv[0]); + return 0; // intentionally return success + } + + const std::string fname = argv[1]; + + std::thread([&fname]() { + llama_backend_init(false); + auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params()); + auto * ctx = llama_new_context_with_model(model, llama_context_default_params()); + llama_free(ctx); + llama_free_model(model); + llama_backend_free(); + }).join(); + + return 0; +} From 44a1a4a41a4c0b03afaa7d9e06bcbc7cf95aa1e6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 18:39:41 +0200 Subject: [PATCH 007/101] backend : add eval callback (#4935) * backend : add eval callback ggml-ci * backend : group nodes in a single compute when user don't need them * backend : clean-up the implementation ggml-ci * simple : do not perform tensor data copy if not needed * simple : fix * simple : no need for ggml_is_contiguous + fix bool parse * llama : fix callback placement in llama_context_params * backend : avoid double-ask callback calls * simple : restore examples, imatrix will serve as a demo --- ggml-backend.c | 42 ++++++++++++++++++++++++++++++++++++++++-- ggml-backend.h | 11 +++++++++++ llama.cpp | 9 +++++++++ llama.h | 4 ++++ 4 files changed, 64 insertions(+), 2 deletions(-) diff --git a/ggml-backend.c b/ggml-backend.c index f5424fb90..4266250f9 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -802,6 +802,9 @@ struct ggml_backend_sched { __attribute__((aligned(GGML_MEM_ALIGN))) #endif char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)]; + + ggml_backend_sched_eval_callback callback_eval; + void * callback_eval_user_data; }; #define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node) @@ -1324,9 +1327,38 @@ static void sched_compute_splits(ggml_backend_sched_t sched) { ggml_graph_dump_dot(split->graph, NULL, split_filename); #endif + uint64_t compute_start_us = ggml_time_us(); - ggml_backend_graph_compute(split_backend, &split->graph); - //ggml_backend_synchronize(split_backend); // necessary to measure compute time + if (!sched->callback_eval) { + ggml_backend_graph_compute(split_backend, &split->graph); + //ggml_backend_synchronize(split_backend); // necessary to measure compute time + } else { + // similar to ggml_backend_compare_graph_backend + for (int j0 = 0; j0 < split->graph.n_nodes; j0++) { + struct ggml_tensor * t = split->graph.nodes[j0]; + + // check if the user needs data from this node + bool need = sched->callback_eval(t, true, sched->callback_eval_user_data); + + int j1 = j0; + + // determine the range [j0, j1] of nodes that can be computed together + while (!need && j1 < split->graph.n_nodes - 1) { + t = split->graph.nodes[++j1]; + need = sched->callback_eval(t, true, sched->callback_eval_user_data); + } + + struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1); + + ggml_backend_graph_compute(split_backend, &gv); + + if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) { + break; + } + + j0 = j1; + } + } uint64_t compute_end_us = ggml_time_us(); compute_us[split_backend_id] += compute_end_us - compute_start_us; } @@ -1431,6 +1463,12 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) { sched_reset(sched); } + +void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) { + sched->callback_eval = callback; + sched->callback_eval_user_data = user_data; +} + int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) { return sched->n_splits; } diff --git a/ggml-backend.h b/ggml-backend.h index 12b4b4ab7..ab4ad773f 100644 --- a/ggml-backend.h +++ b/ggml-backend.h @@ -148,6 +148,14 @@ extern "C" { struct ggml_backend_sched; typedef struct ggml_backend_sched * ggml_backend_sched_t; + // when ask == true, the scheduler wants to know if the user wants to observe this node + // this allows the scheduler to batch nodes together in order to evaluate them in a single call + // + // when ask == false, the scheduler is passing the node tensor to the user for observation + // if the user returns false, the scheduler will cancel the graph compute + // + typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data); + // Initialize a backend scheduler GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size); GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); @@ -168,6 +176,9 @@ extern "C" { // Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs GGML_API void ggml_backend_sched_reset(ggml_backend_sched_t sched); + // Set a callback to be called for each resulting node during graph compute + GGML_API void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data); + // // Utils // diff --git a/llama.cpp b/llama.cpp index 2c5983c67..81829b13e 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1393,6 +1393,9 @@ struct llama_cparams { bool mul_mat_q; bool offload_kqv; + + ggml_backend_sched_eval_callback cb_eval; + void * cb_eval_user_data; }; struct llama_layer { @@ -6254,6 +6257,7 @@ static int llama_decode_internal( //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); ggml_backend_sched_reset(lctx.sched); + ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data); ggml_cgraph * gf = llama_build_graph(lctx, batch); @@ -9276,6 +9280,8 @@ struct llama_context_params llama_context_default_params() { /*.yarn_beta_fast =*/ 32.0f, /*.yarn_beta_slow =*/ 1.0f, /*.yarn_orig_ctx =*/ 0, + /*.cb_eval =*/ nullptr, + /*.cb_eval_user_data =*/ nullptr, /*.type_k =*/ GGML_TYPE_F16, /*.type_v =*/ GGML_TYPE_F16, /*.mul_mat_q =*/ true, @@ -9416,6 +9422,9 @@ struct llama_context * llama_new_context_with_model( hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx : hparams.n_ctx_train; + cparams.cb_eval = params.cb_eval; + cparams.cb_eval_user_data = params.cb_eval_user_data; + auto rope_scaling_type = params.rope_scaling_type; if (rope_scaling_type == LLAMA_ROPE_SCALING_UNSPECIFIED) { rope_scaling_type = hparams.rope_scaling_type_train; diff --git a/llama.h b/llama.h index a570b0d69..e268d7a1d 100644 --- a/llama.h +++ b/llama.h @@ -2,6 +2,7 @@ #define LLAMA_H #include "ggml.h" +#include "ggml-backend.h" #ifdef GGML_USE_CUBLAS #include "ggml-cuda.h" #define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES @@ -231,6 +232,9 @@ extern "C" { float yarn_beta_slow; // YaRN high correction dim uint32_t yarn_orig_ctx; // YaRN original context size + ggml_backend_sched_eval_callback cb_eval; + void * cb_eval_user_data; + enum ggml_type type_k; // data type for K cache enum ggml_type type_v; // data type for V cache From ba69bbc84ced580fe4fdb0713ca2d95634325b7a Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 18:46:30 +0200 Subject: [PATCH 008/101] imatrix : offload to GPU support (#4957) * backend : add eval callback ggml-ci * backend : group nodes in a single compute when user don't need them * backend : clean-up the implementation ggml-ci * simple : do not perform tensor data copy if not needed * simple : fix * imatrix : offload to GPU support * imatrix : fix ggml_mul_mat_id hanlding ggml-ci * ci : add imatrix test ggml-ci * ci : rearrange output ggml-ci --- ci/run.sh | 11 ++- examples/imatrix/imatrix.cpp | 152 +++++++++++++++++++++++++++-------- ggml.c | 14 ---- ggml.h | 6 -- 4 files changed, 129 insertions(+), 54 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index 86293f0db..f3a8ff774 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -216,6 +216,8 @@ function gg_run_open_llama_3b_v2 { (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log + (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { @@ -243,6 +245,8 @@ function gg_run_open_llama_3b_v2 { check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log + cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log + # lora function compare_ppl { qnt="$1" @@ -284,7 +288,6 @@ function gg_run_open_llama_3b_v2 { (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log - set +e } @@ -294,6 +297,7 @@ function gg_sum_open_llama_3b_v2 { gg_printf 'OpenLLaMA 3B-v2:\n' gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)" + gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)" gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)" gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)" gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)" @@ -393,6 +397,8 @@ function gg_run_open_llama_7b_v2 { (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log + (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log + (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { @@ -420,6 +426,8 @@ function gg_run_open_llama_7b_v2 { check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log + cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log + # lora function compare_ppl { qnt="$1" @@ -471,6 +479,7 @@ function gg_sum_open_llama_7b_v2 { gg_printf 'OpenLLaMA 7B-v2:\n' gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)" + gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)" gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)" gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)" gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)" diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 1461bc963..af78711c5 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -33,43 +33,120 @@ class IMatrixCollector { public: IMatrixCollector() = default; void set_parameters(StatParams&& params) { m_params = std::move(params); } - void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1); + bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data); void save_imatrix() const; private: std::unordered_map m_stats; StatParams m_params; std::mutex m_mutex; int m_last_call = 0; + std::vector m_src1_data; + std::vector m_ids; // the expert ids from ggml_mul_mat_id }; -void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) { - if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return; - if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return; +bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) { + GGML_UNUSED(user_data); + + const struct ggml_tensor * src0 = t->src[0]; + const struct ggml_tensor * src1 = t->src[1]; + + // when ask is true, the scheduler wants to know if we are interested in data from this tensor + // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection + if (ask) { + if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications + if (t->op != GGML_OP_MUL_MAT) return false; + if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false; + if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false; + return true; + } + std::lock_guard lock(m_mutex); - auto& e = m_stats[src0->name]; - if (e.values.empty()) { - e.values.resize(src1->ne[0], 0); + + // copy the data from the GPU memory if needed + const bool is_host = ggml_backend_buffer_is_host(src1->buffer); + + if (!is_host) { + m_src1_data.resize(ggml_nelements(src1)); + ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1)); } - else if (e.values.size() != (size_t)src1->ne[0]) { - fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]); - exit(1); //GGML_ASSERT(false); - } - ++e.ncall; - if (m_params.verbosity > 1) { - printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type); - } - for (int row = 0; row < (int)src1->ne[1]; ++row) { - const float * x = (const float *)src1->data + row * src1->ne[0]; - for (int j = 0; j < (int)src1->ne[0]; ++j) { - e.values[j] += x[j]*x[j]; - } - } - if (e.ncall > m_last_call) { - m_last_call = e.ncall; - if (m_last_call % m_params.n_output_frequency == 0) { - save_imatrix(); + + const float * data = is_host ? (const float *) src1->data : m_src1_data.data(); + + if (t->op == GGML_OP_MUL_MAT_ID) { + const int idx = ((int32_t *) t->op_params)[0]; + const int n_as = ((int32_t *) t->op_params)[1]; + + // the top-k selected expert ids are stored in the src0 tensor + // for simplicity, always copy src0 to host, because it is small + // take into account that src0 is not contiguous! + GGML_ASSERT(src0->ne[1] == src1->ne[1]); + GGML_ASSERT(n_as*ggml_nrows(src0)); + m_ids.resize(ggml_nbytes(src0)/sizeof(int)); + ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0)); + + // loop over all possible experts, regardless if they are used or not in the batch + // this is necessary to guarantee equal number of "ncall" for each tensor + for (int ex = 0; ex < n_as; ++ex) { + src0 = t->src[2 + ex]; + auto& e = m_stats[src0->name]; + if (e.values.empty()) { + e.values.resize(src1->ne[0], 0); + } + else if (e.values.size() != (size_t)src1->ne[0]) { + fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]); + exit(1); //GGML_ASSERT(false); + } + // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger + // using the following line, we can correct for that if needed + //if (idx == t->src[0]->ne[0] - 1) ++e.ncall; + ++e.ncall; + if (m_params.verbosity > 1) { + printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type); + } + for (int row = 0; row < (int)src1->ne[1]; ++row) { + const int excur = m_ids[row*n_as + idx]; + GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check + if (excur != ex) continue; + const float * x = data + row * src1->ne[0]; + for (int j = 0; j < (int)src1->ne[0]; ++j) { + e.values[j] += x[j]*x[j]; + } + } + if (e.ncall > m_last_call) { + m_last_call = e.ncall; + if (m_last_call % m_params.n_output_frequency == 0) { + save_imatrix(); + } + } + } + } else { + auto& e = m_stats[src0->name]; + if (e.values.empty()) { + e.values.resize(src1->ne[0], 0); + } + else if (e.values.size() != (size_t)src1->ne[0]) { + fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]); + exit(1); //GGML_ASSERT(false); + } + ++e.ncall; + if (m_params.verbosity > 1) { + printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type); + } + for (int row = 0; row < (int)src1->ne[1]; ++row) { + const float * x = data + row * src1->ne[0]; + for (int j = 0; j < (int)src1->ne[0]; ++j) { + e.values[j] += x[j]*x[j]; + } + } + if (e.ncall > m_last_call) { + m_last_call = e.ncall; + if (m_last_call % m_params.n_output_frequency == 0) { + save_imatrix(); + } } } + + return true; } void IMatrixCollector::save_imatrix() const { @@ -93,8 +170,8 @@ void IMatrixCollector::save_imatrix() const { static IMatrixCollector g_collector; -static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) { - g_collector.collect_imatrix(src0, src1); +static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) { + return g_collector.collect_imatrix(t, ask, user_data); } @@ -320,8 +397,6 @@ int main(int argc, char ** argv) { g_collector.set_parameters(std::move(sparams)); - ggml_set_imatrix_collection(ik_collect_imatrix); - params.logits_all = true; params.n_batch = std::min(params.n_batch, params.n_ctx); @@ -340,16 +415,27 @@ int main(int argc, char ** argv) { llama_backend_init(params.numa); - llama_model * model; - llama_context * ctx; + llama_model_params mparams = llama_model_params_from_gpt_params(params); - // load the model and apply lora adapter, if any - std::tie(model, ctx) = llama_init_from_gpt_params(params); + llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams); if (model == NULL) { fprintf(stderr, "%s: error: unable to load model\n", __func__); return 1; } + llama_context_params cparams = llama_context_params_from_gpt_params(params); + + // pass the callback to the backend scheduler + // it will be executed for each node during the graph computation + cparams.cb_eval = ik_collect_imatrix; + cparams.cb_eval_user_data = NULL; + + llama_context * ctx = llama_new_context_with_model(model, cparams); + if (ctx == NULL) { + fprintf(stderr, "%s: error: unable to create context\n", __func__); + return 1; + } + const int n_ctx_train = llama_n_ctx_train(model); if (params.n_ctx > n_ctx_train) { fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n", diff --git a/ggml.c b/ggml.c index d7e01b81f..35fd29a9e 100644 --- a/ggml.c +++ b/ggml.c @@ -394,12 +394,6 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y); static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y); -ggml_collect_imatrix_t g_imatrix_collect = NULL; - -void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect) { - g_imatrix_collect = imatrix_collect; -} - static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { [GGML_TYPE_I8] = { .type_name = "i8", @@ -9790,10 +9784,6 @@ static void ggml_compute_forward_mul_mat( const int ith = params->ith; const int nth = params->nth; - if (ith == 1 && g_imatrix_collect) { - g_imatrix_collect(src0, src1); - } - const enum ggml_type type = src0->type; const bool src1_cont = ggml_is_contiguous(src1); @@ -10097,10 +10087,6 @@ static void ggml_compute_forward_mul_mat_id( const struct ggml_tensor * src0_cur = dst->src[cur_a + 2]; - if (ith == 1 && g_imatrix_collect) { - g_imatrix_collect(src0_cur, src1); - } - const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata; const size_t row_size = ggml_row_size(vec_dot_type, ne10); diff --git a/ggml.h b/ggml.h index 837c52e68..27daf6fd1 100644 --- a/ggml.h +++ b/ggml.h @@ -2085,12 +2085,6 @@ extern "C" { GGML_API void ggml_init_iq2_quantization(enum ggml_type type); GGML_API void ggml_deinit_iq2_quantization(enum ggml_type type); - // - // Importance matrix - // - typedef void(*ggml_collect_imatrix_t)(const struct ggml_tensor * src0, const struct ggml_tensor * src1); - GGML_API void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect); - // // gguf // From 38566680cdfe982a495562332c25b9227de9cf8d Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 18:54:56 +0200 Subject: [PATCH 009/101] ggml : add IQ2 to test-backend-ops + refactoring (#4990) * ggml : add IQ2 to test-backend-ops + refactoring ggml-ci * cuda : update supports_op for IQ2 ggml-ci * ci : enable LLAMA_CUBLAS=1 for CUDA nodes ggml-ci * cuda : fix out-of-bounds-access in `mul_mat_vec_q` ggml-ci * tests : avoid creating RNGs for each Q tensor ggml-ci * tests : avoid creating RNGs for each tensor ggml-ci --- ci/run.sh | 12 ++++--- ggml-backend.c | 2 ++ ggml-cuda.cu | 12 +++++-- ggml-quants.c | 74 ++++++++++++++++---------------------- ggml-quants.h | 3 ++ ggml.c | 34 ++++++++++++++++-- ggml.h | 20 ++++++++--- llama.cpp | 12 +------ tests/test-backend-ops.cpp | 46 ++++++++++++++---------- 9 files changed, 128 insertions(+), 87 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index f3a8ff774..791b17a19 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -36,6 +36,10 @@ if [ ! -z ${GG_BUILD_METAL} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON" fi +if [ ! -z ${GG_BUILD_CUDA} ]; then + CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1" +fi + ## helpers # download a file if it does not exist or if it is outdated @@ -160,8 +164,8 @@ function gg_run_open_llama_3b_v2 { set -e - (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log + (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log python3 ../convert.py ${path_models} @@ -343,8 +347,8 @@ function gg_run_open_llama_7b_v2 { set -e - (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log - (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log + (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log + (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log python3 ../convert.py ${path_models} diff --git a/ggml-backend.c b/ggml-backend.c index 4266250f9..ef518dae0 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -692,6 +692,8 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) { switch (op->op) { + case GGML_OP_CPY: + return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS; // missing type_traits.from_float case GGML_OP_MUL_MAT: return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type; default: diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 568c411af..b2211d858 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -5131,10 +5131,10 @@ static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * const block_q_t * x = (const block_q_t *) vx; const block_q8_1 * y = (const block_q8_1 *) vy; - for (int i = 0; i < blocks_per_row; i += blocks_per_warp) { - const int ibx = row*blocks_per_row + i + threadIdx.x / (qi/vdr); // x block index + for (int i = threadIdx.x / (qi/vdr); i < blocks_per_row; i += blocks_per_warp) { + const int ibx = row*blocks_per_row + i; // x block index - const int iby = (i + threadIdx.x / (qi/vdr)) * (qk/QK8_1); // y block index that aligns with ibx + const int iby = i * (qk/QK8_1); // y block index that aligns with ibx const int iqs = vdr * (threadIdx.x % (qi/vdr)); // x block quant index when casting the quants to int @@ -10918,6 +10918,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons if (a->ne[3] != b->ne[3]) { return false; } + ggml_type a_type = a->type; + if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS) { + if (b->ne[1] == 1 && ggml_nrows(b) > 1) { + return false; + } + } return true; } break; case GGML_OP_GET_ROWS: diff --git a/ggml-quants.c b/ggml-quants.c index 31b053e33..7d2f033e9 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -1274,7 +1274,12 @@ static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * } float sumlx = 0; float suml2 = 0; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 0; i < n; ++i) { +#else for (int i = 0; i < n; ++i) { +#endif int l = nearest_int(iscale * x[i]); l = MAX(-nmax, MIN(nmax-1, l)); L[i] = l + nmax; @@ -1649,7 +1654,12 @@ static float make_qkx3_quants(int n, int nmax, const float * restrict x, const f float max = x[0]; float sum_w = weights ? weights[0] : x[0]*x[0]; float sum_x = sum_w * x[0]; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 1; i < n; ++i) { +#else for (int i = 1; i < n; ++i) { +#endif if (x[i] < min) min = x[i]; if (x[i] > max) max = x[i]; float w = weights ? weights[i] : x[i]*x[i]; @@ -1660,7 +1670,7 @@ static float make_qkx3_quants(int n, int nmax, const float * restrict x, const f min = 0; } if (max <= min) { - for (int i = 0; i < n; ++i) L[i] = 0; + memset(L, 0, n); *the_min = -min; return 0.f; } @@ -1862,7 +1872,7 @@ static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restri size_t quantize_q2_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { (void)hist; - int row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row); if (!quant_weights) { quantize_row_q2_K_reference(src, dst, nrow*n_per_row); } @@ -2181,7 +2191,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri size_t quantize_q3_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { (void)hist; - int row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row); if (!quant_weights) { quantize_row_q3_K_reference(src, dst, nrow*n_per_row); } @@ -2448,7 +2458,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri size_t quantize_q4_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { (void)hist; - int row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row); if (!quant_weights) { quantize_row_q4_K_reference(src, dst, nrow*n_per_row); } @@ -2771,7 +2781,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri size_t quantize_q5_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { (void)hist; - int row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row); if (!quant_weights) { quantize_row_q5_K_reference(src, dst, nrow*n_per_row); } @@ -3025,7 +3035,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri size_t quantize_q6_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { (void)hist; - int row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row); if (!quant_weights) { quantize_row_q6_K_reference(src, dst, nrow*n_per_row); } @@ -3072,7 +3082,7 @@ size_t quantize_q4_0(const float * src, void * dst, int nrow, int n_per_row, int if (!quant_weights) { return ggml_quantize_q4_0(src, dst, nrow*n_per_row, n_per_row, hist); } - int row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row); char * qrow = (char *)dst; for (int row = 0; row < nrow; ++row) { quantize_row_q4_0_impl(src, (block_q4_0*)qrow, n_per_row, quant_weights); @@ -3116,7 +3126,7 @@ size_t quantize_q4_1(const float * src, void * dst, int nrow, int n_per_row, int if (!quant_weights) { return ggml_quantize_q4_1(src, dst, nrow*n_per_row, n_per_row, hist); } - int row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row); char * qrow = (char *)dst; for (int row = 0; row < nrow; ++row) { quantize_row_q4_1_impl(src, (block_q4_1*)qrow, n_per_row, quant_weights); @@ -3169,7 +3179,7 @@ size_t quantize_q5_0(const float * src, void * dst, int nrow, int n_per_row, int if (!quant_weights) { return ggml_quantize_q5_0(src, dst, nrow*n_per_row, n_per_row, hist); } - int row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row); char * qrow = (char *)dst; for (int row = 0; row < nrow; ++row) { quantize_row_q5_0_impl(src, (block_q5_0*)qrow, n_per_row, quant_weights); @@ -3221,7 +3231,7 @@ size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int if (!quant_weights) { return ggml_quantize_q5_1(src, dst, nrow*n_per_row, n_per_row, hist); } - int row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row); + size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row); char * qrow = (char *)dst; for (int row = 0; row < nrow; ++row) { quantize_row_q5_1_impl(src, (block_q5_1*)qrow, n_per_row, quant_weights); @@ -8565,7 +8575,7 @@ static int iq2_compare_func(const void * left, const void * right) { return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0; } -static void q2xs_init_impl(int grid_size) { +void iq2xs_init_impl(int grid_size) { const int gindex = iq2_data_index(grid_size); if (iq2_data[gindex].grid) { return; @@ -8720,19 +8730,7 @@ static void q2xs_init_impl(int grid_size) { free(dist2); } -void ggml_init_iq2_quantization(enum ggml_type type) { - if (type == GGML_TYPE_IQ2_XXS) { - q2xs_init_impl(256); - } - else if (type == GGML_TYPE_IQ2_XS) { - q2xs_init_impl(512); - } - else { - fprintf(stderr, "======================== Why are you calling %s with type %d?\n", __func__, (int)type); - } -} - -static void q2xs_deinit_impl(int grid_size) { +void iq2xs_free_impl(int grid_size) { GGML_ASSERT(grid_size == 256 || grid_size == 512 || grid_size == 1024); const int gindex = iq2_data_index(grid_size); if (iq2_data[gindex].grid) { @@ -8742,18 +8740,6 @@ static void q2xs_deinit_impl(int grid_size) { } } -void ggml_deinit_iq2_quantization(enum ggml_type type) { - if (type == GGML_TYPE_IQ2_XXS) { - q2xs_deinit_impl(256); - } - else if (type == GGML_TYPE_IQ2_XS) { - q2xs_deinit_impl(512); - } - else { - fprintf(stderr, "======================== Why are you calling %s with type %d?\n", __func__, (int)type); - } -} - static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid, const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) { int num_neighbors = neighbours[0]; @@ -8786,10 +8772,10 @@ static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict const int * kmap_q2xs = iq2_data[gindex].map; const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; - GGML_ASSERT(quant_weights); - GGML_ASSERT(kgrid_q2xs); - GGML_ASSERT(kmap_q2xs); - GGML_ASSERT(kneighbors_q2xs); + GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); GGML_ASSERT(n%QK_K == 0); const int kMaxQ = 3; @@ -9005,10 +8991,10 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v const int * kmap_q2xs = iq2_data[gindex].map; const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours; - GGML_ASSERT(quant_weights); - GGML_ASSERT(kmap_q2xs); - GGML_ASSERT(kgrid_q2xs); - GGML_ASSERT(kneighbors_q2xs); + GGML_ASSERT(quant_weights && "missing quantization weights"); + GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?"); + GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?"); GGML_ASSERT(n%QK_K == 0); const int kMaxQ = 3; diff --git a/ggml-quants.h b/ggml-quants.h index d7fefdb54..7d7cf9178 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -257,3 +257,6 @@ size_t quantize_q4_0 (const float * src, void * dst, int nrows, int n_per_row, size_t quantize_q4_1 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_q5_0 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); size_t quantize_q5_1 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); + +void iq2xs_init_impl(int grid_size); +void iq2xs_free_impl(int grid_size); diff --git a/ggml.c b/ggml.c index 35fd29a9e..cbf2d4bdd 100644 --- a/ggml.c +++ b/ggml.c @@ -18524,6 +18524,28 @@ enum ggml_opt_result ggml_opt_resume_g( //////////////////////////////////////////////////////////////////////////////// +void ggml_quantize_init(enum ggml_type type) { + ggml_critical_section_start(); + + switch (type) { + case GGML_TYPE_IQ2_XXS: iq2xs_init_impl(256); break; + case GGML_TYPE_IQ2_XS: iq2xs_init_impl(512); break; + default: // nothing + break; + } + + ggml_critical_section_end(); +} + +void ggml_quantize_free(void) { + ggml_critical_section_start(); + + iq2xs_free_impl(256); + iq2xs_free_impl(512); + + ggml_critical_section_end(); +} + size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) { assert(k % QK4_0 == 0); const int nb = k / QK4_0; @@ -18651,9 +18673,15 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * return (n/QK8_0*sizeof(block_q8_0)); } +bool ggml_quantize_requires_imatrix(enum ggml_type type) { + return + type == GGML_TYPE_IQ2_XXS || + type == GGML_TYPE_IQ2_XS; +} + size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix) { - (void)imatrix; + ggml_quantize_init(type); // this is noop if already initialized size_t result = 0; int n = nrows * n_per_row; switch (type) { @@ -18766,13 +18794,13 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i } break; case GGML_TYPE_F16: { - int elemsize = sizeof(ggml_fp16_t); + size_t elemsize = sizeof(ggml_fp16_t); ggml_fp32_to_fp16_row(src + start, (ggml_fp16_t *)dst + start, n); result = n * elemsize; } break; case GGML_TYPE_F32: { - int elemsize = sizeof(float); + size_t elemsize = sizeof(float); result = n * elemsize; memcpy((uint8_t *)dst + start * elemsize, src + start, result); } break; diff --git a/ggml.h b/ggml.h index 27daf6fd1..de8162b81 100644 --- a/ggml.h +++ b/ggml.h @@ -2065,6 +2065,18 @@ extern "C" { // quantization // + // - ggml_quantize_init can be called multiple times with the same type + // it will only initialize the quantization tables for the first call or after ggml_quantize_free + // automatically called by ggml_quantize_chunk for convenience + // + // - ggml_quantize_free will free any memory allocated by ggml_quantize_init + // call this at the end of the program to avoid memory leaks + // + // note: these are thread-safe + // + GGML_API void ggml_quantize_init(enum ggml_type type); + GGML_API void ggml_quantize_free(void); + // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist); @@ -2078,13 +2090,13 @@ extern "C" { GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist); GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist); + // some quantization type cannot be used without an importance matrix + GGML_API bool ggml_quantize_requires_imatrix(enum ggml_type type); + + // calls ggml_quantize_init internally (i.e. can allocate memory) GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix); - // These are needed for IQ2_XS and IQ2_XXS quantizations - GGML_API void ggml_init_iq2_quantization(enum ggml_type type); - GGML_API void ggml_deinit_iq2_quantization(enum ggml_type type); - // // gguf // diff --git a/llama.cpp b/llama.cpp index 81829b13e..d28382f7d 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8747,8 +8747,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s // placeholder for the meta data ::zeros(fout, meta_size); - std::set used_iq2; - for (int i = 0; i < ml.n_tensors; ++i) { struct ggml_tensor * tensor = ml.get_tensor_meta(i); @@ -8801,11 +8799,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s } else { const size_t nelements = ggml_nelements(tensor); - if ((new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XS) && used_iq2.find(new_type) == used_iq2.end()) { - ggml_init_iq2_quantization(new_type); - used_iq2.insert(new_type); - } - const float * imatrix = nullptr; if (imatrix_data) { auto it = imatrix_data->find(tensor->name); @@ -8931,10 +8924,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s fout.close(); - for (auto type : used_iq2) { - ggml_deinit_iq2_quantization(type); - } - gguf_free(ctx_out); LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); @@ -9342,6 +9331,7 @@ void llama_backend_free(void) { #ifdef GGML_USE_MPI ggml_mpi_backend_free(); #endif + ggml_quantize_free(); } int64_t llama_time_us(void) { diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 22a7856d4..55ce14e0d 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -16,39 +16,37 @@ #include static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) { + // static RNG initialization (revisit if n_threads stops being constant) + static const size_t n_threads = std::thread::hardware_concurrency(); + static std::vector generators = []() { + std::random_device rd; + std::vector vec; + vec.reserve(n_threads); + //for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(1234 + i); } // fixed seed + for (size_t i = 0; i < n_threads; i++) { vec.emplace_back(rd()); } + return vec; + }(); + size_t size = ggml_nelements(tensor); std::vector data(size); -#if 0 - static std::default_random_engine generator(1234); - std::uniform_real_distribution distribution(min, max); - - for (size_t i = 0; i < size; i++) { - data[i] = distribution(generator); - } -#else - auto init_thread = [&](size_t start, size_t end) { - std::random_device rd; - std::default_random_engine generator(rd()); + auto init_thread = [&](size_t ith, size_t start, size_t end) { std::uniform_real_distribution distribution(min, max); - for (size_t i = start; i < end; i++) { - data[i] = distribution(generator); + data[i] = distribution(generators[ith]); } }; - size_t n_threads = std::thread::hardware_concurrency(); std::vector threads; threads.reserve(n_threads); for (size_t i = 0; i < n_threads; i++) { size_t start = i*size/n_threads; size_t end = (i+1)*size/n_threads; - threads.emplace_back(init_thread, start, end); + threads.emplace_back(init_thread, i, start, end); } for (auto & t : threads) { t.join(); } -#endif if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) { ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float)); @@ -56,7 +54,16 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0); std::vector dataq(ggml_row_size(tensor->type, size)); int64_t hist[16]; - ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], hist, nullptr); + std::vector imatrix(tensor->ne[0], 1.0f); // dummy importance matrix + const float * im = imatrix.data(); + if (!ggml_quantize_requires_imatrix(tensor->type)) { + // when the imatrix is optional, we want to test both quantization with and without imatrix + // use one of the random numbers to decide + if (data[0] > 0.5f*(min + max)) { + im = nullptr; + } + } + ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], hist, im); ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size()); } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) { // This is going to create some weird integers though. @@ -1472,7 +1479,8 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op GGML_TYPE_Q8_0, GGML_TYPE_Q2_K, GGML_TYPE_Q3_K, GGML_TYPE_Q4_K, GGML_TYPE_Q5_K, - GGML_TYPE_Q6_K + GGML_TYPE_Q6_K, + GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, }; // unary ops @@ -1752,6 +1760,8 @@ int main(int argc, char ** argv) { return 1; } + ggml_quantize_free(); + printf("\033[1;32mOK\033[0m\n"); return 0; } From 6b6916b215251e09bd57cdbf870dc8a73345edc2 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 17 Jan 2024 20:54:50 +0200 Subject: [PATCH 010/101] sync : ggml --- scripts/sync-ggml.last | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index be9e408fb..4d52d946b 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -b306d6e996ec0ace77118fa5098822cdc7f9c88f +6c1ce0bd591a430c1d3f6797d905194581c878c1 From 1e605f4102c7ea8dc0dff82f5eaa6a71973d549f Mon Sep 17 00:00:00 2001 From: Paul Tsochantaris Date: Thu, 18 Jan 2024 08:47:24 +0000 Subject: [PATCH 011/101] metal : fix memory leak, dangling pointer and unused autorel (#5007) * Metal memory: Small memory leak on init, dangling pointer, and unused autorelease pool in graph compute * SPM header potential fix * Reverting symlinks --- ggml-metal.m | 16 ++++++---------- 1 file changed, 6 insertions(+), 10 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 66d4d675e..912ddc83f 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -238,21 +238,19 @@ static void * ggml_metal_host_malloc(size_t n) { static struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_LOG_INFO("%s: allocating\n", __func__); - id device; - NSString * s; - -#if TARGET_OS_OSX +#if TARGET_OS_OSX && !GGML_METAL_NDEBUG // Show all the Metal device instances in the system NSArray * devices = MTLCopyAllDevices(); - for (device in devices) { - s = [device name]; + for (id device in devices) { + NSString * s = [device name]; GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]); } + [devices release]; // since it was created by a *Copy* C method #endif // Pick and show default Metal device - device = MTLCreateSystemDefaultDevice(); - s = [device name]; + id device = MTLCreateSystemDefaultDevice(); + NSString * s = [device name]; GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]); // Configure context @@ -712,7 +710,6 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const static bool ggml_metal_graph_compute( struct ggml_metal_context * ctx, struct ggml_cgraph * gf) { - @autoreleasepool { MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor; edesc.dispatchType = MTLDispatchTypeSerial; @@ -2255,7 +2252,6 @@ static bool ggml_metal_graph_compute( } return true; - } } //////////////////////////////////////////////////////////////////////////////// From dcad445d0c83ad49bca1b58cf9c139cfcebee5d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 11:44:49 +0200 Subject: [PATCH 012/101] scritps : add helper script to get hellaswag data in txt format --- scripts/get-hellaswag.sh | 10 ++++++++++ 1 file changed, 10 insertions(+) create mode 100755 scripts/get-hellaswag.sh diff --git a/scripts/get-hellaswag.sh b/scripts/get-hellaswag.sh new file mode 100755 index 000000000..ef8dcceb0 --- /dev/null +++ b/scripts/get-hellaswag.sh @@ -0,0 +1,10 @@ +#!/bin/bash + +wget https://raw.githubusercontent.com/klosax/hellaswag_text_data/main/hellaswag_val_full.txt + +echo "Usage:" +echo "" +echo " ./perplexity --hellaswag --hellaswag-tasks N -f hellaswag_val_full.txt -m modelfile.gguf" +echo "" + +exit 0 From 682986a08eb5cb04865d2e713449f17304d266d8 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 18 Jan 2024 13:46:27 +0200 Subject: [PATCH 013/101] Add Winogrande evaluation (#5015) * winogrande: simple implementation It doesn't look like it is working - why? For Mistral-7B it is barely better than random chance (score ~60% for 1267 tasks), while I see Mistral-7B scoring 78.4% on the HF leader board. 1-sigma statistical uncertainty for 1267 tasks is ~1.4, so no way the difference is due to statistics. * winogrande: somewhat better Score for Mistrali7-B is now 68.9 on the validation set of winogrande_debiased. Still far from the reported 78.4, but better than what I had before. * winogrande: improving Mistral-7B score is now 73.56. Still not quite 78.4 but getting there. We are also getting a lower score on HellaSwag compared to HF leader board, so I'm not expecting we will get up to 78.4 anyway. It looks like it is better to skip the choice word(s) when evaluating the average log-likelihood. This kind of makes sense because a more common word (in Winogrande this is often a name) will have a higher probability without knowing about the follow up context, and this will skew the log-likelihood towards the more common word. We can only do this if the choice words are not last in the sentence. It also looks like it is better to skip the punctuation at the end of the sentence, provided the choice words are not last. * winogrande: add dataset instructions --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 10 ++ common/common.h | 3 + examples/perplexity/perplexity.cpp | 243 ++++++++++++++++++++++++++++- 3 files changed, 251 insertions(+), 5 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 2b0865fff..ce20360a4 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -681,6 +681,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.hellaswag_tasks = std::stoi(argv[i]); + } else if (arg == "--winogrande") { + params.winogrande = true; + } else if (arg == "--winogrande-tasks") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.winogrande_tasks = std::stoi(argv[i]); } else if (arg == "--ignore-eos") { params.ignore_eos = true; } else if (arg == "--no-penalize-nl") { @@ -926,6 +934,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --logits-all return logits for all tokens in the batch (default: disabled)\n"); printf(" --hellaswag compute HellaSwag score over random tasks from datafile supplied with -f\n"); printf(" --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks); + printf(" --winogrande compute Winogrande score over random tasks from datafile supplied with -f\n"); + printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); diff --git a/common/common.h b/common/common.h index 1f43e6282..0ae9c18b3 100644 --- a/common/common.h +++ b/common/common.h @@ -105,6 +105,9 @@ struct gpt_params { bool hellaswag = false; // compute HellaSwag score over random tasks from datafile supplied in prompt size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score + bool winogrande = false; // compute Winogrande score over random tasks from datafile supplied in prompt + size_t winogrande_tasks= 0; // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b4fedf803..57eaa713e 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -9,6 +9,9 @@ #include #include #include +#include +#include +#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -419,9 +422,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par return {tokens, ppl, logit_history, prob_history}; } -static std::vector hellaswag_evaluate_tokens( - llama_context * ctx, std::vector & tokens, int n_past, int n_batch, int n_vocab -) { +static std::vector evaluate_tokens(llama_context * ctx, std::vector & tokens, + int n_past, int n_batch, int n_vocab) { std::vector result; result.reserve(tokens.size() * n_vocab); size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch; @@ -573,7 +575,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // clear the KV cache llama_kv_cache_clear(ctx); - auto logits = hellaswag_evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab); + auto logits = evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab); if (logits.empty()) { fprintf(stderr, "%s : failed to eval\n", __func__); return; @@ -622,7 +624,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { //} // Evaluate the query - logits = hellaswag_evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab); + logits = evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab); if (logits.empty()) { fprintf(stderr, "%s : failed to eval\n", __func__); return; @@ -676,6 +678,235 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { printf("\n"); } +struct winogrande_entry { + std::string first; + std::string second; + std::array choices; + int answer; +}; + +static std::vector load_winogrande_from_csv(const std::string& prompt) { + std::vector result; + std::istringstream in(prompt); + std::string line; + std::array comma_pos; + while (true) { + std::getline(in, line); + if (in.fail() || in.eof()) break; + int ipos = 0; + bool quote_open = false; + for (int i = 0; i < int(line.size()); ++i) { + if (!quote_open) { + if (line[i] == ',') { + comma_pos[ipos++] = i; + if (ipos == 4) break; + } + else if (line[i] == '"') { + quote_open = true; + } + } + else { + if (line[i] == '"') { + quote_open = false; + } + } + } + if (ipos != 4) { + printf("%s: failed to find comma separators in <%s>\n", __func__, line.c_str()); + continue; + } + auto sentence = line[comma_pos[0]+1] == '"' ? line.substr(comma_pos[0]+2, comma_pos[1] - comma_pos[0] - 3) + : line.substr(comma_pos[0]+1, comma_pos[1] - comma_pos[0] - 1); + auto choice1 = line.substr(comma_pos[1]+1, comma_pos[2] - comma_pos[1] - 1); + auto choice2 = line.substr(comma_pos[2]+1, comma_pos[3] - comma_pos[2] - 1); + auto answer = line.substr(comma_pos[3]+1, line.size() - comma_pos[3] - 1); + auto index = line.substr(0, comma_pos[0]); + int where = 0; + for ( ; where < int(sentence.size()); ++where) { + if (sentence[where] == '_') break; + } + if (where == int(sentence.size())) { + printf("%s: no _ in <%s>\n", __func__, sentence.c_str()); + continue; + } + std::istringstream stream(answer.c_str()); + int i_answer; stream >> i_answer; + if (stream.fail() || i_answer < 1 || i_answer > 2) { + printf("%s: failed to parse answer <%s>\n", __func__, answer.c_str()); + continue; + } + result.emplace_back(); + auto& wg = result.back(); + wg.first = sentence.substr(0, where); + wg.second = sentence.substr(where + 1, sentence.size() - where - 1); + wg.choices[0] = std::move(choice1); + wg.choices[1] = std::move(choice2); + wg.answer = i_answer; + } + return result; +} + +/* + * Evaluates the Winogrande score. + * Uses a CSV containing task index, dentence, choice 1, choice 2, answer (1 or 2) + * You can get one such dataset from e.g. https://huggingface.co/datasets/ikawrakow/winogrande-eval-for-llama.cpp + * As an example, the 1st row in the above dataset is + * + * 0,Sarah was a much better surgeon than Maria so _ always got the easier cases.,Sarah,Maria,2 + * + */ +static void winogrande_score(llama_context * ctx, const gpt_params & params) { + + constexpr int k_min_trailing_ctx = 3; + + auto data = load_winogrande_from_csv(params.prompt); + if (data.empty()) { + fprintf(stderr, "%s: no tasks\n", __func__); + return; + } + + fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, data.size()); + + if (params.winogrande_tasks > 0 && params.winogrande_tasks < data.size()) { + fprintf(stderr, "%s : selecting %zu random tasks\n", __func__, params.winogrande_tasks); + std::mt19937 rng(1); + std::vector aux(data.size()); + for (int i = 0; i < int(data.size()); ++i) { + aux[i] = i; + } + float scale = 1/(1.f + (float)rng.max()); + std::vector selected; + selected.reserve(params.winogrande_tasks); + for (int i = 0; i < int(params.winogrande_tasks); ++i) { + int j = int(scale*rng()*aux.size()); + selected[i] = std::move(data[aux[j]]); + aux[j] = aux.back(); + aux.pop_back(); + } + data = std::move(selected); + } + + // This is needed as usual for LLaMA models + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + + fprintf(stderr, "%s : calculating winogrande score over selected tasks.\n", __func__); + + const int n_vocab = llama_n_vocab(llama_get_model(ctx)); + const int n_ctx = llama_n_ctx(ctx); + + std::vector tok_logits(n_vocab); + + int n_correct = 0; + int n_done = 0; + + for (size_t task_idx = 0; task_idx < data.size(); task_idx++) { + const auto& task = data[task_idx]; + + auto base_context = ::llama_tokenize(ctx, task.first, add_bos); + auto base_ctx_1st = ::llama_tokenize(ctx, task.first + task.choices[0], add_bos); + auto base_ctx_2nd = ::llama_tokenize(ctx, task.first + task.choices[1], add_bos); + + auto sentence_1st = task.first + task.choices[0] + task.second; + auto sentence_2nd = task.first + task.choices[1] + task.second; + auto query_1st = ::llama_tokenize(ctx, sentence_1st, add_bos); + auto query_2nd = ::llama_tokenize(ctx, sentence_2nd, add_bos); + + if (query_1st.size() > (size_t)n_ctx || query_2nd.size() > (size_t)n_ctx) { + fprintf(stderr, "%s : number of tokens in queries %zu, %zu > n_ctxl\n", __func__, query_1st.size(), query_2nd.size()); + return; + } + + auto query_1st_size = query_1st.size(); + auto query_2nd_size = query_2nd.size(); + + // Speedup small evaluations by evaluating atleast 32 tokens + // For Winogrande this seems to slow it down rather than speed it up. + //if (query_1st.size() < 32) query_1st.resize(32); + //if (query_2nd.size() < 32) query_2nd.resize(32); + + llama_kv_cache_clear(ctx); + auto logits_1st = evaluate_tokens(ctx, query_1st, 0, params.n_batch, n_vocab); + + llama_kv_cache_clear(ctx); + auto logits_2nd = evaluate_tokens(ctx, query_2nd, 0, params.n_batch, n_vocab); + + if (logits_1st.empty() || logits_2nd.empty()) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return; + } + + bool skip_choice = query_1st_size - base_ctx_1st.size() > k_min_trailing_ctx && + query_2nd_size - base_ctx_2nd.size() > k_min_trailing_ctx; + + float score_1st = 0; + bool is_nan_1st = false; + const auto& base_1 = skip_choice ? base_ctx_1st : base_context; + const int last_1st = query_1st_size - base_1.size() > 1 ? 1 : 0; + for (size_t j = base_1.size()-1; j < query_1st_size-1-last_1st; ++j) { + std::memcpy(tok_logits.data(), logits_1st.data() + j*n_vocab, n_vocab*sizeof(float)); + const float prob = softmax(tok_logits)[query_1st[j+1]]; + if (std::isnan(prob) || !prob) { + fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, + prob, j, sentence_1st.c_str(), base_context.size()); + is_nan_1st = true; + break; + } + score_1st += std::log(prob); + } + score_1st /= (query_1st_size - base_1.size() - last_1st); + + float score_2nd = 0; + bool is_nan_2nd = false; + const auto& base_2 = skip_choice ? base_ctx_2nd : base_context; + const int last_2nd = query_2nd_size - base_2.size() > 1 ? 1 : 0; + for (size_t j = base_2.size()-1; j < query_2nd_size-1-last_2nd; ++j) { + std::memcpy(tok_logits.data(), logits_2nd.data() + j*n_vocab, n_vocab*sizeof(float)); + const float prob = softmax(tok_logits)[query_2nd[j+1]]; + if (std::isnan(prob) || !prob) { + fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, + prob, j, sentence_2nd.c_str(), base_context.size()); + is_nan_2nd = true; + break; + } + score_2nd += std::log(prob); + } + score_2nd /= (query_2nd_size - base_2.size() - last_2nd); + + if (is_nan_1st || is_nan_2nd) { + continue; + } + + if (std::isnan(score_1st) || std::isnan(score_2nd)) { + printf("================== NaN score %g, %g) for:\n", score_1st, score_2nd); + printf("Q1: <%s> - %zu tokens\n", sentence_1st.c_str(), query_1st_size); + printf("Q2: <%s> - %zu tokens\n", sentence_2nd.c_str(), query_2nd_size); + printf("B : <%s> - %zu tokens\n", task.first.c_str(), base_context.size()); + printf("base_1 has %zu tokens, base_2 has %zu tokens, skip_choice = %d\n", base_1.size(), base_2.size(), skip_choice); + continue; + } + + int result = score_1st > score_2nd ? 1 : 2; + + if (result == task.answer) { + ++n_correct; + } + ++n_done; + + // Print the accumulated accuracy mean x 100 + printf("%zu\t%.4lf\t%10.6f %10.6f %d %d\n",task_idx+1, 100.0 * n_correct/n_done,score_1st,score_2nd,result,task.answer); + fflush(stdout); + } + + printf("\n"); + + if (n_done < 100) return; + + const float p = 1.f*n_correct/n_done; + const float sigma = 100.f*sqrt(p*(1-p)/(n_done-1)); + printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma); +} + + int main(int argc, char ** argv) { gpt_params params; @@ -733,6 +964,8 @@ int main(int argc, char ** argv) { struct results_perplexity results; if (params.hellaswag) { hellaswag_score(ctx, params); + } else if (params.winogrande) { + winogrande_score(ctx, params); } else { results = perplexity(ctx, params); } From ad19812cda4062c9f154ef16315df41fbe6a770a Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 15:33:01 +0200 Subject: [PATCH 014/101] perplexity : faster HellaSwag via batching (#5017) * perplexity : faster HellaSwag ggml-ci * perplexity : clean-up ggml-ci * perplexity : no need for decode_helper ggml-ci * perplexity : add comments * perplexity : option to specify max batched tasks via `n_parallel` * perplexity : remove HellaSwag restruction for n_batch --- examples/perplexity/perplexity.cpp | 261 ++++++++++++++++------------- 1 file changed, 149 insertions(+), 112 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 57eaa713e..ea2c8026c 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -470,7 +470,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { prompt_lines.push_back(line); } - if( prompt_lines.size() % 6 != 0) { + if (prompt_lines.size() % 6 != 0) { fprintf(stderr, "%s : number of lines in prompt not a multiple of 6.\n", __func__); return; } @@ -485,7 +485,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); // Number of tasks to use when computing the score - if ( params.hellaswag_tasks < hs_task_count ) { + if (params.hellaswag_tasks < hs_task_count) { hs_task_count = params.hellaswag_tasks; } @@ -502,27 +502,54 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { std::string ending[4]; size_t ending_logprob_count[4]; double ending_logprob[4]; + + size_t i_batch; // starting index in the llama_batch + size_t common_prefix; // max number of initial tokens that are the same in all sentences + size_t required_tokens; // needed number of tokens to evaluate all 4 endings + std::vector seq_tokens[4]; }; fprintf(stderr, "%s : selecting %zu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first") ); // Select and read data from prompt lines - hs_data_t *hs_data = new hs_data_t[hs_task_count]; - for (size_t i=0; i < hs_task_count; i++) { + std::vector hs_data(hs_task_count); + for (size_t i = 0; i < hs_task_count; i++) { size_t idx = i; + auto & hs_cur = hs_data[i]; + // Select a random example of those left in the prompt if (randomize_tasks) { std::uniform_int_distribution dist(0, prompt_lines.size()/6-1 ) ; idx = dist(rng); } - hs_data[i].context = prompt_lines[idx*6]; - hs_data[i].gold_ending_idx = std::stoi( prompt_lines[idx*6+1] ); - for (size_t j=0; j < 4; j++) { - hs_data[i].ending[j] = prompt_lines[idx*6+2+j]; + hs_cur.context = prompt_lines[idx*6]; + hs_cur.gold_ending_idx = std::stoi( prompt_lines[idx*6+1] ); + for (size_t j = 0; j < 4; j++) { + hs_cur.ending[j] = prompt_lines[idx*6+2+j]; + hs_cur.seq_tokens[j] = ::llama_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], add_bos); } + // determine the common prefix of the endings + hs_cur.common_prefix = 0; + hs_cur.required_tokens = 0; + for (size_t k = 0; k < hs_cur.seq_tokens[0].size(); k++) { + if (hs_cur.seq_tokens[0][k] != hs_cur.seq_tokens[1][k] || + hs_cur.seq_tokens[0][k] != hs_cur.seq_tokens[2][k] || + hs_cur.seq_tokens[0][k] != hs_cur.seq_tokens[3][k]) { + break; + } + hs_cur.common_prefix++; + } + hs_cur.required_tokens = hs_cur.common_prefix + + hs_cur.seq_tokens[0].size() - hs_cur.common_prefix + + hs_cur.seq_tokens[1].size() - hs_cur.common_prefix + + hs_cur.seq_tokens[2].size() - hs_cur.common_prefix + + hs_cur.seq_tokens[3].size() - hs_cur.common_prefix; + + //GGML_ASSERT(hs_cur.common_prefix >= ::llama_tokenize(ctx, hs_cur.context, add_bos).size()); + // Delete the selected random example from the prompt if (randomize_tasks) { prompt_lines.erase( std::next(prompt_lines.begin(),idx*6) , std::next(prompt_lines.begin(),idx*6+6) ); @@ -530,150 +557,160 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { } fprintf(stderr, "%s : calculating hellaswag score over selected tasks.\n", __func__); + printf("\ntask\tacc_norm\n"); double acc = 0.0f; - const int n_vocab = llama_n_vocab(llama_get_model(ctx)); - const int n_ctx = llama_n_ctx(ctx); - std::vector> ending_tokens(4); + const int n_vocab = llama_n_vocab(llama_get_model(ctx)); + const int n_ctx = llama_n_ctx(ctx); + const int n_batch = params.n_batch; + + const int max_tasks_per_batch = params.n_parallel; + const int max_seq = 4*max_tasks_per_batch; + + llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); std::vector tok_logits(n_vocab); + std::vector batch_logits(n_ctx*n_vocab); - for (size_t task_idx = 0; task_idx < hs_task_count; task_idx++) { - // Tokenize the context to count tokens - std::vector context_embd = ::llama_tokenize(ctx, hs_data[task_idx].context, add_bos); - size_t context_size = context_embd.size(); + auto decode_helper = [&](llama_context * ctx, llama_batch & batch, int32_t n_batch) { + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { + const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); - for (int i = 0; i < 4; ++i) { - ending_tokens[i] = ::llama_tokenize(ctx, hs_data[task_idx].context + " " + hs_data[task_idx].ending[i], add_bos); - for (int k = 0; k < int(context_size); ++k) { - if (ending_tokens[i][k] != context_embd[k]) { - fprintf(stderr, "Oops: ending %d of task %d differs from context at position %d\n",i,int(task_idx),k); - break; + llama_batch batch_view = { + n_tokens, + batch.token + i, + nullptr, + batch.pos + i, + batch.n_seq_id + i, + batch.seq_id + i, + batch.logits + i, + 0, 0, 0, // unused + }; + + const int ret = llama_decode(ctx, batch_view); + if (ret != 0) { + LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret); + return false; + } + + memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float)); + } + + return true; + }; + + for (size_t i0 = 0; i0 < hs_task_count; i0++) { + int n_cur = 0; + + size_t i1 = i0; + size_t i_batch = 0; // this tells us where in `llama_batch` we are currently + + llama_batch_clear(batch); + + // batch as much tasks as possible into the available context + // each task has 4 unique seuqnce ids - one for each ending + // the common prefix is shared among the 4 sequences to save tokens + // we extract logits only from the last common token and from all ending tokens of each sequence + while (n_cur + (int) hs_data[i1].required_tokens <= n_ctx) { + auto & hs_cur = hs_data[i1]; + + const int s0 = 4*(i1 - i0); + if (s0 + 4 > max_seq) { + break; + } + + for (size_t i = 0; i < hs_cur.common_prefix; ++i) { + llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false); + } + batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix + + for (int s = 0; s < 4; ++s) { + for (size_t i = hs_cur.common_prefix; i < hs_cur.seq_tokens[s].size(); ++i) { + llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, true); } } + + hs_cur.i_batch = i_batch; + i_batch += hs_cur.required_tokens; + + n_cur += hs_data[i1].required_tokens; + if (++i1 == hs_task_count) { + break; + } } - // Do the 1st ending - // In this case we include the context when evaluating - //auto query_embd = ::llama_tokenize(ctx, hs_data[task_idx].context + hs_data[task_idx].ending[0], add_bos); - auto query_embd = ending_tokens[0]; - auto query_size = query_embd.size(); - - // Stop if query wont fit the ctx window - if (query_size > (size_t)n_ctx) { - fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size); + if (i0 == i1) { + fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0); return; } - // Speedup small evaluations by evaluating atleast 32 tokens - if (query_size < 32) { - query_embd.resize(32); - } - - // clear the KV cache llama_kv_cache_clear(ctx); - auto logits = evaluate_tokens(ctx, query_embd, 0, params.n_batch, n_vocab); - if (logits.empty()) { - fprintf(stderr, "%s : failed to eval\n", __func__); + // decode all tasks [i0, i1) + if (!decode_helper(ctx, batch, n_batch)) { + fprintf(stderr, "%s: llama_decode() failed\n", __func__); return; } - std::memcpy(tok_logits.data(), logits.data() + (context_size-1)*n_vocab, n_vocab*sizeof(float)); - const auto first_probs = softmax(tok_logits); + // compute the logprobs for each ending of the decoded tasks + for (size_t i = i0; i < i1; ++i) { + auto & hs_cur = hs_data[i]; - hs_data[task_idx].ending_logprob_count[0] = 1; - hs_data[task_idx].ending_logprob[0] = std::log(first_probs[query_embd[context_size]]); + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + hs_cur.common_prefix - 1), n_vocab*sizeof(float)); - // Calculate the logprobs over the ending - for (size_t j = context_size; j < query_size - 1; j++) { + const auto first_probs = softmax(tok_logits); - std::memcpy(tok_logits.data(), logits.data() + j*n_vocab, n_vocab*sizeof(float)); + size_t li = hs_cur.common_prefix; // logits index in the batch - const float prob = softmax(tok_logits)[query_embd[j + 1]]; + for (int s = 0; s < 4; ++s) { + hs_cur.ending_logprob_count[s] = 1; + hs_cur.ending_logprob[s] = std::log(first_probs[hs_cur.seq_tokens[s][hs_cur.common_prefix]]); - hs_data[task_idx].ending_logprob[0] += std::log(prob); - hs_data[task_idx].ending_logprob_count[0]++; - } + // Calculate the logprobs over the ending + for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) { + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + li++), n_vocab*sizeof(float)); - // Calculate the mean token logprob for acc_norm - hs_data[task_idx].ending_logprob[0] /= hs_data[task_idx].ending_logprob_count[0]; + const float prob = softmax(tok_logits)[hs_cur.seq_tokens[s][j + 1]]; - // Do the remaining endings - // For these, we use the bare ending with n_past = context_size - // - for (size_t ending_idx = 1; ending_idx < 4; ending_idx++) { + hs_cur.ending_logprob[s] += std::log(prob); + hs_cur.ending_logprob_count[s]++; + } - // Tokenize the query - query_embd.resize(ending_tokens[ending_idx].size() - context_size); - std::memcpy(query_embd.data(), ending_tokens[ending_idx].data() + context_size, query_embd.size()*sizeof(int)); - query_size = query_embd.size(); + // account that we skip the last token in the ending + ++li; - // Stop if query wont fit the ctx window - if (context_size + query_size > (size_t)n_ctx) { - fprintf(stderr, "%s : number of tokens in query %zu > n_ctxl\n", __func__, query_size); - return; + // Calculate the mean token logprob for acc_norm + hs_cur.ending_logprob[s] /= hs_cur.ending_logprob_count[s]; } - // Speedup small evaluations by evaluating atleast 32 tokens - // No, resizing to 32 is actually slightly slower (at least on CUDA) - //if (query_size < 32) { - // query_embd.resize(32); - //} - - // Evaluate the query - logits = evaluate_tokens(ctx, query_embd, context_size, params.n_batch, n_vocab); - if (logits.empty()) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return; + // Find the ending with maximum logprob + size_t ending_logprob_max_idx = 0; + double ending_logprob_max_val = hs_cur.ending_logprob[0]; + for (size_t s = 1; s < 4; s++) { + if (hs_cur.ending_logprob[s] > ending_logprob_max_val) { + ending_logprob_max_idx = s; + ending_logprob_max_val = hs_cur.ending_logprob[s]; + } } - hs_data[task_idx].ending_logprob_count[ending_idx] = 1; - hs_data[task_idx].ending_logprob[ending_idx] = std::log(first_probs[query_embd[0]]); + //printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_cur.gold_ending_idx); - // Calculate the logprobs over the ending - for (size_t j = 0; j < query_size - 1; j++) { - std::memcpy(tok_logits.data(), logits.data() + j*n_vocab, n_vocab*sizeof(float)); - - const float prob = softmax(tok_logits)[query_embd[j + 1]]; - - hs_data[task_idx].ending_logprob[ending_idx] += std::log(prob); - hs_data[task_idx].ending_logprob_count[ending_idx]++; + // If the gold ending got the maximum logprobe add one accuracy point + if (ending_logprob_max_idx == hs_cur.gold_ending_idx) { + acc += 1.0; } - // Calculate the mean token logprob for acc_norm - hs_data[task_idx].ending_logprob[ending_idx] /= hs_data[task_idx].ending_logprob_count[ending_idx]; - - -// printf("task %lu, ending %lu, whole_len %lu, context_len %lu, ending_logprob_count %lu, ending_logprob %.4f\n", -// task_idx,ending_idx,whole_size,context_size, hs_data[task_idx].ending_logprob_count[ending_idx], hs_data[task_idx].ending_logprob[ending_idx] ); + // Print the accumulated accuracy mean x 100 + printf("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0); + fflush(stdout); } - // Find the ending with maximum logprob - size_t ending_logprob_max_idx = 0; - double ending_logprob_max_val = hs_data[task_idx].ending_logprob[0]; - for (size_t j = 1; j < 4; j++) { - if (hs_data[task_idx].ending_logprob[j] > ending_logprob_max_val) { - ending_logprob_max_idx = j; - ending_logprob_max_val = hs_data[task_idx].ending_logprob[j]; - } - } - -// printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_data[task_idx].gold_ending_idx); - - // If the gold ending got the maximum logprobe add one accuracy point - if (ending_logprob_max_idx == hs_data[task_idx].gold_ending_idx) { - acc += 1.0; - } - - // Print the accumulated accuracy mean x 100 - printf("%zu\t%.8lf\n",task_idx+1, acc/double(task_idx+1)*100.0); - fflush(stdout); + i0 = i1 - 1; } - delete [] hs_data; + llama_batch_free(batch); printf("\n"); } From 3e945cc1e9c06d2001031360e4e303e9548fb02c Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 18 Jan 2024 19:18:21 +0200 Subject: [PATCH 015/101] HellaSwag: speed up by parallelizing log-prob evaluation (#5020) For Mistral-7B and fp16, time on my system goes down from 536 seconds to 423 seconds for the full evaluation dataset (10042 tasks). Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 80 ++++++++++++++++++++++++------ 1 file changed, 66 insertions(+), 14 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index ea2c8026c..9498dd535 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -8,6 +8,7 @@ #include #include #include +#include #include #include #include @@ -444,6 +445,48 @@ static std::vector evaluate_tokens(llama_context * ctx, std::vector return result; } +static void hellaswag_compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, + const std::vector>& eval_pairs, std::vector& eval_results) { + constexpr int k_token_chunk = 4; + if (eval_results.size() != eval_pairs.size()) { + eval_results.resize(eval_pairs.size()); + } + if (eval_pairs.empty()) return; + + size_t max_threads = std::min((eval_pairs.size() + k_token_chunk - 1)/k_token_chunk, workers.size()); + + std::atomic counter(0); + auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () { + float local_logprobs[k_token_chunk]; + while (true) { + size_t first = counter.fetch_add(k_token_chunk, std::memory_order_relaxed); + if (first >= eval_results.size()) break; + size_t last = std::min(first + k_token_chunk, eval_results.size()); + for (size_t i = first; i < last; ++i) { + auto logits = batch_logits + eval_pairs[i].first * n_vocab; + float max_logit = logits[0]; + for (int j = 1; j < n_vocab; ++j) { + max_logit = std::max(max_logit, logits[j]); + } + float sum_p = 0.f; + for (int j = 0; j < n_vocab; ++j) { + sum_p += expf(logits[j] - max_logit); + } + local_logprobs[i - first] = logits[eval_pairs[i].second] - max_logit - std::log(sum_p); + } + std::memcpy(eval_results.data() + first, local_logprobs, (last - first)*sizeof(float)); + } + }; + + for (size_t it = 0; it < max_threads; ++it) { + workers[it] = std::thread(compute); + } + for (size_t it = 0; it < max_threads; ++it) { + workers[it].join(); + } + +} + static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // Calculates hellaswag score (acc_norm) from prompt // @@ -574,6 +617,10 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { std::vector tok_logits(n_vocab); std::vector batch_logits(n_ctx*n_vocab); + std::vector> eval_pairs; + std::vector eval_results; + std::vector workers(std::thread::hardware_concurrency()); + auto decode_helper = [&](llama_context * ctx, llama_batch & batch, int32_t n_batch) { for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); @@ -654,6 +701,24 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { return; } + // Compute log-probs in parallel + // First we collect all tasks + eval_pairs.clear(); + for (size_t i = i0; i < i1; ++i) { + auto & hs_cur = hs_data[i]; + size_t li = hs_cur.common_prefix; + for (int s = 0; s < 4; ++s) { + for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) { + eval_pairs.push_back(std::make_pair(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1])); + } + ++li; + } + } + // Then we do the actual calculation + hellaswag_compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); + + size_t ir = 0; + // compute the logprobs for each ending of the decoded tasks for (size_t i = i0; i < i1; ++i) { auto & hs_cur = hs_data[i]; @@ -662,26 +727,13 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { const auto first_probs = softmax(tok_logits); - size_t li = hs_cur.common_prefix; // logits index in the batch - for (int s = 0; s < 4; ++s) { hs_cur.ending_logprob_count[s] = 1; hs_cur.ending_logprob[s] = std::log(first_probs[hs_cur.seq_tokens[s][hs_cur.common_prefix]]); - - // Calculate the logprobs over the ending for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) { - std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + li++), n_vocab*sizeof(float)); - - const float prob = softmax(tok_logits)[hs_cur.seq_tokens[s][j + 1]]; - - hs_cur.ending_logprob[s] += std::log(prob); + hs_cur.ending_logprob[s] += eval_results[ir++]; hs_cur.ending_logprob_count[s]++; } - - // account that we skip the last token in the ending - ++li; - - // Calculate the mean token logprob for acc_norm hs_cur.ending_logprob[s] /= hs_cur.ending_logprob_count[s]; } From b46757735d30f5c6ed4f20ebeccc684e02d4f3bf Mon Sep 17 00:00:00 2001 From: David Sommers <12738+databyte@users.noreply.github.com> Date: Thu, 18 Jan 2024 12:20:59 -0500 Subject: [PATCH 016/101] convert.py : fix llama/llama2 conversion due to vocab_size=-1 (#5019) PR #4818 (merged last week) reintroduced a config check for vocab_size that was addressed in PR #4258 (merged 2023-11-30). Without the fix, llama2 models can't be converted. The error is: `ValueError: The model's vocab size is set to -1 in params.json. Please update it manually. Maybe 32000?` --- convert.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/convert.py b/convert.py index e38ee5315..980e6fc72 100755 --- a/convert.py +++ b/convert.py @@ -348,7 +348,7 @@ class Params: f_rope_freq_base = 1e6 return Params( - n_vocab=config.get("vocab_size", model["tok_embeddings.weight"].shape[0]), + n_vocab=model["tok_embeddings.weight"].shape[0], n_embd=config["dim"], n_layer=config["n_layers"], n_ctx=n_ctx, From e9240cdfa06a50c1b5dbafa367cb8cd698e65103 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 20:45:39 +0200 Subject: [PATCH 017/101] scripts : add get-winogrande.sh --- scripts/get-hellaswag.sh | 2 +- scripts/get-wikitext-2.sh | 7 +++++++ scripts/get-winogrande.sh | 10 ++++++++++ 3 files changed, 18 insertions(+), 1 deletion(-) create mode 100755 scripts/get-winogrande.sh diff --git a/scripts/get-hellaswag.sh b/scripts/get-hellaswag.sh index ef8dcceb0..121979fe2 100755 --- a/scripts/get-hellaswag.sh +++ b/scripts/get-hellaswag.sh @@ -4,7 +4,7 @@ wget https://raw.githubusercontent.com/klosax/hellaswag_text_data/main/hellaswag echo "Usage:" echo "" -echo " ./perplexity --hellaswag --hellaswag-tasks N -f hellaswag_val_full.txt -m modelfile.gguf" +echo " ./perplexity -m model.gguf -f hellaswag_val_full.txt --hellaswag [--hellaswag-tasks N] [other params]" echo "" exit 0 diff --git a/scripts/get-wikitext-2.sh b/scripts/get-wikitext-2.sh index 98aec3e3e..ff96f331e 100755 --- a/scripts/get-wikitext-2.sh +++ b/scripts/get-wikitext-2.sh @@ -1,3 +1,10 @@ #!/bin/bash wget https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip + +echo "Usage:" +echo "" +echo " ./perplexity -m model.gguf -f wiki.test.raw [other params]" +echo "" + +exit 0 diff --git a/scripts/get-winogrande.sh b/scripts/get-winogrande.sh new file mode 100755 index 000000000..5f234468e --- /dev/null +++ b/scripts/get-winogrande.sh @@ -0,0 +1,10 @@ +#!/bin/bash + +wget https://huggingface.co/datasets/ikawrakow/winogrande-eval-for-llama.cpp/raw/main/winogrande-debiased-eval.csv + +echo "Usage:" +echo "" +echo " ./perplexity -m model.gguf -f winogrande-debiased-eval.csv --winogrande [--winogrande-tasks N] [other params]" +echo "" + +exit 0 From d391ae9b4919e24624cc963d82162450848beaf4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 20:49:00 +0200 Subject: [PATCH 018/101] perplexity : fix winogrande N tasks option --- examples/perplexity/perplexity.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9498dd535..f72ea6d1c 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -865,7 +865,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { } float scale = 1/(1.f + (float)rng.max()); std::vector selected; - selected.reserve(params.winogrande_tasks); + selected.resize(params.winogrande_tasks); for (int i = 0; i < int(params.winogrande_tasks); ++i) { int j = int(scale*rng()*aux.size()); selected[i] = std::move(data[aux[j]]); From 2d5419d08ab1131623e6a1d554607b7663435e87 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 21:45:51 +0200 Subject: [PATCH 019/101] imatrix : fix assert for src0 non-cont check --- examples/imatrix/imatrix.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index af78711c5..5a3d30b88 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -80,7 +80,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * // for simplicity, always copy src0 to host, because it is small // take into account that src0 is not contiguous! GGML_ASSERT(src0->ne[1] == src1->ne[1]); - GGML_ASSERT(n_as*ggml_nrows(src0)); + GGML_ASSERT(n_as*ggml_nrows(src0)*sizeof(int) == GGML_PAD(ggml_nbytes(src0), n_as*sizeof(int))); m_ids.resize(ggml_nbytes(src0)/sizeof(int)); ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0)); From 96d7f56d2918ffde1995dbb32392571deb76d7fc Mon Sep 17 00:00:00 2001 From: slaren Date: Thu, 18 Jan 2024 21:12:15 +0100 Subject: [PATCH 020/101] llama : fix mlock with no-mmap with Metal (#5025) --- llama.cpp | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/llama.cpp b/llama.cpp index d28382f7d..f1d00a96c 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1599,7 +1599,7 @@ struct llama_model { std::unique_ptr mapping; // objects representing data potentially being locked in memory - llama_mlock mlock_buf; + std::vector> mlock_bufs; llama_mlock mlock_mmap; // for quantize-stats only @@ -3815,8 +3815,10 @@ static bool llm_load_tensors( else { buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (buf != nullptr && use_mlock && ggml_backend_buffer_is_host(buf)) { - model.mlock_buf.init (ggml_backend_buffer_get_base(buf)); - model.mlock_buf.grow_to(ggml_backend_buffer_get_size(buf)); + model.mlock_bufs.emplace_back(new llama_mlock); + auto & mlock_buf = model.mlock_bufs.back(); + mlock_buf->init (ggml_backend_buffer_get_base(buf)); + mlock_buf->grow_to(ggml_backend_buffer_get_size(buf)); } } if (buf == nullptr) { From 821f0a271e7c9ee737945245dd7abfa22cc9b5b0 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Thu, 18 Jan 2024 21:33:05 +0100 Subject: [PATCH 021/101] server : defer tasks when "slot unavailable" (#5018) * server: defer task when no slot is available * remove unnecessary log --------- Co-authored-by: Xuan Son Nguyen --- examples/server/server.cpp | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 93f999298..0462fbd24 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1558,6 +1558,7 @@ struct llama_server_context void process_tasks() { std::unique_lock lock(mutex_tasks); + std::vector deferred_tasks; while (!queue_tasks.empty()) { task_server task = queue_tasks.front(); @@ -1568,9 +1569,8 @@ struct llama_server_context llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); if (slot == nullptr) { - LOG_TEE("slot unavailable\n"); - // send error result - send_error(task, "slot unavailable"); + // if no slot is available, we defer this task for processing later + deferred_tasks.push_back(task); break; } @@ -1616,6 +1616,12 @@ struct llama_server_context } } + // add all the deferred tasks back the the queue + for (task_server &task : deferred_tasks) + { + queue_tasks.push_back(task); + } + // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue std::vector agg_results; auto queue_iterator = queue_multitasks.begin(); From 9b6ea4263ab45e02ff905bf7a29dc143ca1facc3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 18 Jan 2024 23:36:07 +0200 Subject: [PATCH 022/101] cmake : add ggml public headers (#5011) --- CMakeLists.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 7bd640966..3fc65eaf2 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -846,7 +846,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.cmake DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama) -set(GGML_PUBLIC_HEADERS "ggml.h" +set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h" "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}" "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}") From 57e2a7a52a819883f40dada8a2edc24ecf48186b Mon Sep 17 00:00:00 2001 From: John <78893154+cmp-nct@users.noreply.github.com> Date: Thu, 18 Jan 2024 23:12:15 +0100 Subject: [PATCH 023/101] llama : fix falcon arch for tied output embeddings (#4978) * falcon arch fix for tied output embeddings * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp --------- Co-authored-by: Georgi Gerganov --- llama.cpp | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index f1d00a96c..47b4384a8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3438,7 +3438,12 @@ static bool llm_load_tensors( { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + if (gguf_find_tensor(ml.ctx_gguf, tn(LLM_TENSOR_OUTPUT, "weight").c_str()) >= 0) { + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } else { + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU + ml.n_created--; // artificial tensor + } } for (int i = 0; i < n_layer; ++i) { From 8b20858e5e9c44b99b4b31ae9c40b8f20d01d94f Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 Jan 2024 10:45:06 +0200 Subject: [PATCH 024/101] perplexity : faster Winogrande via batching (#5024) * perplexity : faster Winogrande via batching ggml-ci * perplexity : remove unused function * perplexity : only tokenize selected tasks for Winogrande --- examples/perplexity/perplexity.cpp | 287 ++++++++++++++++------------- 1 file changed, 160 insertions(+), 127 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index f72ea6d1c..df902fb1c 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -423,26 +423,31 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par return {tokens, ppl, logit_history, prob_history}; } -static std::vector evaluate_tokens(llama_context * ctx, std::vector & tokens, - int n_past, int n_batch, int n_vocab) { - std::vector result; - result.reserve(tokens.size() * n_vocab); - size_t n_chunk = (tokens.size() + n_batch - 1)/n_batch; - for (size_t i_chunk = 0; i_chunk < n_chunk; ++i_chunk) { - size_t n_tokens = tokens.size() - i_chunk * n_batch; - n_tokens = std::min(n_tokens, size_t(n_batch)); - llama_kv_cache_seq_rm(ctx, 0, n_past, -1); - if (llama_decode(ctx, llama_batch_get_one(tokens.data() + i_chunk * n_batch, n_tokens, n_past, 0))) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return {}; +static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector & batch_logits, int32_t n_batch, int32_t n_vocab) { + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { + const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); + + llama_batch batch_view = { + n_tokens, + batch.token + i, + nullptr, + batch.pos + i, + batch.n_seq_id + i, + batch.seq_id + i, + batch.logits + i, + 0, 0, 0, // unused + }; + + const int ret = llama_decode(ctx, batch_view); + if (ret != 0) { + LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret); + return false; } - const auto logits = llama_get_logits(ctx); - result.insert(result.end(), logits, logits + n_tokens * n_vocab); - - n_past += n_tokens; + memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float)); } - return result; + + return true; } static void hellaswag_compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, @@ -576,7 +581,6 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // determine the common prefix of the endings hs_cur.common_prefix = 0; - hs_cur.required_tokens = 0; for (size_t k = 0; k < hs_cur.seq_tokens[0].size(); k++) { if (hs_cur.seq_tokens[0][k] != hs_cur.seq_tokens[1][k] || hs_cur.seq_tokens[0][k] != hs_cur.seq_tokens[2][k] || @@ -609,45 +613,18 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { const int n_ctx = llama_n_ctx(ctx); const int n_batch = params.n_batch; - const int max_tasks_per_batch = params.n_parallel; + const int max_tasks_per_batch = 32; const int max_seq = 4*max_tasks_per_batch; llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); std::vector tok_logits(n_vocab); - std::vector batch_logits(n_ctx*n_vocab); + std::vector batch_logits(n_vocab*n_ctx); std::vector> eval_pairs; std::vector eval_results; std::vector workers(std::thread::hardware_concurrency()); - auto decode_helper = [&](llama_context * ctx, llama_batch & batch, int32_t n_batch) { - for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { - const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); - - llama_batch batch_view = { - n_tokens, - batch.token + i, - nullptr, - batch.pos + i, - batch.n_seq_id + i, - batch.seq_id + i, - batch.logits + i, - 0, 0, 0, // unused - }; - - const int ret = llama_decode(ctx, batch_view); - if (ret != 0) { - LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret); - return false; - } - - memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float)); - } - - return true; - }; - for (size_t i0 = 0; i0 < hs_task_count; i0++) { int n_cur = 0; @@ -696,7 +673,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { llama_kv_cache_clear(ctx); // decode all tasks [i0, i1) - if (!decode_helper(ctx, batch, n_batch)) { + if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { fprintf(stderr, "%s: llama_decode() failed\n", __func__); return; } @@ -772,6 +749,13 @@ struct winogrande_entry { std::string second; std::array choices; int answer; + + size_t i_batch; + size_t common_prefix; + size_t required_tokens; + size_t n_base1; // number of tokens for context + choice 1 + size_t n_base2; // number of tokens for context + choice 2 + std::vector seq_tokens[2]; }; static std::vector load_winogrande_from_csv(const std::string& prompt) { @@ -875,115 +859,164 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { data = std::move(selected); } + fprintf(stderr, "%s : tokenizing selected tasks\n", __func__); + // This is needed as usual for LLaMA models const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + for (auto & task : data) { + task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, add_bos); + task.seq_tokens[1] = ::llama_tokenize(ctx, task.first + task.choices[1] + task.second, add_bos); + + task.common_prefix = 0; + for (size_t k = 0; k < task.seq_tokens[0].size(); k++) { + if (task.seq_tokens[0][k] != task.seq_tokens[1][k]) { + break; + } + task.common_prefix++; + } + + task.required_tokens = task.common_prefix + + task.seq_tokens[0].size() - task.common_prefix + + task.seq_tokens[1].size() - task.common_prefix; + + task.n_base1 = ::llama_tokenize(ctx, task.first + task.choices[0], add_bos).size(); + task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], add_bos).size(); + } + fprintf(stderr, "%s : calculating winogrande score over selected tasks.\n", __func__); const int n_vocab = llama_n_vocab(llama_get_model(ctx)); - const int n_ctx = llama_n_ctx(ctx); + const int n_ctx = llama_n_ctx(ctx); + const int n_batch = params.n_batch; + + const int max_tasks_per_batch = 128; + const int max_seq = 2*max_tasks_per_batch; + + llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); std::vector tok_logits(n_vocab); + std::vector batch_logits(n_vocab*n_ctx); int n_correct = 0; int n_done = 0; - for (size_t task_idx = 0; task_idx < data.size(); task_idx++) { - const auto& task = data[task_idx]; + for (size_t i0 = 0; i0 < data.size(); i0++) { + int n_cur = 0; - auto base_context = ::llama_tokenize(ctx, task.first, add_bos); - auto base_ctx_1st = ::llama_tokenize(ctx, task.first + task.choices[0], add_bos); - auto base_ctx_2nd = ::llama_tokenize(ctx, task.first + task.choices[1], add_bos); + size_t i1 = i0; + size_t i_batch = 0; - auto sentence_1st = task.first + task.choices[0] + task.second; - auto sentence_2nd = task.first + task.choices[1] + task.second; - auto query_1st = ::llama_tokenize(ctx, sentence_1st, add_bos); - auto query_2nd = ::llama_tokenize(ctx, sentence_2nd, add_bos); + llama_batch_clear(batch); - if (query_1st.size() > (size_t)n_ctx || query_2nd.size() > (size_t)n_ctx) { - fprintf(stderr, "%s : number of tokens in queries %zu, %zu > n_ctxl\n", __func__, query_1st.size(), query_2nd.size()); + while (n_cur + (int) data[i1].required_tokens <= n_ctx) { + const int s0 = 2*(i1 - i0); + if (s0 + 2 > max_seq) { + break; + } + + for (size_t i = 0; i < data[i1].common_prefix; ++i) { + llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1}, false); + } + batch.logits[batch.n_tokens - 1] = true; + + for (int s = 0; s < 2; ++s) { + for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) { + llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true); + } + } + + data[i1].i_batch = i_batch; + i_batch += data[i1].required_tokens; + + n_cur += data[i1].required_tokens; + if (++i1 == data.size()) { + break; + } + } + + if (i0 == i1) { + fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0); return; } - auto query_1st_size = query_1st.size(); - auto query_2nd_size = query_2nd.size(); - - // Speedup small evaluations by evaluating atleast 32 tokens - // For Winogrande this seems to slow it down rather than speed it up. - //if (query_1st.size() < 32) query_1st.resize(32); - //if (query_2nd.size() < 32) query_2nd.resize(32); - llama_kv_cache_clear(ctx); - auto logits_1st = evaluate_tokens(ctx, query_1st, 0, params.n_batch, n_vocab); - llama_kv_cache_clear(ctx); - auto logits_2nd = evaluate_tokens(ctx, query_2nd, 0, params.n_batch, n_vocab); - - if (logits_1st.empty() || logits_2nd.empty()) { - fprintf(stderr, "%s : failed to eval\n", __func__); + // decode all tasks [i0, i1) + if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { + fprintf(stderr, "%s: llama_decode() failed\n", __func__); return; } - bool skip_choice = query_1st_size - base_ctx_1st.size() > k_min_trailing_ctx && - query_2nd_size - base_ctx_2nd.size() > k_min_trailing_ctx; + for (size_t i = i0; i < i1; ++i) { + auto & task = data[i]; - float score_1st = 0; - bool is_nan_1st = false; - const auto& base_1 = skip_choice ? base_ctx_1st : base_context; - const int last_1st = query_1st_size - base_1.size() > 1 ? 1 : 0; - for (size_t j = base_1.size()-1; j < query_1st_size-1-last_1st; ++j) { - std::memcpy(tok_logits.data(), logits_1st.data() + j*n_vocab, n_vocab*sizeof(float)); - const float prob = softmax(tok_logits)[query_1st[j+1]]; - if (std::isnan(prob) || !prob) { - fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, - prob, j, sentence_1st.c_str(), base_context.size()); - is_nan_1st = true; - break; + const bool skip_choice = + task.seq_tokens[0].size() - task.common_prefix > k_min_trailing_ctx && + task.seq_tokens[1].size() - task.common_prefix > k_min_trailing_ctx; + + float score_1st = 0; + bool is_nan_1st = false; + const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix; + const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0; + size_t li = n_base1 - 1; + for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) { + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(task.i_batch + li++), n_vocab*sizeof(float)); + const float prob = softmax(tok_logits)[task.seq_tokens[0][j+1]]; + if (std::isnan(prob) || !prob) { + fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, + prob, j, (task.first + task.choices[0] + task.second).c_str(), n_base1); + is_nan_1st = true; + break; + } + score_1st += std::log(prob); } - score_1st += std::log(prob); - } - score_1st /= (query_1st_size - base_1.size() - last_1st); + score_1st /= (task.seq_tokens[0].size() - n_base1 - last_1st); - float score_2nd = 0; - bool is_nan_2nd = false; - const auto& base_2 = skip_choice ? base_ctx_2nd : base_context; - const int last_2nd = query_2nd_size - base_2.size() > 1 ? 1 : 0; - for (size_t j = base_2.size()-1; j < query_2nd_size-1-last_2nd; ++j) { - std::memcpy(tok_logits.data(), logits_2nd.data() + j*n_vocab, n_vocab*sizeof(float)); - const float prob = softmax(tok_logits)[query_2nd[j+1]]; - if (std::isnan(prob) || !prob) { - fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, - prob, j, sentence_2nd.c_str(), base_context.size()); - is_nan_2nd = true; - break; + float score_2nd = 0; + bool is_nan_2nd = false; + const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix; + const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0; + li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1; + for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) { + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(task.i_batch + li++), n_vocab*sizeof(float)); + const float prob = softmax(tok_logits)[task.seq_tokens[1][j+1]]; + if (std::isnan(prob) || !prob) { + fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, + prob, j, (task.first + task.choices[1] + task.second).c_str(), n_base2); + is_nan_2nd = true; + break; + } + score_2nd += std::log(prob); } - score_2nd += std::log(prob); - } - score_2nd /= (query_2nd_size - base_2.size() - last_2nd); + score_2nd /= (task.seq_tokens[1].size() - n_base2 - last_2nd); - if (is_nan_1st || is_nan_2nd) { - continue; + if (is_nan_1st || is_nan_2nd) { + continue; + } + + if (std::isnan(score_1st) || std::isnan(score_2nd)) { + printf("================== NaN score %g, %g) for:\n", score_1st, score_2nd); + printf("Q1: <%s> - %zu tokens\n", (task.first + task.choices[0] + task.second).c_str(), task.seq_tokens[0].size()); + printf("Q2: <%s> - %zu tokens\n", (task.first + task.choices[1] + task.second).c_str(), task.seq_tokens[1].size()); + printf("B : <%s> - %zu tokens\n", task.first.c_str(), task.common_prefix); + printf("base_1 has %zu tokens, base_2 has %zu tokens, skip_choice = %d\n", n_base1, n_base2, skip_choice); + continue; + } + + int result = score_1st > score_2nd ? 1 : 2; + + if (result == task.answer) { + ++n_correct; + } + ++n_done; + + // Print the accumulated accuracy mean x 100 + printf("%zu\t%.4lf\t%10.6f %10.6f %d %d\n", i+1, 100.0 * n_correct/n_done, score_1st, score_2nd, result, task.answer); + fflush(stdout); } - if (std::isnan(score_1st) || std::isnan(score_2nd)) { - printf("================== NaN score %g, %g) for:\n", score_1st, score_2nd); - printf("Q1: <%s> - %zu tokens\n", sentence_1st.c_str(), query_1st_size); - printf("Q2: <%s> - %zu tokens\n", sentence_2nd.c_str(), query_2nd_size); - printf("B : <%s> - %zu tokens\n", task.first.c_str(), base_context.size()); - printf("base_1 has %zu tokens, base_2 has %zu tokens, skip_choice = %d\n", base_1.size(), base_2.size(), skip_choice); - continue; - } - - int result = score_1st > score_2nd ? 1 : 2; - - if (result == task.answer) { - ++n_correct; - } - ++n_done; - - // Print the accumulated accuracy mean x 100 - printf("%zu\t%.4lf\t%10.6f %10.6f %d %d\n",task_idx+1, 100.0 * n_correct/n_done,score_1st,score_2nd,result,task.answer); - fflush(stdout); + i0 = i1 - 1; } printf("\n"); From 993fba81807e55d27b570945af8e416d535eced1 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 19 Jan 2024 11:02:39 +0200 Subject: [PATCH 025/101] perplexity: avoid unnecessary alloocations and logit copies (#5035) Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 22 +++++++++++++++------- 1 file changed, 15 insertions(+), 7 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index df902fb1c..292502f87 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -325,6 +325,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par double nll = 0.0; double nll2 = 0.0; + const int num_batches = (n_ctx + n_batch - 1) / n_batch; + + std::vector logits; + if (num_batches > 1) { + logits.reserve((size_t)n_ctx * n_vocab); + } + fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch); std::vector workers(std::thread::hardware_concurrency() - 1); @@ -333,10 +340,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par const int start = i * n_ctx; const int end = start + n_ctx; - const int num_batches = (n_ctx + n_batch - 1) / n_batch; - - std::vector logits; - const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache @@ -362,8 +365,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // restore the original token in case it was set to BOS tokens[batch_start] = token_org; - const auto * batch_logits = llama_get_logits(ctx); - logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + if (num_batches > 1) { + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } } const auto t_end = std::chrono::high_resolution_clock::now(); @@ -392,7 +397,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // last 256 tokens. Then, we split the input up into context window size chunks to // process the entire prompt. const int first = n_ctx/2; - process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); count += n_ctx - first - 1; @@ -406,6 +412,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par printf("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2); } fflush(stdout); + + logits.clear(); } printf("\n"); From 2b3b999cacc7ad1207c32fbdf3479a19c06e1a34 Mon Sep 17 00:00:00 2001 From: chiranko <96988916+chiranko@users.noreply.github.com> Date: Fri, 19 Jan 2024 17:07:27 +0800 Subject: [PATCH 026/101] llama : add CodeShell support (#5016) * llama: add codeshell support * llama.cpp: fix codeshell with NeoX rope Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- convert-hf-to-gguf.py | 67 ++++++++++++ gguf-py/gguf/constants.py | 19 ++++ gguf-py/gguf/tensor_mapping.py | 1 + llama.cpp | 181 +++++++++++++++++++++++++++++++++ 4 files changed, 268 insertions(+) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 1178d63a2..aae3a5e87 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -197,6 +197,8 @@ class Model: return Phi2Model if model_architecture == "PlamoForCausalLM": return PlamoModel + if model_architecture == "CodeShellForCausalLM": + return CodeShellModel return Model def _is_model_safetensors(self) -> bool: @@ -242,6 +244,8 @@ class Model: return gguf.MODEL_ARCH.PHI2 if arch == "PlamoForCausalLM": return gguf.MODEL_ARCH.PLAMO + if arch == "CodeShellForCausalLM": + return gguf.MODEL_ARCH.CODESHELL raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -1175,6 +1179,69 @@ class PlamoModel(Model): self.gguf_writer.add_tensor(new_name, data) +class CodeShellModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams["n_layer"] + + self.gguf_writer.add_name("CodeShell") + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_head_count_kv(self.hparams["num_query_groups"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_rope_freq_base(10000.0) + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(1.0) + + def write_tensors(self): + block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer"))) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + tensors = dict(self.get_tensors()) + has_lm_head = "lm_head.weight" in tensors.keys() or "output.weight" in tensors.keys() + for name, data_torch in tensors.items(): + # we don't need these + if name.endswith((".attn.rotary_emb.inv_freq")): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + if not has_lm_head and name == "transformer.wte.weight": + self.gguf_writer.add_tensor("output.weight", data) + print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") ###### CONVERSION LOGIC ###### diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 972b4e9a7..95c58b419 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -99,6 +99,7 @@ class MODEL_ARCH(IntEnum): QWEN = auto() PHI2 = auto() PLAMO = auto() + CODESHELL = auto() class MODEL_TENSOR(IntEnum): @@ -147,6 +148,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.QWEN: "qwen", MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PLAMO: "plamo", + MODEL_ARCH.CODESHELL: "codeshell", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -396,6 +398,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_NORM, MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.CODESHELL: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, ] # TODO } @@ -417,6 +432,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.CODESHELL: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index e5b146106..de177af13 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -154,6 +154,7 @@ class TensorNameMap: "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth "model.layers.layers.{bid}.self_attn.rotary_emb.inv_freq", # plamo + "transformer.h.{bid}.attn.rotary_emb.inv_freq", # codeshell ), # Feed-forward norm diff --git a/llama.cpp b/llama.cpp index 47b4384a8..1cee5a791 100644 --- a/llama.cpp +++ b/llama.cpp @@ -194,6 +194,7 @@ enum llm_arch { LLM_ARCH_QWEN, LLM_ARCH_PHI2, LLM_ARCH_PLAMO, + LLM_ARCH_CODESHELL, LLM_ARCH_UNKNOWN, }; @@ -213,6 +214,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN, "qwen" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PLAMO, "plamo" }, + { LLM_ARCH_CODESHELL, "codeshell" }, }; enum llm_kv { @@ -600,6 +602,26 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_CODESHELL, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, @@ -2877,6 +2899,14 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_CODESHELL: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 42: model.type = e_model::MODEL_SMALL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3784,6 +3814,42 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_CODESHELL: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}); + layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}); + + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}); + layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}); + + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -5965,6 +6031,117 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_codeshell() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(tmpq, "tmpq", il); + cb(tmpk, "tmpk", il); + cb(Vcur, "Vcur", il); + + struct ggml_tensor * Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } }; static struct ggml_cgraph * llama_build_graph( @@ -6159,6 +6336,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_gpt2(); } break; + case LLM_ARCH_CODESHELL: + { + result = llm.build_codeshell(); + } break; default: GGML_ASSERT(false); } From 7051aacfac0057fa5fac9ea46c55bffc3892d810 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 19 Jan 2024 11:39:11 +0200 Subject: [PATCH 027/101] winogrande: evaluate log-probs in parallel (#5036) This is a relatively minor performance tweak resulting in ~10% speedup on my system. Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 71 ++++++++++++++---------------- 1 file changed, 33 insertions(+), 38 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 292502f87..b07320190 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -458,7 +458,7 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector< return true; } -static void hellaswag_compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, +static void compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, const std::vector>& eval_pairs, std::vector& eval_results) { constexpr int k_token_chunk = 4; if (eval_results.size() != eval_pairs.size()) { @@ -700,7 +700,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { } } // Then we do the actual calculation - hellaswag_compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); + compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); size_t ir = 0; @@ -906,6 +906,10 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { std::vector tok_logits(n_vocab); std::vector batch_logits(n_vocab*n_ctx); + std::vector> eval_pairs; + std::vector eval_results; + std::vector workers(std::thread::hardware_concurrency()); + int n_correct = 0; int n_done = 0; @@ -956,6 +960,30 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { return; } + eval_pairs.clear(); + for (size_t i = i0; i < i1; ++i) { + auto & task = data[i]; + + const bool skip_choice = + task.seq_tokens[0].size() - task.common_prefix > k_min_trailing_ctx && + task.seq_tokens[1].size() - task.common_prefix > k_min_trailing_ctx; + + const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix; + const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0; + size_t li = n_base1 - 1; + for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) { + eval_pairs.push_back(std::make_pair(task.i_batch + li++, task.seq_tokens[0][j+1])); + } + const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix; + const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0; + li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1; + for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) { + eval_pairs.push_back(std::make_pair(task.i_batch + li++, task.seq_tokens[1][j+1])); + } + } + compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); + + size_t ir = 0; for (size_t i = i0; i < i1; ++i) { auto & task = data[i]; @@ -964,54 +992,21 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { task.seq_tokens[1].size() - task.common_prefix > k_min_trailing_ctx; float score_1st = 0; - bool is_nan_1st = false; const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix; const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0; - size_t li = n_base1 - 1; for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) { - std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(task.i_batch + li++), n_vocab*sizeof(float)); - const float prob = softmax(tok_logits)[task.seq_tokens[0][j+1]]; - if (std::isnan(prob) || !prob) { - fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, - prob, j, (task.first + task.choices[0] + task.second).c_str(), n_base1); - is_nan_1st = true; - break; - } - score_1st += std::log(prob); + score_1st += eval_results[ir++]; } score_1st /= (task.seq_tokens[0].size() - n_base1 - last_1st); float score_2nd = 0; - bool is_nan_2nd = false; const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix; const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0; - li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1; for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) { - std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(task.i_batch + li++), n_vocab*sizeof(float)); - const float prob = softmax(tok_logits)[task.seq_tokens[1][j+1]]; - if (std::isnan(prob) || !prob) { - fprintf(stderr, "%s: %g probability for token %zu when evaluating <%s>. Base context has %zu tokens\n", __func__, - prob, j, (task.first + task.choices[1] + task.second).c_str(), n_base2); - is_nan_2nd = true; - break; - } - score_2nd += std::log(prob); + score_2nd += eval_results[ir++]; } score_2nd /= (task.seq_tokens[1].size() - n_base2 - last_2nd); - if (is_nan_1st || is_nan_2nd) { - continue; - } - - if (std::isnan(score_1st) || std::isnan(score_2nd)) { - printf("================== NaN score %g, %g) for:\n", score_1st, score_2nd); - printf("Q1: <%s> - %zu tokens\n", (task.first + task.choices[0] + task.second).c_str(), task.seq_tokens[0].size()); - printf("Q2: <%s> - %zu tokens\n", (task.first + task.choices[1] + task.second).c_str(), task.seq_tokens[1].size()); - printf("B : <%s> - %zu tokens\n", task.first.c_str(), task.common_prefix); - printf("base_1 has %zu tokens, base_2 has %zu tokens, skip_choice = %d\n", n_base1, n_base2, skip_choice); - continue; - } - int result = score_1st > score_2nd ? 1 : 2; if (result == task.answer) { @@ -1019,7 +1014,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { } ++n_done; - // Print the accumulated accuracy mean x 100 + // print the accumulated accuracy mean x 100 printf("%zu\t%.4lf\t%10.6f %10.6f %d %d\n", i+1, 100.0 * n_correct/n_done, score_1st, score_2nd, result, task.answer); fflush(stdout); } From de9a147df14e62f54f879d2d15e6c4793107f4fc Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 Jan 2024 13:52:22 +0200 Subject: [PATCH 028/101] py : fix flake8 lint --- convert-hf-to-gguf.py | 1 + 1 file changed, 1 insertion(+) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index aae3a5e87..d2d6948d8 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -1179,6 +1179,7 @@ class PlamoModel(Model): self.gguf_writer.add_tensor(new_name, data) + class CodeShellModel(Model): def set_gguf_parameters(self): block_count = self.hparams["n_layer"] From 9b75cb2b3ccbed3df2e14c1202168db3e5145095 Mon Sep 17 00:00:00 2001 From: Shijie <821898965@qq.com> Date: Fri, 19 Jan 2024 19:53:13 +0800 Subject: [PATCH 029/101] llama : support upcoming Qwen2 (#5037) --- convert-hf-to-gguf.py | 4 + gguf-py/gguf/constants.py | 16 ++++ llama.cpp | 191 ++++++++++++++++++++++++++++++++++++++ 3 files changed, 211 insertions(+) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index d2d6948d8..5cb3e63fb 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -189,6 +189,8 @@ class Model: return StableLMModel if model_architecture == "QWenLMHeadModel": return QwenModel + if model_architecture == "Qwen2ForCausalLM": + return Model if model_architecture == "MixtralForCausalLM": return MixtralModel if model_architecture == "GPT2LMHeadModel": @@ -236,6 +238,8 @@ class Model: return gguf.MODEL_ARCH.STABLELM if arch == "QWenLMHeadModel": return gguf.MODEL_ARCH.QWEN + if arch == "Qwen2ForCausalLM": + return gguf.MODEL_ARCH.QWEN2 if arch == "MixtralForCausalLM": return gguf.MODEL_ARCH.LLAMA if arch == "GPT2LMHeadModel": diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 95c58b419..2d9c33c7d 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -97,6 +97,7 @@ class MODEL_ARCH(IntEnum): BLOOM = auto() STABLELM = auto() QWEN = auto() + QWEN2 = auto() PHI2 = auto() PLAMO = auto() CODESHELL = auto() @@ -146,6 +147,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.BLOOM: "bloom", MODEL_ARCH.STABLELM: "stablelm", MODEL_ARCH.QWEN: "qwen", + MODEL_ARCH.QWEN2: "qwen2", MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PLAMO: "plamo", MODEL_ARCH.CODESHELL: "codeshell", @@ -358,6 +360,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.QWEN2: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], MODEL_ARCH.PLAMO: [ MODEL_TENSOR.TOKEN_EMBD, MODEL_TENSOR.OUTPUT_NORM, diff --git a/llama.cpp b/llama.cpp index 1cee5a791..90579ac85 100644 --- a/llama.cpp +++ b/llama.cpp @@ -192,6 +192,7 @@ enum llm_arch { LLM_ARCH_BLOOM, LLM_ARCH_STABLELM, LLM_ARCH_QWEN, + LLM_ARCH_QWEN2, LLM_ARCH_PHI2, LLM_ARCH_PLAMO, LLM_ARCH_CODESHELL, @@ -212,6 +213,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_BLOOM, "bloom" }, { LLM_ARCH_STABLELM, "stablelm" }, { LLM_ARCH_QWEN, "qwen" }, + { LLM_ARCH_QWEN2, "qwen2" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PLAMO, "plamo" }, { LLM_ARCH_CODESHELL, "codeshell" }, @@ -568,6 +570,23 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_QWEN2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_PHI2, { @@ -2869,6 +2888,17 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_QWEN2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; + case 32: model.type = e_model::MODEL_7B; break; + case 40: model.type = e_model::MODEL_13B; break; + case 80: model.type = e_model::MODEL_70B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; case LLM_ARCH_PHI2: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); @@ -3704,6 +3734,41 @@ static bool llm_load_tensors( layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff/2}); } } break; + case LLM_ARCH_QWEN2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + + // output + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + // optional bias tensors + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; case LLM_ARCH_PHI2: { model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); @@ -5698,6 +5763,128 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_qwen2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(gf, Qcur); + ggml_build_forward_expand(gf, Kcur); + ggml_build_forward_expand(gf, Vcur); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, model, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + struct ggml_cgraph * build_phi2() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -6324,6 +6511,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_qwen(); } break; + case LLM_ARCH_QWEN2: + { + result = llm.build_qwen2(); + } break; case LLM_ARCH_PHI2: { result = llm.build_phi2(); From a5cacb22b2114fd9adf61c00cbb237384d86bced Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 19 Jan 2024 15:24:47 +0200 Subject: [PATCH 030/101] imatrix : add README.md --- examples/imatrix/README.md | 32 ++++++++++++++++++++++++++++++++ 1 file changed, 32 insertions(+) create mode 100644 examples/imatrix/README.md diff --git a/examples/imatrix/README.md b/examples/imatrix/README.md new file mode 100644 index 000000000..578e8fc27 --- /dev/null +++ b/examples/imatrix/README.md @@ -0,0 +1,32 @@ +# llama.cpp/examples/imatrix + +Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models. +More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861 + +## Usage + +``` +./imatrix -m -f [-o ] [--verbosity ] + [-ofreq num_chunks] [-ow <0 or 1>] [other common params] +``` + +Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory. +The parameters in square brackets are optional and have the following meaning: +* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used. +* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`. +* `-ofreq` (or `--output-frequency`) specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks) +* `-ow` (or `--output-weight`) specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default. + +For faster computation, make sure to use GPU offloading via the `-ngl` argument + +## Example + +```bash +LLAMA_CUBLAS=1 make -j + +# generate importance matrix (imatrix.dat) +./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99 + +# use the imatrix to perform a Q4_K_M quantization +./quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m +``` From 381ee195721d8e747ee31a60c0751822b3072f02 Mon Sep 17 00:00:00 2001 From: Uzo Nweke Date: Fri, 19 Jan 2024 13:20:50 -0500 Subject: [PATCH 031/101] finetune : fix ggml_allocr lifetimes (tmp workaround) (#5033) * Fix issue with alloc causing max_compute_size to be calculated * remove ggml_allocr_free as suggested in issue #4791 --- examples/train-text-from-scratch/train-text-from-scratch.cpp | 4 ---- 1 file changed, 4 deletions(-) diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp index 4a9a2340b..eee9d4de3 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -263,7 +263,6 @@ static void init_model(struct my_llama_model * model) { model->data.resize(size + tensor_alignment); alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment); alloc_model(alloc, model); - ggml_allocr_free(alloc); } static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) { @@ -1102,7 +1101,6 @@ int main(int argc, char ** argv) { alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment); ggml_allocr_alloc(alloc, tokens_input); ggml_allocr_alloc(alloc, target_probs); - ggml_allocr_free(alloc); // context for compute tensors without their data const size_t estimated_compute_size_wo_data = ( @@ -1149,7 +1147,6 @@ int main(int argc, char ** argv) { best_compute_size = max_compute_size; best_order = gf->order; } - ggml_allocr_free(alloc); ggml_free(ctx_compute); } size_t max_compute_size = best_compute_size; @@ -1177,7 +1174,6 @@ int main(int argc, char ** argv) { params.common.use_flash, params.common.use_checkpointing ); - ggml_allocr_free(alloc); std::vector train_tokens; std::vector train_samples_begin; From cca894f16a5eade15afd07b015e4cb3d8658943f Mon Sep 17 00:00:00 2001 From: Kylin <56434533+KyL0N@users.noreply.github.com> Date: Sat, 20 Jan 2024 15:01:46 +0800 Subject: [PATCH 032/101] cuda : fix compile error in jetson platform (#4975) * cuda: fix compile error in jetson platform * cuda: update comment in ggml-cuda.cu * cuda: update ggml-cuda.cu comment --- ggml-cuda.cu | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index b2211d858..ec3837fb8 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -12,9 +12,6 @@ #include #include #include -#include "ggml-cuda.h" -#include "ggml.h" -#include "ggml-backend-impl.h" #if defined(GGML_USE_HIPBLAS) #include @@ -118,6 +115,11 @@ #endif // defined(GGML_USE_HIPBLAS) +// ggml-cuda need half type so keep ggml headers include at last +#include "ggml-cuda.h" +#include "ggml.h" +#include "ggml-backend-impl.h" + #define CUDART_HMAX 11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed) #define CC_PASCAL 600 From 48e2b1337257bb77573bf76cfffcff3a5efa8704 Mon Sep 17 00:00:00 2001 From: adel boussaken Date: Sat, 20 Jan 2024 09:05:43 +0100 Subject: [PATCH 033/101] Add a dart/flutter binding to README.md (#4882) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 866aa87b4..cbfba01bc 100644 --- a/README.md +++ b/README.md @@ -128,6 +128,7 @@ as the main playground for developing new features for the [ggml](https://github - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn) - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp) - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig) +- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart) **UI:** From 77bc1bbd05f0c31cb45773eb5eb59b9ff2b07e1b Mon Sep 17 00:00:00 2001 From: Herman Semenov Date: Sat, 20 Jan 2024 08:11:31 +0000 Subject: [PATCH 034/101] cmake : add support for ccache (#5002) * Added support ccache for speedup recompilation * cmake : option to disable ccache --------- Co-authored-by: Georgi Gerganov --- CMakeLists.txt | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 3fc65eaf2..6b3b1396b 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -47,6 +47,7 @@ option(BUILD_SHARED_LIBS "build shared libraries" option(LLAMA_STATIC "llama: static link libraries" OFF) option(LLAMA_NATIVE "llama: enable -march=native flag" ON) option(LLAMA_LTO "llama: enable link time optimization" OFF) +option(LLAMA_CCACHE "llama: use ccache if available" ON) # debug option(LLAMA_ALL_WARNINGS "llama: enable all compiler warnings" ON) @@ -561,6 +562,17 @@ if (LLAMA_LTO) endif() endif() +if (LLAMA_CCACHE) + find_program(LLAMA_CCACHE_FOUND ccache) + if (LLAMA_CCACHE_FOUND) + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache) + set(ENV{CCACHE_SLOPPINESS} time_macros) + message(STATUS "Using ccache") + else() + message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF") + endif () +endif() + # this version of Apple ld64 is buggy execute_process( COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v From 6df465a91d402370bcba6676b19fad85b06ce7e0 Mon Sep 17 00:00:00 2001 From: slaren Date: Sat, 20 Jan 2024 16:05:49 +0100 Subject: [PATCH 035/101] llama : run all KQV ops on the CPU with no KV offload (#5049) ggml-ci --- ggml-backend.c | 34 +++++++----- llama.cpp | 145 +++++++++++++++++++++++++++---------------------- 2 files changed, 99 insertions(+), 80 deletions(-) diff --git a/ggml-backend.c b/ggml-backend.c index ef518dae0..423512def 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -1191,6 +1191,24 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g ggml_tallocr_t src_allocr = node_allocr(src); GGML_ASSERT(src_allocr != NULL); // all inputs should be assigned by now if (src_allocr != node_allocr) { + // create a copy of the input in the split's backend + size_t id = hash_id(src); + if (sched->node_copies[id][cur_backend_id] == NULL) { + ggml_backend_t backend = get_allocr_backend(sched, cur_allocr); + struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); + + sched->node_copies[id][cur_backend_id] = tensor_copy; + node_allocr(tensor_copy) = cur_allocr; + SET_CAUSE(tensor_copy, "4.cpy"); + + int n_inputs = sched->splits[cur_split].n_inputs++; + GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); + sched->splits[cur_split].inputs[n_inputs] = src; + } + node->src[j] = sched->node_copies[id][cur_backend_id]; + +#if 0 // check if the input is already in the split bool found = false; for (int k = 0; k < sched->splits[cur_split].n_inputs; k++) { @@ -1206,19 +1224,7 @@ static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * g GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); sched->splits[cur_split].inputs[n_inputs] = src; } - - // create a copy of the input in the split's backend - size_t id = hash_id(src); - if (sched->node_copies[id][cur_backend_id] == NULL) { - ggml_backend_t backend = get_allocr_backend(sched, cur_allocr); - struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); - ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); - - sched->node_copies[id][cur_backend_id] = tensor_copy; - node_allocr(tensor_copy) = cur_allocr; - SET_CAUSE(tensor_copy, "4.cpy"); - } - node->src[j] = sched->node_copies[id][cur_backend_id]; +#endif } } } @@ -1333,7 +1339,7 @@ static void sched_compute_splits(ggml_backend_sched_t sched) { uint64_t compute_start_us = ggml_time_us(); if (!sched->callback_eval) { ggml_backend_graph_compute(split_backend, &split->graph); - //ggml_backend_synchronize(split_backend); // necessary to measure compute time + //ggml_backend_synchronize(split_backend); // necessary to measure compute time } else { // similar to ggml_backend_compare_graph_backend for (int j0 = 0; j0 < split->graph.n_nodes; j0++) { diff --git a/llama.cpp b/llama.cpp index 90579ac85..909ad4ad8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4315,6 +4315,7 @@ static struct ggml_tensor * llm_build_kqv( const llama_model & model, const llama_hparams & hparams, const llama_kv_cache & kv, + struct ggml_cgraph * graph, struct ggml_tensor * wo, struct ggml_tensor * wo_b, struct ggml_tensor * q_cur, @@ -4393,6 +4394,8 @@ static struct ggml_tensor * llm_build_kqv( struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd_head_k*n_head, n_tokens); cb(cur, "kqv_merged_cont", il); + ggml_build_forward_expand(graph, cur); + cur = ggml_mul_mat(ctx, wo, cur); if (wo_b) { cb(cur, "kqv_wo", il); @@ -4405,6 +4408,44 @@ static struct ggml_tensor * llm_build_kqv( return cur; } +static struct ggml_tensor * llm_build_kv( + struct ggml_context * ctx, + const llama_model & model, + const llama_hparams & hparams, + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + struct ggml_tensor * wo, + struct ggml_tensor * wo_b, + struct ggml_tensor * k_cur, + struct ggml_tensor * v_cur, + struct ggml_tensor * q_cur, + struct ggml_tensor * kq_mask, + int64_t n_ctx, + int32_t n_tokens, + int32_t kv_head, + int32_t n_kv, + float max_alibi_bias, + float kq_scale, + const llm_build_cb & cb, + int il) { + + // these nodes are added to the graph together so that they are not reordered + // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(graph, k_cur); + ggml_build_forward_expand(graph, v_cur); + ggml_build_forward_expand(graph, q_cur); + + llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); + + struct ggml_tensor * cur; + cur = llm_build_kqv(ctx, model, hparams, kv, graph, + wo, wo_b, + q_cur, kq_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, kq_scale, cb, il); + cb(cur, "kqv_out", il); + + return cur; +} + struct llm_build_context { const llama_model & model; const llama_hparams & hparams; @@ -4562,12 +4603,6 @@ struct llm_build_context { cb(Vcur, "Vcur", il); } - // these nodes are added to the graph together so that they are not reordered - // by doing so, the number of splits in the graph is reduced - ggml_build_forward_expand(gf, Qcur); - ggml_build_forward_expand(gf, Kcur); - ggml_build_forward_expand(gf, Vcur); - Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale, @@ -4582,11 +4617,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4763,14 +4796,13 @@ struct llm_build_context { cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); // apply ALiBi for 13B model const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4892,11 +4924,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -4993,11 +5023,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5200,12 +5228,9 @@ struct llm_build_context { ); cb(Vcur, "Vcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - // TODO: not tested, could be broken - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Q, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Q, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5292,11 +5317,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); cb(Qcur, "Qcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5390,11 +5413,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, 8.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5485,11 +5506,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, hparams.f_max_alibi_bias, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5597,11 +5616,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5714,11 +5731,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5837,11 +5852,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -5966,11 +5979,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f, cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f, cb, il); cb(cur, "kqv_out", il); } @@ -6071,11 +6082,9 @@ struct llm_build_context { ext_factor, attn_factor, beta_fast, beta_slow); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, NULL, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } struct ggml_tensor * sa_out = cur; @@ -6172,11 +6181,9 @@ struct llm_build_context { Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -6283,11 +6290,9 @@ struct llm_build_context { ); cb(Kcur, "Kcur", il); - llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - - cur = llm_build_kqv(ctx0, model, hparams, kv_self, + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); cb(cur, "kqv_out", il); } @@ -6355,6 +6360,14 @@ static struct ggml_cgraph * llama_build_graph( ggml_set_name(cur, name); } + + if (!lctx.cparams.offload_kqv) { + if (strcmp(name, "kqv_merged_cont") == 0) { + // all nodes between the KV store and the attention output are run on the CPU + ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu); + } + } + // // allocate input tensors and set input data // From 97c1549808d2742d37584a3c9df28154bdf34417 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Sat, 20 Jan 2024 10:08:08 -0500 Subject: [PATCH 036/101] perplexity : fix MSVC build after #5020 (#5043) * perplexity : fix MSVC build after #5020 * try a differerent fix --- examples/perplexity/perplexity.cpp | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b07320190..f91f5795a 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -458,23 +458,24 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector< return true; } +#define K_TOKEN_CHUNK 4 + static void compute_logprobs(const float * batch_logits, int n_vocab, std::vector& workers, const std::vector>& eval_pairs, std::vector& eval_results) { - constexpr int k_token_chunk = 4; if (eval_results.size() != eval_pairs.size()) { eval_results.resize(eval_pairs.size()); } if (eval_pairs.empty()) return; - size_t max_threads = std::min((eval_pairs.size() + k_token_chunk - 1)/k_token_chunk, workers.size()); + size_t max_threads = std::min((eval_pairs.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK, workers.size()); std::atomic counter(0); auto compute = [&counter, &eval_pairs, &eval_results, batch_logits, n_vocab] () { - float local_logprobs[k_token_chunk]; + float local_logprobs[K_TOKEN_CHUNK]; while (true) { - size_t first = counter.fetch_add(k_token_chunk, std::memory_order_relaxed); + size_t first = counter.fetch_add(K_TOKEN_CHUNK, std::memory_order_relaxed); if (first >= eval_results.size()) break; - size_t last = std::min(first + k_token_chunk, eval_results.size()); + size_t last = std::min(first + K_TOKEN_CHUNK, eval_results.size()); for (size_t i = first; i < last; ++i) { auto logits = batch_logits + eval_pairs[i].first * n_vocab; float max_logit = logits[0]; @@ -497,7 +498,6 @@ static void compute_logprobs(const float * batch_logits, int n_vocab, std::vecto for (size_t it = 0; it < max_threads; ++it) { workers[it].join(); } - } static void hellaswag_score(llama_context * ctx, const gpt_params & params) { From b43ebde3b0ccbc42d9dd782b32e2fd8eb35b43b5 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Sat, 20 Jan 2024 18:14:18 -0500 Subject: [PATCH 037/101] convert : partially revert PR #4818 (#5041) --- convert-hf-to-gguf.py | 9 +- convert-llama-ggml-to-gguf.py | 14 +- convert-lora-to-ggml.py | 3 +- convert-persimmon-to-gguf.py | 12 +- convert.py | 627 ++++++++++++---------------------- mypy.ini | 1 + 6 files changed, 237 insertions(+), 429 deletions(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5cb3e63fb..4d995ef78 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -10,7 +10,7 @@ import re import sys from enum import IntEnum from pathlib import Path -from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast, Optional +from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast import numpy as np import torch @@ -487,7 +487,8 @@ class MPTModel(Model): # map tensor names if "scales" in name: new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias", ".scales")) - new_name = new_name.replace("scales", "act.scales") + if new_name is not None: + new_name = new_name.replace("scales", "act.scales") else: new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) if new_name is None: @@ -904,7 +905,7 @@ class QwenModel(Model): return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')]) @staticmethod - def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: Optional[int] = None) -> list[bytes]: + def bpe(mergeable_ranks: dict[bytes, int], token: bytes, max_rank: int | None = None) -> list[bytes]: parts = [bytes([b]) for b in token] while True: min_idx = None @@ -1285,7 +1286,7 @@ def main() -> None: if args.awq_path: sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) - from awq.apply_awq import add_scale_weights + from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] tmp_model_path = args.model / "weighted_model" dir_model = tmp_model_path if tmp_model_path.is_dir(): diff --git a/convert-llama-ggml-to-gguf.py b/convert-llama-ggml-to-gguf.py index e359330af..b33108062 100755 --- a/convert-llama-ggml-to-gguf.py +++ b/convert-llama-ggml-to-gguf.py @@ -2,6 +2,7 @@ from __future__ import annotations import argparse +import os import struct import sys from enum import IntEnum @@ -9,7 +10,6 @@ from pathlib import Path import numpy as np -import os if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf @@ -371,15 +371,11 @@ def handle_metadata(cfg, hp): params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path) else: raise ValueError('Unable to load metadata') - vocab = convert.load_vocab( - cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, - cfg.vocabtype) - # FIXME: Respect cfg.vocab_dir? - svocab = gguf.SpecialVocab(cfg.model_metadata_dir, - load_merges = cfg.vocabtype == 'bpe', - n_vocab = vocab.vocab_size) + vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir) + vocab_factory = convert.VocabFactory(vocab_path) + vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir) convert.check_vocab_size(params, vocab) - return (params, vocab, svocab) + return params, vocab, special_vocab def handle_args(): diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 35ce152f4..4904bf128 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -5,17 +5,16 @@ import json import os import struct import sys +from pathlib import Path from typing import Any, BinaryIO, Sequence import numpy as np import torch -from pathlib import Path if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) import gguf - NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1} diff --git a/convert-persimmon-to-gguf.py b/convert-persimmon-to-gguf.py index 1ba5864dc..d2be805d1 100755 --- a/convert-persimmon-to-gguf.py +++ b/convert-persimmon-to-gguf.py @@ -1,11 +1,13 @@ #!/usr/bin/env python3 -import torch -import os -from pprint import pprint -import sys import argparse +import os +import sys from pathlib import Path +from pprint import pprint + +import torch from sentencepiece import SentencePieceProcessor + if 'NO_LOCAL_GGUF' not in os.environ: sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf @@ -69,7 +71,7 @@ def main(): persimmon_model = torch.load(args.ckpt_path) hparams = persimmon_model['args'] pprint(hparams) - tensors = {} + tensors: dict[str, torch.Tensor] = {} _flatten_dict(persimmon_model['model'], tensors, None) arch = gguf.MODEL_ARCH.PERSIMMON diff --git a/convert.py b/convert.py index 980e6fc72..06768033d 100755 --- a/convert.py +++ b/convert.py @@ -17,58 +17,28 @@ import signal import struct import sys import time -import warnings import zipfile from abc import ABCMeta, abstractmethod -from argparse import ArgumentParser from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import ( - IO, - TYPE_CHECKING, - Any, - Callable, - Iterable, - Literal, - Optional, - Tuple, - TypeVar, -) +from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar import numpy as np from sentencepiece import SentencePieceProcessor -try: - from transformers import AutoTokenizer -except ModuleNotFoundError as e: - warnings.warn(f"Could not import AutoTokenizer from transformers: {e}") +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf -# If NO_LOCAL_GGUF is not set, try to import gguf from the local gguf-py directory -if "NO_LOCAL_GGUF" not in os.environ: - # Use absolute path to the gguf-py directory - gguf_py_dir = str(Path(__file__).resolve().parent / "gguf-py") - print(gguf_py_dir) # NOTE: Remove this once path is verified after changes are completed - if gguf_py_dir not in sys.path: - sys.path.insert(1, gguf_py_dir) +if TYPE_CHECKING: + from typing import TypeAlias -# Import gguf module -try: - import gguf -except ModuleNotFoundError as e: - print(f"Could not import gguf: {e}") - sys.exit(1) - -if TYPE_CHECKING: # NOTE: This isn't necessary. - from typing import TypeAlias # This can technically be omitted. - -if hasattr(faulthandler, "register") and hasattr(signal, "SIGUSR1"): +if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'): faulthandler.register(signal.SIGUSR1) -# NOTE: n-dimensional arrays should be directly referenced -NDArray: TypeAlias = "np.ndarray[Any, Any]" +NDArray: TypeAlias = 'np.ndarray[Any, Any]' -# Why is this here? LLAMA and GPT are technically the only compatible ARCHs. ARCH = gguf.MODEL_ARCH.LLAMA DEFAULT_CONCURRENCY = 8 @@ -78,7 +48,6 @@ DEFAULT_CONCURRENCY = 8 # -# TODO: Clean up and refactor data types @dataclass(frozen=True) class DataType: name: str @@ -183,85 +152,65 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { @dataclass class Params: - n_vocab: int - n_embd: int - n_layer: int - n_ctx: int - n_ff: int - n_head: int - n_head_kv: int - f_norm_eps: Optional[float] = None - n_experts: Optional[int] = None - n_experts_used: Optional[int] = None + n_vocab: int + n_embd: int + n_layer: int + n_ctx: int + n_ff: int + n_head: int + n_head_kv: int + n_experts: int | None = None + n_experts_used: int | None = None + f_norm_eps: float | None = None - rope_scaling_type: Optional[gguf.RopeScalingType] = None - f_rope_freq_base: Optional[float] = None - f_rope_scale: Optional[float] = None - n_orig_ctx: Optional[int] = None - rope_finetuned: Optional[bool] = None + rope_scaling_type: gguf.RopeScalingType | None = None + f_rope_freq_base: float | None = None + f_rope_scale: float | None = None + n_orig_ctx: int | None = None + rope_finetuned: bool | None = None - ftype: Optional[GGMLFileType] = None + ftype: GGMLFileType | None = None # path to the directory containing the model files - path_model: Optional[Path] = None + path_model: Path | None = None @staticmethod - def guessed(model: LazyModel) -> "Params": + def guessed(model: LazyModel) -> Params: # try transformer naming first - n_vocab, n_embd = ( - model["model.embed_tokens.weight"].shape - if "model.embed_tokens.weight" in model - else model["tok_embeddings.weight"].shape - ) + n_vocab, n_embd = model["model.embed_tokens.weight"].shape if "model.embed_tokens.weight" in model else model["tok_embeddings.weight"].shape # try transformer naming first if "model.layers.0.self_attn.q_proj.weight" in model: - n_layer = next( - i - for i in itertools.count() - if f"model.layers.{i}.self_attn.q_proj.weight" not in model - ) - elif ( - "model.layers.0.self_attn.W_pack.weight" in model - ): # next: try baichuan naming - n_layer = next( - i - for i in itertools.count() - if f"model.layers.{i}.self_attn.W_pack.weight" not in model - ) + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) + elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) else: - n_layer = next( - i - for i in itertools.count() - if f"layers.{i}.attention.wq.weight" not in model - ) + n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) if n_layer < 1: - raise Exception( - "failed to guess 'n_layer'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files." - ) + raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files.") - n_head = n_embd // 128 # guessed - n_mult = 256 # guessed + n_head = n_embd // 128 # guessed + n_mult = 256 # guessed # TODO: verify this n_ff = int(2 * (4 * n_embd) / 3) n_ff = n_mult * ((n_ff + n_mult - 1) // n_mult) return Params( - n_vocab=n_vocab, - n_embd=n_embd, - n_layer=n_layer, - n_ctx=-1, - n_ff=n_ff, - n_head=n_head, - n_head_kv=n_head, - f_norm_eps=1e-5, + n_vocab = n_vocab, + n_embd = n_embd, + n_layer = n_layer, + n_ctx = -1, + n_ff = n_ff, + n_head = n_head, + n_head_kv = n_head, + f_norm_eps = 1e-5, ) @staticmethod - def load_transformers_config(model: LazyModel, config_path: Path) -> "Params": + def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None @@ -274,22 +223,20 @@ class Params: rope_scaling_type = gguf.RopeScalingType.LINEAR elif typ == "yarn": rope_scaling_type = gguf.RopeScalingType.YARN - n_orig_ctx = rope_scaling["original_max_position_embeddings"] - rope_finetuned = rope_scaling["finetuned"] + n_orig_ctx = rope_scaling['original_max_position_embeddings'] + rope_finetuned = rope_scaling['finetuned'] else: - raise NotImplementedError(f"Unknown rope scaling type: {typ}") + raise NotImplementedError(f'Unknown rope scaling type: {typ}') if "max_sequence_length" in config: n_ctx = config["max_sequence_length"] elif "max_position_embeddings" in config: n_ctx = config["max_position_embeddings"] else: - raise Exception( - "failed to guess 'n_ctx'. This model is unknown or unsupported.\n" - "Suggestion: provide 'config.json' of the model in the same directory containing model files." - ) + raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n" + "Suggestion: provide 'config.json' of the model in the same directory containing model files.") - n_experts = None + n_experts = None n_experts_used = None if "num_local_experts" in config: @@ -297,30 +244,30 @@ class Params: n_experts_used = config["num_experts_per_tok"] return Params( - n_vocab=config["vocab_size"], - n_embd=config["hidden_size"], - n_layer=config["num_hidden_layers"], - n_ctx=n_ctx, - n_ff=config["intermediate_size"], - n_head=(n_head := config["num_attention_heads"]), - n_head_kv=config.get("num_key_value_heads", n_head), - n_experts=n_experts, - n_experts_used=n_experts_used, - f_norm_eps=config["rms_norm_eps"], - f_rope_freq_base=config.get("rope_theta"), - rope_scaling_type=rope_scaling_type, - f_rope_scale=f_rope_scale, - n_orig_ctx=n_orig_ctx, - rope_finetuned=rope_finetuned, + n_vocab = config["vocab_size"], + n_embd = config["hidden_size"], + n_layer = config["num_hidden_layers"], + n_ctx = n_ctx, + n_ff = config["intermediate_size"], + n_head = (n_head := config["num_attention_heads"]), + n_head_kv = config.get("num_key_value_heads", n_head), + n_experts = n_experts, + n_experts_used = n_experts_used, + f_norm_eps = config["rms_norm_eps"], + f_rope_freq_base = config.get("rope_theta"), + rope_scaling_type = rope_scaling_type, + f_rope_scale = f_rope_scale, + n_orig_ctx = n_orig_ctx, + rope_finetuned = rope_finetuned, ) # LLaMA v2 70B params.json # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1} @staticmethod - def load_torch_params(model: LazyModel, config_path: Path) -> "Params": + def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) - n_experts = None + n_experts = None n_experts_used = None f_rope_freq_base = None @@ -343,50 +290,50 @@ class Params: if config.get("moe"): n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0] - n_experts = config["moe"]["num_experts"] + n_experts = config["moe"]["num_experts"] n_experts_used = config["moe"]["num_experts_per_tok"] f_rope_freq_base = 1e6 return Params( - n_vocab=model["tok_embeddings.weight"].shape[0], - n_embd=config["dim"], - n_layer=config["n_layers"], - n_ctx=n_ctx, - n_ff=n_ff, - n_head=(n_head := config["n_heads"]), - n_head_kv=config.get("n_kv_heads", n_head), - n_experts=n_experts, - n_experts_used=n_experts_used, - f_norm_eps=config["norm_eps"], - f_rope_freq_base=config.get("rope_theta", f_rope_freq_base), + n_vocab = model["tok_embeddings.weight"].shape[0], + n_embd = config["dim"], + n_layer = config["n_layers"], + n_ctx = n_ctx, + n_ff = n_ff, + n_head = (n_head := config["n_heads"]), + n_head_kv = config.get("n_kv_heads", n_head), + n_experts = n_experts, + n_experts_used = n_experts_used, + f_norm_eps = config["norm_eps"], + f_rope_freq_base = config.get("rope_theta", f_rope_freq_base), ) @staticmethod - def load(model_plus: ModelPlus) -> "Params": - hf_config_path = model_plus.paths[0].parent / "config.json" + def load(model_plus: ModelPlus) -> Params: + hf_config_path = model_plus.paths[0].parent / "config.json" orig_config_path = model_plus.paths[0].parent / "params.json" if hf_config_path.exists(): - params = Params.load_transformers_config(model_plus.model, hf_config_path) + params = Params.loadHFTransformerJson(model_plus.model, hf_config_path) elif orig_config_path.exists(): - params = Params.load_torch_params(model_plus.model, orig_config_path) - elif model_plus.format != "none": + params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path) + elif model_plus.format != 'none': params = Params.guessed(model_plus.model) else: - raise ValueError("Cannot guess params when model format is none") + raise ValueError('Cannot guess params when model format is none') params.path_model = model_plus.paths[0].parent return params -class BpeVocab: # GPT - def __init__( - self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] - ) -> None: - self.bpe_tokenizer = json.loads( - open(str(fname_tokenizer), encoding="utf-8").read() - ) +# +# vocab +# + +class BpeVocab: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: + self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read()) self.vocab = self.bpe_tokenizer["model"]["vocab"] added_tokens: dict[str, int] if fname_added_tokens is not None: @@ -394,34 +341,31 @@ class BpeVocab: # GPT added_tokens = json.load(open(fname_added_tokens, encoding="utf-8")) else: # Fall back to trying to find the added tokens in tokenizer.json - tokenizer_json_file = fname_tokenizer.parent / "tokenizer.json" + tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json' if not tokenizer_json_file.is_file(): added_tokens = {} else: tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8")) added_tokens = dict( - (item["content"], item["id"]) - for item in tokenizer_json.get("added_tokens", []) + (item['content'], item['id']) + for item in tokenizer_json.get('added_tokens', []) # Added tokens here can be duplicates of the main vocabulary. - if item["content"] not in self.bpe_tokenizer - ) + if item['content'] not in self.bpe_tokenizer) vocab_size: int = len(self.vocab) - expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) - actual_ids = sorted(added_tokens.values()) + expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) + actual_ids = sorted(added_tokens.values()) if expected_ids != actual_ids: expected_end_id = vocab_size + len(actual_ids) - 1 - raise Exception( - f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}" - ) + raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}") items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1]) - self.added_tokens_dict = added_tokens - self.added_tokens_list = [text for (text, idx) in items] + self.added_tokens_dict = added_tokens + self.added_tokens_list = [text for (text, idx) in items] self.vocab_size_base: int = vocab_size - self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer - self.fname_added_tokens = fname_added_tokens + self.vocab_size: int = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer + self.fname_added_tokens = fname_added_tokens def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()} @@ -442,10 +386,8 @@ class BpeVocab: # GPT return f"" -class SentencePieceVocab: # LlaMa - def __init__( - self, fname_tokenizer: Path, fname_added_tokens: Optional[Path] - ) -> None: +class SentencePieceVocab: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None: self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer)) added_tokens: dict[str, int] if fname_added_tokens is not None: @@ -455,23 +397,19 @@ class SentencePieceVocab: # LlaMa vocab_size: int = self.sentencepiece_tokenizer.vocab_size() - new_tokens = { - id: piece for piece, id in added_tokens.items() if id >= vocab_size - } + new_tokens = {id: piece for piece, id in added_tokens.items() if id >= vocab_size} expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens))) - actual_new_ids = sorted(new_tokens.keys()) + actual_new_ids = sorted(new_tokens.keys()) if expected_new_ids != actual_new_ids: - raise ValueError( - f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}" - ) + raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}") # Token pieces that were added to the base vocabulary. self.added_tokens_dict = added_tokens - self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] - self.vocab_size_base = vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer + self.added_tokens_list = [new_tokens[id] for id in actual_new_ids] + self.vocab_size_base = vocab_size + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: @@ -512,11 +450,15 @@ class SentencePieceVocab: # LlaMa class HfVocab: - def __init__( - self, - fname_tokenizer: Path, - fname_added_tokens: Optional[Path] = None, - ) -> None: + def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None: + try: + from transformers import AutoTokenizer + except ImportError as e: + raise ImportError( + "To use HfVocab, please install the `transformers` package. " + "You can install it with `pip install transformers`." + ) from e + print("fname_tokenizer:", fname_tokenizer) # Allow the tokenizer to default to slow or fast versions. # Explicitly set tokenizer to use local paths. @@ -529,7 +471,7 @@ class HfVocab: # Initialize lists and dictionaries for added tokens self.added_tokens_list = [] self.added_tokens_dict = dict() - self.added_tokens_ids = set() + self.added_tokens_ids = set() # Process added tokens for tok, tokidx in sorted( @@ -550,12 +492,12 @@ class HfVocab: # Set vocabulary sizes self.vocab_size_base = self.tokenizer.vocab_size - self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) + self.vocab_size = self.vocab_size_base + len(self.added_tokens_list) - self.fname_tokenizer = fname_tokenizer + self.fname_tokenizer = fname_tokenizer self.fname_added_tokens = fname_added_tokens - def hf_tokens(self) -> Iterable[Tuple[bytes, float, gguf.TokenType]]: + def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: reverse_vocab = { id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items() } @@ -573,11 +515,9 @@ class HfVocab: token_id, self.special_ids # Reuse already stored special IDs ) - def get_token_type(self, token_id: int, special_ids: set) -> gguf.TokenType: + def get_token_type(self, token_id: int, special_ids: set[int]) -> gguf.TokenType: # Determine token type based on whether it's a special token - return ( - gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL - ) + return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL def get_token_score(self, token_id: int) -> float: # Placeholder for actual logic to determine the token's score @@ -589,7 +529,6 @@ class HfVocab: if text in self.specials: toktype = self.get_token_type(self.specials[text], self.special_ids) score = self.get_token_score(self.specials[text]) - else: toktype = gguf.TokenType.USER_DEFINED score = -1000.0 @@ -783,7 +722,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus: else: model = merge_sharded([mp.model for mp in models_plus]) - return ModelPlus(model, paths, format, vocab) + return ModelPlus(model, paths, format, vocab) # pytype: disable=wrong-arg-types def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor: @@ -871,17 +810,13 @@ class LazyUnpickler(pickle.Unpickler): CLASSES: dict[tuple[str, str], Any] = { # getattr used here as a workaround for mypy not being smart enough to determine # the staticmethods have a __func__ attribute. - ("torch._tensor", "_rebuild_from_type_v2"): getattr( - rebuild_from_type_v2, "__func__" - ), - ("torch._utils", "_rebuild_tensor_v2"): getattr( - lazy_rebuild_tensor_v2, "__func__" - ), - ("torch", "BFloat16Storage"): LazyStorageKind(DT_BF16), - ("torch", "HalfStorage"): LazyStorageKind(DT_F16), - ("torch", "FloatStorage"): LazyStorageKind(DT_F32), - ("torch", "IntStorage"): LazyStorageKind(DT_I32), - ("torch", "Tensor"): LazyTensor, + ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'), + ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'), + ('torch', 'BFloat16Storage'): LazyStorageKind(DT_BF16), + ('torch', 'HalfStorage'): LazyStorageKind(DT_F16), + ('torch', 'FloatStorage'): LazyStorageKind(DT_F32), + ('torch', 'IntStorage'): LazyStorageKind(DT_I32), + ('torch', 'Tensor'): LazyTensor, } def find_class(self, module: str, name: str) -> Any: @@ -968,7 +903,7 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc executor_class = ProcessPoolExecutor else: executor_class = ThreadPoolExecutor - with executor_class(max_workers = max_workers) as executor: + with executor_class(max_workers=max_workers) as executor: futures: list[concurrent.futures.Future[Out]] = [] done = False for _ in range(concurrency): @@ -1022,12 +957,8 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N class OutputFile: - def __init__( - self, fname_out: Path, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE - ) -> None: - self.gguf = gguf.GGUFWriter( - fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess - ) + def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: + self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) def add_meta_arch(self, params: Params) -> None: name = "LLaMA" @@ -1036,21 +967,16 @@ class OutputFile: if params.n_ctx == 4096: name = "LLaMA v2" elif params.path_model is not None: - name = str(params.path_model.parent).split("/")[-1] + name = str(params.path_model.parent).split('/')[-1] - self.gguf.add_name(name) - self.gguf.add_context_length(params.n_ctx) - self.gguf.add_embedding_length(params.n_embd) - self.gguf.add_block_count(params.n_layer) - self.gguf.add_feed_forward_length(params.n_ff) + self.gguf.add_name (name) + self.gguf.add_context_length (params.n_ctx) + self.gguf.add_embedding_length (params.n_embd) + self.gguf.add_block_count (params.n_layer) + self.gguf.add_feed_forward_length (params.n_ff) self.gguf.add_rope_dimension_count(params.n_embd // params.n_head) - self.gguf.add_head_count(params.n_head) - self.gguf.add_head_count_kv(params.n_head_kv) - - if params.f_norm_eps is None: - raise ValueError("f_norm_eps is None") - - self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) + self.gguf.add_head_count (params.n_head) + self.gguf.add_head_count_kv (params.n_head_kv) if params.n_experts: self.gguf.add_expert_count(params.n_experts) @@ -1058,6 +984,11 @@ class OutputFile: if params.n_experts_used: self.gguf.add_expert_used_count(params.n_experts_used) + if params.f_norm_eps: + self.gguf.add_layer_norm_rms_eps(params.f_norm_eps) + else: + raise ValueError('f_norm_eps is None') + if params.f_rope_freq_base is not None: self.gguf.add_rope_freq_base(params.f_rope_freq_base) @@ -1089,7 +1020,7 @@ class OutputFile: return tokenizer_model - def extract_vocabulary_from_model(self, vocab: Vocab) -> Tuple[list, list, list]: + def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]: tokens = [] scores = [] toktypes = [] @@ -1124,14 +1055,10 @@ class OutputFile: def add_tensor_info(self, name: str, tensor: LazyTensor) -> None: n_elements = int(np.prod(tensor.shape)) - raw_dtype = getattr(tensor.data_type, "ggml_type", None) - data_type = ( - getattr(tensor.data_type, "quantized_type", None) or tensor.data_type.dtype - ) + raw_dtype = getattr(tensor.data_type, 'ggml_type', None) + data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype data_nbytes = tensor.data_type.elements_to_bytes(n_elements) - self.gguf.add_tensor_info( - name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype - ) + self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype) def write_meta(self) -> None: self.gguf.write_header_to_file() @@ -1145,14 +1072,10 @@ class OutputFile: @staticmethod def write_vocab_only( - fname_out: Path, - params: Params, - vocab: Vocab, - svocab: gguf.SpecialVocab, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, - pad_vocab: bool = False, + fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, + endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: - check_vocab_size(params, vocab, pad_vocab=pad_vocab) + check_vocab_size(params, vocab, pad_vocab = pad_vocab) of = OutputFile(fname_out, endianess=endianess) @@ -1180,14 +1103,8 @@ class OutputFile: @staticmethod def write_all( - fname_out: Path, - ftype: GGMLFileType, - params: Params, - model: LazyModel, - vocab: Vocab, - svocab: gguf.SpecialVocab, - concurrency: int = DEFAULT_CONCURRENCY, - endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, + fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, + concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, ) -> None: check_vocab_size(params, vocab, pad_vocab=pad_vocab) @@ -1207,26 +1124,19 @@ class OutputFile: of.write_tensor_info() # tensor data - ndarrays_inner = bounded_parallel_map( - OutputFile.do_item, model.items(), concurrency=concurrency - ) + ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency) if ftype == GGMLFileType.MostlyQ8_0: ndarrays = bounded_parallel_map( - OutputFile.maybe_do_quantize, - ndarrays_inner, - concurrency=concurrency, - max_workers=concurrency, + OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency, use_processpool_executor=True, ) else: ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner) start = time.time() - for i, ((name, lazy_tensor), ndarray) in enumerate( - zip(model.items(), ndarrays) - ): + for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)): elapsed = time.time() - start - size = " x ".join(f"{dim:6d}" for dim in lazy_tensor.shape) + size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape) padi = len(str(len(model))) print( f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}" @@ -1363,7 +1273,7 @@ def load_some_model(path: Path) -> ModelPlus: class VocabFactory: def __init__(self, path: Path): self.path = path - self.files = { + self.files: dict[str, Path | None] = { "tokenizer.model": None, "vocab.json": None, "tokenizer.json": None, @@ -1380,24 +1290,18 @@ class VocabFactory: self.files[file] = parent_file_path print(f"Found vocab files: {self.files}") - def _select_file(self, vocabtype: Optional[str]) -> Path: + def _select_file(self, vocabtype: str | None) -> Path: if vocabtype in ["spm", "bpe"]: for file_key in self.files.keys(): - if self.files[file_key]: - return self.files[file_key] + if (file := self.files[file_key]) is not None: + return file raise FileNotFoundError(f"{vocabtype} vocab not found.") - elif vocabtype == "hfft": + if vocabtype == "hfft": # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file return self.path - else: - raise ValueError(f"Unsupported vocabulary type {vocabtype}") + raise ValueError(f"Unsupported vocabulary type {vocabtype}") - def _create_special_vocab( - self, - vocab: Vocab, - vocabtype: str, - model_parent_path: Path, - ) -> gguf.SpecialVocab: + def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab: load_merges = vocabtype == "bpe" n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None return gguf.SpecialVocab( @@ -1407,13 +1311,12 @@ class VocabFactory: n_vocab=n_vocab, ) - def load_vocab( - self, vocabtype: str, model_parent_path: Path - ) -> Tuple[Vocab, gguf.SpecialVocab]: + def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]: path = self._select_file(vocabtype) print(f"Loading vocab file '{path}', type '{vocabtype}'") added_tokens_path = path.parent / "added_tokens.json" + vocab: Vocab if vocabtype == "bpe": vocab = BpeVocab( path, added_tokens_path if added_tokens_path.exists() else None @@ -1428,6 +1331,7 @@ class VocabFactory: ) else: raise ValueError(f"Unsupported vocabulary type {vocabtype}") + # FIXME: Respect --vocab-dir? special_vocab = self._create_special_vocab( vocab, vocabtype, @@ -1436,18 +1340,17 @@ class VocabFactory: return vocab, special_vocab -def default_output_file(model_paths: list[Path], file_type: GGMLFileType) -> Path: +def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path: namestr = { - GGMLFileType.AllF32: "f32", + GGMLFileType.AllF32: "f32", GGMLFileType.MostlyF16: "f16", - GGMLFileType.MostlyQ8_0: "q8_0", + GGMLFileType.MostlyQ8_0:"q8_0", }[file_type] ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf" if ret in model_paths: sys.stderr.write( f"Error: Default output path ({ret}) would overwrite the input. " - "Please explicitly specify a path using --outfile.\n" - ) + "Please explicitly specify a path using --outfile.\n") sys.exit(1) return ret @@ -1457,111 +1360,34 @@ def do_dump_model(model_plus: ModelPlus) -> None: print(f"model_plus.format = {model_plus.format!r}") print(f"model_plus.vocab = {model_plus.vocab!r}") for name, lazy_tensor in model_plus.model.items(): - print( - f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}" - ) + print(f"{name}: shape={lazy_tensor.shape} type={lazy_tensor.data_type}; {lazy_tensor.description}") -def get_argument_parser() -> ArgumentParser: +def main(args_in: list[str] | None = None) -> None: output_choices = ["f32", "f16"] if np.uint32(1) == np.uint32(1).newbyteorder("<"): # We currently only support Q8_0 output on little endian systems. output_choices.append("q8_0") + vocab_types = ["spm", "bpe", "hfft"] + parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") + parser.add_argument("--awq-path", type=Path, help="Path to scale awq cache file", default=None) + parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") + parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") + parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") + parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") + parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") + parser.add_argument("--vocab-type", choices=vocab_types, help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm") + parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") + parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") + parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY) + parser.add_argument("--big-endian", action="store_true", help="model is executed on big endian machine") + parser.add_argument("--pad-vocab", action="store_true", help="add pad tokens when model vocab expects more than tokenizer metadata provides") - parser = argparse.ArgumentParser( - description="Convert a LLaMa model to a GGML compatible file" - ) - - parser.add_argument( - "model", - type=Path, - help="Directory containing the model file or the model file itself (*.pth, *.pt, *.bin)", - ) - - parser.add_argument( - "--awq-path", - type=Path, - help="Path to the Activation-aware Weight Quantization cache file", - default=None, - ) - - parser.add_argument( - "--dump", - action="store_true", - help="Display the model content without converting it", - ) - - parser.add_argument( - "--dump-single", - action="store_true", - help="Display the content of a single model file without conversion", - ) - - parser.add_argument( - "--vocab-only", - action="store_true", - help="Extract and output only the vocabulary", - ) - - parser.add_argument( - "--outtype", - choices=output_choices, - help="Output format - note: q8_0 may be very slow (default: f16 or f32 based on input)", - ) - - parser.add_argument( - "--vocab-dir", - type=Path, - help="Directory containing the tokenizer.model, if separate from the model file", - ) - - parser.add_argument( - "--vocab-type", - choices=["spm", "bpe", "hfft"], # hfft: Hugging Face Fast Tokenizer - default="spm", - help="The vocabulary format used to define the tokenizer model (default: spm)", - ) - - parser.add_argument( - "--pad-vocab", - action="store_true", - help="Add padding tokens when the model's vocabulary size exceeds the tokenizer metadata", - ) - - parser.add_argument( - "--outfile", - type=Path, - help="Specify the path for the output file (default is based on input)", - ) - - parser.add_argument( - "--ctx", type=int, help="Model training context (default is based on input)" - ) - - parser.add_argument( - "--concurrency", - type=int, - help=f"Concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", - default=DEFAULT_CONCURRENCY, - ) - - parser.add_argument( - "--big-endian", - action="store_true", - help="Indicate that the model is executed on a big-endian machine", - ) - - return parser - - -def main(argv: Optional[list[str]] = None) -> None: - parser = get_argument_parser() - args = parser.parse_args(argv) - + args = parser.parse_args(args_in) if args.awq_path: - sys.path.insert(1, str(Path(__file__).resolve().parent / "awq-py")) - from awq.apply_awq import add_scale_weights - + sys.path.insert(1, str(Path(__file__).parent / 'awq-py')) + from awq.apply_awq import add_scale_weights # type: ignore[import-not-found] tmp_model_path = args.model / "weighted_model" if tmp_model_path.is_dir(): print(f"{tmp_model_path} exists as a weighted model.") @@ -1580,14 +1406,11 @@ def main(argv: Optional[list[str]] = None) -> None: if not args.vocab_only: model_plus = load_some_model(args.model) else: - model_plus = ModelPlus( - model={}, paths=[args.model / "dummy"], format="none", vocab=None - ) + model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None) if args.dump: do_dump_model(model_plus) return - endianess = gguf.GGUFEndian.LITTLE if args.big_endian: endianess = gguf.GGUFEndian.BIG @@ -1595,12 +1418,10 @@ def main(argv: Optional[list[str]] = None) -> None: params = Params.load(model_plus) if params.n_ctx == -1: if args.ctx is None: - raise Exception( - "The model doesn't have a context size, and you didn't specify one with --ctx\n" - "Please specify one with --ctx:\n" - " - LLaMA v1: --ctx 2048\n" - " - LLaMA v2: --ctx 4096\n" - ) + raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n" + "Please specify one with --ctx:\n" + " - LLaMA v1: --ctx 2048\n" + " - LLaMA v2: --ctx 4096\n") params.n_ctx = args.ctx if args.outtype: @@ -1621,42 +1442,30 @@ def main(argv: Optional[list[str]] = None) -> None: if not args.outfile: raise ValueError("need --outfile if using --vocab-only") outfile = args.outfile - OutputFile.write_vocab_only( - outfile, - params, - vocab, - special_vocab, - endianess=endianess, - pad_vocab=args.pad_vocab, - ) + OutputFile.write_vocab_only(outfile, params, vocab, special_vocab, + endianess=endianess, pad_vocab=args.pad_vocab) print(f"Wrote {outfile}") return if model_plus.vocab is not None and args.vocab_dir is None: vocab = model_plus.vocab - model = model_plus.model - model = convert_model_names(model, params) - ftype = pick_output_type(model, args.outtype) - model = convert_to_output_type(model, ftype) - outfile = args.outfile or default_output_file(model_plus.paths, ftype) + print(f"Vocab info: {vocab}") + print(f"Special vocab info: {special_vocab}") + + model = model_plus.model + model = convert_model_names(model, params) + ftype = pick_output_type(model, args.outtype) + model = convert_to_output_type(model, ftype) + outfile = args.outfile or default_outfile(model_plus.paths, ftype) params.ftype = ftype print(f"Writing {outfile}, format {ftype}") - OutputFile.write_all( - outfile, - ftype, - params, - model, - vocab, - special_vocab, - concurrency=args.concurrency, - endianess=endianess, - pad_vocab=args.pad_vocab, - ) + OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, + concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab) print(f"Wrote {outfile}") -if __name__ == "__main__": - main(sys.argv[1:]) # Exclude the first element (script name) from sys.argv +if __name__ == '__main__': + main() diff --git a/mypy.ini b/mypy.ini index 7215a05dd..e51910ca7 100644 --- a/mypy.ini +++ b/mypy.ini @@ -4,3 +4,4 @@ allow_untyped_calls = true allow_untyped_defs = true allow_incomplete_defs = true disable_error_code = import-untyped +warn_return_any = false From 942c0107a7301434c0a5e7da46bc4cf2393aa556 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 21 Jan 2024 05:17:27 +0200 Subject: [PATCH 038/101] flake.lock: Update (#5054) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/9b19f5e77dd906cb52dade0b7bd280339d2a1f3d' (2024-01-13) → 'github:NixOS/nixpkgs/bbe7d8f876fbbe7c959c90ba2ae2852220573261' (2024-01-19) Co-authored-by: github-actions[bot] --- flake.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/flake.lock b/flake.lock index cd532ef4f..1b253cb44 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1705133751, - "narHash": "sha256-rCIsyE80jgiOU78gCWN3A0wE0tR2GI5nH6MlS+HaaSQ=", + "lastModified": 1705677747, + "narHash": "sha256-eyM3okYtMgYDgmYukoUzrmuoY4xl4FUujnsv/P6I/zI=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "9b19f5e77dd906cb52dade0b7bd280339d2a1f3d", + "rev": "bbe7d8f876fbbe7c959c90ba2ae2852220573261", "type": "github" }, "original": { From 726c0fa9a2da976e9c5d5c51e185d9dd453fc9e5 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 21 Jan 2024 08:01:20 +0200 Subject: [PATCH 039/101] Slightly faster imatrix (#5050) * imatrix: speedup by avoiding unnecessary allocations and copies * imatrix: add --no-ppl option to skip PPL calculations altogether --------- Co-authored-by: Iwan Kawrakow --- examples/imatrix/imatrix.cpp | 74 ++++++++++++++++++++++++------------ 1 file changed, 49 insertions(+), 25 deletions(-) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 5a3d30b88..5687476cd 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -248,7 +248,7 @@ static void process_logits( } } -static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { +static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const int n_ctx = llama_n_ctx(ctx); @@ -269,10 +269,12 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { } std::vector logit_history; - logit_history.resize(tokens.size()); - std::vector prob_history; - prob_history.resize(tokens.size()); + + if (compute_ppl) { + logit_history.resize(tokens.size()); + prob_history.resize(tokens.size()); + } const int n_chunk_max = tokens.size() / n_ctx; @@ -288,12 +290,17 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { std::vector workers(std::thread::hardware_concurrency() - 1); + const int num_batches = (n_ctx + n_batch - 1) / n_batch; + + std::vector logits; + if (compute_ppl && num_batches > 1) { + logits.reserve((size_t)n_ctx * n_vocab); + } + for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; const int end = start + n_ctx; - const int num_batches = (n_ctx + n_batch - 1) / n_batch; - std::vector logits; const auto t_start = std::chrono::high_resolution_clock::now(); @@ -321,8 +328,10 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { // restore the original token in case it was set to BOS tokens[batch_start] = token_org; - const auto * batch_logits = llama_get_logits(ctx); - logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + if (compute_ppl && num_batches > 1) { + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } } const auto t_end = std::chrono::high_resolution_clock::now(); @@ -338,25 +347,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); } - const int first = n_ctx/2; - process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); - count += n_ctx - first - 1; + if (compute_ppl) { + const int first = n_ctx/2; + const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + count += n_ctx - first - 1; - printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); - fflush(stdout); + printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); + fflush(stdout); + + logits.clear(); + } } printf("\n"); - nll2 /= count; - nll /= count; - const double ppl = exp(nll); - nll2 -= nll * nll; - if (nll2 > 0) { - nll2 = sqrt(nll2/(count-1)); - printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); - } else { - printf("Unexpected negative standard deviation of log(prob)\n"); + if (compute_ppl) { + nll2 /= count; + nll /= count; + const double ppl = exp(nll); + nll2 -= nll * nll; + if (nll2 > 0) { + nll2 = sqrt(nll2/(count-1)); + printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl); + } else { + printf("Unexpected negative standard deviation of log(prob)\n"); + } } return true; @@ -365,6 +381,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) { int main(int argc, char ** argv) { StatParams sparams; + bool compute_ppl = true; std::vector args; args.push_back(argv[0]); int iarg = 1; @@ -381,12 +398,19 @@ int main(int argc, char ** argv) { } else if (arg == "--verbosity") { sparams.verbosity = std::stoi(argv[++iarg]); + } else if (arg == "--no-ppl") { + compute_ppl = false; } else { args.push_back(argv[iarg]); } } if (iarg < argc) { - args.push_back(argv[iarg]); + std::string arg{argv[iarg]}; + if (arg == "--no-ppl") { + compute_ppl = false; + } else { + args.push_back(argv[iarg]); + } } gpt_params params; @@ -448,7 +472,7 @@ int main(int argc, char ** argv) { fprintf(stderr, "%s\n", get_system_info(params).c_str()); } - bool OK = compute_imatrix(ctx, params); + bool OK = compute_imatrix(ctx, params, compute_ppl); if (!OK) { return 1; } From 7dcbe39d36b76389f6c5cd3b151928472b7e22ff Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Sun, 21 Jan 2024 14:42:44 +0200 Subject: [PATCH 040/101] Add ability to evauate multiple choice tasks (#5047) * TruthfulQA: 1st attempt, does not look like it is working The same implementation can be used for HellaSwag as well, so I converted a HellaSwag validation dataset to the binary format used here and tested with that. The score is only around 50, so something is not quite right. * TruthfulQA: works but the result is bad I know it works because if I convert the HellaSwag validation data to the binary format used in the truthful_qa_score() function I get the exact same result as from the hellaswag_score() function. But I guess, the questions are tricky and the way I have done the combination of question + answer is very likely not the best. The TruthfulQA validation dataset contains 817 questions, with random chance result around 19%. With this version I get 29.1% for Mistral-7B and 55.2% for Mistral-7B-Instruct-v0.2. The HF leader board results for these two models are 42.2% and 68.3%, respectively. * TruthfulQA: fix random sample * TruthfulQA: prepare tasks in parallel for large test datasets * Rename truthful_qa to multiple_choice * Make MSVC happy I had forgotten that MSVC does not make constexpr's available inside a lambda. --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 31 +++ common/common.h | 3 + examples/perplexity/perplexity.cpp | 391 ++++++++++++++++++++++++++++- 3 files changed, 422 insertions(+), 3 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index ce20360a4..0e4b8bab2 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -203,6 +203,25 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { params.prompt_cache_all = true; } else if (arg == "--prompt-cache-ro") { params.prompt_cache_ro = true; + } else if (arg == "-bf" || arg == "--binary-file") { + if (++i >= argc) { + invalid_param = true; + break; + } + std::ifstream file(argv[i], std::ios::binary); + if (!file) { + fprintf(stderr, "error: failed to open file '%s'\n", argv[i]); + invalid_param = true; + break; + } + // store the external file name in params + params.prompt_file = argv[i]; + file.seekg(0, std::ios::end); + size_t size = file.tellg(); + file.seekg(0, std::ios::beg); + params.prompt.resize(size); + file.read((char *)params.prompt.data(), size); + fprintf(stderr, "Read %zu bytes from binary file %s\n", size, argv[i]); } else if (arg == "-f" || arg == "--file") { if (++i >= argc) { invalid_param = true; @@ -689,6 +708,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.winogrande_tasks = std::stoi(argv[i]); + } else if (arg == "--multiple-choice") { + params.multiple_choice = true; + } else if (arg == "--multiple-choice-tasks") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.multiple_choice_tasks = std::stoi(argv[i]); } else if (arg == "--ignore-eos") { params.ignore_eos = true; } else if (arg == "--no-penalize-nl") { @@ -888,6 +915,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --in-suffix STRING string to suffix after user inputs with (default: empty)\n"); printf(" -f FNAME, --file FNAME\n"); printf(" prompt file to start generation.\n"); + printf(" -bf FNAME, --binary-file FNAME\n"); + printf(" binary file containing multiple choice tasks.\n"); printf(" -n N, --n-predict N number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict); printf(" -c N, --ctx-size N size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx); printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); @@ -936,6 +965,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --hellaswag-tasks N number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks); printf(" --winogrande compute Winogrande score over random tasks from datafile supplied with -f\n"); printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); + printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n"); + printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); diff --git a/common/common.h b/common/common.h index 0ae9c18b3..c69ad7e94 100644 --- a/common/common.h +++ b/common/common.h @@ -108,6 +108,9 @@ struct gpt_params { bool winogrande = false; // compute Winogrande score over random tasks from datafile supplied in prompt size_t winogrande_tasks= 0; // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed + bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt + size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index f91f5795a..b7ef9a084 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -540,14 +540,14 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { // This is needed as usual for LLaMA models const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + // The tasks should be randomized so the score stabilizes quickly. + bool randomize_tasks = true; + // Number of tasks to use when computing the score if (params.hellaswag_tasks < hs_task_count) { hs_task_count = params.hellaswag_tasks; } - // The tasks should be randomized so the score stabilizes quickly. - bool randomize_tasks = true; - // The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now std::mt19937 rng(1); @@ -1031,6 +1031,389 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma); } +static bool deserialize_string(std::istream& in, std::string& str) { + uint32_t size; + if (!in.read((char *)&size, sizeof(size)).fail()) { + str.resize(size); + if (!in.read((char *)str.data(), size).fail()) return true; + } + return false; +} + +struct multiple_choice_answers { + std::vector answers; + std::vector labels; + bool deserialize(std::istream& in) { + uint32_t n; + in.read((char *)&n, sizeof(n)); + if (in.fail() || n > 100) return false; // 100 as max. number of answers should be good enough for any practical purpose + answers.resize(n); + labels.resize(n); + for (auto& a : answers) { + if (!deserialize_string(in, a)) return false; + } + in.read((char *)labels.data(), n*sizeof(int)); + return !in.fail(); + } +}; + +struct multiple_choice_task { + std::string question; // the question (or context that needs to be continued) + multiple_choice_answers mc1; // possible answers (continuations) with a single correct answer + multiple_choice_answers mc2; // possible answers (continuations) with multiple correct answers - not handled yet + bool deserialize(std::istream& in) { + if (!deserialize_string(in, question)) return false; + return mc1.deserialize(in) && mc2.deserialize(in); + } + + // For evaluation + size_t i_batch; // starting index in the llama_batch + size_t common_prefix; // max number of initial tokens that are the same in all sentences + size_t required_tokens; // needed number of tokens to evaluate all answers + std::vector> seq_tokens; + std::vector log_probs; +}; + +static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos, multiple_choice_task& task, bool log_error) { + if (task.question.empty() || task.mc1.answers.empty()) { + if (log_error) { + printf("%s: found bad task with empty question and/or answers\n", __func__); + } + return false; + } + task.seq_tokens.reserve(task.mc1.answers.size()); + for (auto& answer : task.mc1.answers) { + if (answer.empty()) { + if (log_error) { + printf("%s: found empty answer\n", __func__); + } + return false; + } + task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, add_bos)); + } + auto min_len = task.seq_tokens.front().size(); + for (auto& seq : task.seq_tokens) { + min_len = std::min(min_len, seq.size()); + } + task.common_prefix = 0; + for (size_t k = 0; k < min_len; ++k) { + auto token = task.seq_tokens[0][k]; + bool all_same = true; + for (size_t i = 1; i < task.seq_tokens.size(); ++i) { + if (task.seq_tokens[i][k] != token) { + all_same = false; + break; + } + } + if (!all_same) { + break; + } + ++task.common_prefix; + } + task.required_tokens = task.common_prefix; + for (auto& seq : task.seq_tokens) { + task.required_tokens += seq.size() - task.common_prefix; + } + return true; +} + +// +// Calculates score for multiple choice tasks with single correct answer from prompt. +// Commonly used LLM evaluation metrics of this type are +// * ARC +// * HellaSwag +// * MMLU +// * TruthfulQA +// +// Validation datasets for these 4 tests can be found at +// https://huggingface.co/datasets/ikawrakow/validation-datasets-for-llama.cpp +// The data for these datasets was extracted from +// git@hf.co:datasets/allenai/ai2_arc +// https://github.com/rowanz/hellaswag/blob/master/data/hellaswag_val.jsonl +// git@hf.co:datasets/Stevross/mmlu +// https://huggingface.co/datasets/truthful_qa +// +static void multiple_choice_score(llama_context * ctx, const gpt_params & params) { + + std::istringstream strstream(params.prompt); + uint32_t n_task; + strstream.read((char *)&n_task, sizeof(n_task)); + if (strstream.fail() || n_task == 0) { + printf("%s: no tasks\n", __func__); + return; + } + printf("%s: there are %u tasks in prompt\n", __func__, n_task); + std::vector task_pos(n_task); + strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t)); + if (strstream.fail()) { + printf("%s: failed to raad task positions from prompt\n", __func__); + return; + } + + std::vector tasks; + if (params.multiple_choice_tasks == 0 || params.multiple_choice_tasks >= (size_t)n_task) { + // Use all tasks + tasks.resize(n_task); + printf("%s: reading tasks", __func__); + int n_dot = n_task/100; + int i = 0; + for (auto& task : tasks) { + ++i; + if (!task.deserialize(strstream)) { + printf("%s: failed to read task %d of %u\n", __func__, i, n_task); + return; + } + if (i%n_dot == 0) printf("."); + } + printf("done\n"); + } + else { + printf("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task); + std::mt19937 rng(1); + std::vector aux(n_task); + for (uint32_t i = 0; i < n_task; ++i) aux[i] = i; + float scale = 1.f/(1.f + (float)std::mt19937::max()); + tasks.resize(params.multiple_choice_tasks); + for (auto& task : tasks) { + int j = (int)(scale * rng() * aux.size()); + int idx = aux[j]; + aux[j] = aux.back(); + aux.pop_back(); + strstream.seekg(task_pos[idx], std::ios::beg); + if (!task.deserialize(strstream)) { + printf("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]); + return; + } + } + n_task = params.multiple_choice_tasks; + } + + // This is needed as usual for LLaMA models + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + + printf("%s: preparing task data", __func__); + fflush(stdout); + if (n_task > 500) { + printf("..."); + fflush(stdout); + std::atomic counter(0); + std::atomic n_bad(0); + auto prepare = [&counter, &n_bad, &tasks, ctx, add_bos] () { + int num_tasks = tasks.size(); + int n_bad_local = 0; + while (true) { + int first = counter.fetch_add(K_TOKEN_CHUNK); + if (first >= num_tasks) { + if (n_bad_local > 0) n_bad += n_bad_local; + break; + } + int last = std::min(first + K_TOKEN_CHUNK, num_tasks); + for (int i = first; i < last; ++i) { + if (!multiple_choice_prepare_one_task(ctx, add_bos, tasks[i], false)) ++n_bad_local; + } + } + }; + size_t max_thread = std::thread::hardware_concurrency(); + max_thread = std::min(max_thread, (tasks.size() + K_TOKEN_CHUNK - 1)/K_TOKEN_CHUNK); + std::vector workers(max_thread-1); + for (auto& w : workers) w = std::thread(prepare); + prepare(); + for (auto& w : workers) w.join(); + printf("done\n"); + fflush(stdout); + int nbad = n_bad; + if (nbad > 0) { + printf("%s: found %d malformed tasks\n", __func__, nbad); + return; + } + } else { + int n_dot = n_task/100; + int i_task = 0; + for (auto& task : tasks) { + ++i_task; + if (!multiple_choice_prepare_one_task(ctx, add_bos, task, true)) { + return; + } + if (i_task%n_dot == 0) { + printf("."); + fflush(stdout); + } + } + printf("done\n"); + } + + printf("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size()); + + printf("\ntask\tacc_norm\n"); + + const int n_vocab = llama_n_vocab(llama_get_model(ctx)); + const int n_ctx = llama_n_ctx(ctx); + const int n_batch = params.n_batch; + + const int max_tasks_per_batch = 32; + const int max_seq = 4*max_tasks_per_batch; + + llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); + + std::vector tok_logits(n_vocab); + std::vector batch_logits(n_vocab*n_ctx); + + std::vector> eval_pairs; + std::vector eval_results; + std::vector workers(std::thread::hardware_concurrency()); + std::vector batch_indeces; + + int n_done = 0; + int n_correct = 0; + int n_tot_answers = 0; + + for (size_t i0 = 0; i0 < tasks.size(); i0++) { + int n_cur = 0; + + size_t i1 = i0; + size_t i_batch = 0; // this tells us where in `llama_batch` we are currently + + llama_batch_clear(batch); + + // batch as much tasks as possible into the available context + // each task has 4 unique seuqnce ids - one for each ending + // the common prefix is shared among the 4 sequences to save tokens + // we extract logits only from the last common token and from all ending tokens of each sequence + int s0 = 0; + while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) { + auto& cur_task = tasks[i1]; + + int num_answers = cur_task.seq_tokens.size(); + if (s0 + num_answers > max_seq) { + break; + } + + if (int(batch_indeces.size()) != num_answers) { + batch_indeces.resize(num_answers); + } + for (int s = 0; s < num_answers; ++s) batch_indeces[s] = s0 + s; + + for (size_t i = 0; i < cur_task.common_prefix; ++i) { + //llama_batch_add(batch, cur_task.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false); + llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false); + } + batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix + + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) { + llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true); + } + } + + s0 += num_answers; + + cur_task.i_batch = i_batch; + i_batch += cur_task.required_tokens; + + n_cur += cur_task.required_tokens; + if (++i1 == tasks.size()) { + break; + } + } + + if (i0 == i1) { + fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0); + return; + } + + llama_kv_cache_clear(ctx); + + // decode all tasks [i0, i1) + if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) { + fprintf(stderr, "%s: llama_decode() failed\n", __func__); + return; + } + + // Compute log-probs in parallel + // First we collect all tasks + eval_pairs.clear(); + for (size_t i = i0; i < i1; ++i) { + auto& cur_task = tasks[i]; + size_t li = cur_task.common_prefix; + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) { + eval_pairs.push_back(std::make_pair(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1])); + } + ++li; + } + } + // Then we do the actual calculation + compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results); + + size_t ir = 0; + + // compute the logprobs for each ending of the decoded tasks + for (size_t i = i0; i < i1; ++i) { + auto & cur_task = tasks[i]; + //printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str()); + //for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) { + // if (cur_task.mc1.labels[j] == 1) { + // printf("%d", j+1); + // } + //} + //printf("\n common_prefix: %zu\n", cur_task.common_prefix); + + std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float)); + + const auto first_probs = softmax(tok_logits); + + cur_task.log_probs.resize(cur_task.seq_tokens.size()); + for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) { + size_t count = 1; + float log_prob = std::log(first_probs[cur_task.seq_tokens[s][cur_task.common_prefix]]); + for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) { + //printf(" %zu %g\n", ir, eval_results[ir]); + ++count; + log_prob += eval_results[ir++]; + } + cur_task.log_probs[s] = log_prob / count; + //printf(" Final: %g\n", log_prob / count); + //printf(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count); + } + + // Find the ending with maximum logprob + size_t logprob_max_idx = 0; + float logprob_max_val = cur_task.log_probs[0]; + for (size_t s = 1; s < cur_task.log_probs.size(); s++) { + if (cur_task.log_probs[s] > logprob_max_val) { + logprob_max_val = cur_task.log_probs[s]; + logprob_max_idx = s; + } + } + + n_tot_answers += cur_task.log_probs.size(); + if (cur_task.mc1.labels[logprob_max_idx] == 1) { + ++n_correct; + } + ++n_done; + + // Print the accumulated accuracy mean x 100 + printf("%d\t%.8lf\n", n_done, 100.*n_correct/n_done); + fflush(stdout); + } + + i0 = i1 - 1; + } + + llama_batch_free(batch); + + if (n_done < 100) return; + + float p = 1.f*n_correct/n_done; + float sigma = sqrt(p*(1-p)/(n_done-1)); + printf("\n Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma); + p = 1.f*n_done/n_tot_answers; + sigma = sqrt(p*(1-p)/(n_done-1)); + printf("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma); + + printf("\n"); +} + int main(int argc, char ** argv) { gpt_params params; @@ -1091,6 +1474,8 @@ int main(int argc, char ** argv) { hellaswag_score(ctx, params); } else if (params.winogrande) { winogrande_score(ctx, params); + } else if (params.multiple_choice) { + multiple_choice_score(ctx, params); } else { results = perplexity(ctx, params); } From 6c5629d4d2d15557c38a0e609b30c1a42abad80d Mon Sep 17 00:00:00 2001 From: bobqianic <129547291+bobqianic@users.noreply.github.com> Date: Sun, 21 Jan 2024 15:17:35 +0000 Subject: [PATCH 041/101] add `#include ` to unicode.h (#5051) Co-authored-by: Jared Van Bortel --- unicode.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/unicode.h b/unicode.h index aeca879ea..844eff3da 100644 --- a/unicode.h +++ b/unicode.h @@ -2,8 +2,9 @@ #include #include -#include +#include #include +#include static const std::vector> digit_ranges = { {0x30, 0x39}, {0xB2, 0xB3}, {0xB9, 0xB9}, {0x660, 0x669}, {0x6F0, 0x6F9}, {0x7C0, 0x7C9}, {0x966, 0x96F}, {0x9E6, 0x9EF}, {0xA66, 0xA6F}, {0xAE6, 0xAEF}, {0xB66, 0xB6F}, {0xBE6, 0xBEF}, {0xC66, 0xC6F}, From 05490fad7f7f60ff2bed9ad05cd81b44e82ccde3 Mon Sep 17 00:00:00 2001 From: kuronekosaiko Date: Mon, 22 Jan 2024 00:28:14 +0800 Subject: [PATCH 042/101] add safetensors support to convert-lora-to-ggml.py (#5062) * add safetensors support to convert-lora-to-ggml.py * Update convert-lora-to-ggml.py Remove white space in line 69. --- convert-lora-to-ggml.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/convert-lora-to-ggml.py b/convert-lora-to-ggml.py index 4904bf128..9a9936dec 100755 --- a/convert-lora-to-ggml.py +++ b/convert-lora-to-ggml.py @@ -59,7 +59,14 @@ if __name__ == '__main__': input_model = os.path.join(sys.argv[1], "adapter_model.bin") output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin") - model = torch.load(input_model, map_location="cpu") + if os.path.exists(input_model): + model = torch.load(input_model, map_location="cpu") + else: + input_model = os.path.join(sys.argv[1], "adapter_model.safetensors") + # lazy import load_file only if lora is in safetensors format. + from safetensors.torch import load_file + model = load_file(input_model, device="cpu") + arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama" if arch_name not in gguf.MODEL_ARCH_NAMES.values(): From 504dc37be8446fb09b1ede70300250ad41be32a2 Mon Sep 17 00:00:00 2001 From: iSma Date: Sun, 21 Jan 2024 22:37:13 +0100 Subject: [PATCH 043/101] Revert LLAMA_NATIVE to OFF in flake.nix (#5066) --- .devops/nix/package.nix | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 43bdbd755..1ef2c6cd9 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -159,7 +159,7 @@ effectiveStdenv.mkDerivation ( cmakeFlags = [ - (cmakeBool "LLAMA_NATIVE" true) + (cmakeBool "LLAMA_NATIVE" false) (cmakeBool "LLAMA_BUILD_SERVER" true) (cmakeBool "BUILD_SHARED_LIBS" true) (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true) From 3466c6ebcf668cac453f891b493ead19283347a8 Mon Sep 17 00:00:00 2001 From: Shijie <821898965@qq.com> Date: Mon, 22 Jan 2024 15:33:19 +0800 Subject: [PATCH 044/101] llama : add more qwen2 models (#5071) --- llama.cpp | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/llama.cpp b/llama.cpp index 909ad4ad8..9ad74d735 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1325,8 +1325,10 @@ static llama_state g_state; // available llama models enum e_model { MODEL_UNKNOWN, + MODEL_0_5B, MODEL_1B, MODEL_3B, + MODEL_4B, MODEL_7B, MODEL_8B, MODEL_13B, @@ -2892,9 +2894,9 @@ static void llm_load_hparams( { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { - case 24: model.type = e_model::MODEL_1B; break; + case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_7B; break; - case 40: model.type = e_model::MODEL_13B; break; + case 40: model.type = hparams.n_head == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break; case 80: model.type = e_model::MODEL_70B; break; default: model.type = e_model::MODEL_UNKNOWN; } From 57744932c64266359ee905518de7e096c0295d8c Mon Sep 17 00:00:00 2001 From: bobqianic <129547291+bobqianic@users.noreply.github.com> Date: Mon, 22 Jan 2024 08:55:05 +0000 Subject: [PATCH 045/101] ci : fix Windows CI by updating Intel SDE version (#5053) --- .github/workflows/build.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 367df07a7..c3aa6f992 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -295,7 +295,7 @@ jobs: OPENBLAS_VERSION: 0.3.23 OPENCL_VERSION: 2023.04.17 CLBLAST_VERSION: 1.6.0 - SDE_VERSION: 9.21.1-2023-04-24 + SDE_VERSION: 9.33.0-2024-01-07 strategy: matrix: @@ -400,7 +400,7 @@ jobs: id: cmake_test_sde if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation run: | - curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz" + curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz" # for some weird reason windows tar doesn't like sde tar.xz 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar From 66d575c45c5a370d668f9c3283cdf348e2329fa2 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 12:43:33 +0200 Subject: [PATCH 046/101] llama : add Q3_K_XS (#5060) * Add Q3_K_XS - intermediate size between Q2_K and Q3_K_S * Q3_K_XS: quanize first 1/8 of ffn_down layers with Q4_K Together with an importance matrix, this brings perplexity for LLaMA-v2-70B below the perplexity of the former Q2_K with a 800 MB smaller quantized model size. --------- Co-authored-by: Iwan Kawrakow --- examples/quantize/quantize.cpp | 1 + llama.cpp | 62 +++++++++++++++++++++++++--------- llama.h | 1 + 3 files changed, 48 insertions(+), 16 deletions(-) diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index 2ae046933..f4786157e 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -26,6 +26,7 @@ static const std::vector QUANT_OPTIONS = { { "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", }, { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", }, { "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" }, + { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization" , }, { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", }, { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", }, { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", }, diff --git a/llama.cpp b/llama.cpp index 9ad74d735..c56d31163 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2661,6 +2661,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) { case LLAMA_FTYPE_MOSTLY_Q6_K: return "Q6_K"; case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw"; case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw"; + case LLAMA_FTYPE_MOSTLY_Q3_K_XS:return "Q3_K - Extra small"; default: return "unknown, may not work"; } @@ -8765,9 +8766,13 @@ struct quantize_state_internal { const llama_model_quantize_params * params; int n_attention_wv = 0; - int n_feed_forward_w2 = 0; + int n_ffn_down = 0; + int n_ffn_gate = 0; + int n_ffn_up = 0; int i_attention_wv = 0; - int i_feed_forward_w2 = 0; + int i_ffn_down = 0; + int i_ffn_gate = 0; + int i_ffn_up = 0; int n_k_quantized = 0; int n_fallback = 0; @@ -8870,8 +8875,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty ++qs.i_attention_wv; } else if (name.find("ffn_down") != std::string::npos) { - if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q2_K; - ++qs.i_feed_forward_w2; + if (qs.i_ffn_down < qs.n_ffn_down/8) new_type = GGML_TYPE_Q2_K; + ++qs.i_ffn_down; } else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K; } else if (name.find("attn_v.weight") != std::string::npos) { @@ -8908,18 +8913,21 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // TODO: explore better strategies new_type = GGML_TYPE_Q8_0; } + else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { + new_type = GGML_TYPE_Q2_K; + } } else if (name.find("ffn_down") != std::string::npos) { const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); int i_layer, n_layer; if (n_expert == 1) { - i_layer = qs.i_feed_forward_w2; - n_layer = qs.n_feed_forward_w2; + i_layer = qs.i_ffn_down; + n_layer = qs.n_ffn_down; } else { // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly - // sprinkled in the model. Hence, simply dividing i_feed_forward_w2 by n_expert does not work + // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work // for getting the current layer as I initially thought, and we need to resort to parsing the // tensor name. - n_layer = qs.n_feed_forward_w2 / n_expert; + n_layer = qs.n_ffn_down / n_expert; if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) { throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str())); } @@ -8928,7 +8936,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty } } if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; - else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) { + else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; } else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) { @@ -8958,11 +8966,12 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // same quantization as before imatrix stuff, and b) Q4_1/Q5_1 do go crazy on ffn_down without an imatrix. new_type = ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ? GGML_TYPE_Q4_1 : GGML_TYPE_Q5_1; } - ++qs.i_feed_forward_w2; + ++qs.i_ffn_down; } else if (name.find("attn_output.weight") != std::string::npos) { if (arch != LLM_ARCH_FALCON) { if (qs.model.hparams.n_expert == 8) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || + if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS || + ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) { new_type = GGML_TYPE_Q5_K; } @@ -8980,6 +8989,20 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } + else if (name.find("ffn_gate") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) { + new_type = GGML_TYPE_Q2_K; + } + ++qs.i_ffn_gate; + } + else if (name.find("ffn_up") != std::string::npos) { + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) { + new_type = GGML_TYPE_Q2_K; + } + ++qs.i_ffn_up; + } + // if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; + //} // IK: let's remove this, else Q2_K is almost the same as Q3_K_S //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) { // if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; @@ -9034,8 +9057,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break; // K-quants + case LLAMA_FTYPE_MOSTLY_Q2_K_S: case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break; - case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break; + case LLAMA_FTYPE_MOSTLY_Q3_K_XS: case LLAMA_FTYPE_MOSTLY_Q3_K_S: case LLAMA_FTYPE_MOSTLY_Q3_K_M: case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break; @@ -9103,12 +9127,18 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s ++qs.n_attention_wv; } else if (name.find("ffn_down") != std::string::npos) { - ++qs.n_feed_forward_w2; + ++qs.n_ffn_down; + } + else if (name.find("ffn_gate") != std::string::npos) { + ++qs.n_ffn_gate; + } + else if (name.find("ffn_up") != std::string::npos) { + ++qs.n_ffn_up; } } - if (qs.n_attention_wv != qs.n_feed_forward_w2 || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { - LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n", - __func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer); + if (qs.n_attention_wv != qs.n_ffn_down || (uint32_t)qs.n_attention_wv != model.hparams.n_layer) { + LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_ffn_down = %d, hparams.n_layer = %d\n", + __func__, qs.n_attention_wv, qs.n_ffn_down, model.hparams.n_layer); } size_t total_size_org = 0; diff --git a/llama.h b/llama.h index e268d7a1d..bb6054557 100644 --- a/llama.h +++ b/llama.h @@ -107,6 +107,7 @@ extern "C" { LLAMA_FTYPE_MOSTLY_IQ2_XXS = 19, // except 1d tensors LLAMA_FTYPE_MOSTLY_IQ2_XS = 20, // except 1d tensors LLAMA_FTYPE_MOSTLY_Q2_K_S = 21, // except 1d tensors + LLAMA_FTYPE_MOSTLY_Q3_K_XS = 22, // except 1d tensors LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; From 152d9d05e097e35f1cac21262946d57faec7542a Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Mon, 22 Jan 2024 12:11:01 +0100 Subject: [PATCH 047/101] finetune : print sample-start/include-sample-start (#5072) This commit adds `--sample-start` and `--include-sample-start` to the output from the main function in finetune.cpp. The motivation for this is that even though these are set explicitly by the user via the command line, if one forgets to set them then it is useful to have their values printed out. Otherwise it is possible to go through the whole training process before realizing that the values are not what one expected. Signed-off-by: Daniel Bevenius --- examples/finetune/finetune.cpp | 2 ++ 1 file changed, 2 insertions(+) diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 11fcbf443..b7e19c5fe 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -1800,6 +1800,8 @@ int main(int argc, char ** argv) { std::vector train_samples_begin; std::vector train_samples_size; printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data); + printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str()); + printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false"); tokenize_file(lctx, params.common.fn_train_data, params.common.sample_start, From d6bd4d46ddb6926087c11e0f6633ab1c81da58c3 Mon Sep 17 00:00:00 2001 From: compilade <113953597+compilade@users.noreply.github.com> Date: Mon, 22 Jan 2024 06:21:52 -0500 Subject: [PATCH 048/101] llama : support StableLM 2 1.6B (#5052) * llama : support StableLM 2 1.6B * convert : fix Qwen's set_vocab wrongly naming all special tokens [PAD{id}] * convert : refactor Qwen's set_vocab to use it for StableLM 2 too * nix : add tiktoken to llama-python-extra * convert : use presence of tokenizer.json to determine StableLM tokenizer loader It's a less arbitrary heuristic than the vocab size. --- .devops/nix/package.nix | 1 + convert-hf-to-gguf.py | 106 +++++++++++++++++++++++----------------- llama.cpp | 18 +++++++ 3 files changed, 79 insertions(+), 46 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 1ef2c6cd9..e8534956f 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -73,6 +73,7 @@ let ps: [ ps.numpy ps.sentencepiece + ps.tiktoken ps.torchWithoutCuda ps.transformers ] diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 4d995ef78..7a0a8c3db 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -289,6 +289,58 @@ class Model: special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_qwen(self): + dir_model = self.dir_model + hparams = self.hparams + tokens: list[bytearray] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) + vocab_size = hparams["vocab_size"] + assert max(tokenizer.get_vocab().values()) < vocab_size + + merges = [] + vocab = {} + mergeable_ranks = tokenizer.mergeable_ranks + for token, rank in mergeable_ranks.items(): + vocab[QwenModel.token_bytes_to_string(token)] = rank + if len(token) == 1: + continue + merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank) + assert len(merged) == 2 + merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged))) + + # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined + added_vocab = tokenizer.special_tokens + reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in (vocab | added_vocab).items()} + + for i in range(vocab_size): + if i not in reverse_vocab: + pad_token = f"[PAD{i}]".encode("utf-8") + tokens.append(bytearray(pad_token)) + toktypes.append(gguf.TokenType.USER_DEFINED) + elif reverse_vocab[i] in added_vocab: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.CONTROL) + else: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.NORMAL) + + self.gguf_writer.add_tokenizer_model("gpt2") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) + special_vocab.merges = merges + # only add special tokens when they were not already loaded from config.json + if len(special_vocab.special_token_ids) == 0: + special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"]) + special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"]) + # this one is usually not in config.json anyway + special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) + special_vocab.add_to_gguf(self.gguf_writer) + def _set_vocab_sentencepiece(self): from sentencepiece import SentencePieceProcessor @@ -877,6 +929,13 @@ class PersimmonModel(Model): class StableLMModel(Model): + def set_vocab(self): + if (self.dir_model / "tokenizer.json").is_file(): + self._set_vocab_gpt2() + else: + # StableLM 2 1.6B uses a vocab in a similar format to Qwen's vocab + self._set_vocab_qwen() + def set_gguf_parameters(self): hparams = self.hparams block_count = hparams["num_hidden_layers"] @@ -922,52 +981,7 @@ class QwenModel(Model): return parts def set_vocab(self): - dir_model = self.dir_model - hparams = self.hparams - tokens: list[bytearray] = [] - toktypes: list[int] = [] - - from transformers import AutoTokenizer - tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) - vocab_size = hparams["vocab_size"] - assert max(tokenizer.get_vocab().values()) < vocab_size - - merges = [] - vocab = {} - mergeable_ranks = tokenizer.mergeable_ranks - for token, rank in mergeable_ranks.items(): - vocab[self.token_bytes_to_string(token)] = rank - if len(token) == 1: - continue - merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank) - assert len(merged) == 2 - merges.append(' '.join(map(self.token_bytes_to_string, merged))) - - reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in vocab.items()} - added_vocab = tokenizer.special_tokens - - for i in range(vocab_size): - if i not in reverse_vocab: - pad_token = f"[PAD{i}]".encode("utf-8") - tokens.append(bytearray(pad_token)) - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.CONTROL) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - - self.gguf_writer.add_tokenizer_model("gpt2") - self.gguf_writer.add_token_list(tokens) - self.gguf_writer.add_token_types(toktypes) - - special_vocab = gguf.SpecialVocab(dir_model, load_merges=False) - special_vocab.merges = merges - special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"]) - special_vocab.add_to_gguf(self.gguf_writer) + self._set_vocab_qwen() def set_gguf_parameters(self): self.gguf_writer.add_name("Qwen") diff --git a/llama.cpp b/llama.cpp index c56d31163..8c906a22f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2877,6 +2877,7 @@ static void llm_load_hparams( ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); switch (hparams.n_layer) { + case 24: model.type = e_model::MODEL_1B; break; case 32: model.type = e_model::MODEL_3B; break; default: model.type = e_model::MODEL_UNKNOWN; } @@ -3700,6 +3701,11 @@ static bool llm_load_tensors( layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + // optional bias tensors, present in Stable LM 2 1.6B + layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, false); + layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, false); + layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, false); + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); @@ -5598,12 +5604,24 @@ struct llm_build_context { // compute Q and K and RoPE them struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } Qcur = ggml_rope_custom( ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, From 15bceec2d73d4166340b46b27677c45ac1b4cdad Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 14:18:43 +0200 Subject: [PATCH 049/101] imatrix : keep intermediate imatrix results (#5077) Co-authored-by: Iwan Kawrakow --- examples/imatrix/imatrix.cpp | 25 ++++++++++++++++++++++++- 1 file changed, 24 insertions(+), 1 deletion(-) diff --git a/examples/imatrix/imatrix.cpp b/examples/imatrix/imatrix.cpp index 5687476cd..ea06fcdbf 100644 --- a/examples/imatrix/imatrix.cpp +++ b/examples/imatrix/imatrix.cpp @@ -26,6 +26,7 @@ struct StatParams { std::string ofile = "imatrix.dat"; int n_output_frequency = 10; int verbosity = 1; + int keep_every = 0; bool collect_output_weight = false; }; @@ -42,6 +43,9 @@ private: int m_last_call = 0; std::vector m_src1_data; std::vector m_ids; // the expert ids from ggml_mul_mat_id + // + void save_imatrix(const char * file_name) const; + void keep_imatrix(int ncall) const; }; bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) { @@ -117,6 +121,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * if (m_last_call % m_params.n_output_frequency == 0) { save_imatrix(); } + if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) { + keep_imatrix(m_last_call); + } } } } else { @@ -143,6 +150,9 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * if (m_last_call % m_params.n_output_frequency == 0) { save_imatrix(); } + if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) { + keep_imatrix(m_last_call); + } } } @@ -150,7 +160,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * } void IMatrixCollector::save_imatrix() const { - const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(); + save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str()); +} + +void IMatrixCollector::keep_imatrix(int ncall) const { + auto file_name = m_params.ofile; + if (file_name.empty()) file_name = "imatrix.dat"; + file_name += ".at_"; + file_name += std::to_string(ncall); + save_imatrix(file_name.c_str()); +} + +void IMatrixCollector::save_imatrix(const char * fname) const { std::ofstream out(fname, std::ios::binary); int n_entries = m_stats.size(); out.write((const char*)&n_entries, sizeof(n_entries)); @@ -400,6 +421,8 @@ int main(int argc, char ** argv) { sparams.verbosity = std::stoi(argv[++iarg]); } else if (arg == "--no-ppl") { compute_ppl = false; + } else if (arg == "--keep-imatrix") { + sparams.keep_every = std::stoi(argv[++iarg]); } else { args.push_back(argv[iarg]); } From f7276f7500f7ea588836dd1fc6f126334c517878 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:10:19 +0000 Subject: [PATCH 050/101] workflows: nix-ci: rebuild on flake.lock updates --- .github/workflows/nix-ci.yml | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 845b93bfb..4592b9607 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -5,10 +5,10 @@ on: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] jobs: nix-eval: From f4dd059259d0234913b9e9780e1662811744c09d Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:16:54 +0000 Subject: [PATCH 051/101] workflows: nix-build-aarch64: rate limit --- .github/workflows/nix-ci-aarch64.yml | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/.github/workflows/nix-ci-aarch64.yml b/.github/workflows/nix-ci-aarch64.yml index be7c26d40..0c6cf5f09 100644 --- a/.github/workflows/nix-ci-aarch64.yml +++ b/.github/workflows/nix-ci-aarch64.yml @@ -2,13 +2,20 @@ name: Nix aarch64 builds on: workflow_dispatch: # allows manual triggering + schedule: + # Rebuild daily rather than on every push because QEMU is expensive (e.g. + # 1.5h instead of minutes with the cold cache). + # + # randint(0, 59), randint(0, 23) + - cron: '26 12 * * *' + # But also rebuild if we touched any of the Nix expressions: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix'] + paths: ['**/*.nix', 'flake.lock'] jobs: nix-build-aarch64: From fe8b3c0d4b0d806e8b46660e24eaf4b90b8b385f Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:38:32 +0000 Subject: [PATCH 052/101] workflows: nix-ci: drop the redundant "paths" filter --- .github/workflows/nix-ci.yml | 2 -- 1 file changed, 2 deletions(-) diff --git a/.github/workflows/nix-ci.yml b/.github/workflows/nix-ci.yml index 4592b9607..d19c7a576 100644 --- a/.github/workflows/nix-ci.yml +++ b/.github/workflows/nix-ci.yml @@ -5,10 +5,8 @@ on: push: branches: - master - paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] pull_request: types: [opened, synchronize, reopened] - paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix', 'flake.lock'] jobs: nix-eval: From 7251870780e2d572dd6f239d7a0bfe438c82fa74 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sat, 13 Jan 2024 17:45:01 +0000 Subject: [PATCH 053/101] nix: refactor the cleanSource rules --- .devops/nix/package.nix | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index e8534956f..c25d99f01 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -115,14 +115,22 @@ effectiveStdenv.mkDerivation ( pname = "llama-cpp${pnameSuffix}"; version = llamaVersion; + # Note: none of the files discarded here are visible in the sandbox or + # affect the output hash. This also means they can be modified without + # triggering a rebuild. src = lib.cleanSourceWith { filter = name: type: - !(builtins.any (_: _) [ + let + noneOf = builtins.all (x: !x); + baseName = baseNameOf name; + in + noneOf [ (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths - (name == "README.md") # Ignore *.md changes whe computing outPaths - (lib.hasPrefix "." name) # Skip hidden files and directories - ]); + (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths + (lib.hasPrefix "." baseName) # Skip hidden files and directories + (baseName == "flake.lock") + ]; src = lib.cleanSource ../../.; }; From 5e97ec91ae3038720a5b15cde4c52d2a53ec2137 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:15:13 +0000 Subject: [PATCH 054/101] nix: add a comment about makeScope --- .devops/nix/scope.nix | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/.devops/nix/scope.nix b/.devops/nix/scope.nix index 7932ac1e8..d295995a4 100644 --- a/.devops/nix/scope.nix +++ b/.devops/nix/scope.nix @@ -4,6 +4,10 @@ llamaVersion ? "0.0.0", }: +# We're using `makeScope` instead of just writing out an attrset +# because it allows users to apply overlays later using `overrideScope'`. +# Cf. https://noogle.dev/f/lib/makeScope + lib.makeScope newScope ( self: { inherit llamaVersion; From 28603cd2833cedd4434f398d847f87fc83546dbb Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:29:38 +0000 Subject: [PATCH 055/101] nix: add a comment on the many nixpkgs-with-cuda instances --- .devops/nix/nixpkgs-instances.nix | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/.devops/nix/nixpkgs-instances.nix b/.devops/nix/nixpkgs-instances.nix index 6e9872b28..4a2f81c4b 100644 --- a/.devops/nix/nixpkgs-instances.nix +++ b/.devops/nix/nixpkgs-instances.nix @@ -7,6 +7,18 @@ { system, ... }: { _module.args = { + # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs + # again, the below creates several nixpkgs instances which the + # flake-centric CLI will be forced to evaluate e.g. on `nix flake show`. + # + # This is currently "slow" and "expensive", on a certain scale. + # This also isn't "right" in that this hinders dependency injection at + # the level of flake inputs. This might get removed in the foreseeable + # future. + # + # Note that you can use these expressions without Nix + # (`pkgs.callPackage ./devops/nix/scope.nix { }` is the entry point). + pkgsCuda = import inputs.nixpkgs { inherit system; # Ensure dependencies use CUDA consistently (e.g. that openmpi, ucc, From b2d80e105a59b54822edf7ce7f3ed5f317e96e21 Mon Sep 17 00:00:00 2001 From: Someone Serge Date: Sun, 21 Jan 2024 03:41:37 +0000 Subject: [PATCH 056/101] flake.nix: add a comment about flakes vs nix --- flake.nix | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/flake.nix b/flake.nix index ec62c773a..a776ba024 100644 --- a/flake.nix +++ b/flake.nix @@ -1,3 +1,17 @@ +# The flake interface to llama.cpp's Nix expressions. The flake is used as a +# more discoverable entry-point, as well as a way to pin the dependencies and +# expose default outputs, including the outputs built by the CI. + +# For more serious applications involving some kind of customization you may +# want to consider consuming the overlay, or instantiating `llamaPackages` +# directly: +# +# ```nix +# pkgs.callPackage ${llama-cpp-root}/.devops/nix/scope.nix { }` +# ``` + +# Cf. https://jade.fyi/blog/flakes-arent-real/ for a more detailed exposition +# of the relation between Nix and the Nix Flakes. { description = "Port of Facebook's LLaMA model in C/C++"; From 3ce7e8f8e7ccfce07e5947ac5f1f3f4628cf68ea Mon Sep 17 00:00:00 2001 From: XiaotaoChen Date: Mon, 22 Jan 2024 21:09:35 +0800 Subject: [PATCH 057/101] llava : MobileVLM support (#4954) * MobileVLM native implementation * delete depthwise_conv_2d and permute_cpy relative code, replace the two by the existed functions, and opt ldp definition, support LLAMA_PERF option for CMake * move android script to example/llava directory * Fix the editor config checks --------- Co-authored-by: Chenxiaotao03 --- CMakeLists.txt | 7 + examples/llava/MobileVLM-README.md | 131 ++++++ examples/llava/android/adb_run.sh | 53 +++ examples/llava/android/build_64.sh | 8 + examples/llava/clip.cpp | 391 +++++++++++++++++- .../llava/convert-image-encoder-to-gguf.py | 6 +- ggml.c | 141 ++++++- ggml.h | 24 ++ 8 files changed, 737 insertions(+), 24 deletions(-) create mode 100644 examples/llava/MobileVLM-README.md create mode 100755 examples/llava/android/adb_run.sh create mode 100755 examples/llava/android/build_64.sh diff --git a/CMakeLists.txt b/CMakeLists.txt index 6b3b1396b..5a333ff52 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -108,6 +108,13 @@ option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STA option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_SERVER "llama: build server example" ON) + +# add perf arguments +option(LLAMA_PERF "llama: enable perf" OFF) +if (LLAMA_PERF) + add_definitions(-DGGML_PERF) +endif() + # Required for relocatable CMake package include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake) diff --git a/examples/llava/MobileVLM-README.md b/examples/llava/MobileVLM-README.md new file mode 100644 index 000000000..c6258eba6 --- /dev/null +++ b/examples/llava/MobileVLM-README.md @@ -0,0 +1,131 @@ +# MobileVLM + +Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants. + +for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM) + +The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava. + +## Usage +Build with cmake or run `make llava-cli` to build it. + +After building, run: `./llava-cli` to see the usage. For example: + +```sh +./llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \ + --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \ + --image path/to/an/image.jpg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:" +``` + +## Model conversion + +- Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally: + +```sh +git clone https://huggingface.co/mtgv/MobileVLM-1.7B + +git clone https://huggingface.co/openai/clip-vit-large-patch14-336 +``` + +2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents: + +```sh +python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B +``` + +3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF: + +```sh +python ./examples/llava/convert-image-encoder-to-gguf \ + -m path/to/clip-vit-large-patch14-336 \ + --llava-projector path/to/MobileVLM-1.7B/llava.projector \ + --output-dir path/to/MobileVLM-1.7B \ + --projector-type ldp +``` + +4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF: + +```sh +python ./convert.py path/to/MobileVLM-1.7B +``` + +5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k` +```sh +./quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s +``` + +Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory. + +## Android compile and run +### compile +refer to `examples/llava/android/build_64.sh` +```sh +mkdir examples/llava/android/build_64 +cd examples/llava/android/build_64 +../build_64.sh +``` +### run on Android +refer to `android/adb_run.sh`, modify resources' `name` and `path` + +## some result on Android with `Snapdragon 888` chip +### case 1 +**input** +```sh +/data/local/tmp/llava-cli \ + -m /data/local/tmp/ggml-model-q4_k.gguf \ + --mmproj /data/local/tmp/mmproj-model-f16.gguf \ + -t 4 \ + --image /data/local/tmp/demo.jpg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" +``` +**output** +```sh +encode_image_with_clip: image encoded in 21148.71 ms by CLIP ( 146.87 ms per image patch) + Susan Wise Bauer +llama_print_timings: load time = 23574.72 ms +llama_print_timings: sample time = 1.24 ms / 6 runs ( 0.21 ms per token, 4850.44 tokens per second) +llama_print_timings: prompt eval time = 12460.15 ms / 246 tokens ( 50.65 ms per token, 19.74 tokens per second) +llama_print_timings: eval time = 424.86 ms / 6 runs ( 70.81 ms per token, 14.12 tokens per second) +llama_print_timings: total time = 34731.93 ms +``` +### case 2 +**input** +```sh +/data/local/tmp/llava-cli \ + -m /data/local/tmp/ggml-model-q4_k.gguf \ + --mmproj /data/local/tmp/mmproj-model-f16.gguf \ + -t 4 \ + --image /data/local/tmp/cat.jpeg \ + -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:" +``` + +**output** +```sh +encode_image_with_clip: image encoded in 21149.51 ms by CLIP ( 146.87 ms per image patch) + The image depicts a cat sitting in the grass near some tall green plants. +llama_print_timings: load time = 23257.32 ms +llama_print_timings: sample time = 5.25 ms / 18 runs ( 0.29 ms per token, 3430.53 tokens per second) +llama_print_timings: prompt eval time = 11900.73 ms / 232 tokens ( 51.30 ms per token, 19.49 tokens per second) +llama_print_timings: eval time = 1279.03 ms / 18 runs ( 71.06 ms per token, 14.07 tokens per second) +llama_print_timings: total time = 34570.79 ms +``` + +## Minor shortcomings +The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost. + +## TODO + +- [ ] Support non-CPU backend for the new operators, such as `depthwise`, `hardswish`, `hardsigmoid` +- [ ] Optimize LDP projector performance + + - Optimize the structure definition to avoid unnecessary memory rearrangements, to reduce the use of `ggml_permute_cpy`; + - Optimize operator implementation (ARM CPU/NVIDIA GPU): such as depthwise conv, hardswish, hardsigmoid, etc. +- [ ] run MobileVLM on `Jetson Orin` +- [ ] Support more model variants, such as `MobileVLM-3B`. + + +## contributor +```sh +zhangjidong05, yangyang260, huyiming03, chenxiaotao03 +``` diff --git a/examples/llava/android/adb_run.sh b/examples/llava/android/adb_run.sh new file mode 100755 index 000000000..f73623ae3 --- /dev/null +++ b/examples/llava/android/adb_run.sh @@ -0,0 +1,53 @@ +#!/bin/bash + +model_dir="/Users/cxt/model/llm/mobileVLM/MobileVLM-1.7B_processed" +projector_name="mmproj-model-f16.gguf" +llama_name="ggml-model-q4_k.gguf" +img_dir="/Users/cxt/model/llm" +img_name="demo.jpg" +prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWho is the author of this book? \nAnswer the question using a single word or phrase. ASSISTANT:" +# img_name="cat.jpeg" +# prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: \nWhat is in the image? ASSISTANT:" + +program_dir="build_64/bin" +binName="llava-cli" +n_threads=4 + + +deviceDir="/data/local/tmp" +saveDir="output" +if [ ! -d ${saveDir} ]; then + mkdir ${saveDir} +fi + + +function android_run() { + # # copy resource into device + # adb push ${model_dir}/${projector_name} ${deviceDir}/${projector_name} + # adb push ${model_dir}/${llama_name} ${deviceDir}/${llama_name} + adb push ${img_dir}/${img_name} ${deviceDir}/${img_name} + # copy program into device + adb push ${program_dir}/${binName} ${deviceDir}/${binName} + adb shell "chmod 0777 ${deviceDir}/${binName}" + + # run + adb shell "echo cd ${deviceDir} ${deviceDir}/${binName} \ + -m ${deviceDir}/${llama_name} \ + --mmproj ${deviceDir}/${projector_name} \ + -t ${n_threads} \ + --image ${deviceDir}/${img_name} \ + -p \"${prompt}\" \ + > ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt" + adb shell "cd ${deviceDir}; pwd; ${deviceDir}/${binName} \ + -m ${deviceDir}/${llama_name} \ + --mmproj ${deviceDir}/${projector_name} \ + -t ${n_threads} \ + --image ${deviceDir}/${img_name} \ + -p \"${prompt}\" \ + >> ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt 2>&1" + adb pull ${deviceDir}/${modelName}_${projector_name}_${n_threads}_${img_name}.txt ${saveDir} +} + +android_run + +echo "android_run is Done!" diff --git a/examples/llava/android/build_64.sh b/examples/llava/android/build_64.sh new file mode 100755 index 000000000..71b6fd3f7 --- /dev/null +++ b/examples/llava/android/build_64.sh @@ -0,0 +1,8 @@ +#!/bin/bash +cmake ../../../../ \ +-DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \ +-DCMAKE_BUILD_TYPE=Release \ +-DANDROID_ABI="arm64-v8a" \ +-DANDROID_PLATFORM=android-23 $1 + +make -j4 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 2ae8853d3..6161fd858 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -12,6 +12,7 @@ #include #include #include +#include #include "clip.h" #include "ggml.h" @@ -67,6 +68,7 @@ static std::string format(const char * fmt, ...) { #define KEY_PATCH_SIZE "clip.vision.patch_size" #define KEY_IMAGE_MEAN "clip.vision.image_mean" #define KEY_IMAGE_STD "clip.vision.image_std" +#define KEY_PROJ_TYPE "clip.projector_type" // // tensor name constants @@ -89,6 +91,21 @@ static std::string format(const char * fmt, ...) { #define TN_TEXT_PROJ "text_projection.weight" #define TN_VIS_PROJ "visual_projection.weight" #define TN_LLAVA_PROJ "mm.%d.%s" +#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s" +#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s" + + +enum projector_type { + PROJECTOR_TYPE_MLP, + PROJECTOR_TYPE_LDP, + PROJECTOR_TYPE_UNKNOWN, +}; + +static std::map PROJECTOR_TYPE_NAMES = { + { PROJECTOR_TYPE_MLP, "mlp" }, + { PROJECTOR_TYPE_LDP, "ldp" }, +}; + // // utilities to get data from a gguf file @@ -129,6 +146,91 @@ static std::string get_ftype(int ftype) { return ggml_type_name(static_cast(ftype)); } +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + + +static void replace_all(std::string & s, const std::string & search, const std::string & replace) { + std::string result; + for (size_t pos = 0; ; pos += search.length()) { + auto new_pos = s.find(search, pos); + if (new_pos == std::string::npos) { + result += s.substr(pos, s.size() - pos); + break; + } + result += s.substr(pos, new_pos - pos) + replace; + pos = new_pos; + } + s = std::move(result); +} + +static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + +static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") { + size_t tensor_size = ggml_nbytes(tensor); + printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%d, %d, %d, %d], type: %d\n", + prefix, ggml_n_dims(tensor), tensor->name, tensor_size, + tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->type); +} + +static projector_type clip_projector_type_from_string(const std::string & name) { + for (const auto & kv : PROJECTOR_TYPE_NAMES) { // NOLINT + if (kv.second == name) { + return kv.first; + } + } + return PROJECTOR_TYPE_UNKNOWN; +} + // // image data // @@ -205,6 +307,32 @@ struct clip_vision_model { struct ggml_tensor * mm_0_b; struct ggml_tensor * mm_2_w; struct ggml_tensor * mm_2_b; + + // MobileVLM projection + struct ggml_tensor * mm_model_mlp_1_w; + struct ggml_tensor * mm_model_mlp_1_b; + struct ggml_tensor * mm_model_mlp_3_w; + struct ggml_tensor * mm_model_mlp_3_b; + struct ggml_tensor * mm_model_block_1_block_0_0_w; + struct ggml_tensor * mm_model_block_1_block_0_1_w; + struct ggml_tensor * mm_model_block_1_block_0_1_b; + struct ggml_tensor * mm_model_block_1_block_1_fc1_w; + struct ggml_tensor * mm_model_block_1_block_1_fc1_b; + struct ggml_tensor * mm_model_block_1_block_1_fc2_w; + struct ggml_tensor * mm_model_block_1_block_1_fc2_b; + struct ggml_tensor * mm_model_block_1_block_2_0_w; + struct ggml_tensor * mm_model_block_1_block_2_1_w; + struct ggml_tensor * mm_model_block_1_block_2_1_b; + struct ggml_tensor * mm_model_block_2_block_0_0_w; + struct ggml_tensor * mm_model_block_2_block_0_1_w; + struct ggml_tensor * mm_model_block_2_block_0_1_b; + struct ggml_tensor * mm_model_block_2_block_1_fc1_w; + struct ggml_tensor * mm_model_block_2_block_1_fc1_b; + struct ggml_tensor * mm_model_block_2_block_1_fc2_w; + struct ggml_tensor * mm_model_block_2_block_1_fc2_b; + struct ggml_tensor * mm_model_block_2_block_2_0_w; + struct ggml_tensor * mm_model_block_2_block_2_1_w; + struct ggml_tensor * mm_model_block_2_block_2_1_b; }; struct clip_ctx { @@ -213,6 +341,7 @@ struct clip_ctx { bool has_llava_projector = false; struct clip_vision_model vision_model; + projector_type proj_type = PROJECTOR_TYPE_MLP; float image_mean[3]; float image_std[3]; @@ -430,16 +559,135 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 free(patches_data); } + // shape [1, 576, 1024] + // ne is whcn, ne = [1024, 576, 1, 1] embeddings = ggml_get_rows(ctx0, embeddings, patches); - // mm projection 0 - embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + // print_tensor_info(embeddings, "embeddings"); - embeddings = ggml_gelu(ctx0, embeddings); + // llava projector + if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); - embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); - embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + embeddings = ggml_gelu(ctx0, embeddings); + + embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + } + else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + // MobileVLM projector + int n_patch = 24; + struct ggml_tensor * mlp_1 = ggml_mul_mat(ctx0, model.mm_model_mlp_1_w, embeddings); + mlp_1 = ggml_add(ctx0, mlp_1, model.mm_model_mlp_1_b); + mlp_1 = ggml_gelu(ctx0, mlp_1); + struct ggml_tensor * mlp_3 = ggml_mul_mat(ctx0, model.mm_model_mlp_3_w, mlp_1); + mlp_3 = ggml_add(ctx0, mlp_3, model.mm_model_mlp_3_b); + // mlp_3 shape = [1, 576, 2048], ne = [2048, 576, 1, 1] + + // block 1 + struct ggml_tensor * block_1 = nullptr; + { + // transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24] + mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); + mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); + // stride = 1, padding = 1, bias is nullptr + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, nullptr, 1, 1, 1, 1, 1, 1); + + // layer norm + // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_0_1_w), model.mm_model_block_1_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + + // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // hardswish + struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_1_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + // block_1_hw shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1], block_1 shape = [1, 2048], ne = [2048, 1, 1, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_1_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + // block_1 shape = [1, 24, 24, 2048], ne = [2048, 24, 24, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_1_block_2_1_w), model.mm_model_block_1_block_2_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] + // residual + block_1 = ggml_add(ctx0, mlp_3, block_1); + } + + // block_2 + { + // stride = 2 + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, nullptr, 2, 2, 1, 1, 1, 1); + + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // layer norm + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 2, 0, 3)); + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_0_1_w), model.mm_model_block_2_block_0_1_b); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 2, 0, 1, 3)); + // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] + // hardswish + struct ggml_tensor * block_1_hw = ggml_hardswish(ctx0, block_1); + + // not sure the parameters is right for globalAvgPooling + block_1 = ggml_pool_2d(ctx0, block_1_hw, GGML_OP_POOL_AVG, block_1_hw->ne[0], block_1_hw->ne[1], block_1_hw->ne[0], block_1_hw->ne[1], 0, 0); + // block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + // pointwise conv + block_1 = ggml_reshape_2d(ctx0, block_1, block_1->ne[0]*block_1->ne[1]*block_1->ne[2], block_1->ne[3]); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc1_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc1_b); + block_1 = ggml_relu(ctx0, block_1); + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_1_fc2_w, block_1); + block_1 = ggml_add(ctx0, block_1, model.mm_model_block_2_block_1_fc2_b); + block_1 = ggml_hardsigmoid(ctx0, block_1); + + // block_1_hw shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1], block_1 shape = [1, 2048, 1, 1], ne = [1, 1, 2048, 1] + block_1 = ggml_reshape_4d(ctx0, block_1, 1, 1, block_1->ne[0], block_1->ne[1]); + block_1 = ggml_mul(ctx0, block_1_hw, block_1); + + int w = block_1->ne[0], h = block_1->ne[1]; + block_1 = ggml_reshape_3d(ctx0, block_1, w*h, block_1->ne[2], block_1->ne[3]); + block_1 = ggml_cont(ctx0, ggml_permute(ctx0, block_1, 1, 0, 2, 3)); + // block_1 shape = [1, 24*24, 2048], ne = [24*24, 2048, 1] + block_1 = ggml_mul_mat(ctx0, model.mm_model_block_2_block_2_0_w, block_1); + block_1 = ggml_reshape_4d(ctx0, block_1, block_1->ne[0], w, h, block_1->ne[3]); + + + // block_1 shape = [1, 12, 12, 2048], ne = [2048, 12, 12, 1] + block_1 = ggml_norm(ctx0, block_1, eps); + block_1 = ggml_add(ctx0, ggml_mul(ctx0, block_1, model.mm_model_block_2_block_2_1_w), model.mm_model_block_2_block_2_1_b); + block_1 = ggml_reshape_3d(ctx0, block_1, block_1->ne[0], block_1->ne[1] * block_1->ne[2], block_1->ne[3]); + // block_1 shape = [1, 144, 2048], ne = [2048, 144, 1] + } + embeddings = block_1; + } + else { + GGML_ASSERT(false); + } } // build the graph @@ -485,16 +733,55 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { printf("\n"); } const int n_tensors = gguf_get_n_tensors(ctx); + // kv - if (verbosity >= 3) { - const int n_kv = gguf_get_n_kv(ctx); + const int n_kv = gguf_get_n_kv(ctx); + printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n", + __func__, n_kv, n_tensors, fname); + { + std::map n_type; - for (int i = 0; i < n_kv; ++i) { - const char * key = gguf_get_key(ctx, i); + uint32_t n_type_max = 0; + enum ggml_type type_max = GGML_TYPE_F32; - printf("%s: kv[%d]: key = %s\n", __func__, i, key); + for (int i = 0; i < n_tensors; i++) { + enum ggml_type type = gguf_get_tensor_type(ctx, i); + + n_type[type]++; + + if (n_type_max < n_type[type]) { + n_type_max = n_type[type]; + type_max = type; + } + } + + printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); + for (int i = 0; i < n_kv; i++) { + const char * name = gguf_get_key(ctx, i); + const enum gguf_type type = gguf_get_kv_type(ctx, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx, i)), gguf_get_arr_n(ctx, i)) + : gguf_type_name(type); + + std::string value = gguf_kv_to_str(ctx, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); + + printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); + } + + // print type counts + for (auto & kv : n_type) { + if (kv.second == 0) { + continue; + } + + printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second); } - printf("\n"); } // data @@ -503,20 +790,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { for (int i = 0; i < n_tensors; ++i) { const char * name = gguf_get_tensor_name(ctx, i); const size_t offset = gguf_get_tensor_offset(ctx, i); + enum ggml_type type = gguf_get_tensor_type(ctx, i); struct ggml_tensor * cur = ggml_get_tensor(meta, name); size_t tensor_size = ggml_nbytes(cur); buffer_size += tensor_size; if (verbosity >= 3) { - printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i, - ggml_n_dims(cur), cur->name, tensor_size, offset); + printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%d, %d, %d, %d], type: %d\n", __func__, i, + ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], type); } } } + + buffer_size += n_tensors * 128 /* CLIP PADDING */; clip_ctx * new_clip = new clip_ctx; + // update projector type + { + int idx = gguf_find_key(ctx, KEY_PROJ_TYPE); + if (idx != -1) { + const std::string proj_type = gguf_get_val_str(ctx, idx); + new_clip->proj_type = clip_projector_type_from_string(proj_type); + } + else { + new_clip->proj_type = PROJECTOR_TYPE_MLP; + } + } + #ifdef GGML_USE_CUBLAS new_clip->backend = ggml_backend_cuda_init(0); printf("%s: CLIP using CUDA backend\n", __func__); @@ -661,10 +963,45 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v")); vision_model.pre_ln_w = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight")); vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); - vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); - vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + + // LLaVA projection + if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); + vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); + vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + } + else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { + // MobileVLM projection + vision_model.mm_model_mlp_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "weight")); + vision_model.mm_model_mlp_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 1, "bias")); + vision_model.mm_model_mlp_3_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "weight")); + vision_model.mm_model_mlp_3_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 3, "bias")); + vision_model.mm_model_block_1_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "0.weight")); + vision_model.mm_model_block_1_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.weight")); + vision_model.mm_model_block_1_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 0, "1.bias")); + vision_model.mm_model_block_1_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.weight")); + vision_model.mm_model_block_1_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc1.bias")); + vision_model.mm_model_block_1_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.weight")); + vision_model.mm_model_block_1_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 1, "fc2.bias")); + vision_model.mm_model_block_1_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "0.weight")); + vision_model.mm_model_block_1_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.weight")); + vision_model.mm_model_block_1_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 1, 2, "1.bias")); + vision_model.mm_model_block_2_block_0_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "0.weight")); + vision_model.mm_model_block_2_block_0_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.weight")); + vision_model.mm_model_block_2_block_0_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 0, "1.bias")); + vision_model.mm_model_block_2_block_1_fc1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.weight")); + vision_model.mm_model_block_2_block_1_fc1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc1.bias")); + vision_model.mm_model_block_2_block_1_fc2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.weight")); + vision_model.mm_model_block_2_block_1_fc2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 1, "fc2.bias")); + vision_model.mm_model_block_2_block_2_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight")); + vision_model.mm_model_block_2_block_2_1_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight")); + vision_model.mm_model_block_2_block_2_1_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias")); + } + else { + std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type]; + throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); + } vision_model.layers.resize(hparams.n_layer); for (int il = 0; il < hparams.n_layer; ++il) { @@ -1100,13 +1437,25 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i } int clip_n_mmproj_embd(const struct clip_ctx * ctx) { - return ctx->vision_model.mm_2_b->ne[0]; + if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0]; + } + else if (ctx->proj_type == PROJECTOR_TYPE_MLP) { + return ctx->vision_model.mm_2_b->ne[0]; + } + else { + std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; + throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str())); + } } int clip_n_patches(const struct clip_ctx * ctx) { auto & params = ctx->vision_model.hparams; - - return (params.image_size / params.patch_size) * (params.image_size / params.patch_size); + int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size); + if (ctx->proj_type == PROJECTOR_TYPE_LDP) { + n_patches /= 4; + } + return n_patches; } size_t clip_embd_nbytes(const struct clip_ctx * ctx) { diff --git a/examples/llava/convert-image-encoder-to-gguf.py b/examples/llava/convert-image-encoder-to-gguf.py index 03688e0ea..f5a3c9b46 100644 --- a/examples/llava/convert-image-encoder-to-gguf.py +++ b/examples/llava/convert-image-encoder-to-gguf.py @@ -81,6 +81,7 @@ ap.add_argument("--vision-only", action="store_true", required=False, ap.add_argument("--clip_model_is_vision", action="store_true", required=False, help="The clip model is a pure vision model (ShareGPT4V vision extract for example)") ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.") +ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp") ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values") ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values") ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None) @@ -174,6 +175,8 @@ elif args.vision_only and not has_llava_projector: fout.add_description("vision-only CLIP model") elif has_llava_projector: fout.add_description("image encoder for LLaVA") + # add projector type + fout.add_string("clip.projector_type", args.projector_type) else: fout.add_description("two-tower CLIP model") @@ -218,7 +221,8 @@ if has_llava_projector: projector = torch.load(args.llava_projector) for name, data in projector.items(): name = get_tensor_name(name) - if data.ndim == 2: + # pw and dw conv ndim==4 + if data.ndim == 2 or data.ndim == 4: data = data.squeeze().numpy().astype(np.float16) else: data = data.squeeze().numpy().astype(np.float32) diff --git a/ggml.c b/ggml.c index cbf2d4bdd..a7a88e382 100644 --- a/ggml.c +++ b/ggml.c @@ -1418,6 +1418,9 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); } +// TODO: optimize performance +inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } +inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); } static const float GELU_COEF_A = 0.044715f; static const float GELU_QUICK_COEF = -1.702f; @@ -1776,9 +1779,11 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = { "GELU", "GELU_QUICK", "SILU", + "HARDSWISH", + "HARDSIGMOID", }; -static_assert(GGML_UNARY_OP_COUNT == 10, "GGML_UNARY_OP_COUNT != 10"); +static_assert(GGML_UNARY_OP_COUNT == 12, "GGML_UNARY_OP_COUNT != 12"); static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN"); @@ -3945,6 +3950,20 @@ struct ggml_tensor * ggml_silu_back( return result; } +// ggml hardswish +struct ggml_tensor * ggml_hardswish( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSWISH); +} + +// ggml hardsigmoid +struct ggml_tensor * ggml_hardsigmoid( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID); +} + // ggml_norm static struct ggml_tensor * ggml_norm_impl( @@ -5344,6 +5363,33 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( return result; } +// ggml_conv_depthwise +struct ggml_tensor * ggml_conv_depthwise_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + + struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); + struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, + ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), + s0, s1, p0, p1, d0, d1, true); // [N * IC, OH, OW, KH * KW] + + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2], new_a->ne[3], 1), // [OC，1, KH, KW] => [1, OC, 1, KH * KW] + ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3])); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW] + + result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW] + + return result; +} // ggml_conv_2d // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW] @@ -9333,6 +9379,87 @@ static void ggml_compute_forward_silu_back( } } + +static void ggml_compute_forward_hardswish_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_hardswish_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} +static void ggml_compute_forward_hardswish( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardswish_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +static void ggml_compute_forward_hardsigmoid_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_hardsigmoid_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_hardsigmoid( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_hardsigmoid_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + + // ggml_compute_forward_norm static void ggml_compute_forward_norm_f32( @@ -12349,6 +12476,7 @@ static void ggml_compute_forward_im2col( } } + // ggml_compute_forward_conv_transpose_2d static void ggml_compute_forward_conv_transpose_2d( @@ -13917,6 +14045,14 @@ static void ggml_compute_forward_unary( { ggml_compute_forward_silu(params, src0, dst); } break; + case GGML_UNARY_OP_HARDSWISH: + { + ggml_compute_forward_hardswish(params, src0, dst); + } break; + case GGML_UNARY_OP_HARDSIGMOID: + { + ggml_compute_forward_hardsigmoid(params, src0, dst); + } break; default: { GGML_ASSERT(false); @@ -16330,6 +16466,8 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { case GGML_UNARY_OP_TANH: case GGML_UNARY_OP_ELU: case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_HARDSWISH: // to opt for multiple threads + case GGML_UNARY_OP_HARDSIGMOID: // to opt for multiple threads { n_tasks = 1; } break; @@ -16562,7 +16700,6 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { // distribute new work or execute it direct if 1T while (++node_n < cgraph->n_nodes) { GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); - struct ggml_tensor * node = cgraph->nodes[node_n]; const int n_tasks = ggml_get_n_tasks(node, n_threads); diff --git a/ggml.h b/ggml.h index de8162b81..dca7bd9ce 100644 --- a/ggml.h +++ b/ggml.h @@ -489,6 +489,8 @@ extern "C" { GGML_UNARY_OP_GELU, GGML_UNARY_OP_GELU_QUICK, GGML_UNARY_OP_SILU, + GGML_UNARY_OP_HARDSWISH, + GGML_UNARY_OP_HARDSIGMOID, GGML_UNARY_OP_COUNT, }; @@ -1032,6 +1034,16 @@ extern "C" { struct ggml_tensor * a, struct ggml_tensor * b); + // hardswish(x) = x * relu6(x + 3) / 6 + GGML_API struct ggml_tensor * ggml_hardswish( + struct ggml_context * ctx, + struct ggml_tensor * a); + + // hardsigmoid(x) = relu6(x + 3) / 6 + GGML_API struct ggml_tensor * ggml_hardsigmoid( + struct ggml_context * ctx, + struct ggml_tensor * a); + // normalize along rows GGML_API struct ggml_tensor * ggml_norm( struct ggml_context * ctx, @@ -1483,6 +1495,18 @@ extern "C" { int d1, bool is_2D); + GGML_API struct ggml_tensor * ggml_conv_depthwise_2d( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + struct ggml_tensor * c, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1); + GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, From 780e24a22eb595b705cbe8284771e9ceff1c4dd2 Mon Sep 17 00:00:00 2001 From: Reinforce-II Date: Mon, 22 Jan 2024 21:15:08 +0800 Subject: [PATCH 058/101] ggml : parallelize FP32 conversion when using BLAS (#5045) * make GGML_TASK_INIT phase can be run in multithread * multithreaded dequantize in mul_mat when using blas library * minor fixes * update outdated comment * fix coding style * simplify code Co-authored-by: Georgi Gerganov --------- Co-authored-by: Georgi Gerganov --- ggml.c | 198 +++++++++++++++++++++++++++++++++++++++++++-------------- 1 file changed, 150 insertions(+), 48 deletions(-) diff --git a/ggml.c b/ggml.c index a7a88e382..f85045c9c 100644 --- a/ggml.c +++ b/ggml.c @@ -7810,6 +7810,9 @@ static void ggml_compute_forward_acc_f32( bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { + if (params->ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase memcpy( @@ -9952,11 +9955,30 @@ static void ggml_compute_forward_mul_mat( #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(dst)) { - if (params->ith != 0) { - return; - } + const int64_t ne_plane = ne01*ne00; + const int64_t desired_wsize = ne13*ne12*ne_plane*sizeof(float); + UNUSED(desired_wsize); if (params->type == GGML_TASK_INIT) { + if (type != GGML_TYPE_F32) { + assert(params->wsize >= desired_wsize); + // parallelize by src0 rows + for (int64_t i13 = 0; i13 < ne13; i13++) { + for (int64_t i12 = 0; i12 < ne12; i12++) { + // broadcast src0 into src1 across 2nd,3rd dimension + const int64_t i03 = i13/r3; + const int64_t i02 = i12/r2; + + const void * x = (char *) src0->data + i02*nb02 + i03*nb03; + float * const wdata = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane; + ggml_to_float_t const to_float = type_traits[type].to_float; + + for (int64_t i01 = ith; i01 < ne01; i01 += nth) { + to_float((const char *) x + i01*nb01, wdata + i01*ne00, ne00); + } + } + } + } return; } @@ -9964,9 +9986,14 @@ static void ggml_compute_forward_mul_mat( return; } + // perform sgemm, parallelization controlled by blas lib + if (ith != 0) { + return; + } + + const int64_t tgemm0 = ggml_perf_time_us(); for (int64_t i13 = 0; i13 < ne13; i13++) { for (int64_t i12 = 0; i12 < ne12; i12++) { - // broadcast src0 into src1 across 2nd,3rd dimension const int64_t i03 = i13/r3; const int64_t i02 = i12/r2; @@ -9975,17 +10002,7 @@ static void ggml_compute_forward_mul_mat( float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3); if (type != GGML_TYPE_F32) { - float * const wdata = params->wdata; - ggml_to_float_t const to_float = type_traits[type].to_float; - - size_t id = 0; - for (int64_t i01 = 0; i01 < ne01; ++i01) { - to_float((const char *) x + i01*nb01, wdata + id, ne00); - id += ne00; - } - - assert(id*sizeof(float) <= params->wsize); - x = wdata; + x = (float *) params->wdata + i13*ne12*ne_plane + i12*ne_plane; } cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans, @@ -9995,6 +10012,7 @@ static void ggml_compute_forward_mul_mat( 0.0f, d, ne01); } } + //printf("cblas_sgemm = %.3f ms, %lld flops\n", (ggml_perf_time_us() - tgemm0)/1000.0, ne13*ne12*ne1*ne01*ne10*2); //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3); @@ -10003,6 +10021,9 @@ static void ggml_compute_forward_mul_mat( #endif if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } if (src1->type != vec_dot_type) { char * wdata = params->wdata; const size_t row_size = ggml_row_size(vec_dot_type, ne10); @@ -10167,6 +10188,9 @@ static void ggml_compute_forward_mul_mat_id( #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)] if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } char * wdata = params->wdata; if (src1->type != vec_dot_type) { const size_t row_size = ggml_row_size(vec_dot_type, ne10); @@ -10352,6 +10376,9 @@ static void ggml_compute_forward_out_prod_f32( return; } #endif + if (ith != 0) { + return; + } ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -10535,6 +10562,9 @@ static void ggml_compute_forward_out_prod_q_f32( // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -10719,6 +10749,9 @@ static void ggml_compute_forward_set_f32( bool inplace = (bool) ((int32_t *) dst->op_params)[4]; if (!inplace && (params->type == GGML_TASK_INIT)) { + if (params->ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase memcpy( @@ -11043,6 +11076,9 @@ static void ggml_compute_forward_get_rows_back_f32_f16( // ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memset(dst->data, 0, ggml_nbytes(dst)); } @@ -11077,6 +11113,9 @@ static void ggml_compute_forward_get_rows_back_f32( // ggml_compute_forward_dup_same_cont(params, opt0, dst); if (params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memset(dst->data, 0, ggml_nbytes(dst)); } @@ -11214,6 +11253,9 @@ static void ggml_compute_forward_diag_mask_f32( GGML_ASSERT(n_past >= 0); if (!inplace && (params->type == GGML_TASK_INIT)) { + if (ith != 0) { + return; + } // memcpy needs to be synchronized across threads to avoid race conditions. // => do it in INIT phase GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0)); @@ -12184,6 +12226,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) @@ -12278,6 +12323,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout) @@ -12502,6 +12550,9 @@ static void ggml_compute_forward_conv_transpose_2d( GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { + if (ith != 0) { + return; + } memset(params->wdata, 0, params->wsize); // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout) @@ -14116,6 +14167,9 @@ static void ggml_compute_forward_add_rel_pos_f32( const bool inplace = (bool) ((int32_t *) dst->op_params)[0]; if (!inplace && params->type == GGML_TASK_INIT) { + if (params->ith != 0) { + return; + } memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst)); return; } @@ -16409,8 +16463,9 @@ struct ggml_compute_state_shared { const int n_threads; // synchronization primitives - atomic_int n_active; // num active threads - atomic_int node_n; // active graph node + atomic_int n_active; // num active threads + atomic_int node_n; // active graph node + atomic_int node_task; // active graph node task phase bool (*abort_callback)(void * data); // abort ggml_graph_compute when true void * abort_callback_data; @@ -16658,6 +16713,34 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { return n_tasks; } +static void ggml_graph_compute_thread_sync_node(int * node_n, struct ggml_compute_state * state, const bool do_yield) { + // wait for other threads to finish + const int last_node_n = * node_n; + + while (true) { + if (do_yield) { + sched_yield(); + } + + * node_n = atomic_load(&state->shared->node_n); + if (* node_n != last_node_n) break; + } +} + +static void ggml_graph_compute_thread_sync_task(int * task_phase, struct ggml_compute_state * state, const bool do_yield) { + // wait for other threads to finish + const int last_task_phase = * task_phase; + + while (true) { + if (do_yield) { + sched_yield(); + } + + * task_phase = atomic_load(&state->shared->node_task); + if (* task_phase != last_task_phase) break; + } +} + static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; @@ -16668,7 +16751,8 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { set_numa_thread_affinity(state->ith, n_threads); - int node_n = -1; + int node_n = -1; + int task_phase = GGML_TASK_FINALIZE; while (true) { if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) { @@ -16708,13 +16792,13 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { params.nth = n_tasks; - /* INIT */ - if (GGML_OP_HAS_INIT[node->op]) { - params.type = GGML_TASK_INIT; - ggml_compute_forward(¶ms, node); - } - if (n_tasks == 1) { + /* INIT */ + if (GGML_OP_HAS_INIT[node->op]) { + params.type = GGML_TASK_INIT; + ggml_compute_forward(¶ms, node); + } + // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1, // they do something more efficient than spinning (?) params.type = GGML_TASK_COMPUTE; @@ -16735,38 +16819,24 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { } } - atomic_store(&state->shared->n_active, n_threads); - atomic_store(&state->shared->node_n, node_n); + task_phase = GGML_TASK_INIT; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_n, node_n); + atomic_store(&state->shared->node_task, task_phase); } else { - // wait for other threads to finish - const int last = node_n; - - const bool do_yield = last < 0 || cgraph->nodes[last]->op == GGML_OP_MUL_MAT; - - while (true) { - // TODO: this sched_yield can have significant impact on the performance - either positive or negative - // depending on the workload and the operating system. - // since it is not clear what is the best approach, it should potentially become user-configurable - // ref: https://github.com/ggerganov/ggml/issues/291 - // UPD: adding the do_yield flag seems to resolve the issue universally - if (do_yield) { - sched_yield(); - } - - node_n = atomic_load(&state->shared->node_n); - if (node_n != last) break; - }; + ggml_graph_compute_thread_sync_node(&node_n, state, false); + ggml_graph_compute_thread_sync_task(&task_phase, state, false); } // check if we should stop if (node_n >= cgraph->n_nodes) break; - /* COMPUTE */ + /* INIT & COMPUTE */ struct ggml_tensor * node = cgraph->nodes[node_n]; const int n_tasks = ggml_get_n_tasks(node, n_threads); struct ggml_compute_params params = { - /*.type =*/ GGML_TASK_COMPUTE, + /*.type =*/ GGML_TASK_INIT, /*.ith =*/ state->ith, /*.nth =*/ n_tasks, /*.wsize =*/ cplan->work_size, @@ -16774,8 +16844,39 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { }; if (state->ith < n_tasks) { + if (GGML_OP_HAS_INIT[node->op]) { + ggml_compute_forward(¶ms, node); + } + } + + if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { + task_phase = GGML_TASK_COMPUTE; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_task, task_phase); + } + else { + // TODO: this sched_yield can have significant impact on the performance - either positive or negative + // depending on the workload and the operating system. + // since it is not clear what is the best approach, it should potentially become user-configurable + // ref: https://github.com/ggerganov/ggml/issues/291 + // UPD: adding the do_yield flag seems to resolve the issue universally + const bool do_yield = node_n < 0 || cgraph->nodes[node_n]->op == GGML_OP_MUL_MAT; + ggml_graph_compute_thread_sync_task(&task_phase, state, do_yield); + } + + if (state->ith < n_tasks) { + params.type = GGML_TASK_COMPUTE; ggml_compute_forward(¶ms, node); } + + if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) { + task_phase = GGML_TASK_FINALIZE; + atomic_store(&state->shared->n_active, n_threads); + atomic_store(&state->shared->node_task, task_phase); + } + else { + ggml_graph_compute_thread_sync_task(&task_phase, state, false); + } } return GGML_EXIT_SUCCESS; @@ -16832,8 +16933,8 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(node)) { if (node->src[0]->type != GGML_TYPE_F32) { - // here we need memory just for single 2D matrix from src0 - cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]); + // here we need memory for fully dequantized matrix from src0 + cur = ggml_type_size(GGML_TYPE_F32)*ggml_nelements(node->src[0]); } } else #endif @@ -16987,6 +17088,7 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { /*.n_threads =*/ n_threads, /*.n_active =*/ n_threads, /*.node_n =*/ -1, + /*.node_task =*/ GGML_TASK_FINALIZE, /*.abort_callback =*/ NULL, /*.abort_callback_data =*/ NULL, }; From 6f9939d119b2d004c264952eb510bd106455531e Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Mon, 22 Jan 2024 16:10:14 +0200 Subject: [PATCH 059/101] KL-divergence (#5076) * kl-divergence: be able to save all logits to a file * Add ability to compute KL-divergence --------- Co-authored-by: Iwan Kawrakow --- common/common.cpp | 9 + common/common.h | 3 + examples/perplexity/perplexity.cpp | 319 ++++++++++++++++++++++++++++- 3 files changed, 329 insertions(+), 2 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 0e4b8bab2..0a7096171 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -672,6 +672,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { if (params.logdir.back() != DIRECTORY_SEPARATOR) { params.logdir += DIRECTORY_SEPARATOR; } + } else if (arg == "--save-all-logits" || arg == "--kl-divergence-base") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.logits_file = argv[i]; } else if (arg == "--perplexity" || arg == "--all-logits") { params.logits_all = true; } else if (arg == "--ppl-stride") { @@ -716,6 +722,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.multiple_choice_tasks = std::stoi(argv[i]); + } else if (arg == "--kl-divergence") { + params.kl_divergence = true; } else if (arg == "--ignore-eos") { params.ignore_eos = true; } else if (arg == "--no-penalize-nl") { @@ -967,6 +975,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --winogrande-tasks N number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks); printf(" --multiple-choice compute multiple choice score over random tasks from datafile supplied with -f\n"); printf(" --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks); + printf(" --kl-divergence computes KL-divergence to logits provided via --kl-divergence-base"); printf(" --keep N number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep); printf(" --draft N number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft); printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); diff --git a/common/common.h b/common/common.h index c69ad7e94..214a379b5 100644 --- a/common/common.h +++ b/common/common.h @@ -91,6 +91,7 @@ struct gpt_params { std::string input_suffix = ""; // string to suffix user inputs with std::vector antiprompt; // string upon seeing which more user input is prompted std::string logdir = ""; // directory in which to save YAML log files + std::string logits_file = ""; // file for saving *all* logits std::vector kv_overrides; @@ -111,6 +112,8 @@ struct gpt_params { bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed + bool kl_divergence = false; // compute KL-divergence + bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index b7ef9a084..1b7f85f49 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -112,6 +112,43 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp}; } +static inline int nearest_int(float fval) { + //assert(fval <= 4194303.f); + float val = fval + 12582912.f; + int i; memcpy(&i, &val, sizeof(int)); + return (i & 0x007fffff) - 0x00400000; +} + +static double log_softmax(int n_vocab, const float * logits, uint16_t * log_prob, int tok) { + float max_logit = logits[0]; + float min_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) { + max_logit = std::max(max_logit, logits[i]); + min_logit = std::min(min_logit, logits[i]); + } + min_logit = std::max(min_logit, max_logit - 16); + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) { + sum_exp += expf(logits[i] - max_logit); + } + const float log_sum_exp = log(sum_exp); + const float min_log_prob = min_logit - max_logit - log_sum_exp; + const float scale = (max_logit - min_logit)/65535.f; + float * d = (float *)log_prob; + d[0] = scale; + d[1] = min_log_prob; + log_prob += 4; + if (scale) { + const float inv_scale = 1/scale; + for (int i = 0; i < n_vocab; ++i) { + log_prob[i] = logits[i] > min_logit ? nearest_int(inv_scale*(logits[i] - min_logit)) : 0; + } + } else { + std::memset(log_prob, 0, n_vocab*sizeof(uint16_t)); + } + return max_logit + log_sum_exp - logits[tok]; +} + static void process_logits( int n_vocab, const float * logits, const int * tokens, int n_token, std::vector & workers, double & nll, double & nll2, float * logit_history, float * prob_history @@ -147,6 +184,114 @@ static void process_logits( } } +static void process_logits(std::ostream& out, int n_vocab, const float * logits, const int * tokens, int n_token, + std::vector & workers, std::vector & log_probs, double & nll, double & nll2) { + std::mutex mutex; + const int nv = 2*((n_vocab + 1)/2) + 4; + int counter = 0; + auto compute = [&mutex, &counter, &log_probs, &nll, &nll2, n_vocab, logits, tokens, n_token, nv] () { + double local_nll = 0; + double local_nll2 = 0; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + nll += local_nll; nll2 += local_nll2; + break; + } + lock.unlock(); + const double v = log_softmax(n_vocab, logits + i*n_vocab, log_probs.data() + i*nv, tokens[i+1]); + local_nll += v; + local_nll2 += v*v; + } + }; + for (auto & w : workers) { + w = std::thread(compute); + } + compute(); + for (auto & w : workers) { + w.join(); + } + out.write((const char *)log_probs.data(), n_token*nv*sizeof(uint16_t)); +} + +struct kl_divergence_result { + double sum_nll = 0; + double sum_nll2 = 0; + double sum_kld = 0; + double sum_kld2 = 0; + double sum_nll_diff = 0; + double sum_nll_diff2 = 0; + size_t count = 0; +}; + +static void log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { + float max_logit = logits[0]; + for (int i = 1; i < n_vocab; ++i) { + max_logit = std::max(max_logit, logits[i]); + } + double sum_exp = 0.0; + for (int i = 0; i < n_vocab; ++i) { + sum_exp += expf(logits[i] - max_logit); + } + const float log_sum_exp = log(sum_exp); + const float * d = (const float *)base_log_prob; + const float scale = d[0]; + const float min_log_prob = d[1]; + base_log_prob += 4; + float nll = max_logit + log_sum_exp - logits[tok]; + kld.sum_nll += nll; + kld.sum_nll2 += nll*nll; + nll += (scale*base_log_prob[tok] + min_log_prob); + kld.sum_nll_diff += nll; + kld.sum_nll_diff2 += nll*nll; + max_logit += log_sum_exp; + double sum = 0; + for (int i = 0; i < n_vocab; ++i) { + const float p_log_base = scale*base_log_prob[i] + min_log_prob; + if (p_log_base > -16.f) { + const float p_base = expf(p_log_base); + sum += p_base * (p_log_base - logits[i] + max_logit); + } + } + kld.sum_kld += sum; + kld.sum_kld2 += sum*sum; + ++kld.count; +} + +static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, + std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld) { + std::mutex mutex; + const int nv = 2*((n_vocab + 1)/2) + 4; + int counter = 0; + auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv] () { + kl_divergence_result local_kld; + while (true) { + std::unique_lock lock(mutex); + int i = counter++; + if (i >= n_token) { + kld.sum_nll += local_kld.sum_nll; + kld.sum_nll2 += local_kld.sum_nll2; + kld.sum_kld += local_kld.sum_kld; + kld.sum_kld2 += local_kld.sum_kld2; + kld.sum_nll_diff += local_kld.sum_nll_diff; + kld.sum_nll_diff2 += local_kld.sum_nll_diff2; + kld.count += local_kld.count; + break; + } + lock.unlock(); + log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + } + }; + for (auto & w : workers) { + w = std::thread(compute); + } + compute(); + for (auto & w : workers) { + w.join(); + } +} + static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) { // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw` @@ -294,6 +439,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const int n_ctx = llama_n_ctx(ctx); + std::ofstream logits_stream; + if (!params.logits_file.empty()) { + logits_stream.open(params.logits_file.c_str()); + if (!logits_stream.is_open()) { + fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str()); + return {}; + } + fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str()); + logits_stream.write("_logits_", 8); + logits_stream.write((const char *)&n_ctx, sizeof(n_ctx)); + } + auto tim1 = std::chrono::high_resolution_clock::now(); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); @@ -336,6 +493,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par std::vector workers(std::thread::hardware_concurrency() - 1); + std::vector log_probs; + if (!params.logits_file.empty()) { + logits_stream.write((const char *)&n_vocab, sizeof(n_vocab)); + logits_stream.write((const char *)&n_chunk, sizeof(n_chunk)); + logits_stream.write((const char *)tokens.data(), n_chunk*n_ctx*sizeof(tokens[0])); + const int nv = 2*((n_vocab + 1)/2) + 4; + log_probs.resize(n_ctx * nv); + } + for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; const int end = start + n_ctx; @@ -398,8 +564,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // process the entire prompt. const int first = n_ctx/2; const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); - process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + if (!params.logits_file.empty()) { + process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, log_probs, nll, nll2); + } else { + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first); + } count += n_ctx - first - 1; // perplexity is e^(average negative log-likelihood) @@ -1414,6 +1585,148 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params printf("\n"); } +static void kl_divergence(llama_context * ctx, const gpt_params & params) { + if (params.logits_file.empty()) { + fprintf(stderr, "%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__); + return; + } + std::ifstream in(params.logits_file.c_str(), std::ios::binary); + if (!in) { + fprintf(stderr, "%s: failed to open %s\n", __func__, params.logits_file.c_str()); + return; + } + { + char check[9]; check[8] = 0; + in.read(check, 8); + if (in.fail() || strncmp("_logits_", check, 8) != 0) { + fprintf(stderr, "%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str()); + return; + } + } + + uint32_t n_ctx; + in.read((char *)&n_ctx, sizeof(n_ctx)); + if (n_ctx > llama_n_ctx(ctx)) { + fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n", + __func__, params.logits_file.c_str(), n_ctx, params.n_ctx); + } + + int n_vocab, n_chunk; + in.read((char *)&n_vocab, sizeof(n_vocab)); + in.read((char *)&n_chunk, sizeof(n_chunk)); + if (in.fail()) { + fprintf(stderr, "%s: failed rwading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); + return; + } + if (n_vocab != llama_n_vocab(llama_get_model(ctx))) { + fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx))); + } + + std::vector tokens(n_ctx * n_chunk); + if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) { + fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str()); + return; + } + + const int n_batch = params.n_batch; + const int num_batches = (n_ctx + n_batch - 1)/n_batch; + const int nv = 2*((n_vocab + 1)/2) + 4; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + + std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); + std::vector logits; + if (num_batches > 1) { + logits.reserve(n_ctx * n_vocab); + } + + std::vector workers(std::thread::hardware_concurrency() - 1); + + auto mean_and_uncertainty = [] (double sum, double sum2, size_t count) { + if (count < 1) { + return std::make_pair(0., 0.); + } + double f = sum/count; + double df = sum2/count - f*f; + df = df > 0 && count > 10 ? sqrt(df/(count-1)) : 0.; + return std::make_pair(f, df); + }; + + kl_divergence_result kld; + + for (int i = 0; i < n_chunk; ++i) { + const int start = i * n_ctx; + const int end = start + n_ctx; + + const auto t_start = std::chrono::high_resolution_clock::now(); + + if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) { + fprintf(stderr, "%s: failed reading log-probs for chunk %d\n", __func__, i); + return; + } + + // clear the KV cache + llama_kv_cache_clear(ctx); + + for (int j = 0; j < num_batches; ++j) { + const int batch_start = start + j * n_batch; + const int batch_size = std::min(end - batch_start, n_batch); + + // save original token and restore it after eval + const auto token_org = tokens[batch_start]; + + // add BOS token for the first batch of each chunk + if (add_bos && j == 0) { + tokens[batch_start] = llama_token_bos(llama_get_model(ctx)); + } + + if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return; + } + + // restore the original token in case it was set to BOS + tokens[batch_start] = token_org; + + if (num_batches > 1) { + const auto * batch_logits = llama_get_logits(ctx); + logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab); + } + } + + const auto t_end = std::chrono::high_resolution_clock::now(); + + if (i == 0) { + const float t_total = std::chrono::duration(t_end - t_start).count(); + fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total); + int total_seconds = (int)(t_total * n_chunk); + if (total_seconds >= 60*60) { + fprintf(stderr, "%d hours ", total_seconds / (60*60)); + total_seconds = total_seconds % (60*60); + } + fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); + + printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence\n"); + } + + const int first = n_ctx/2; + const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); + process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, + workers, log_probs_uint16, kld); + + auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); + auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count); + auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + + printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf\n", i+1, exp(ppl.first), + log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second); + + fflush(stdout); + + logits.clear(); + } + printf("\n"); + +} int main(int argc, char ** argv) { gpt_params params; @@ -1476,6 +1789,8 @@ int main(int argc, char ** argv) { winogrande_score(ctx, params); } else if (params.multiple_choice) { multiple_choice_score(ctx, params); + } else if (params.kl_divergence) { + kl_divergence(ctx, params); } else { results = perplexity(ctx, params); } From 011e8ec577fd135cbc02993d3ea9840c516d6a1c Mon Sep 17 00:00:00 2001 From: slaren Date: Mon, 22 Jan 2024 23:42:41 +0100 Subject: [PATCH 060/101] llama : fix not enough space in buffer with Qwen (#5086) --- llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 8c906a22f..f6f1ec0f4 100644 --- a/llama.cpp +++ b/llama.cpp @@ -4440,9 +4440,9 @@ static struct ggml_tensor * llm_build_kv( // these nodes are added to the graph together so that they are not reordered // by doing so, the number of splits in the graph is reduced + ggml_build_forward_expand(graph, q_cur); ggml_build_forward_expand(graph, k_cur); ggml_build_forward_expand(graph, v_cur); - ggml_build_forward_expand(graph, q_cur); llm_build_kv_store(ctx, hparams, kv, graph, k_cur, v_cur, n_ctx, n_tokens, kv_head, cb, il); From 125d03a5036a02a983c8e98c2cdc126e061afb8e Mon Sep 17 00:00:00 2001 From: Michael Coppola Date: Tue, 23 Jan 2024 01:51:27 -0500 Subject: [PATCH 061/101] llama.vim : added api key support (#5090) Co-authored-by: Michael Coppola --- examples/llama.vim | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/examples/llama.vim b/examples/llama.vim index f03fadfb7..1b5ad6ba0 100644 --- a/examples/llama.vim +++ b/examples/llama.vim @@ -6,7 +6,7 @@ " Similarly, you could add an insert mode keybind with " inoremap call llama#doLlamaGen() " -" g:llama_api_url and g:llama_overrides can be configured in your .vimrc +" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc " let g:llama_api_url = "192.168.1.10:8080" " llama_overrides can also be set through buffer/window scopes. For instance " autocmd filetype python let b:llama_overrides = {"temp": 0.2} @@ -82,6 +82,9 @@ func llama#doLlamaGen() endif let l:querydata.prompt = join(l:buflines, "\n") let l:curlcommand = copy(s:curlcommand) + if exists("g:llama_api_key") + call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key]) + endif let l:curlcommand[2] = json_encode(l:querydata) let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])}) endfunction From 2bed4aa3f37cb4e39e16e9ec7b595a7738fd5faf Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Tue, 23 Jan 2024 08:11:39 +0100 Subject: [PATCH 062/101] devops : add intel oneapi dockerfile (#5068) Co-authored-by: Xuan Son Nguyen --- .devops/main-intel.Dockerfile | 26 ++++++++++++++++++++++++++ .github/workflows/docker.yml | 1 + CMakeLists.txt | 5 +++++ 3 files changed, 32 insertions(+) create mode 100644 .devops/main-intel.Dockerfile diff --git a/.devops/main-intel.Dockerfile b/.devops/main-intel.Dockerfile new file mode 100644 index 000000000..e1e6acc24 --- /dev/null +++ b/.devops/main-intel.Dockerfile @@ -0,0 +1,26 @@ +ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 +ARG UBUNTU_VERSION=22.04 + +FROM intel/hpckit:$ONEAPI_VERSION as build + +RUN apt-get update && \ + apt-get install -y git + +WORKDIR /app + +COPY . . + +# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance +RUN mkdir build && \ + cd build && \ + cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \ + cmake --build . --config Release --target main server + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/build/bin/main /main +COPY --from=build /app/build/bin/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/main" ] diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 87904b75e..825b8f503 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -35,6 +35,7 @@ jobs: - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" } steps: - name: Check out the repo uses: actions/checkout@v3 diff --git a/CMakeLists.txt b/CMakeLists.txt index 5a333ff52..af3665129 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -478,6 +478,11 @@ function(get_flags CCID CCVER) if (CCVER VERSION_GREATER_EQUAL 8.1.0) set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi) endif() + elseif (CCID MATCHES "Intel") + # enable max optimization level when using Intel compiler + set(C_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + add_link_options(-fuse-ld=lld -static-intel) endif() set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE) From 89758723c75ba594e401f6513751beeba7ca1d28 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 23 Jan 2024 14:12:57 +0200 Subject: [PATCH 063/101] minor : clean-up some warnings and style (#5094) * minor : clean-up some warnings and style ggml-ci * ggml : add comment --- common/common.cpp | 10 +++---- examples/llava/clip.cpp | 47 ++++++++++++------------------ examples/perplexity/perplexity.cpp | 4 +-- ggml.c | 9 +++--- ggml.h | 1 - llama.cpp | 24 +++++++-------- 6 files changed, 42 insertions(+), 53 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index 0a7096171..6b07f1197 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -216,12 +216,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { } // store the external file name in params params.prompt_file = argv[i]; - file.seekg(0, std::ios::end); - size_t size = file.tellg(); - file.seekg(0, std::ios::beg); - params.prompt.resize(size); - file.read((char *)params.prompt.data(), size); - fprintf(stderr, "Read %zu bytes from binary file %s\n", size, argv[i]); + std::ostringstream ss; + ss << file.rdbuf(); + params.prompt = ss.str(); + fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), argv[i]); } else if (arg == "-f" || arg == "--file") { if (++i >= argc) { invalid_param = true; diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 6161fd858..4a0338a37 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -2,18 +2,6 @@ // so there might be still unnecessary artifacts hanging around // I'll gradually clean and extend it -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include - #include "clip.h" #include "ggml.h" #include "ggml-alloc.h" @@ -30,6 +18,19 @@ #define STB_IMAGE_IMPLEMENTATION #include "stb_image.h" +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + static std::string format(const char * fmt, ...) { va_list ap; va_list ap2; @@ -217,9 +218,9 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) { static void print_tensor_info(const ggml_tensor* tensor, const char* prefix = "") { size_t tensor_size = ggml_nbytes(tensor); - printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%d, %d, %d, %d], type: %d\n", + printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n", prefix, ggml_n_dims(tensor), tensor->name, tensor_size, - tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->type); + tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type)); } static projector_type clip_projector_type_from_string(const std::string & name) { @@ -592,7 +593,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3)); mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]); // stride = 1, padding = 1, bias is nullptr - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, nullptr, 1, 1, 1, 1, 1, 1); + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1); // layer norm // // block_1 shape = [1, 2048, 24, 24], ne = [24, 24, 2048, 1] @@ -640,7 +641,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // block_2 { // stride = 2 - block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, nullptr, 2, 2, 1, 1, 1, 1); + block_1 = ggml_conv_depthwise_2d(ctx0, model.mm_model_block_2_block_0_0_w, block_1, 2, 2, 1, 1, 1, 1); // block_1 shape = [1, 2048, 12, 12], ne = [12, 12, 2048, 1] // layer norm @@ -741,18 +742,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { { std::map n_type; - uint32_t n_type_max = 0; - enum ggml_type type_max = GGML_TYPE_F32; - for (int i = 0; i < n_tensors; i++) { enum ggml_type type = gguf_get_tensor_type(ctx, i); n_type[type]++; - - if (n_type_max < n_type[type]) { - n_type_max = n_type[type]; - type_max = type; - } } printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__); @@ -795,14 +788,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { size_t tensor_size = ggml_nbytes(cur); buffer_size += tensor_size; if (verbosity >= 3) { - printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%d, %d, %d, %d], type: %d\n", __func__, i, - ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], type); + printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n", + __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type)); } } } - - buffer_size += n_tensors * 128 /* CLIP PADDING */; clip_ctx * new_clip = new clip_ctx; diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 1b7f85f49..de6d3eb41 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1202,11 +1202,11 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma); } -static bool deserialize_string(std::istream& in, std::string& str) { +static bool deserialize_string(std::istream & in, std::string & str) { uint32_t size; if (!in.read((char *)&size, sizeof(size)).fail()) { str.resize(size); - if (!in.read((char *)str.data(), size).fail()) return true; + if (!in.read((char *)&str[0], size).fail()) return true; } return false; } diff --git a/ggml.c b/ggml.c index f85045c9c..ca98fde8a 100644 --- a/ggml.c +++ b/ggml.c @@ -5368,14 +5368,12 @@ struct ggml_tensor * ggml_conv_depthwise_2d( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - struct ggml_tensor * c, int s0, int s1, int p0, int p1, int d0, int d1) { - struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]); struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]), @@ -9991,7 +9989,7 @@ static void ggml_compute_forward_mul_mat( return; } - const int64_t tgemm0 = ggml_perf_time_us(); + //const int64_t tgemm0 = ggml_perf_time_us(); for (int64_t i13 = 0; i13 < ne13; i13++) { for (int64_t i12 = 0; i12 < ne12; i12++) { const int64_t i03 = i13/r3; @@ -16934,7 +16932,10 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa if (ggml_compute_forward_mul_mat_use_blas(node)) { if (node->src[0]->type != GGML_TYPE_F32) { // here we need memory for fully dequantized matrix from src0 - cur = ggml_type_size(GGML_TYPE_F32)*ggml_nelements(node->src[0]); + // take into account that src0 can be broadcasted into src1[2,3] + cur = ggml_type_size(GGML_TYPE_F32) + * node->src[0]->ne[0]*node->src[0]->ne[1] + * node->src[1]->ne[2]*node->src[1]->ne[3]; } } else #endif diff --git a/ggml.h b/ggml.h index dca7bd9ce..1c4976271 100644 --- a/ggml.h +++ b/ggml.h @@ -1499,7 +1499,6 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, - struct ggml_tensor * c, int s0, int s1, int p0, diff --git a/llama.cpp b/llama.cpp index f6f1ec0f4..582e82260 100644 --- a/llama.cpp +++ b/llama.cpp @@ -2300,18 +2300,18 @@ struct llama_model_loader { } switch (type_max) { - case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; - case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; - case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; - case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; - case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; - case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1; break; - case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break; - case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; - case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; - case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; - case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; - case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; + case GGML_TYPE_F32: ftype = LLAMA_FTYPE_ALL_F32; break; + case GGML_TYPE_F16: ftype = LLAMA_FTYPE_MOSTLY_F16; break; + case GGML_TYPE_Q4_0: ftype = LLAMA_FTYPE_MOSTLY_Q4_0; break; + case GGML_TYPE_Q4_1: ftype = LLAMA_FTYPE_MOSTLY_Q4_1; break; + case GGML_TYPE_Q5_0: ftype = LLAMA_FTYPE_MOSTLY_Q5_0; break; + case GGML_TYPE_Q5_1: ftype = LLAMA_FTYPE_MOSTLY_Q5_1; break; + case GGML_TYPE_Q8_0: ftype = LLAMA_FTYPE_MOSTLY_Q8_0; break; + case GGML_TYPE_Q2_K: ftype = LLAMA_FTYPE_MOSTLY_Q2_K; break; + case GGML_TYPE_Q3_K: ftype = LLAMA_FTYPE_MOSTLY_Q3_K_M; break; + case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break; + case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; + case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break; case GGML_TYPE_IQ2_XS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS; break; default: From 9ecdd12e95aee20d6dfaf5f5a0f0ce5ac1fb2747 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Tue, 23 Jan 2024 13:31:56 +0100 Subject: [PATCH 064/101] CUDA: more info when no device code (#5088) --- ggml-cuda.cu | 89 +++++++++++++++++++++++++++++++--------------------- 1 file changed, 54 insertions(+), 35 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index ec3837fb8..7f460449e 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -13,6 +13,10 @@ #include #include +// stringize macro for converting __CUDA_ARCH_LIST__ (list of integers) to string +#define STRINGIZE_IMPL(...) #__VA_ARGS__ +#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__) + #if defined(GGML_USE_HIPBLAS) #include #include @@ -584,13 +588,28 @@ static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0, static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr}; [[noreturn]] -static __device__ void bad_arch() { - printf("ERROR: ggml-cuda was compiled without support for the current GPU architecture.\n"); +static __device__ void no_device_code( + const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) { + +#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) + printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n", + file_name, line, function_name, arch); + (void) arch_list; +#else + printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n", + file_name, line, function_name, arch, arch_list); +#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) __trap(); - (void) bad_arch; // suppress unused function warning + (void) no_device_code; // suppress unused function warning } +#ifdef __CUDA_ARCH__ +#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__)) +#else +#define NO_DEVICE_CODE GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.") +#endif // __CUDA_ARCH__ + static __device__ __forceinline__ float warp_reduce_sum(float x) { #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { @@ -617,7 +636,7 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) { return a; #else (void) a; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL } @@ -638,7 +657,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) { return x; #else (void) x; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } @@ -2421,7 +2440,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h } #else (void) vx; (void) y; (void) k; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_PASCAL } @@ -2452,7 +2471,7 @@ template static __device__ __forceinline__ float vec_dot_q4_0_q8_1_imp // second part effectively subtracts 8 from each quant value return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2489,7 +2508,7 @@ template static __device__ __forceinline__ float vec_dot_q4_1_q8_1_imp // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2524,7 +2543,7 @@ template static __device__ __forceinline__ float vec_dot_q5_0_q8_1_imp // second part effectively subtracts 16 from each quant value return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2569,7 +2588,7 @@ template static __device__ __forceinline__ float vec_dot_q5_1_q8_1_imp return sumi*d5d8 + m5s8 / (QI5_1 / vdr); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2590,7 +2609,7 @@ template static __device__ __forceinline__ float vec_dot_q8_0_q8_1_imp return d8_0*d8_1 * sumi; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2620,7 +2639,7 @@ template static __device__ __forceinline__ float vec_dot_q8_1_q8_1_imp // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it return sumi*d8d8 + m8s8 / (QI8_1 / vdr); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2655,7 +2674,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq( return dm2f.x*sumf_d - dm2f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2692,7 +2711,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq( return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m); #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2732,7 +2751,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq( return d3 * sumf; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2757,7 +2776,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq( return d3*d8 * sumi; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2790,7 +2809,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2823,7 +2842,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2863,7 +2882,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq( return dm5f.x*sumf_d - dm5f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2896,7 +2915,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq( return dm4f.x*sumf_d - dm4f.y*sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2926,7 +2945,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq( return d*sumf; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -2957,7 +2976,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq( return d6 * sumf_d; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A } @@ -3823,7 +3842,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( return dall * sumf_d - dmin * sumf_m; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A #endif @@ -4006,7 +4025,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( return d * sumf_d; #else - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= MIN_CC_DP4A #endif @@ -4501,7 +4520,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4570,7 +4589,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_1_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4637,7 +4656,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4704,7 +4723,7 @@ mul_mat_q5_1( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_1_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4771,7 +4790,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q8_0_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4838,7 +4857,7 @@ mul_mat_q2_K( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q2_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4907,7 +4926,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q3_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -4976,7 +4995,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q4_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5043,7 +5062,7 @@ mul_mat_q5_K( (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q5_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5112,7 +5131,7 @@ template static __global__ void (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst); #else (void) vec_dot_q6_K_q8_1_mul_mat; - bad_arch(); + NO_DEVICE_CODE; #endif // __CUDA_ARCH__ >= CC_VOLTA } @@ -5835,7 +5854,7 @@ static __global__ void soft_max_f16(const float * x, const float * y, float * ds } #else (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale; - bad_arch(); + NO_DEVICE_CODE; #endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL && CUDART_VERSION >= CUDART_HMAX } From 44879ee885f48ecf4675dd216b373dce0a6f3690 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Tue, 23 Jan 2024 15:17:20 +0200 Subject: [PATCH 065/101] Additional KL-divergence statistics (#5081) * perplexity: add top-token probability * perplexity: add additional KL-divergence statistics * perplexity: a better organized KL-divergence statistics output --------- Co-authored-by: Iwan Kawrakow --- examples/perplexity/perplexity.cpp | 71 +++++++++++++++++++++++++----- 1 file changed, 61 insertions(+), 10 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index de6d3eb41..8d2204969 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -222,13 +222,18 @@ struct kl_divergence_result { double sum_kld2 = 0; double sum_nll_diff = 0; double sum_nll_diff2 = 0; + size_t n_same_top = 0; size_t count = 0; }; -static void log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { +static double log_softmax(int n_vocab, const float * logits, const uint16_t * base_log_prob, int tok, kl_divergence_result & kld) { float max_logit = logits[0]; + int imax = 0; for (int i = 1; i < n_vocab; ++i) { - max_logit = std::max(max_logit, logits[i]); + if (logits[i] > max_logit) { + max_logit = logits[i]; + imax = i; + } } double sum_exp = 0.0; for (int i = 0; i < n_vocab; ++i) { @@ -247,8 +252,14 @@ static void log_softmax(int n_vocab, const float * logits, const uint16_t * base kld.sum_nll_diff2 += nll*nll; max_logit += log_sum_exp; double sum = 0; + int imax_base = -1; + float p_log_base_max = 0; for (int i = 0; i < n_vocab; ++i) { const float p_log_base = scale*base_log_prob[i] + min_log_prob; + if (i == 0 || p_log_base > p_log_base_max) { + p_log_base_max = p_log_base; + imax_base = i; + } if (p_log_base > -16.f) { const float p_base = expf(p_log_base); sum += p_base * (p_log_base - logits[i] + max_logit); @@ -257,14 +268,17 @@ static void log_softmax(int n_vocab, const float * logits, const uint16_t * base kld.sum_kld += sum; kld.sum_kld2 += sum*sum; ++kld.count; + if (imax == imax_base) ++kld.n_same_top; + return sum; } static void process_logits(int n_vocab, const float * logits, const int * tokens, int n_token, - std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld) { + std::vector & workers, const std::vector & base_log_probs, kl_divergence_result & kld, + float * kld_values) { std::mutex mutex; const int nv = 2*((n_vocab + 1)/2) + 4; int counter = 0; - auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv] () { + auto compute = [&mutex, &counter, &base_log_probs, &kld, n_vocab, logits, tokens, n_token, nv, kld_values] () { kl_divergence_result local_kld; while (true) { std::unique_lock lock(mutex); @@ -276,11 +290,13 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens kld.sum_kld2 += local_kld.sum_kld2; kld.sum_nll_diff += local_kld.sum_nll_diff; kld.sum_nll_diff2 += local_kld.sum_nll_diff2; + kld.n_same_top += local_kld.n_same_top; kld.count += local_kld.count; break; } lock.unlock(); - log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + double v = log_softmax(n_vocab, logits + i*n_vocab, base_log_probs.data() + i*nv, tokens[i+1], local_kld); + kld_values[i] = (float)v; } }; for (auto & w : workers) { @@ -1615,7 +1631,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { in.read((char *)&n_vocab, sizeof(n_vocab)); in.read((char *)&n_chunk, sizeof(n_chunk)); if (in.fail()) { - fprintf(stderr, "%s: failed rwading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); + fprintf(stderr, "%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str()); return; } if (n_vocab != llama_n_vocab(llama_get_model(ctx))) { @@ -1634,6 +1650,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); std::vector log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv); + std::vector kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk); std::vector logits; if (num_batches > 1) { logits.reserve(n_ctx * n_vocab); @@ -1652,6 +1669,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { }; kl_divergence_result kld; + auto kld_ptr = kld_values.data(); for (int i = 0; i < n_chunk; ++i) { const int start = i * n_ctx; @@ -1705,20 +1723,24 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { } fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); - printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence\n"); + printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL-Divergence Same top\n"); } const int first = n_ctx/2; const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first, - workers, log_probs_uint16, kld); + workers, log_probs_uint16, kld, kld_ptr); + kld_ptr += n_ctx - 1 - first; auto ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count); auto log_ppl_ratio = mean_and_uncertainty(kld.sum_nll_diff, kld.sum_nll_diff2, kld.count); auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + auto p_top = 1.*kld.n_same_top/kld.count; + auto d_p_top = sqrt(p_top*(1 - p_top)/(kld.count - 1)); - printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf\n", i+1, exp(ppl.first), - log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second); + printf("%4d %10.4lf %10.5lf ± %10.5f %10.5f ± %10.5lf %.5f ± %.5f\n", i+1, exp(ppl.first), + log_ppl_ratio.first, log_ppl_ratio.second, kl_div.first, kl_div.second, + p_top, d_p_top); fflush(stdout); @@ -1726,6 +1748,35 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { } printf("\n"); + if (kld.count < 100) return; // we do not wish to do statistics on so few values + + std::sort(kld_values.begin(), kld_values.end()); + + printf("===== KL-divergence statistics\n"); + auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count); + printf("Average: %10.6f ±%10.6lf\n", kl_div.first, kl_div.second); + auto kld_median = kld_values.size()%2 == 0 ? 0.5f*(kld_values[kld_values.size()/2] + kld_values[kld_values.size()/2-1]) + : kld_values[kld_values.size()/2]; + printf("Median : %10.6f\n", kld_median); + + auto percentile = [&kld_values] (float fraction) { + if (fraction <= 0) return kld_values.front(); + if (fraction >= 1) return kld_values.back(); + float p = fraction*(kld_values.size() - 1); + size_t ip = size_t(p); p -= ip; + return (1 - p)*kld_values[ip] + p*kld_values[std::min(ip+1, kld_values.size()-1)]; + }; + + printf("Maximum: %10.6f\n", kld_values.back()); + printf("KLD_99 : %10.6f\n", percentile(0.99f)); + printf("KLD_95 : %10.6f\n", percentile(0.95f)); + printf("KLD_90 : %10.6f\n", percentile(0.90f)); + + printf("Minimum: %10.6f\n", kld_values.front()); + printf("KLD_01 : %10.6f\n", percentile(0.01f)); + printf("KLD_05 : %10.6f\n", percentile(0.05f)); + printf("KLD_10 : %10.6f\n", percentile(0.10f)); + } int main(int argc, char ** argv) { From 26d607608d794efa56df3bdb6043a2f94c1d632c Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Tue, 23 Jan 2024 15:50:56 +0200 Subject: [PATCH 066/101] metal : disable support for MUL_MAT F32 x F16 --- ggml-metal.m | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml-metal.m b/ggml-metal.m index 912ddc83f..4b3eb4914 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -668,7 +668,8 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const return true; case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT_ID: - return ctx->support_simdgroup_reduction; + return ctx->support_simdgroup_reduction && + (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32); case GGML_OP_CPY: case GGML_OP_DUP: case GGML_OP_CONT: From 1387ea21178f9f154944013d4dd9764b54c69deb Mon Sep 17 00:00:00 2001 From: slaren Date: Wed, 24 Jan 2024 12:48:14 +0100 Subject: [PATCH 067/101] llama : pre-allocate input tensors in a separate buffer (#5100) --- ggml-alloc.c | 4 +- llama.cpp | 345 ++++++++++++++++++++++++--------------------------- 2 files changed, 167 insertions(+), 182 deletions(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 89b85d348..60141a34d 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -109,8 +109,8 @@ void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { if (block->size >= size) { best_fit_block = alloc->n_free_blocks - 1; } else { - fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n", - __func__, size, max_avail); + fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, largest block available %zu)\n", + __func__, tensor->name, size, max_avail); GGML_ASSERT(!"not enough space in the buffer"); return; } diff --git a/llama.cpp b/llama.cpp index 582e82260..114046db9 100644 --- a/llama.cpp +++ b/llama.cpp @@ -1669,6 +1669,9 @@ struct llama_context { for (ggml_backend_t backend : backends) { ggml_backend_free(backend); } + + ggml_backend_buffer_free(buf_input); + ggml_free(ctx_input); } llama_cparams cparams; @@ -1715,8 +1718,14 @@ struct llama_context { // allocator for the input tensors ggml_tallocr * alloc = nullptr; - // temporary buffer for copying data to/from the backend - std::vector> buf_copy; + // input tensors + ggml_backend_buffer_t buf_input = nullptr; + ggml_context * ctx_input = nullptr; + struct ggml_tensor * inp_tokens; // I32 [n_batch] + struct ggml_tensor * inp_embd; // F32 [n_embd, n_batch] + struct ggml_tensor * inp_pos; // I32 [n_batch] + struct ggml_tensor * inp_KQ_mask; // F32 [n_ctx, n_batch] + struct ggml_tensor * inp_K_shift; // I32 [n_ctx] #ifdef GGML_USE_MPI ggml_mpi_context * ctx_mpi = NULL; @@ -4089,22 +4098,24 @@ static struct ggml_tensor * llm_build_inp_embd( const llama_hparams & hparams, const llama_batch & batch, struct ggml_tensor * tok_embd, + struct ggml_tensor * inp_tokens, + struct ggml_tensor * inp_embd, const llm_build_cb & cb) { const int64_t n_embd = hparams.n_embd; struct ggml_tensor * inpL; if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, batch.n_tokens); + struct ggml_tensor * inp_tokens_v = ggml_view_1d(ctx, inp_tokens, batch.n_tokens, 0); cb(inp_tokens, "inp_tokens", -1); - inpL = ggml_get_rows(ctx, tok_embd, inp_tokens); + inpL = ggml_get_rows(ctx, tok_embd, inp_tokens_v); } else { #ifdef GGML_USE_MPI GGML_ASSERT(false && "not implemented"); #endif - inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens); + inpL = ggml_view_2d(ctx, inp_embd, n_embd, batch.n_tokens, inp_embd->nb[1], 0); } return inpL; @@ -4118,6 +4129,7 @@ static void llm_build_k_shift( const llama_cparams & cparams, const llama_kv_cache & kv, struct ggml_cgraph * graph, + struct ggml_tensor * K_shift, llm_rope_type type, int64_t n_ctx, float freq_base, @@ -4134,9 +4146,6 @@ static void llm_build_k_shift( const float beta_fast = cparams.yarn_beta_fast; const float beta_slow = cparams.yarn_beta_slow; - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx); - cb(K_shift, "K_shift", -1); - int rope_type = 0; switch (type) { @@ -4457,6 +4466,7 @@ static struct ggml_tensor * llm_build_kv( struct llm_build_context { const llama_model & model; + const llama_context & lctx; const llama_hparams & hparams; const llama_cparams & cparams; const llama_batch & batch; @@ -4503,6 +4513,7 @@ struct llm_build_context { const llm_build_cb & cb, bool worst_case) : model (lctx.model), + lctx (lctx), hparams (model.hparams), cparams (lctx.cparams), batch (batch), @@ -4563,20 +4574,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4747,20 +4758,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4868,20 +4879,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -4990,15 +5001,15 @@ struct llm_build_context { struct ggml_tensor * pos; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); @@ -5087,19 +5098,19 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5294,11 +5305,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); for (int il = 0; il < n_layer; ++il) { @@ -5384,11 +5395,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); inpL = llm_build_norm(ctx0, inpL, hparams, @@ -5477,11 +5488,11 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); for (int il = 0; il < n_layer; ++il) { @@ -5573,20 +5584,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5696,20 +5707,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5810,20 +5821,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -5931,20 +5942,20 @@ struct llm_build_context { struct ggml_tensor * ffn_output; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6053,20 +6064,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6160,15 +6171,15 @@ struct llm_build_context { struct ggml_tensor * pos; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); @@ -6258,20 +6269,20 @@ struct llm_build_context { struct ggml_tensor * cur; struct ggml_tensor * inpL; - inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); cb(inpL, "inp_embd", -1); // inp_pos - contains the positions - struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); cb(inp_pos, "inp_pos", -1); // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); cb(KQ_mask, "KQ_mask", -1); // shift the entire K-cache if needed if (do_rope_shift) { - llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); } for (int il = 0; il < n_layer; ++il) { @@ -6365,15 +6376,7 @@ static struct ggml_cgraph * llama_build_graph( // check if we should build the worst-case graph (for memory measurement) const bool worst_case = ggml_tallocr_is_measure(lctx.alloc); - // keep track of the input that has already been allocated - bool alloc_inp_tokens = false; - bool alloc_inp_embd = false; - bool alloc_inp_pos = false; - bool alloc_inp_KQ_mask = false; - bool alloc_inp_K_shift = false; - // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) - // TODO: improve handling of input and output tensors, then replace this with ggml_set_name llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) { if (il >= 0) { ggml_format_name(cur, "%s-%d", name, il); @@ -6381,127 +6384,79 @@ static struct ggml_cgraph * llama_build_graph( ggml_set_name(cur, name); } - if (!lctx.cparams.offload_kqv) { if (strcmp(name, "kqv_merged_cont") == 0) { // all nodes between the KV store and the attention output are run on the CPU ggml_backend_sched_set_node_backend(lctx.sched, cur, lctx.backend_cpu); } } - - // - // allocate input tensors and set input data - // - - if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.token) { - const int64_t n_tokens = cur->ne[0]; - - ggml_backend_tensor_set(cur, batch.token, 0, n_tokens*ggml_element_size(cur)); - } - - alloc_inp_tokens = true; - } - - if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0 && batch.embd) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.embd) { - const int64_t n_embd = cur->ne[0]; - const int64_t n_tokens = cur->ne[1]; - - ggml_backend_tensor_set(cur, batch.embd, 0, n_tokens*n_embd*ggml_element_size(cur)); - } - - alloc_inp_embd = true; - } - - if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc) && batch.pos) { - const int64_t n_tokens = cur->ne[0]; - - static_assert(std::is_same::value, "llama_pos must be int32_t"); - ggml_backend_tensor_set(cur, batch.pos, 0, n_tokens*ggml_element_size(cur)); - } - - alloc_inp_pos = true; - } - - if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc)) { - const int64_t n_kv = cur->ne[0]; - const int64_t n_tokens = cur->ne[1]; - - float * data; - if (ggml_backend_buffer_is_host(cur->buffer)) { - data = (float *) cur->data; - } else { - lctx.buf_copy.resize(ggml_nbytes(cur)); - data = (float *) lctx.buf_copy.data(); - } - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - float f; - if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { - f = -INFINITY; - } else { - f = 0; - } - data[h*(n_kv*n_tokens) + j*n_kv + i] = f; - } - } - } - - if (data != cur->data) { - ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); - } - } - - alloc_inp_KQ_mask = true; - } - - if (!alloc_inp_K_shift && strcmp(name, "K_shift") == 0) { - ggml_tallocr_alloc(lctx.alloc, cur); - - if (!ggml_tallocr_is_measure(lctx.alloc)) { - const int64_t n_ctx = cur->ne[0]; - - int32_t * data; - if (ggml_backend_buffer_is_host(cur->buffer)) { - data = (int32_t *) cur->data; - } else { - lctx.buf_copy.resize(ggml_nbytes(cur)); - data = (int32_t *) lctx.buf_copy.data(); - } - - for (int i = 0; i < n_ctx; ++i) { - data[i] = lctx.kv_self.cells[i].delta; - } - - if (data != cur->data) { - ggml_backend_tensor_set(cur, data, 0, ggml_nbytes(cur)); - } - } - - alloc_inp_K_shift = true; - } }; struct ggml_cgraph * result = NULL; struct llm_build_context llm(lctx, batch, cb, worst_case); + // + // set input data + // + + if (!ggml_tallocr_is_measure(lctx.alloc)) { + if (batch.token) { + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_tokens, batch.token, 0, n_tokens*ggml_element_size(lctx.inp_tokens)); + } + + if (batch.embd) { + const int64_t n_embd = llm.n_embd; + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_embd, batch.embd, 0, n_tokens*n_embd*ggml_element_size(lctx.inp_embd)); + } + + if (batch.pos) { + const int64_t n_tokens = batch.n_tokens; + + ggml_backend_tensor_set(lctx.inp_pos, batch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos)); + } + + { + const int64_t n_kv = llm.n_kv; + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_mask->buffer)); + float * data = (float *) lctx.inp_KQ_mask->data; + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + const llama_pos pos = batch.pos[j]; + const llama_seq_id seq_id = batch.seq_id[j][0]; + + for (int i = 0; i < n_kv; ++i) { + float f; + if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { + f = -INFINITY; + } else { + f = 0; + } + data[h*(n_kv*n_tokens) + j*n_kv + i] = f; + } + } + } + } + + if (llm.do_rope_shift) { + const int64_t n_ctx = llm.n_ctx; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_K_shift->buffer)); + int32_t * data = (int32_t *) lctx.inp_K_shift->data; + + for (int i = 0; i < n_ctx; ++i) { + data[i] = lctx.kv_self.cells[i].delta; + } + } + } + llm.init(); switch (model.arch) { @@ -9964,6 +9919,35 @@ struct llama_context * llama_new_context_with_model( ctx->embedding.resize(hparams.n_embd); } + // graph inputs + { + ggml_init_params init_params = { + /* .mem_size */ ggml_tensor_overhead()*5, + /* .mem_buffer */ nullptr, + /* .no_alloc */ true, + }; + ctx->ctx_input = ggml_init(init_params); + + ctx->inp_tokens = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch); + ctx->inp_embd = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, hparams.n_embd, cparams.n_batch); + ctx->inp_pos = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch); + ctx->inp_KQ_mask = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx, cparams.n_batch); + ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_ctx); + + ggml_set_name(ctx->inp_tokens, "inp_tokens"); + ggml_set_name(ctx->inp_embd, "inp_embd"); + ggml_set_name(ctx->inp_pos, "inp_pos"); + ggml_set_name(ctx->inp_KQ_mask, "inp_KQ_mask"); + ggml_set_name(ctx->inp_K_shift, "inp_K_shift"); + + ctx->buf_input = ggml_backend_alloc_ctx_tensors_from_buft(ctx->ctx_input, llama_default_buffer_type_cpu(true)); + + LLAMA_LOG_INFO("%s: %10s input buffer size = %8.2f MiB\n", __func__, + ggml_backend_buffer_name(ctx->buf_input), + ggml_backend_buffer_get_size(ctx->buf_input) / 1024.0 / 1024.0); + } + + // scheduler and compute buffers { // buffer types used for the compute buffer of each backend std::vector backend_buft; @@ -9990,9 +9974,6 @@ struct llama_context * llama_new_context_with_model( // initialize scheduler with the worst-case graph ggml_backend_sched_init_measure(ctx->sched, gf); - // note: the number of splits during measure is higher than during inference due to the kv shift - int n_splits = ggml_backend_sched_get_n_splits(ctx->sched); - LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits); ctx->alloc = ggml_backend_sched_get_tallocr(ctx->sched, ctx->backend_cpu); for (ggml_backend_t backend : ctx->backends) { @@ -10001,6 +9982,10 @@ struct llama_context * llama_new_context_with_model( ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf) / 1024.0 / 1024.0); } + + // note: the number of splits during measure is higher than during inference due to the kv shift + int n_splits = ggml_backend_sched_get_n_splits(ctx->sched); + LLAMA_LOG_INFO("%s: graph splits (measure): %d\n", __func__, n_splits); } } From bf63d695b804b1c995c7ae4427a8a86936ea6d25 Mon Sep 17 00:00:00 2001 From: Michael Hueschen Date: Mon, 22 Jan 2024 03:17:05 -0700 Subject: [PATCH 068/101] nix: add cc to devShell LD_LIBRARY_PATH this fixes the error I encountered when trying to run the convert.py script in a venv: ``` $ nix develop [...]$ source .venv/bin/activate (.venv) [...]$ pip3 install -r requirements.txt <... clipped ...> [...]$ python3 ./convert.py Traceback (most recent call last): File "/home/mhueschen/projects-reference/llama.cpp/./convert.py", line 40, in from sentencepiece import SentencePieceProcessor File "/home/mhueschen/projects-reference/llama.cpp/.venv/lib/python3.11/site-packages/sentencepiece/__init__.py", line 13, in from . import _sentencepiece ImportError: libstdc++.so.6: cannot open shared object file: No such file or directory ``` however, I am not sure this is the cleanest way to address this linker issue... --- .devops/nix/package.nix | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index c25d99f01..91ddb8890 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -221,10 +221,16 @@ effectiveStdenv.mkDerivation ( ; shell = mkShell { + NIX_LD_LIBRARY_PATH = lib.makeLibraryPath [ + effectiveStdenv.cc.cc + ]; name = "shell-${finalAttrs.finalPackage.name}"; description = "contains numpy and sentencepiece"; buildInputs = [ llama-python ]; inputsFrom = [ finalAttrs.finalPackage ]; + shellHook = '' + export LD_LIBRARY_PATH=$NIX_LD_LIBRARY_PATH + ''; }; shell-extra = mkShell { From c9b316c78fba31e65879a2ec91cbafd341b88cce Mon Sep 17 00:00:00 2001 From: Michael Hueschen Date: Mon, 22 Jan 2024 16:44:10 -0700 Subject: [PATCH 069/101] nix-shell: use addToSearchPath thx to @SomeoneSerge for the suggestion! --- .devops/nix/package.nix | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/.devops/nix/package.nix b/.devops/nix/package.nix index 91ddb8890..a868a9a61 100644 --- a/.devops/nix/package.nix +++ b/.devops/nix/package.nix @@ -221,15 +221,12 @@ effectiveStdenv.mkDerivation ( ; shell = mkShell { - NIX_LD_LIBRARY_PATH = lib.makeLibraryPath [ - effectiveStdenv.cc.cc - ]; name = "shell-${finalAttrs.finalPackage.name}"; description = "contains numpy and sentencepiece"; buildInputs = [ llama-python ]; inputsFrom = [ finalAttrs.finalPackage ]; shellHook = '' - export LD_LIBRARY_PATH=$NIX_LD_LIBRARY_PATH + addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib" ''; }; From cd4fddb29f81d6a1f6d51a0c016bc6b486d68def Mon Sep 17 00:00:00 2001 From: Engininja2 <139037756+Engininja2@users.noreply.github.com> Date: Wed, 24 Jan 2024 16:18:15 -0600 Subject: [PATCH 070/101] cuda : fix 2-bit quants on amd hip (#5105) * cuda : fix 2-bit quants on amd hip * use __low2float intrinsic function for new quants --- ggml-cuda.cu | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 7f460449e..05e5d18ab 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -4283,7 +4283,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( q8 += 8; aux32 >>= 7; } - const float d = (float)bq2->d * (0.5f + aux32) * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * (0.5f + aux32) * __low2float(bq8_1[ib32].ds) * 0.25f; return d * sumi; #else // iqs is 0...15 @@ -4294,7 +4294,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1( const uint8_t * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]); const uint8_t * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]); const uint32_t aux32 = q2[2] | (q2[3] << 16); - const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * __low2float(bq8_1[ib32].ds) * 0.25f; const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127]; const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127]; const int8_t * q8 = bq8_1[ib32].qs + 16*il; @@ -4339,7 +4339,7 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1( } q8 += 8; } - const float d = (float)bq2->d * (float)bq8_1[ib32].ds.x * 0.25f; + const float d = (float)bq2->d * __low2float(bq8_1[ib32].ds) * 0.25f; return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2); #else assert(false); From ddc5a5033f948dc7ab0a3a6ec2d914d13c274077 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 25 Jan 2024 11:26:17 +0200 Subject: [PATCH 071/101] metal : show compile log messages --- ggml-metal.m | 13 ++++++++----- 1 file changed, 8 insertions(+), 5 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 4b3eb4914..60fef1a19 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -277,6 +277,10 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { NSURL * libURL = [NSURL fileURLWithPath:libPath]; GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]); ctx->library = [ctx->device newLibraryWithURL:libURL error:&error]; + if (error) { + GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } } else { GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); @@ -315,13 +319,12 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { //[options setFastMathEnabled:false]; ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error]; + if (error) { + GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); + return NULL; + } } } - - if (error) { - GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]); - return NULL; - } } // print MTL GPU family: From faa3526a1eba458120987ed8269e5616385a76f4 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Thu, 25 Jan 2024 17:58:53 +0200 Subject: [PATCH 072/101] Fix Q3_K_XS for MoE models (#5113) Co-authored-by: Iwan Kawrakow --- llama.cpp | 45 +++++++++++++++++++++++++-------------------- 1 file changed, 25 insertions(+), 20 deletions(-) diff --git a/llama.cpp b/llama.cpp index 114046db9..6a7506e85 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8829,6 +8829,23 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty auto use_more_bits = [](int i_layer, int num_layers) -> bool { return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2; }; + const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); + auto layer_info = [n_expert] (int i_layer, int n_layer, const char * name) { + if (n_expert > 1) { + // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly + // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work + // for getting the current layer as I initially thought, and we need to resort to parsing the + // tensor name. + n_layer /= n_expert; + if (sscanf(name, "blk.%d.", &i_layer) != 1) { + throw std::runtime_error(format("Failed to determine layer for tensor %s", name)); + } + if (i_layer < 0 || i_layer >= n_layer) { + throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name, n_layer)); + } + } + return std::make_pair(i_layer, n_layer); + }; if (name == tn(LLM_TENSOR_OUTPUT, "weight")) { int nx = tensor->ne[0]; @@ -8890,24 +8907,8 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty new_type = GGML_TYPE_Q2_K; } } else if (name.find("ffn_down") != std::string::npos) { - const int n_expert = std::max(1, (int)qs.model.hparams.n_expert); - int i_layer, n_layer; - if (n_expert == 1) { - i_layer = qs.i_ffn_down; - n_layer = qs.n_ffn_down; - } else { - // Believe it or not, "experts" in the FFN of Mixtral-8x7B are not consecutive, but iccasionally randomly - // sprinkled in the model. Hence, simply dividing i_ffn_down by n_expert does not work - // for getting the current layer as I initially thought, and we need to resort to parsing the - // tensor name. - n_layer = qs.n_ffn_down / n_expert; - if (sscanf(name.c_str(), "blk.%d.ffn_down", &i_layer) != 1) { - throw std::runtime_error(format("Failed to determine layer for tensor %s", name.c_str())); - } - if (i_layer < 0 || i_layer >= n_layer) { - throw std::runtime_error(format("Bad layer %d for tensor %s. Must be in [0, %d)", i_layer, name.c_str(), n_layer)); - } - } + auto info = layer_info(qs.i_ffn_down, qs.n_ffn_down, name.c_str()); + int i_layer = info.first, n_layer = info.second; if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K; else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS) { if (i_layer < n_layer/8) new_type = GGML_TYPE_Q4_K; @@ -8963,13 +8964,17 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K; } else if (name.find("ffn_gate") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_gate, qs.n_ffn_gate)) { + auto info = layer_info(qs.i_ffn_gate, qs.n_ffn_gate, name.c_str()); + int i_layer = info.first, n_layer = info.second; + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) { new_type = GGML_TYPE_Q2_K; } ++qs.i_ffn_gate; } else if (name.find("ffn_up") != std::string::npos) { - if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(qs.i_ffn_up, qs.n_ffn_up)) { + auto info = layer_info(qs.i_ffn_up, qs.n_ffn_up, name.c_str()); + int i_layer = info.first, n_layer = info.second; + if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_XS && !use_more_bits(i_layer, n_layer)) { new_type = GGML_TYPE_Q2_K; } ++qs.i_ffn_up; From 256d1bb0ddce6a0a21f5a7503019bdd5c1933cba Mon Sep 17 00:00:00 2001 From: Valentin Konovalov Date: Thu, 25 Jan 2024 12:05:51 -0500 Subject: [PATCH 073/101] android : use release cmake build type by default (#5123) --- examples/llama.android/app/build.gradle.kts | 1 + 1 file changed, 1 insertion(+) diff --git a/examples/llama.android/app/build.gradle.kts b/examples/llama.android/app/build.gradle.kts index 7815a8025..aadbe22c9 100644 --- a/examples/llama.android/app/build.gradle.kts +++ b/examples/llama.android/app/build.gradle.kts @@ -30,6 +30,7 @@ android { } externalNativeBuild { cmake { + arguments += "-DCMAKE_BUILD_TYPE=Release" cppFlags += listOf() arguments += listOf() } From d292f4f2047963f558dd516f1baaa71793e9acf2 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Thu, 25 Jan 2024 14:51:24 -0500 Subject: [PATCH 074/101] examples : make pydantic scripts pass mypy and support py3.8 (#5099) --- .../pydantic-models-to-grammar-examples.py | 43 +---- examples/pydantic_models_to_grammar.py | 166 +++++++++--------- 2 files changed, 88 insertions(+), 121 deletions(-) diff --git a/examples/pydantic-models-to-grammar-examples.py b/examples/pydantic-models-to-grammar-examples.py index cbf376652..160966649 100644 --- a/examples/pydantic-models-to-grammar-examples.py +++ b/examples/pydantic-models-to-grammar-examples.py @@ -1,14 +1,14 @@ # Function calling example using pydantic models. import datetime +import importlib import json from enum import Enum -from typing import Union, Optional +from typing import Optional, Union import requests from pydantic import BaseModel, Field - -import importlib -from pydantic_models_to_grammar import generate_gbnf_grammar_and_documentation, convert_dictionary_to_pydantic_model, add_run_method_to_dynamic_model, create_dynamic_model_from_function +from pydantic_models_to_grammar import (add_run_method_to_dynamic_model, convert_dictionary_to_pydantic_model, + create_dynamic_model_from_function, generate_gbnf_grammar_and_documentation) # Function to get completion on the llama.cpp server with grammar. @@ -35,7 +35,7 @@ class SendMessageToUser(BaseModel): print(self.message) -# Enum for the calculator function. +# Enum for the calculator tool. class MathOperation(Enum): ADD = "add" SUBTRACT = "subtract" @@ -43,7 +43,7 @@ class MathOperation(Enum): DIVIDE = "divide" -# Very simple calculator tool for the agent. +# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. class Calculator(BaseModel): """ Perform a math operation on two numbers. @@ -148,37 +148,6 @@ def get_current_datetime(output_format: Optional[str] = None): return datetime.datetime.now().strftime(output_format) -# Enum for the calculator tool. -class MathOperation(Enum): - ADD = "add" - SUBTRACT = "subtract" - MULTIPLY = "multiply" - DIVIDE = "divide" - - - -# Simple pydantic calculator tool for the agent that can add, subtract, multiply, and divide. Docstring and description of fields will be used in system prompt. -class Calculator(BaseModel): - """ - Perform a math operation on two numbers. - """ - number_one: Union[int, float] = Field(..., description="First number.") - operation: MathOperation = Field(..., description="Math operation to perform.") - number_two: Union[int, float] = Field(..., description="Second number.") - - def run(self): - if self.operation == MathOperation.ADD: - return self.number_one + self.number_two - elif self.operation == MathOperation.SUBTRACT: - return self.number_one - self.number_two - elif self.operation == MathOperation.MULTIPLY: - return self.number_one * self.number_two - elif self.operation == MathOperation.DIVIDE: - return self.number_one / self.number_two - else: - raise ValueError("Unknown operation.") - - # Example function to get the weather def get_current_weather(location, unit): """Get the current weather in a given location""" diff --git a/examples/pydantic_models_to_grammar.py b/examples/pydantic_models_to_grammar.py index 848c1c367..9acc7cc6d 100644 --- a/examples/pydantic_models_to_grammar.py +++ b/examples/pydantic_models_to_grammar.py @@ -1,15 +1,21 @@ +from __future__ import annotations + import inspect import json +import re from copy import copy -from inspect import isclass, getdoc -from types import NoneType +from enum import Enum +from inspect import getdoc, isclass +from typing import TYPE_CHECKING, Any, Callable, List, Optional, Union, get_args, get_origin, get_type_hints from docstring_parser import parse -from pydantic import BaseModel, create_model, Field -from typing import Any, Type, List, get_args, get_origin, Tuple, Union, Optional, _GenericAlias -from enum import Enum -from typing import get_type_hints, Callable -import re +from pydantic import BaseModel, Field, create_model + +if TYPE_CHECKING: + from types import GenericAlias +else: + # python 3.8 compat + from typing import _GenericAlias as GenericAlias class PydanticDataType(Enum): @@ -43,7 +49,7 @@ class PydanticDataType(Enum): SET = "set" -def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str: +def map_pydantic_type_to_gbnf(pydantic_type: type[Any]) -> str: if isclass(pydantic_type) and issubclass(pydantic_type, str): return PydanticDataType.STRING.value elif isclass(pydantic_type) and issubclass(pydantic_type, bool): @@ -57,22 +63,22 @@ def map_pydantic_type_to_gbnf(pydantic_type: Type[Any]) -> str: elif isclass(pydantic_type) and issubclass(pydantic_type, BaseModel): return format_model_and_field_name(pydantic_type.__name__) - elif get_origin(pydantic_type) == list: + elif get_origin(pydantic_type) is list: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-list" - elif get_origin(pydantic_type) == set: + elif get_origin(pydantic_type) is set: element_type = get_args(pydantic_type)[0] return f"{map_pydantic_type_to_gbnf(element_type)}-set" - elif get_origin(pydantic_type) == Union: + elif get_origin(pydantic_type) is Union: union_types = get_args(pydantic_type) union_rules = [map_pydantic_type_to_gbnf(ut) for ut in union_types] return f"union-{'-or-'.join(union_rules)}" - elif get_origin(pydantic_type) == Optional: + elif get_origin(pydantic_type) is Optional: element_type = get_args(pydantic_type)[0] return f"optional-{map_pydantic_type_to_gbnf(element_type)}" elif isclass(pydantic_type): return f"{PydanticDataType.CUSTOM_CLASS.value}-{format_model_and_field_name(pydantic_type.__name__)}" - elif get_origin(pydantic_type) == dict: + elif get_origin(pydantic_type) is dict: key_type, value_type = get_args(pydantic_type) return f"custom-dict-key-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(key_type))}-value-type-{format_model_and_field_name(map_pydantic_type_to_gbnf(value_type))}" else: @@ -106,7 +112,6 @@ def get_members_structure(cls, rule_name): return f"{cls.__name__.lower()} ::= " + " | ".join(members) if cls.__annotations__ and cls.__annotations__ != {}: result = f'{rule_name} ::= "{{"' - type_list_rules = [] # Modify this comprehension members = [ f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param_type)}' @@ -116,27 +121,25 @@ def get_members_structure(cls, rule_name): result += '"," '.join(members) result += ' "}"' - return result, type_list_rules - elif rule_name == "custom-class-any": + return result + if rule_name == "custom-class-any": result = f"{rule_name} ::= " result += "value" - type_list_rules = [] - return result, type_list_rules - else: - init_signature = inspect.signature(cls.__init__) - parameters = init_signature.parameters - result = f'{rule_name} ::= "{{"' - type_list_rules = [] - # Modify this comprehension too - members = [ - f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param.annotation)}' - for name, param in parameters.items() - if name != "self" and param.annotation != inspect.Parameter.empty - ] + return result - result += '", "'.join(members) - result += ' "}"' - return result, type_list_rules + init_signature = inspect.signature(cls.__init__) + parameters = init_signature.parameters + result = f'{rule_name} ::= "{{"' + # Modify this comprehension too + members = [ + f' "\\"{name}\\"" ":" {map_pydantic_type_to_gbnf(param.annotation)}' + for name, param in parameters.items() + if name != "self" and param.annotation != inspect.Parameter.empty + ] + + result += '", "'.join(members) + result += ' "}"' + return result def regex_to_gbnf(regex_pattern: str) -> str: @@ -269,7 +272,7 @@ def generate_gbnf_float_rules(max_digit=None, min_digit=None, max_precision=None def generate_gbnf_rule_for_type( model_name, field_name, field_type, is_optional, processed_models, created_rules, field_info=None -) -> Tuple[str, list]: +) -> tuple[str, list[str]]: """ Generate GBNF rule for a given field type. @@ -283,7 +286,7 @@ def generate_gbnf_rule_for_type( :param field_info: Additional information about the field (optional). :return: Tuple containing the GBNF type and a list of additional rules. - :rtype: Tuple[str, list] + :rtype: tuple[str, list] """ rules = [] @@ -321,8 +324,7 @@ def generate_gbnf_rule_for_type( gbnf_type, rules = model_name + "-" + field_name, rules elif gbnf_type.startswith("custom-class-"): - nested_model_rules, field_types = get_members_structure(field_type, gbnf_type) - rules.append(nested_model_rules) + rules.append(get_members_structure(field_type, gbnf_type)) elif gbnf_type.startswith("custom-dict-"): key_type, value_type = get_args(field_type) @@ -341,14 +343,14 @@ def generate_gbnf_rule_for_type( union_rules = [] for union_type in union_types: - if isinstance(union_type, _GenericAlias): + if isinstance(union_type, GenericAlias): union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type( model_name, field_name, union_type, False, processed_models, created_rules ) union_rules.append(union_gbnf_type) rules.extend(union_rules_list) - elif not issubclass(union_type, NoneType): + elif not issubclass(union_type, type(None)): union_gbnf_type, union_rules_list = generate_gbnf_rule_for_type( model_name, field_name, union_type, False, processed_models, created_rules ) @@ -424,14 +426,10 @@ def generate_gbnf_rule_for_type( else: gbnf_type, rules = gbnf_type, [] - if gbnf_type not in created_rules: - return gbnf_type, rules - else: - if gbnf_type in created_rules: - return gbnf_type, rules + return gbnf_type, rules -def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created_rules: dict) -> (list, bool, bool): +def generate_gbnf_grammar(model: type[BaseModel], processed_models: set[type[BaseModel]], created_rules: dict[str, list[str]]) -> tuple[list[str], bool]: """ Generate GBnF Grammar @@ -452,7 +450,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created ``` """ if model in processed_models: - return [] + return [], False processed_models.add(model) model_name = format_model_and_field_name(model.__name__) @@ -518,7 +516,7 @@ def generate_gbnf_grammar(model: Type[BaseModel], processed_models: set, created def generate_gbnf_grammar_from_pydantic_models( - models: List[Type[BaseModel]], outer_object_name: str = None, outer_object_content: str = None, + models: list[type[BaseModel]], outer_object_name: str | None = None, outer_object_content: str | None = None, list_of_outputs: bool = False ) -> str: """ @@ -528,7 +526,7 @@ def generate_gbnf_grammar_from_pydantic_models( * grammar. Args: - models (List[Type[BaseModel]]): A list of Pydantic models to generate the grammar from. + models (list[type[BaseModel]]): A list of Pydantic models to generate the grammar from. outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling. outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling. list_of_outputs (str, optional): Allows a list of output objects @@ -543,9 +541,9 @@ def generate_gbnf_grammar_from_pydantic_models( # root ::= UserModel | PostModel # ... """ - processed_models = set() + processed_models: set[type[BaseModel]] = set() all_rules = [] - created_rules = {} + created_rules: dict[str, list[str]] = {} if outer_object_name is None: for model in models: model_rules, _ = generate_gbnf_grammar(model, processed_models, created_rules) @@ -608,7 +606,7 @@ def get_primitive_grammar(grammar): Returns: str: GBNF primitive grammar string. """ - type_list = [] + type_list: list[type[object]] = [] if "string-list" in grammar: type_list.append(str) if "boolean-list" in grammar: @@ -666,14 +664,14 @@ triple-quotes ::= "'''" """ def generate_markdown_documentation( - pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields", + pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields", documentation_with_field_description=True ) -> str: """ Generate markdown documentation for a list of Pydantic models. Args: - pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes. + pydantic_models (list[type[BaseModel]]): list of Pydantic model classes. model_prefix (str): Prefix for the model section. fields_prefix (str): Prefix for the fields section. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -731,7 +729,7 @@ def generate_markdown_documentation( def generate_field_markdown( - field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1, + field_name: str, field_type: type[Any], model: type[BaseModel], depth=1, documentation_with_field_description=True ) -> str: """ @@ -739,8 +737,8 @@ def generate_field_markdown( Args: field_name (str): Name of the field. - field_type (Type[Any]): Type of the field. - model (Type[BaseModel]): Pydantic model class. + field_type (type[Any]): Type of the field. + model (type[BaseModel]): Pydantic model class. depth (int): Indentation depth in the documentation. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -798,7 +796,7 @@ def generate_field_markdown( return field_text -def format_json_example(example: dict, depth: int) -> str: +def format_json_example(example: dict[str, Any], depth: int) -> str: """ Format a JSON example into a readable string with indentation. @@ -819,14 +817,14 @@ def format_json_example(example: dict, depth: int) -> str: def generate_text_documentation( - pydantic_models: List[Type[BaseModel]], model_prefix="Model", fields_prefix="Fields", + pydantic_models: list[type[BaseModel]], model_prefix="Model", fields_prefix="Fields", documentation_with_field_description=True ) -> str: """ Generate text documentation for a list of Pydantic models. Args: - pydantic_models (List[Type[BaseModel]]): List of Pydantic model classes. + pydantic_models (list[type[BaseModel]]): List of Pydantic model classes. model_prefix (str): Prefix for the model section. fields_prefix (str): Prefix for the fields section. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -885,7 +883,7 @@ def generate_text_documentation( def generate_field_text( - field_name: str, field_type: Type[Any], model: Type[BaseModel], depth=1, + field_name: str, field_type: type[Any], model: type[BaseModel], depth=1, documentation_with_field_description=True ) -> str: """ @@ -893,8 +891,8 @@ def generate_field_text( Args: field_name (str): Name of the field. - field_type (Type[Any]): Type of the field. - model (Type[BaseModel]): Pydantic model class. + field_type (type[Any]): Type of the field. + model (type[BaseModel]): Pydantic model class. depth (int): Indentation depth in the documentation. documentation_with_field_description (bool): Include field descriptions in the documentation. @@ -1017,8 +1015,8 @@ def generate_and_save_gbnf_grammar_and_documentation( pydantic_model_list, grammar_file_path="./generated_grammar.gbnf", documentation_file_path="./generated_grammar_documentation.md", - outer_object_name: str = None, - outer_object_content: str = None, + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1053,8 +1051,8 @@ def generate_and_save_gbnf_grammar_and_documentation( def generate_gbnf_grammar_and_documentation( pydantic_model_list, - outer_object_name: str = None, - outer_object_content: str = None, + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1086,9 +1084,9 @@ def generate_gbnf_grammar_and_documentation( def generate_gbnf_grammar_and_documentation_from_dictionaries( - dictionaries: List[dict], - outer_object_name: str = None, - outer_object_content: str = None, + dictionaries: list[dict[str, Any]], + outer_object_name: str | None = None, + outer_object_content: str | None = None, model_prefix: str = "Output Model", fields_prefix: str = "Output Fields", list_of_outputs: bool = False, @@ -1098,7 +1096,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries( Generate GBNF grammar and documentation from a list of dictionaries. Args: - dictionaries (List[dict]): List of dictionaries representing Pydantic models. + dictionaries (list[dict]): List of dictionaries representing Pydantic models. outer_object_name (str): Outer object name for the GBNF grammar. If None, no outer object will be generated. Eg. "function" for function calling. outer_object_content (str): Content for the outer rule in the GBNF grammar. Eg. "function_parameters" or "params" for function calling. model_prefix (str): Prefix for the model section in the documentation. @@ -1120,7 +1118,7 @@ def generate_gbnf_grammar_and_documentation_from_dictionaries( return grammar, documentation -def create_dynamic_model_from_function(func: Callable): +def create_dynamic_model_from_function(func: Callable[..., Any]): """ Creates a dynamic Pydantic model from a given function's type hints and adds the function as a 'run' method. @@ -1135,6 +1133,7 @@ def create_dynamic_model_from_function(func: Callable): sig = inspect.signature(func) # Parse the docstring + assert func.__doc__ is not None docstring = parse(func.__doc__) dynamic_fields = {} @@ -1157,7 +1156,6 @@ def create_dynamic_model_from_function(func: Callable): f"Parameter '{param.name}' in function '{func.__name__}' lacks a description in the docstring") # Add parameter details to the schema - param_doc = next((d for d in docstring.params if d.arg_name == param.name), None) param_docs.append((param.name, param_doc)) if param.default == inspect.Parameter.empty: default_value = ... @@ -1166,10 +1164,10 @@ def create_dynamic_model_from_function(func: Callable): dynamic_fields[param.name] = ( param.annotation if param.annotation != inspect.Parameter.empty else str, default_value) # Creating the dynamic model - dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) + dynamic_model = create_model(f"{func.__name__}", **dynamic_fields) # type: ignore[call-overload] - for param_doc in param_docs: - dynamic_model.model_fields[param_doc[0]].description = param_doc[1].description + for name, param_doc in param_docs: + dynamic_model.model_fields[name].description = param_doc.description dynamic_model.__doc__ = docstring.short_description @@ -1182,16 +1180,16 @@ def create_dynamic_model_from_function(func: Callable): return dynamic_model -def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable): +def add_run_method_to_dynamic_model(model: type[BaseModel], func: Callable[..., Any]): """ Add a 'run' method to a dynamic Pydantic model, using the provided function. Args: - model (Type[BaseModel]): Dynamic Pydantic model class. + model (type[BaseModel]): Dynamic Pydantic model class. func (Callable): Function to be added as a 'run' method to the model. Returns: - Type[BaseModel]: Pydantic model class with the added 'run' method. + type[BaseModel]: Pydantic model class with the added 'run' method. """ def run_method_wrapper(self): @@ -1204,15 +1202,15 @@ def add_run_method_to_dynamic_model(model: Type[BaseModel], func: Callable): return model -def create_dynamic_models_from_dictionaries(dictionaries: List[dict]): +def create_dynamic_models_from_dictionaries(dictionaries: list[dict[str, Any]]): """ Create a list of dynamic Pydantic model classes from a list of dictionaries. Args: - dictionaries (List[dict]): List of dictionaries representing model structures. + dictionaries (list[dict]): List of dictionaries representing model structures. Returns: - List[Type[BaseModel]]: List of generated dynamic Pydantic model classes. + list[type[BaseModel]]: List of generated dynamic Pydantic model classes. """ dynamic_models = [] for func in dictionaries: @@ -1249,7 +1247,7 @@ def list_to_enum(enum_name, values): return Enum(enum_name, {value: value for value in values}) -def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "CustomModel") -> Type[BaseModel]: +def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name: str = "CustomModel") -> type[Any]: """ Convert a dictionary to a Pydantic model class. @@ -1258,9 +1256,9 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu model_name (str): Name of the generated Pydantic model. Returns: - Type[BaseModel]: Generated Pydantic model class. + type[BaseModel]: Generated Pydantic model class. """ - fields = {} + fields: dict[str, Any] = {} if "properties" in dictionary: for field_name, field_data in dictionary.get("properties", {}).items(): @@ -1277,7 +1275,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict, model_name: str = "Cu if items != {}: array = {"properties": items} array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items") - fields[field_name] = (List[array_type], ...) + fields[field_name] = (List[array_type], ...) # type: ignore[valid-type] else: fields[field_name] = (list, ...) elif field_type == "object": From 5eaf9964fc797d4585c214db32a463d557f3ed33 Mon Sep 17 00:00:00 2001 From: l3utterfly Date: Fri, 26 Jan 2024 05:06:22 +0900 Subject: [PATCH 075/101] llama : dynamic temperature sampling (#4972) * implemented dynamic temperature sampling from koboldcpp * removed trailing whitespace * removed unused temp parameter in llama_sample_entropy * exposed exponent_val in dynamic temp sampler * added debug check for printf statements * use nullptr in llama_sample_softmax call during llama_sample_entropy this avoids counting the time taken stats twice Co-authored-by: Georgi Gerganov * return earlier if there is only 1 candiate (i.e. max_entropy == 0) * reformat 't' case in llama_sample_queue Co-authored-by: Jared Van Bortel * check for one or zero candidates case in llama_sample_entropy --------- Co-authored-by: Georgi Gerganov Co-authored-by: Jared Van Bortel --- common/sampling.cpp | 12 +++++++- common/sampling.h | 2 ++ llama.cpp | 67 +++++++++++++++++++++++++++++++++++++++++++++ llama.h | 8 ++++++ 4 files changed, 88 insertions(+), 1 deletion(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index dd1ffeb1b..efd7eab6e 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -129,6 +129,8 @@ static void sampler_queue( const int n_vocab = llama_n_vocab(llama_get_model(ctx_main)); const float temp = params.temp; + const float dynatemp_range = params.dynatemp_range; + const float dynatemp_exponent = params.dynatemp_exponent; const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k; const float top_p = params.top_p; const float min_p = params.min_p; @@ -143,7 +145,15 @@ static void sampler_queue( case 'y': llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); break; case 'p': llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); break; case 'm': llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); break; - case 't': llama_sample_temp (ctx_main, &cur_p, temp); break; + case 't': + if (dynatemp_range > 0) { + float dynatemp_min = std::max(0.0f, temp - dynatemp_range); + float dynatemp_max = std::max(0.0f, temp + dynatemp_range); + llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent); + } else { + llama_sample_temp(ctx_main, &cur_p, temp); + } + break; default : break; } } diff --git a/common/sampling.h b/common/sampling.h index 2ee180376..88899c094 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -18,6 +18,8 @@ typedef struct llama_sampling_params { float tfs_z = 1.00f; // 1.0 = disabled float typical_p = 1.00f; // 1.0 = disabled float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities + float dynatemp_range = 0.00f; // 0.0 = disabled + float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size) float penalty_repeat = 1.10f; // 1.0 = disabled float penalty_freq = 0.00f; // 0.0 = disabled diff --git a/llama.cpp b/llama.cpp index 6a7506e85..823d42d7f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8151,6 +8151,73 @@ void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * c } } +void llama_sample_entropy(struct llama_context * ctx, llama_token_data_array * candidates_p, float min_temp, float max_temp, float exponent_val) { + const int64_t t_start_sample_us = ggml_time_us(); + + // no need to do anything if there is only one (or zero) candidates + if(candidates_p->size <= 1) { + return; + } + + // Calculate maximum possible entropy + float max_entropy = -logf(1.0f / candidates_p->size); + + llama_sample_softmax(nullptr, candidates_p); + + // Calculate entropy of the softmax probabilities + float entropy = 0.0f; + for (size_t i = 0; i < candidates_p->size; ++i) { + float prob = candidates_p->data[i].p; + if (prob > 0.0f) { // Ensure no log(0) + entropy -= prob * logf(prob); + } + } + + // Normalize the entropy (max_entropy cannot be 0 here because we checked candidates_p->size != 1 above) + float normalized_entropy = entropy / max_entropy; + + // Map the normalized entropy to the desired temperature range using the power function + float dyn_temp = min_temp + (max_temp - min_temp) * powf(normalized_entropy, exponent_val); + +#ifdef DEBUG + LLAMA_LOG_INFO("Your text maxtemp value is: %f\n", max_temp); + LLAMA_LOG_INFO("Entropy: %f\n", entropy); + LLAMA_LOG_INFO("Max Possible Entropy: %f\n", max_entropy); + LLAMA_LOG_INFO("Normalized Entropy: %f\n", normalized_entropy); + LLAMA_LOG_INFO("Exponent: %f\n", exponent_val); + LLAMA_LOG_INFO("Dynamic Temperature (dyn_temp): %f\n", dyn_temp); +#endif + + // Apply the dynamically calculated temperature scaling + for (size_t i = 0; i < candidates_p->size; ++i) { + candidates_p->data[i].logit /= dyn_temp; + } + + // Re-compute softmax probabilities after scaling logits with dynamic temperature + double max_l_double = candidates_p->data[0].logit; + double cum_sum_double = 0.0; + for (size_t i = 0; i < candidates_p->size; ++i) { + double p = exp(candidates_p->data[i].logit - max_l_double); + candidates_p->data[i].p = p; // Store the scaled probability + cum_sum_double += p; + } + for (size_t i = 0; i < candidates_p->size; ++i) { + candidates_p->data[i].p /= cum_sum_double; // Re-normalize the probabilities + } + +#ifdef DEBUG + // Print the updated top 25 probabilities after temperature scaling + LLAMA_LOG_INFO("\nUpdated Top 25 Probabilities After Dynamic Temperature Scaling (in percentages):\n"); + for (size_t i = 0; i < 25 && i < candidates_p->size; ++i) { + LLAMA_LOG_INFO("Token %zu: %f%%\n", i + 1, candidates_p->data[i].p * 100.0f); + } +#endif + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + void llama_sample_temp(struct llama_context * ctx, llama_token_data_array * candidates_p, float temp) { const int64_t t_start_sample_us = ggml_time_us(); diff --git a/llama.h b/llama.h index bb6054557..7b3634aa6 100644 --- a/llama.h +++ b/llama.h @@ -775,6 +775,14 @@ extern "C" { float p, size_t min_keep); + /// @details Dynamic temperature implementation described in the paper https://arxiv.org/abs/2309.02772. + LLAMA_API void llama_sample_entropy( + struct llama_context * ctx, + llama_token_data_array * candidates_p, + float min_temp, + float max_temp, + float exponent_val); + LLAMA_API void llama_sample_temp( struct llama_context * ctx, llama_token_data_array * candidates, From fe54033b69b83164cabb5f3ed92dc0ff7ea47605 Mon Sep 17 00:00:00 2001 From: XiaotaoChen Date: Fri, 26 Jan 2024 04:14:32 +0800 Subject: [PATCH 076/101] readme : add MobileVLM 1.7B/3B to the supported models list (#5107) Co-authored-by: Chenxiaotao03 --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index cbfba01bc..c2a33022c 100644 --- a/README.md +++ b/README.md @@ -112,6 +112,7 @@ as the main playground for developing new features for the [ggml](https://github - [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava) - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5) - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V) +- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM) **Bindings:** From 1182cf4d4f6ee383b92695c2e3fe438086dcdba7 Mon Sep 17 00:00:00 2001 From: Kawrakow <48489457+ikawrakow@users.noreply.github.com> Date: Fri, 26 Jan 2024 09:14:39 +0200 Subject: [PATCH 077/101] Another bucket sort (#5109) * Initial bucket sort * Bucket sort: slightly better version * Bucket sort: another minor improvement --------- Co-authored-by: Iwan Kawrakow --- llama.cpp | 53 ++++++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 50 insertions(+), 3 deletions(-) diff --git a/llama.cpp b/llama.cpp index 823d42d7f..b03b67e16 100644 --- a/llama.cpp +++ b/llama.cpp @@ -7956,10 +7956,57 @@ void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * can auto comp = [](const llama_token_data & a, const llama_token_data & b) { return a.logit > b.logit; }; - if (k == (int) candidates->size) { - std::sort(candidates->data, candidates->data + candidates->size, comp); - } else { + if (k <= 128) { std::partial_sort(candidates->data, candidates->data + k, candidates->data + candidates->size, comp); + } else { + constexpr int nbuckets = 128; + constexpr float bucket_low = -10.0f; + constexpr float bucket_high = 10.0f; + constexpr float bucket_scale = nbuckets/(bucket_high - bucket_low); + constexpr float bucker_inter = -bucket_low * bucket_scale; + + std::vector bucket_idx(candidates->size); + std::vector histo(nbuckets, 0); + + for (int i = 0; i < (int)candidates->size; ++i) { + const float val = candidates->data[i].logit; + int ib = int(bucket_scale * val + bucker_inter); //nbuckets * (val - bucket_low) / (bucket_high - bucket_low); + ib = std::max(0, std::min(nbuckets-1, ib)); + bucket_idx[i] = ib; + ++histo[ib]; + } + int nhave = 0; + int ib = nbuckets - 1; + for ( ; ib >= 0; --ib) { + nhave += histo[ib]; + if (nhave >= k) break; + } + std::vector tmp_tokens(nhave); + auto ptr = tmp_tokens.data(); + std::vector bucket_ptrs; + bucket_ptrs.reserve(nbuckets - ib); + for (int j = nbuckets - 1; j >= ib; --j) { + bucket_ptrs.push_back(ptr); + ptr += histo[j]; + } + for (int i = 0; i < (int)candidates->size; ++i) { + int j = bucket_idx[i]; + if (j >= ib) { + *bucket_ptrs[nbuckets-1-j]++ = candidates->data[i]; + } + } + + ptr = tmp_tokens.data(); + int ndone = 0; + for (int j = nbuckets-1; j > ib; --j) { + std::sort(ptr, ptr + histo[j], comp); + ptr += histo[j]; + ndone += histo[j]; + } + std::partial_sort(ptr, ptr + k - ndone, ptr + histo[ib], comp); + + std::memcpy(candidates->data, tmp_tokens.data(), k*sizeof(llama_token_data)); + } candidates->sorted = true; } From aad0b01d7380a7cdfe0dd42307b18c7b6bac9575 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 10:52:33 +0200 Subject: [PATCH 078/101] readme : update hot topics --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index c2a33022c..76e48ce8a 100644 --- a/README.md +++ b/README.md @@ -10,11 +10,11 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics +- ⚠️ Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138 - New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow - Collecting Apple Silicon performance stats: - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167 - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508 -- Added Mixtral support: https://github.com/ggerganov/llama.cpp/pull/4406 - Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216 ---- From 38b431de232d1b736b5af19b8c7d72f7075a70bc Mon Sep 17 00:00:00 2001 From: Riceball LEE Date: Fri, 26 Jan 2024 17:10:28 +0800 Subject: [PATCH 079/101] gguf : fix "general.alignment" type in gguf_reader.py (#5136) --- gguf-py/gguf/gguf_reader.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py index 8682765ed..5b6d4ba6b 100644 --- a/gguf-py/gguf/gguf_reader.py +++ b/gguf-py/gguf/gguf_reader.py @@ -107,7 +107,7 @@ class GGUFReader: offs, tensors_fields = self._build_tensors_fields(offs, tensor_count) new_align = self.fields.get('general.alignment') if new_align is not None: - if new_align.types != [GGUFValueType.UINT64]: + if new_align.types != [GGUFValueType.UINT32]: raise ValueError('Bad type for general.alignment field') self.alignment = new_align.parts[-1][0] padding = offs % self.alignment From 6dd3c28c9cd1ef74b49d79f47d668759346a3c6c Mon Sep 17 00:00:00 2001 From: Paul Tsochantaris Date: Fri, 26 Jan 2024 12:16:07 +0000 Subject: [PATCH 080/101] metal : remove unused `n_buffers` and `buffers` (#5129) --- ggml-metal.m | 73 ++++++++++++---------------------------------------- 1 file changed, 16 insertions(+), 57 deletions(-) diff --git a/ggml-metal.m b/ggml-metal.m index 60fef1a19..ab3c84f7f 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -26,15 +26,6 @@ #define GGML_METAL_MAX_KERNELS 256 -struct ggml_metal_buffer { - const char * name; - - void * data; - size_t size; - - id metal; -}; - struct ggml_metal_kernel { id function; id pipeline; @@ -172,9 +163,6 @@ struct ggml_metal_context { dispatch_queue_t d_queue; - int n_buffers; - struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS]; - struct ggml_metal_kernel kernels[GGML_METAL_MAX_KERNELS]; bool support_simdgroup_reduction; @@ -242,24 +230,20 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { // Show all the Metal device instances in the system NSArray * devices = MTLCopyAllDevices(); for (id device in devices) { - NSString * s = [device name]; - GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]); + GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]); } [devices release]; // since it was created by a *Copy* C method #endif // Pick and show default Metal device id device = MTLCreateSystemDefaultDevice(); - NSString * s = [device name]; - GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]); + GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]); // Configure context struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context)); ctx->device = device; ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS); ctx->queue = [ctx->device newCommandQueue]; - ctx->n_buffers = 0; - ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT); // load library @@ -534,10 +518,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) { static void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_LOG_INFO("%s: deallocating\n", __func__); - for (int i = 0; i < ctx->n_buffers; ++i) { - [ctx->buffers[i].metal release]; - } - for (int i = 0; i < GGML_METAL_MAX_KERNELS; ++i) { if (ctx->kernels[i].pipeline) { [ctx->kernels[i].pipeline release]; @@ -580,51 +560,30 @@ struct ggml_backend_metal_buffer_context { // the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the // Metal buffer based on the host memory pointer // -static id ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) { +static id ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) { //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach); const int64_t tsize = ggml_nbytes(t); ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer; - // compatibility with ggml-backend - if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) { - struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context; - - // find the view that contains the tensor fully - for (int i = 0; i < buf_ctx->n_buffers; ++i) { - const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data; - - //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size); - if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) { - *offs = (size_t) ioffs; - - //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); - - return buf_ctx->buffers[i].metal; - } - } - - GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); - - return nil; - } + struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context; // find the view that contains the tensor fully - for (int i = 0; i < ctx->n_buffers; ++i) { - const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data; + for (int i = 0; i < buf_ctx->n_buffers; ++i) { + const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data; - //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name); - if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) { + //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size); + if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) { *offs = (size_t) ioffs; - //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs); + //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs); - return ctx->buffers[i].metal; + return buf_ctx->buffers[i].metal; } } - GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__); + GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name); return nil; } @@ -817,9 +776,9 @@ static bool ggml_metal_graph_compute( const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT; const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT; - id id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil; - id id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil; - id id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil; + id id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil; + id id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil; + id id_dst = dst ? ggml_metal_get_buffer(dst, &offs_dst) : nil; //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op)); //if (src0) { @@ -1601,7 +1560,7 @@ static bool ggml_metal_graph_compute( struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)]; size_t offs_src_cur = 0; - id id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur); + id id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur); [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j]; } @@ -1746,7 +1705,7 @@ static bool ggml_metal_graph_compute( struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)]; size_t offs_src_cur = 0; - id id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur); + id id_src_cur = ggml_metal_get_buffer(src_cur, &offs_src_cur); [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j]; } From 413e7b0559f922bd4de5e9eec548829d111651b1 Mon Sep 17 00:00:00 2001 From: crasm Date: Fri, 26 Jan 2024 07:18:00 -0500 Subject: [PATCH 081/101] ci : add model tests + script wrapper (#4586) * scripts : add lib.sh and lib_test.sh * scripts : stub out new ci-run.sh script * scripts : switch to PascalCase for functions This looks a little odd at first, but I find it very useful as a convention to know if a command is part of our code vs a builtin. * scripts : add some fancy conversion from snake_case to PascalCase * Add venv to ci/run.sh * Revert scripts work * scripts : add wrapper script for local use of ci/run.sh * Simplify .gitignore for tests, clang-tidy fixes * Label all ctest tests * ci : ctest uses -L main * Attempt at writing ctest_with_model * Update test-model-load-cancel * ci : add ctest_with_model for debug and release ggml-ci * Fix gg_get_model function ggml-ci * got stuck on CMake * Add get_model.cpp to tests/CMakeLists.txt ggml-ci * Fix README.md output for ctest_with_model ggml-ci * workflows : use `-L main` for all ctest ggml-ci * Fixes * GG_RUN_CTEST_MODELFILE => LLAMACPP_TESTMODELFILE * Always show warning rather than failing if model file variable is not set * scripts : update usage text for ci-run.sh --- .github/workflows/build.yml | 12 ++--- .gitignore | 19 +------- Makefile | 7 ++- ci/run.sh | 81 +++++++++++++++++++++++++++----- scripts/ci-run.sh | 50 ++++++++++++++++++++ tests/.gitignore | 2 + tests/CMakeLists.txt | 14 ++++-- tests/get-model.cpp | 21 +++++++++ tests/get-model.h | 2 + tests/test-autorelease.cpp | 12 ++--- tests/test-model-load-cancel.cpp | 27 +++++++++++ 11 files changed, 199 insertions(+), 48 deletions(-) create mode 100755 scripts/ci-run.sh create mode 100644 tests/.gitignore create mode 100644 tests/get-model.cpp create mode 100644 tests/get-model.h create mode 100644 tests/test-model-load-cancel.cpp diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index c3aa6f992..d22a041a6 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -72,7 +72,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 ubuntu-latest-cmake-sanitizer: runs-on: ubuntu-latest @@ -107,7 +107,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 ubuntu-latest-cmake-mpi: runs-on: ubuntu-latest @@ -141,7 +141,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose + ctest -L main --verbose # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know # how to debug it. @@ -202,7 +202,7 @@ jobs: id: cmake_test run: | cd build - ctest --verbose --timeout 900 + ctest -L main --verbose --timeout 900 macOS-latest-cmake-ios: runs-on: macos-latest @@ -394,7 +394,7 @@ jobs: if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512 run: | cd build - ctest -C Release --verbose --timeout 900 + ctest -L main -C Release --verbose --timeout 900 - name: Test (Intel SDE) id: cmake_test_sde @@ -406,7 +406,7 @@ jobs: 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe) cd build - & $sde -future -- ctest -C Release --verbose --timeout 900 + & $sde -future -- ctest -L main -C Release --verbose --timeout 900 - name: Determine tag name id: tag diff --git a/.gitignore b/.gitignore index 5ab81445d..cb0069bfb 100644 --- a/.gitignore +++ b/.gitignore @@ -27,7 +27,7 @@ lcov-report/ gcovr-report/ -build*/ +build* out/ tmp/ @@ -89,20 +89,3 @@ examples/jeopardy/results.txt poetry.lock poetry.toml - -# Test binaries -/tests/test-grammar-parser -/tests/test-llama-grammar -/tests/test-double-float -/tests/test-grad0 -/tests/test-opt -/tests/test-quantize-fns -/tests/test-quantize-perf -/tests/test-sampling -/tests/test-tokenizer-0-llama -/tests/test-tokenizer-0-falcon -/tests/test-tokenizer-1-llama -/tests/test-tokenizer-1-bpe -/tests/test-rope -/tests/test-backend-ops -/tests/test-autorelease diff --git a/Makefile b/Makefile index a8658a596..82c89e87a 100644 --- a/Makefile +++ b/Makefile @@ -9,7 +9,7 @@ TEST_TARGETS = \ tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \ tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \ tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope \ - tests/test-backend-ops tests/test-autorelease + tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -748,5 +748,8 @@ tests/test-c.o: tests/test-c.c llama.h tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/ci/run.sh b/ci/run.sh index 791b17a19..2427e55a2 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -22,9 +22,9 @@ mkdir -p "$2" OUT=$(realpath "$1") MNT=$(realpath "$2") -rm -v $OUT/*.log -rm -v $OUT/*.exit -rm -v $OUT/*.md +rm -f "$OUT/*.log" +rm -f "$OUT/*.exit" +rm -f "$OUT/*.md" sd=`dirname $0` cd $sd/../ @@ -94,7 +94,7 @@ function gg_run_ctest_debug { (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log - (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log set +e } @@ -123,9 +123,9 @@ function gg_run_ctest_release { (time make -j ) 2>&1 | tee -a $OUT/${ci}-make.log if [ -z ${GG_BUILD_LOW_PERF} ]; then - (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log else - (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log + (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log fi set +e @@ -141,6 +141,61 @@ function gg_sum_ctest_release { gg_printf '```\n' } +function gg_get_model { + local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf" + local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf" + if [[ -s $gguf_3b ]]; then + echo -n "$gguf_3b" + elif [[ -s $gguf_7b ]]; then + echo -n "$gguf_7b" + else + echo >&2 "No model found. Can't run gg_run_ctest_with_model." + exit 1 + fi +} + +function gg_run_ctest_with_model_debug { + cd ${SRC} + + local model; model=$(gg_get_model) + cd build-ci-debug + set -e + (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log + set +e + cd .. +} + +function gg_run_ctest_with_model_release { + cd ${SRC} + + local model; model=$(gg_get_model) + cd build-ci-release + set -e + (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log + set +e + cd .. +} + +function gg_sum_ctest_with_model_debug { + gg_printf '### %s\n\n' "${ci}" + + gg_printf 'Runs ctest with model files in debug mode\n' + gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" + gg_printf '```\n' + gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)" + gg_printf '```\n' +} + +function gg_sum_ctest_with_model_release { + gg_printf '### %s\n\n' "${ci}" + + gg_printf 'Runs ctest with model files in release mode\n' + gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)" + gg_printf '```\n' + gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)" + gg_printf '```\n' +} + # open_llama_3b_v2 function gg_run_open_llama_3b_v2 { @@ -183,8 +238,6 @@ function gg_run_open_llama_3b_v2 { wiki_test_60="${path_wiki}/wiki.test-60.raw" - ./bin/test-autorelease ${model_f16} - ./bin/quantize ${model_f16} ${model_q8_0} q8_0 ./bin/quantize ${model_f16} ${model_q4_0} q4_0 ./bin/quantize ${model_f16} ${model_q4_1} q4_1 @@ -507,14 +560,18 @@ function gg_sum_open_llama_7b_v2 { ## main if [ -z ${GG_BUILD_LOW_PERF} ]; then + # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt rm -rf ${SRC}/models-mnt - mnt_models=${MNT}/models mkdir -p ${mnt_models} ln -sfn ${mnt_models} ${SRC}/models-mnt - python3 -m pip install -r ${SRC}/requirements.txt - python3 -m pip install --editable gguf-py + # Create a fresh python3 venv and enter it + python3 -m venv "$MNT/venv" + source "$MNT/venv/bin/activate" + + pip install -r ${SRC}/requirements.txt --disable-pip-version-check + pip install --editable gguf-py --disable-pip-version-check fi ret=0 @@ -529,6 +586,8 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then else test $ret -eq 0 && gg_run open_llama_7b_v2 fi + test $ret -eq 0 && gg_run ctest_with_model_debug + test $ret -eq 0 && gg_run ctest_with_model_release fi fi diff --git a/scripts/ci-run.sh b/scripts/ci-run.sh new file mode 100755 index 000000000..06b5d9c6e --- /dev/null +++ b/scripts/ci-run.sh @@ -0,0 +1,50 @@ +#!/bin/bash +set -euo pipefail +this=$(realpath "$0"); readonly this +cd "$(dirname "$this")" +shellcheck "$this" + +if (( $# != 1 && $# != 2 )); then + cat >&2 <<'EOF' +usage: + ci-run.sh [] + +This script wraps ci/run.sh: +* If is a ramdisk, you can reduce writes to your SSD. If is not a ramdisk, keep in mind that total writes will increase by the size of . + (openllama_3b_v2: quantized models are about 30GB) +* Persistent model and data files are synced to and from , + excluding generated .gguf files. + (openllama_3b_v2: persistent files are about 6.6GB) +* defaults to ~/.cache/llama.cpp +EOF + exit 1 +fi + +cd .. # => llama.cpp repo root + +tmp="$1" +mkdir -p "$tmp" +tmp=$(realpath "$tmp") +echo >&2 "Using tmp=$tmp" + +cache="${2-$HOME/.cache/llama.cpp}" +mkdir -p "$cache" +cache=$(realpath "$cache") +echo >&2 "Using cache=$cache" + +_sync() { + local from="$1"; shift + local to="$1"; shift + + echo >&2 "Syncing from $from to $to" + mkdir -p "$from" "$to" + rsync -a "$from" "$to" --delete-during "$@" +} + +_sync "$(realpath .)/" "$tmp/llama.cpp" +_sync "$cache/ci-mnt/models/" "$tmp/llama.cpp/ci-mnt/models/" + +cd "$tmp/llama.cpp" +bash ci/run.sh ci-out ci-mnt + +_sync 'ci-mnt/models/' "$cache/ci-mnt/models/" --exclude='*.gguf' -P diff --git a/tests/.gitignore b/tests/.gitignore new file mode 100644 index 000000000..59be43b99 --- /dev/null +++ b/tests/.gitignore @@ -0,0 +1,2 @@ +* +!*.* diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index d7aaab843..3e40a78cd 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -1,6 +1,6 @@ function(llama_build_executable source) get_filename_component(TEST_TARGET ${source} NAME_WE) - add_executable(${TEST_TARGET} ${source}) + add_executable(${TEST_TARGET} ${source} get-model.cpp) install(TARGETS ${TEST_TARGET} RUNTIME) target_link_libraries(${TEST_TARGET} PRIVATE common) endfunction() @@ -8,14 +8,20 @@ endfunction() function(llama_test_executable name source) get_filename_component(TEST_TARGET ${source} NAME_WE) add_test(NAME ${name} COMMAND $ ${ARGN}) + set_property(TEST ${name} PROPERTY LABELS "main") endfunction() function(llama_build_and_test_executable source) + llama_build_and_test_executable_with_label(${source} "main") +endfunction() + +function(llama_build_and_test_executable_with_label source label) get_filename_component(TEST_TARGET ${source} NAME_WE) - add_executable(${TEST_TARGET} ${source}) + add_executable(${TEST_TARGET} ${source} get-model.cpp) install(TARGETS ${TEST_TARGET} RUNTIME) target_link_libraries(${TEST_TARGET} PRIVATE common) add_test(NAME ${TEST_TARGET} COMMAND $ ${ARGN}) + set_property(TEST ${TEST_TARGET} PROPERTY LABELS ${label}) endfunction() # llama_build_and_test_executable(test-double-float.cpp) # SLOW @@ -49,10 +55,12 @@ llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # llama_build_and_test_executable(test-opt.cpp) # SLOW llama_build_and_test_executable(test-backend-ops.cpp) -llama_build_and_test_executable(test-autorelease.cpp) llama_build_and_test_executable(test-rope.cpp) +llama_build_and_test_executable_with_label(test-model-load-cancel.cpp "model") +llama_build_and_test_executable_with_label(test-autorelease.cpp "model") + # dummy executable - not installed get_filename_component(TEST_TARGET test-c.c NAME_WE) add_executable(${TEST_TARGET} test-c.c) diff --git a/tests/get-model.cpp b/tests/get-model.cpp new file mode 100644 index 000000000..4edb685f0 --- /dev/null +++ b/tests/get-model.cpp @@ -0,0 +1,21 @@ +#include +#include +#include + +#include "get-model.h" + +char * get_model_or_exit(int argc, char *argv[]) { + char * model_path; + if (argc > 1) { + model_path = argv[1]; + + } else { + model_path = getenv("LLAMACPP_TEST_MODELFILE"); + if (!model_path || strlen(model_path) == 0) { + fprintf(stderr, "\033[33mWARNING: No model file provided. Skipping this test. Set LLAMACPP_TEST_MODELFILE= to silence this warning and run this test.\n\033[0m"); + exit(EXIT_SUCCESS); + } + } + + return model_path; +} diff --git a/tests/get-model.h b/tests/get-model.h new file mode 100644 index 000000000..81a3a0fef --- /dev/null +++ b/tests/get-model.h @@ -0,0 +1,2 @@ +#pragma once +char * get_model_or_exit(int, char*[]); diff --git a/tests/test-autorelease.cpp b/tests/test-autorelease.cpp index 289c6ba6c..36a23c0bb 100644 --- a/tests/test-autorelease.cpp +++ b/tests/test-autorelease.cpp @@ -5,19 +5,15 @@ #include #include "llama.h" +#include "get-model.h" // This creates a new context inside a pthread and then tries to exit cleanly. int main(int argc, char ** argv) { - if (argc < 2) { - printf("Usage: %s model.gguf\n", argv[0]); - return 0; // intentionally return success - } + auto * model_path = get_model_or_exit(argc, argv); - const std::string fname = argv[1]; - - std::thread([&fname]() { + std::thread([&model_path]() { llama_backend_init(false); - auto * model = llama_load_model_from_file(fname.c_str(), llama_model_default_params()); + auto * model = llama_load_model_from_file(model_path, llama_model_default_params()); auto * ctx = llama_new_context_with_model(model, llama_context_default_params()); llama_free(ctx); llama_free_model(model); diff --git a/tests/test-model-load-cancel.cpp b/tests/test-model-load-cancel.cpp new file mode 100644 index 000000000..7ea4bbacc --- /dev/null +++ b/tests/test-model-load-cancel.cpp @@ -0,0 +1,27 @@ +#include "llama.h" +#include "get-model.h" + +#include + +int main(int argc, char *argv[] ) { + auto * model_path = get_model_or_exit(argc, argv); + auto * file = fopen(model_path, "r"); + if (file == nullptr) { + fprintf(stderr, "no model at '%s' found\n", model_path); + return EXIT_FAILURE; + } + + fprintf(stderr, "using '%s'\n", model_path); + fclose(file); + + llama_backend_init(false); + auto params = llama_model_params{}; + params.use_mmap = false; + params.progress_callback = [](float progress, void * ctx){ + (void) ctx; + return progress > 0.50; + }; + auto * model = llama_load_model_from_file(model_path, params); + llama_backend_free(); + return model == nullptr ? EXIT_SUCCESS : EXIT_FAILURE; +} From 48c857aa10aea73210a4a72da3f1a6f99269e75d Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Fri, 26 Jan 2024 13:42:20 +0100 Subject: [PATCH 082/101] server : refactored the task processing logic (#5065) * server: add llama_server_queue struct * server: add llama_server_response_event * server: add comments * server: move all mutexes away from server.cpp * server: correct multitask response * server: only add back deferred tasks when one slot is available * server: fix a race condition cause by "request_completion" --- Makefile | 2 +- examples/server/CMakeLists.txt | 2 +- examples/server/oai.hpp | 208 ++++++++ examples/server/server.cpp | 849 ++++++--------------------------- examples/server/utils.hpp | 507 ++++++++++++++++++++ 5 files changed, 876 insertions(+), 692 deletions(-) create mode 100644 examples/server/oai.hpp create mode 100644 examples/server/utils.hpp diff --git a/Makefile b/Makefile index 82c89e87a..b8858b412 100644 --- a/Makefile +++ b/Makefile @@ -619,7 +619,7 @@ embedding: examples/embedding/embedding.cpp ggml.o llama.o $(C save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual gguf: examples/gguf/gguf.cpp ggml.o $(OBJS) diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index 81709e448..cc13b2d63 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -1,7 +1,7 @@ set(TARGET server) option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON) include_directories(${CMAKE_CURRENT_SOURCE_DIR}) -add_executable(${TARGET} server.cpp json.hpp httplib.h) +add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h) install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ diff --git a/examples/server/oai.hpp b/examples/server/oai.hpp new file mode 100644 index 000000000..bc5db6eef --- /dev/null +++ b/examples/server/oai.hpp @@ -0,0 +1,208 @@ +#pragma once + +#include +#include +#include +#include +#include +#include + +#include "json.hpp" +#include "utils.hpp" + +#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" + +using json = nlohmann::json; + +inline static json oaicompat_completion_params_parse( + const json &body /* openai api json semantics */) +{ + json llama_params; + + llama_params["__oaicompat"] = true; + + // Map OpenAI parameters to llama.cpp parameters + // + // For parameters that are defined by the OpenAI documentation (e.g. + // temperature), we explicitly specify OpenAI's intended default; we + // need to do that because sometimes OpenAI disagrees with llama.cpp + // + // https://platform.openai.com/docs/api-reference/chat/create + llama_sampling_params default_sparams; + llama_params["model"] = json_value(body, "model", std::string("unknown")); + llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt' + llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); + llama_params["temperature"] = json_value(body, "temperature", 0.0); + llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k); + llama_params["top_p"] = json_value(body, "top_p", 1.0); + llama_params["n_predict"] = json_value(body, "max_tokens", -1); + llama_params["logit_bias"] = json_value(body, "logit_bias",json::object()); + llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0); + llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0); + llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED); + llama_params["stream"] = json_value(body, "stream", false); + llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat); + llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); + llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); + llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl); + llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p); + llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); + llama_params["ignore_eos"] = json_value(body, "ignore_eos", false); + llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z); + + if (body.count("grammar") != 0) { + llama_params["grammar"] = json_value(body, "grammar", json::object()); + } + + // Handle 'stop' field + if (body.contains("stop") && body["stop"].is_string()) { + llama_params["stop"] = json::array({body["stop"].get()}); + } else { + llama_params["stop"] = json_value(body, "stop", json::array()); + } + + // Ensure there is ChatML-specific end sequence among stop words + llama_params["stop"].push_back("<|im_end|>"); + + return llama_params; +} + +inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false) +{ + json result = response.result_json; + + bool stopped_word = result.count("stopped_word") != 0; + bool stopped_eos = json_value(result, "stopped_eos", false); + int num_tokens_predicted = json_value(result, "tokens_predicted", 0); + int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); + std::string content = json_value(result, "content", std::string("")); + + std::string finish_reason = "length"; + if (stopped_word || stopped_eos) { + finish_reason = "stop"; + } + + json choices = + streaming ? json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}) + : json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"message", json{{"content", content}, + {"role", "assistant"}}}}}); + + std::time_t t = std::time(0); + + json res = + json{{"choices", choices}, + {"created", t}, + {"model", + json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, + {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, + {"usage", + json{{"completion_tokens", num_tokens_predicted}, + {"prompt_tokens", num_prompt_tokens}, + {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, + {"id", gen_chatcmplid()}}; + + if (server_verbose) { + res["__verbose"] = result; + } + + if (result.contains("completion_probabilities")) { + res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); + } + + return res; +} + +// return value is vector as there is one case where we might need to generate two responses +inline static std::vector format_partial_response_oaicompat(const task_result &response) { + json result = response.result_json; + + if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { + return std::vector({response.result_json}); + } + + bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; + std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); + + bool stopped_word = json_value(result, "stopped_word", false); + bool stopped_eos = json_value(result, "stopped_eos", false); + bool stopped_limit = json_value(result, "stopped_limit", false); + std::string content = json_value(result, "content", std::string("")); + + std::string finish_reason; + if (stopped_word || stopped_eos) { + finish_reason = "stop"; + } + if (stopped_limit) { + finish_reason = "length"; + } + + std::time_t t = std::time(0); + + json choices; + + if (!finish_reason.empty()) { + choices = json::array({json{{"finish_reason", finish_reason}, + {"index", 0}, + {"delta", json::object()}}}); + } else { + if (first) { + if (content.empty()) { + choices = json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{{"role", "assistant"}}}}}); + } else { + // We have to send this as two updates to conform to openai behavior + json initial_ret = json{{"choices", json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"role", "assistant"} + }}}})}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + json second_ret = json{ + {"choices", json::array({json{{"finish_reason", nullptr}, + {"index", 0}, + {"delta", json{ + {"content", content}}} + }})}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + return std::vector({initial_ret, second_ret}); + } + } else { + // Some idiosyncrasy in task processing logic makes several trailing calls + // with empty content, we ignore these at the calee site. + if (content.empty()) { + return std::vector({json::object()}); + } + + choices = json::array({json{ + {"finish_reason", nullptr}, + {"index", 0}, + {"delta", + json{ + {"content", content}, + }}, + }}); + } + } + + json ret = json{{"choices", choices}, + {"created", t}, + {"id", gen_chatcmplid()}, + {"model", modelname}, + {"object", "chat.completion.chunk"}}; + + return std::vector({ret}); +} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 0462fbd24..392836132 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,8 @@ #include "common.h" #include "llama.h" #include "grammar-parser.h" +#include "utils.hpp" +#include "oai.hpp" #include "../llava/clip.h" @@ -23,17 +25,10 @@ #include #include -#include #include #include #include -#ifndef SERVER_VERBOSE -#define SERVER_VERBOSE 1 -#endif - -#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" - using json = nlohmann::json; struct server_params @@ -46,197 +41,7 @@ struct server_params int32_t write_timeout = 600; }; -static bool server_verbose = false; - -#if SERVER_VERBOSE != 1 -#define LOG_VERBOSE(MSG, ...) -#else -#define LOG_VERBOSE(MSG, ...) \ - do \ - { \ - if (server_verbose) \ - { \ - server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ - } \ - } while (0) -#endif - -#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__) -#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__) -#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) - -json oaicompat_completion_params_parse(const json &body); -std::string format_chatml(std::vector messages); - - -// -// base64 utils (TODO: move to common in the future) -// - -static const std::string base64_chars = - "ABCDEFGHIJKLMNOPQRSTUVWXYZ" - "abcdefghijklmnopqrstuvwxyz" - "0123456789+/"; - -static inline bool is_base64(uint8_t c) -{ - return (isalnum(c) || (c == '+') || (c == '/')); -} - -static std::vector base64_decode(const std::string & encoded_string) -{ - int i = 0; - int j = 0; - int in_ = 0; - - int in_len = encoded_string.size(); - - uint8_t char_array_4[4]; - uint8_t char_array_3[3]; - - std::vector ret; - - while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) - { - char_array_4[i++] = encoded_string[in_]; in_++; - if (i == 4) - { - for (i = 0; i <4; i++) - { - char_array_4[i] = base64_chars.find(char_array_4[i]); - } - - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - - for (i = 0; (i < 3); i++) - { - ret.push_back(char_array_3[i]); - } - i = 0; - } - } - - if (i) - { - for (j = i; j <4; j++) - { - char_array_4[j] = 0; - } - - for (j = 0; j <4; j++) - { - char_array_4[j] = base64_chars.find(char_array_4[j]); - } - - char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); - char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); - char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; - - for (j = 0; (j < i - 1); j++) - { - ret.push_back(char_array_3[j]); - } - } - - return ret; -} - -// -// parallel -// - -enum server_state { - SERVER_STATE_LOADING_MODEL, // Server is starting up, model not fully loaded yet - SERVER_STATE_READY, // Server is ready and model is loaded - SERVER_STATE_ERROR // An error occurred, load_model failed -}; - -enum task_type { - TASK_TYPE_COMPLETION, - TASK_TYPE_CANCEL, -}; - -struct task_server { - int id; - int target_id; - task_type type; - json data; - bool infill_mode = false; - bool embedding_mode = false; - int multitask_id = -1; -}; - -struct task_result { - int id; - int multitask_id = -1; - bool stop; - bool error; - json result_json; -}; - -struct task_multi { - int id; - std::set subtasks_remaining{}; - std::vector results{}; -}; - -// TODO: can become bool if we can't find use of more states -enum slot_state -{ - IDLE, - PROCESSING, -}; - -enum slot_command -{ - NONE, - LOAD_PROMPT, - RELEASE, -}; - -struct slot_params -{ - bool stream = true; - bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt - - uint32_t seed = -1; // RNG seed - int32_t n_keep = 0; // number of tokens to keep from initial prompt - int32_t n_predict = -1; // new tokens to predict - - std::vector antiprompt; - - json input_prefix; - json input_suffix; -}; - -struct slot_image -{ - int32_t id; - - bool request_encode_image = false; - float * image_embedding = nullptr; - int32_t image_tokens = 0; - - clip_image_u8 * img_data; - - std::string prefix_prompt; // before of this image -}; - -// completion token output with probabilities -struct completion_token_output -{ - struct token_prob - { - llama_token tok; - float prob; - }; - - std::vector probs; - llama_token tok; - std::string text_to_send; -}; +bool server_verbose = false; static size_t common_part(const std::vector &a, const std::vector &b) { @@ -292,28 +97,6 @@ static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) return ret; } -static void server_log(const char *level, const char *function, int line, - const char *message, const nlohmann::ordered_json &extra) -{ - nlohmann::ordered_json log - { - {"timestamp", time(nullptr)}, - {"level", level}, - {"function", function}, - {"line", line}, - {"message", message}, - }; - - if (!extra.empty()) - { - log.merge_patch(extra); - } - - const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace); - printf("%.*s\n", (int)str.size(), str.data()); - fflush(stdout); -} - // format incomplete utf-8 multibyte character for output static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token) { @@ -355,15 +138,6 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector -static T json_value(const json &body, const std::string &key, const T &default_value) -{ - // Fallback null to default value - return body.contains(key) && !body.at(key).is_null() - ? body.value(key, default_value) - : default_value; -} - struct llama_client_slot { int id; @@ -491,7 +265,7 @@ struct llama_client_slot } void release() { - if (state == IDLE || state == PROCESSING) + if (state == PROCESSING) { t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; command = RELEASE; @@ -539,7 +313,6 @@ struct llama_server_context bool all_slots_are_idle = false; bool add_bos_token = true; - int32_t id_gen; int32_t n_ctx; // total context for all clients / slots // system prompt @@ -554,13 +327,8 @@ struct llama_server_context // slots / clients std::vector slots; - std::vector queue_tasks; - std::vector queue_results; - std::vector queue_multitasks; - std::mutex mutex_tasks; // also guards id_gen, and queue_multitasks - std::condition_variable condition_tasks; - std::mutex mutex_results; - std::condition_variable condition_results; + llama_server_queue queue_tasks; + llama_server_response queue_results; ~llama_server_context() { @@ -619,8 +387,6 @@ struct llama_server_context } void initialize() { - id_gen = 0; - // create slots all_slots_are_idle = true; @@ -1183,39 +949,13 @@ struct llama_server_context void send_error(task_server& task, const std::string &error) { LOG_TEE("task %i - error: %s\n", task.id, error.c_str()); - std::unique_lock lock(mutex_results); task_result res; res.id = task.id; res.multitask_id = task.multitask_id; res.stop = false; res.error = true; res.result_json = { { "content", error } }; - queue_results.push_back(res); - condition_results.notify_all(); - } - - void add_multi_task(int id, std::vector& sub_ids) - { - std::lock_guard lock(mutex_tasks); - task_multi multi; - multi.id = id; - std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end())); - queue_multitasks.push_back(multi); - condition_tasks.notify_one(); - } - - void update_multi_task(int multitask_id, int subtask_id, task_result& result) - { - std::lock_guard lock(mutex_tasks); - for (auto& multitask : queue_multitasks) - { - if (multitask.id == multitask_id) - { - multitask.subtasks_remaining.erase(subtask_id); - multitask.results.push_back(result); - condition_tasks.notify_one(); - } - } + queue_results.send(res); } json get_model_props() @@ -1261,7 +1001,6 @@ struct llama_server_context void send_partial_response(llama_client_slot &slot, completion_token_output tkn) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1296,13 +1035,11 @@ struct llama_server_context res.result_json["model"] = slot.oaicompat_model; } - queue_results.push_back(res); - condition_results.notify_all(); + queue_results.send(res); } void send_final_response(llama_client_slot &slot) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1351,22 +1088,11 @@ struct llama_server_context res.result_json["model"] = slot.oaicompat_model; } - queue_results.push_back(res); - condition_results.notify_all(); - - // done with results, unlock - lock.unlock(); - - // parent multitask, if any, needs to be updated - if (slot.multitask_id != -1) - { - update_multi_task(slot.multitask_id, slot.task_id, res); - } + queue_results.send(res); } void send_embedding(llama_client_slot &slot) { - std::unique_lock lock(mutex_results); task_result res; res.id = slot.task_id; res.multitask_id = slot.multitask_id; @@ -1393,15 +1119,13 @@ struct llama_server_context {"embedding", embedding }, }; } - queue_results.push_back(res); - condition_results.notify_all(); + queue_results.send(res); } - int request_completion(json data, bool infill, bool embedding, int multitask_id) + void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id) { - std::unique_lock lock(mutex_tasks); task_server task; - task.id = id_gen++; + task.id = task_id; task.target_id = 0; task.data = std::move(data); task.infill_mode = infill; @@ -1412,47 +1136,11 @@ struct llama_server_context // when a completion task's prompt array is not a singleton, we split it into multiple requests if (task.data.count("prompt") && task.data.at("prompt").size() > 1) { - lock.unlock(); // entering new func scope - return split_multiprompt_task(task); + split_multiprompt_task(task_id, task); } // otherwise, it's a single-prompt task, we actually queue it - queue_tasks.push_back(task); - condition_tasks.notify_one(); - return task.id; - } - - task_result next_result(int task_id) - { - while (true) - { - std::unique_lock lock(mutex_results); - condition_results.wait(lock, [&]{ - return !queue_results.empty(); - }); - - for (int i = 0; i < (int) queue_results.size(); i++) - { - // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result - if (queue_results[i].multitask_id == task_id) - { - update_multi_task(task_id, queue_results[i].id, queue_results[i]); - queue_results.erase(queue_results.begin() + i); - continue; - } - - if (queue_results[i].id == task_id) - { - assert(queue_results[i].multitask_id == -1); - task_result res = queue_results[i]; - queue_results.erase(queue_results.begin() + i); - return res; - } - } - } - - // never reached - //return task_result{-1, false, false, {}}; + queue_tasks.post(task); } // for multiple images processing @@ -1525,150 +1213,117 @@ struct llama_server_context void request_cancel(int task_id) { - std::unique_lock lock(mutex_tasks); task_server task; - task.id = id_gen++; task.type = TASK_TYPE_CANCEL; task.target_id = task_id; - queue_tasks.push_back(task); - condition_tasks.notify_one(); + queue_tasks.post(task); } - int split_multiprompt_task(task_server& multiprompt_task) + void split_multiprompt_task(int multitask_id, task_server& multiprompt_task) { int prompt_count = multiprompt_task.data.at("prompt").size(); assert(prompt_count > 1); - int multitask_id = id_gen++; + // generate all the ID for subtask std::vector subtask_ids(prompt_count); for (int i = 0; i < prompt_count; i++) + { + subtask_ids[i] = queue_tasks.get_new_id(); + } + + // queue up the multitask so we can track its subtask progression + queue_tasks.add_multitask(multitask_id, subtask_ids); + + // add subtasks + for (int i = 0; i < prompt_count; i++) { json subtask_data = multiprompt_task.data; subtask_data["prompt"] = subtask_data["prompt"][i]; // subtasks inherit everything else (infill mode, embedding mode, etc.) - subtask_ids[i] = request_completion(subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); + request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); } - - // queue up the multitask so we can track its subtask progression - add_multi_task(multitask_id, subtask_ids); - return multitask_id; } - void process_tasks() + void process_single_task(task_server& task) { - std::unique_lock lock(mutex_tasks); - std::vector deferred_tasks; - while (!queue_tasks.empty()) + switch (task.type) { - task_server task = queue_tasks.front(); - queue_tasks.erase(queue_tasks.begin()); - switch (task.type) - { - case TASK_TYPE_COMPLETION: { - llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); - if (slot == nullptr) - { - // if no slot is available, we defer this task for processing later - deferred_tasks.push_back(task); - break; - } - - if (task.data.contains("system_prompt")) - { - if (!all_slots_are_idle) { - send_error(task, "system prompt can only be updated when all slots are idle"); - break; - } - process_system_prompt_data(task.data["system_prompt"]); - - // reset cache_tokens for all slots - for (llama_client_slot &slot : slots) - { - slot.cache_tokens.clear(); - } - } - - slot->reset(); - - slot->infill = task.infill_mode; - slot->embedding = task.embedding_mode; - slot->task_id = task.id; - slot->multitask_id = task.multitask_id; - - if (!launch_slot_with_data(slot, task.data)) - { - // send error result - send_error(task, "internal_error"); - break; - } - } break; - case TASK_TYPE_CANCEL: { // release slot linked with the task id - for (auto & slot : slots) - { - if (slot.task_id == task.target_id) - { - slot.release(); - break; - } - } - } break; - } - } - - // add all the deferred tasks back the the queue - for (task_server &task : deferred_tasks) - { - queue_tasks.push_back(task); - } - - // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue - std::vector agg_results; - auto queue_iterator = queue_multitasks.begin(); - while (queue_iterator != queue_multitasks.end()) - { - if (queue_iterator->subtasks_remaining.empty()) - { - // all subtasks done == multitask is done - task_result aggregate_result; - aggregate_result.id = queue_iterator->id; - aggregate_result.stop = true; - aggregate_result.error = false; - - // collect json results into one json result - std::vector result_jsons; - for (auto& subres : queue_iterator->results) + case TASK_TYPE_COMPLETION: { + llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); + if (slot == nullptr) { - result_jsons.push_back(subres.result_json); - aggregate_result.error = aggregate_result.error && subres.error; + // if no slot is available, we defer this task for processing later + LOG_VERBOSE("no slot is available", {}); + queue_tasks.defer(task); + break; } - aggregate_result.result_json = json{ "results", result_jsons }; + if (task.data.contains("system_prompt")) + { + if (!all_slots_are_idle) { + send_error(task, "system prompt can only be updated when all slots are idle"); + break; + } + process_system_prompt_data(task.data["system_prompt"]); - agg_results.push_back(aggregate_result); + // reset cache_tokens for all slots + for (llama_client_slot &slot : slots) + { + slot.cache_tokens.clear(); + } + } - condition_results.notify_all(); + slot->reset(); - queue_iterator = queue_multitasks.erase(queue_iterator); - } - else - { - ++queue_iterator; - } + slot->infill = task.infill_mode; + slot->embedding = task.embedding_mode; + slot->task_id = task.id; + slot->multitask_id = task.multitask_id; + + if (!launch_slot_with_data(slot, task.data)) + { + // send error result + send_error(task, "internal_error"); + break; + } + } break; + case TASK_TYPE_CANCEL: { // release slot linked with the task id + for (auto & slot : slots) + { + if (slot.task_id == task.target_id) + { + slot.release(); + break; + } + } + } break; + case TASK_TYPE_NEXT_RESPONSE: { + // do nothing + } break; } + } - // done with tasks, unlock - lock.unlock(); + void on_finish_multitask(task_multi& multitask) + { + // all subtasks done == multitask is done + task_result result; + result.id = multitask.id; + result.stop = true; + result.error = false; - // copy aggregate results of complete multi-tasks to the results queue - std::lock_guard lock_results(mutex_results); - queue_results.insert(queue_results.end(), agg_results.begin(), agg_results.end()); + // collect json results into one json result + std::vector result_jsons; + for (auto& subres : multitask.results) + { + result_jsons.push_back(subres.result_json); + result.error = result.error && subres.error; + } + result.result_json = json{ { "results", result_jsons } }; + queue_results.send(result); } bool update_slots() { - // attend tasks - process_tasks(); - if (system_need_update) { LOG_TEE("updating system prompt\n"); @@ -1684,10 +1339,12 @@ struct llama_server_context LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n"); kv_cache_clear(); } - std::unique_lock lock(mutex_tasks); - condition_tasks.wait(lock, [&]{ - return !queue_tasks.empty(); - }); + return true; + } else { + task_server task; + task.type = TASK_TYPE_NEXT_RESPONSE; + task.target_id = -1; + queue_tasks.post(task); } for (llama_client_slot &slot : slots) @@ -1732,6 +1389,7 @@ struct llama_server_context slot.t_last_used = ggml_time_us(); LOG_TEE("slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size()); + queue_tasks.notify_slot_changed(); continue; } @@ -1997,6 +1655,10 @@ struct llama_server_context } return true; } + + void run_on_all_tasks_finished() { + update_slots(); + } }; static void server_print_usage(const char *argv0, const gpt_params ¶ms, @@ -2541,239 +2203,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } } -static std::string random_string() -{ - static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); - - std::random_device rd; - std::mt19937 generator(rd()); - - std::string result(32, ' '); - - for (int i = 0; i < 32; ++i) { - result[i] = str[generator() % str.size()]; - } - - return result; -} - -static std::string gen_chatcmplid() -{ - std::stringstream chatcmplid; - chatcmplid << "chatcmpl-" << random_string(); - return chatcmplid.str(); -} - -std::string format_chatml(std::vector messages) -{ - std::ostringstream chatml_msgs; - - for (auto it = messages.begin(); it != messages.end(); ++it) { - chatml_msgs << "<|im_start|>" - << json_value(*it, "role", std::string("user")) << '\n'; - chatml_msgs << json_value(*it, "content", std::string("")) - << "<|im_end|>\n"; - } - - chatml_msgs << "<|im_start|>assistant" << '\n'; - - return chatml_msgs.str(); -} - /* llama.cpp completion api semantics */ -json oaicompat_completion_params_parse( - const json &body /* openai api json semantics */) -{ - json llama_params; - - llama_params["__oaicompat"] = true; - - // Map OpenAI parameters to llama.cpp parameters - // - // For parameters that are defined by the OpenAI documentation (e.g. - // temperature), we explicitly specify OpenAI's intended default; we - // need to do that because sometimes OpenAI disagrees with llama.cpp - // - // https://platform.openai.com/docs/api-reference/chat/create - llama_sampling_params default_sparams; - llama_params["model"] = json_value(body, "model", std::string("unknown")); - llama_params["prompt"] = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt' - llama_params["cache_prompt"] = json_value(body, "cache_prompt", false); - llama_params["temperature"] = json_value(body, "temperature", 0.0); - llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k); - llama_params["top_p"] = json_value(body, "top_p", 1.0); - llama_params["n_predict"] = json_value(body, "max_tokens", -1); - llama_params["logit_bias"] = json_value(body, "logit_bias",json::object()); - llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0); - llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0); - llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED); - llama_params["stream"] = json_value(body, "stream", false); - llama_params["mirostat"] = json_value(body, "mirostat", default_sparams.mirostat); - llama_params["mirostat_tau"] = json_value(body, "mirostat_tau", default_sparams.mirostat_tau); - llama_params["mirostat_eta"] = json_value(body, "mirostat_eta", default_sparams.mirostat_eta); - llama_params["penalize_nl"] = json_value(body, "penalize_nl", default_sparams.penalize_nl); - llama_params["typical_p"] = json_value(body, "typical_p", default_sparams.typical_p); - llama_params["repeat_last_n"] = json_value(body, "repeat_last_n", default_sparams.penalty_last_n); - llama_params["ignore_eos"] = json_value(body, "ignore_eos", false); - llama_params["tfs_z"] = json_value(body, "tfs_z", default_sparams.tfs_z); - - if (body.count("grammar") != 0) { - llama_params["grammar"] = json_value(body, "grammar", json::object()); - } - - // Handle 'stop' field - if (body.contains("stop") && body["stop"].is_string()) { - llama_params["stop"] = json::array({body["stop"].get()}); - } else { - llama_params["stop"] = json_value(body, "stop", json::array()); - } - - // Ensure there is ChatML-specific end sequence among stop words - llama_params["stop"].push_back("<|im_end|>"); - - return llama_params; -} - -static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false) -{ - json result = response.result_json; - - bool stopped_word = result.count("stopped_word") != 0; - bool stopped_eos = json_value(result, "stopped_eos", false); - int num_tokens_predicted = json_value(result, "tokens_predicted", 0); - int num_prompt_tokens = json_value(result, "tokens_evaluated", 0); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason = "length"; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - - json choices = - streaming ? json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}) - : json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"message", json{{"content", content}, - {"role", "assistant"}}}}}); - - std::time_t t = std::time(0); - - json res = - json{{"choices", choices}, - {"created", t}, - {"model", - json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", streaming ? "chat.completion.chunk" : "chat.completion"}, - {"usage", - json{{"completion_tokens", num_tokens_predicted}, - {"prompt_tokens", num_prompt_tokens}, - {"total_tokens", num_tokens_predicted + num_prompt_tokens}}}, - {"id", gen_chatcmplid()}}; - - if (server_verbose) { - res["__verbose"] = result; - } - - if (result.contains("completion_probabilities")) { - res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array()); - } - - return res; -} - -// return value is vector as there is one case where we might need to generate two responses -static std::vector format_partial_response_oaicompat(const task_result &response) { - json result = response.result_json; - - if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { - return std::vector({response.result_json}); - } - - bool first = json_value(result, "oaicompat_token_ctr", 0) == 0; - std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); - - bool stopped_word = json_value(result, "stopped_word", false); - bool stopped_eos = json_value(result, "stopped_eos", false); - bool stopped_limit = json_value(result, "stopped_limit", false); - std::string content = json_value(result, "content", std::string("")); - - std::string finish_reason; - if (stopped_word || stopped_eos) { - finish_reason = "stop"; - } - if (stopped_limit) { - finish_reason = "length"; - } - - std::time_t t = std::time(0); - - json choices; - - if (!finish_reason.empty()) { - choices = json::array({json{{"finish_reason", finish_reason}, - {"index", 0}, - {"delta", json::object()}}}); - } else { - if (first) { - if (content.empty()) { - choices = json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{{"role", "assistant"}}}}}); - } else { - // We have to send this as two updates to conform to openai behavior - json initial_ret = json{{"choices", json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"role", "assistant"} - }}}})}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - json second_ret = json{ - {"choices", json::array({json{{"finish_reason", nullptr}, - {"index", 0}, - {"delta", json{ - {"content", content}}} - }})}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({initial_ret, second_ret}); - } - } else { - // Some idiosyncrasy in task processing logic makes several trailing calls - // with empty content, we ignore these at the calee site. - if (content.empty()) { - return std::vector({json::object()}); - } - - choices = json::array({json{ - {"finish_reason", nullptr}, - {"index", 0}, - {"delta", - json{ - {"content", content}, - }}, - }}); - } - } - - json ret = json{{"choices", choices}, - {"created", t}, - {"id", gen_chatcmplid()}, - {"model", modelname}, - {"object", "chat.completion.chunk"}}; - - return std::vector({ret}); -} - static json format_partial_response( llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector &probs ) { @@ -3069,10 +2499,12 @@ int main(int argc, char **argv) return; } json data = json::parse(req.body); - const int task_id = llama.request_completion(data, false, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, false, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); } @@ -3080,14 +2512,14 @@ int main(int argc, char **argv) { res.status = 404; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { while (true) { - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error) { const std::string str = "data: " + @@ -3098,6 +2530,7 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } if (result.stop) { @@ -3113,11 +2546,14 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } break; } } + + llama.queue_results.remove_waiting_task_id(task_id); sink.done(); return true; }; @@ -3126,6 +2562,7 @@ int main(int argc, char **argv) { // cancel llama.request_cancel(task_id); + llama.queue_results.remove_waiting_task_id(task_id); }; res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); @@ -3162,11 +2599,13 @@ int main(int argc, char **argv) } json data = oaicompat_completion_params_parse(json::parse(req.body)); - const int task_id = llama.request_completion(data, false, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, false, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { json oaicompat_result = format_final_response_oaicompat(data, result); @@ -3177,12 +2616,12 @@ int main(int argc, char **argv) } else { res.status = 500; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink &sink) { while (true) { - task_result llama_result = llama.next_result(task_id); + task_result llama_result = llama.queue_results.recv(task_id); if (!llama_result.error) { std::vector result_array = format_partial_response_oaicompat( llama_result); @@ -3195,6 +2634,7 @@ int main(int argc, char **argv) "\n\n"; LOG_VERBOSE("data stream", {{"to_send", str}}); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } } @@ -3210,18 +2650,21 @@ int main(int argc, char **argv) "\n\n"; LOG_VERBOSE("data stream", {{"to_send", str}}); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } break; } } sink.done(); + llama.queue_results.remove_waiting_task_id(task_id); return true; }; auto on_complete = [task_id, &llama](bool) { // cancel request llama.request_cancel(task_id); + llama.queue_results.remove_waiting_task_id(task_id); }; res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); @@ -3235,10 +2678,12 @@ int main(int argc, char **argv) return; } json data = json::parse(req.body); - const int task_id = llama.request_completion(data, true, false, -1); + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, data, true, false, -1); if (!json_value(data, "stream", false)) { std::string completion_text; - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error && result.stop) { res.set_content(result.result_json.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); @@ -3247,13 +2692,13 @@ int main(int argc, char **argv) { res.status = 404; res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - return; } + llama.queue_results.remove_waiting_task_id(task_id); } else { const auto chunked_content_provider = [task_id, &llama](size_t, httplib::DataSink & sink) { while (true) { - task_result result = llama.next_result(task_id); + task_result result = llama.queue_results.recv(task_id); if (!result.error) { const std::string str = "data: " + @@ -3264,6 +2709,7 @@ int main(int argc, char **argv) }); if (!sink.write(str.c_str(), str.size())) { + llama.queue_results.remove_waiting_task_id(task_id); return false; } if (result.stop) @@ -3277,8 +2723,8 @@ int main(int argc, char **argv) } } + llama.queue_results.remove_waiting_task_id(task_id); sink.done(); - return true; }; @@ -3352,23 +2798,46 @@ int main(int argc, char **argv) image_data = ""; } - const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1); - task_result result = llama.next_result(task_id); + // create and queue the task + const int task_id = llama.queue_tasks.get_new_id(); + llama.queue_results.add_waiting_task_id(task_id); + llama.request_completion(task_id, { {"prompt", prompt}, { "n_predict", 0}, {"image_data", image_data} }, false, true, -1); + + // get the result + task_result result = llama.queue_results.recv(task_id); + llama.queue_results.remove_waiting_task_id(task_id); + + // send the result return res.set_content(result.result_json.dump(), "application/json; charset=utf-8"); }); // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!? // "Bus error: 10" - this is on macOS, it does not crash on Linux //std::thread t2([&]() - { + /*{ bool running = true; while (running) { running = llama.update_slots(); } - } + }*/ //); + llama.queue_tasks.on_new_task(std::bind( + &llama_server_context::process_single_task, &llama, std::placeholders::_1)); + llama.queue_tasks.on_finish_multitask(std::bind( + &llama_server_context::on_finish_multitask, &llama, std::placeholders::_1)); + llama.queue_tasks.on_all_tasks_finished(std::bind( + &llama_server_context::run_on_all_tasks_finished, &llama)); + llama.queue_results.on_multitask_update(std::bind( + &llama_server_queue::update_multitask, + &llama.queue_tasks, + std::placeholders::_1, + std::placeholders::_2, + std::placeholders::_3 + )); + llama.queue_tasks.start_loop(); + t.join(); llama_backend_free(); diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp new file mode 100644 index 000000000..e2b6065f7 --- /dev/null +++ b/examples/server/utils.hpp @@ -0,0 +1,507 @@ +#pragma once + +#include +#include +#include +#include +#include +#include + +#include "json.hpp" + +#include "../llava/clip.h" + +using json = nlohmann::json; + +extern bool server_verbose; + +#ifndef SERVER_VERBOSE +#define SERVER_VERBOSE 1 +#endif + +#if SERVER_VERBOSE != 1 +#define LOG_VERBOSE(MSG, ...) +#else +#define LOG_VERBOSE(MSG, ...) \ + do \ + { \ + if (server_verbose) \ + { \ + server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \ + } \ + } while (0) +#endif + +#define LOG_ERROR( MSG, ...) server_log("ERROR", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__) +#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) + +// +// parallel +// + +enum server_state { + SERVER_STATE_LOADING_MODEL, // Server is starting up, model not fully loaded yet + SERVER_STATE_READY, // Server is ready and model is loaded + SERVER_STATE_ERROR // An error occurred, load_model failed +}; + +enum task_type { + TASK_TYPE_COMPLETION, + TASK_TYPE_CANCEL, + TASK_TYPE_NEXT_RESPONSE +}; + +struct task_server { + int id = -1; // to be filled by llama_server_queue + int target_id; + task_type type; + json data; + bool infill_mode = false; + bool embedding_mode = false; + int multitask_id = -1; +}; + +struct task_result { + int id; + int multitask_id = -1; + bool stop; + bool error; + json result_json; +}; + +struct task_multi { + int id; + std::set subtasks_remaining{}; + std::vector results{}; +}; + +// TODO: can become bool if we can't find use of more states +enum slot_state +{ + IDLE, + PROCESSING, +}; + +enum slot_command +{ + NONE, + LOAD_PROMPT, + RELEASE, +}; + +struct slot_params +{ + bool stream = true; + bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt + + uint32_t seed = -1; // RNG seed + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_predict = -1; // new tokens to predict + + std::vector antiprompt; + + json input_prefix; + json input_suffix; +}; + +struct slot_image +{ + int32_t id; + + bool request_encode_image = false; + float * image_embedding = nullptr; + int32_t image_tokens = 0; + + clip_image_u8 * img_data; + + std::string prefix_prompt; // before of this image +}; + +// completion token output with probabilities +struct completion_token_output +{ + struct token_prob + { + llama_token tok; + float prob; + }; + + std::vector probs; + llama_token tok; + std::string text_to_send; +}; + +static inline void server_log(const char *level, const char *function, int line, + const char *message, const nlohmann::ordered_json &extra) +{ + nlohmann::ordered_json log + { + {"timestamp", time(nullptr)}, + {"level", level}, + {"function", function}, + {"line", line}, + {"message", message}, + }; + + if (!extra.empty()) + { + log.merge_patch(extra); + } + + const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace); + printf("%.*s\n", (int)str.size(), str.data()); + fflush(stdout); +} + +// +// server utils +// + +template +static T json_value(const json &body, const std::string &key, const T &default_value) +{ + // Fallback null to default value + return body.contains(key) && !body.at(key).is_null() + ? body.value(key, default_value) + : default_value; +} + +inline std::string format_chatml(std::vector messages) +{ + std::ostringstream chatml_msgs; + + for (auto it = messages.begin(); it != messages.end(); ++it) { + chatml_msgs << "<|im_start|>" + << json_value(*it, "role", std::string("user")) << '\n'; + chatml_msgs << json_value(*it, "content", std::string("")) + << "<|im_end|>\n"; + } + + chatml_msgs << "<|im_start|>assistant" << '\n'; + + return chatml_msgs.str(); +} + +// +// work queue utils +// + +struct llama_server_queue { + int id = 0; + std::mutex mutex_tasks; + // queues + std::vector queue_tasks; + std::vector queue_tasks_deferred; + std::vector queue_multitasks; + std::condition_variable condition_tasks; + // callback functions + std::function callback_new_task; + std::function callback_finish_multitask; + std::function callback_all_task_finished; + + // Add a new task to the end of the queue + int post(task_server task) { + std::unique_lock lock(mutex_tasks); + if (task.id == -1) { + task.id = id++; + } + queue_tasks.push_back(std::move(task)); + condition_tasks.notify_one(); + return task.id; + } + + // Add a new task, but defer until one slot is available + void defer(task_server task) { + std::unique_lock lock(mutex_tasks); + queue_tasks_deferred.push_back(std::move(task)); + } + + // Get the next id for creating anew task + int get_new_id() { + std::unique_lock lock(mutex_tasks); + return id++; + } + + // Register function to process a new task + void on_new_task(std::function callback) { + callback_new_task = callback; + } + + // Register function to process a multitask + void on_finish_multitask(std::function callback) { + callback_finish_multitask = callback; + } + + // Register the function to be called when the batch of tasks is finished + void on_all_tasks_finished(std::function callback) { + callback_all_task_finished = callback; + } + + // Call when the state of one slot is changed + void notify_slot_changed() { + // move deferred tasks back to main loop + std::unique_lock lock(mutex_tasks); + for (auto & task : queue_tasks_deferred) { + queue_tasks.push_back(std::move(task)); + } + queue_tasks_deferred.clear(); + } + + // Start the main loop. This call is blocking + void start_loop() { + while (true) { + // new task arrived + LOG_VERBOSE("have new task", {}); + { + while (true) + { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + lock.unlock(); + break; + } + task_server task = queue_tasks.front(); + queue_tasks.erase(queue_tasks.begin()); + lock.unlock(); + LOG_VERBOSE("callback_new_task", {}); + callback_new_task(task); + } + LOG_VERBOSE("callback_all_task_finished", {}); + // process and update all the multitasks + auto queue_iterator = queue_multitasks.begin(); + while (queue_iterator != queue_multitasks.end()) + { + if (queue_iterator->subtasks_remaining.empty()) + { + // all subtasks done == multitask is done + task_multi current_multitask = *queue_iterator; + callback_finish_multitask(current_multitask); + // remove this multitask + queue_iterator = queue_multitasks.erase(queue_iterator); + } + else + { + ++queue_iterator; + } + } + // all tasks in the current loop is finished + callback_all_task_finished(); + } + LOG_VERBOSE("wait for new task", {}); + // wait for new task + { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + condition_tasks.wait(lock, [&]{ + return !queue_tasks.empty(); + }); + } + } + } + } + + // + // functions to manage multitasks + // + + // add a multitask by specifying the id of all subtask (subtask is a task_server) + void add_multitask(int multitask_id, std::vector& sub_ids) + { + std::lock_guard lock(mutex_tasks); + task_multi multi; + multi.id = multitask_id; + std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end())); + queue_multitasks.push_back(multi); + } + + // updatethe remaining subtasks, while appending results to multitask + void update_multitask(int multitask_id, int subtask_id, task_result& result) + { + std::lock_guard lock(mutex_tasks); + for (auto& multitask : queue_multitasks) + { + if (multitask.id == multitask_id) + { + multitask.subtasks_remaining.erase(subtask_id); + multitask.results.push_back(result); + } + } + } +}; + +struct llama_server_response { + typedef std::function callback_multitask_t; + callback_multitask_t callback_update_multitask; + // for keeping track of all tasks waiting for the result + std::set waiting_task_ids; + // the main result queue + std::vector queue_results; + std::mutex mutex_results; + std::condition_variable condition_results; + + void add_waiting_task_id(int task_id) { + std::unique_lock lock(mutex_results); + waiting_task_ids.insert(task_id); + } + + void remove_waiting_task_id(int task_id) { + std::unique_lock lock(mutex_results); + waiting_task_ids.erase(task_id); + } + + // This function blocks the thread until there is a response for this task_id + task_result recv(int task_id) { + while (true) + { + std::unique_lock lock(mutex_results); + condition_results.wait(lock, [&]{ + return !queue_results.empty(); + }); + LOG_VERBOSE("condition_results unblock", {}); + + for (int i = 0; i < (int) queue_results.size(); i++) + { + if (queue_results[i].id == task_id) + { + assert(queue_results[i].multitask_id == -1); + task_result res = queue_results[i]; + queue_results.erase(queue_results.begin() + i); + return res; + } + } + } + + // should never reach here + } + + // Register the function to update multitask + void on_multitask_update(callback_multitask_t callback) { + callback_update_multitask = callback; + } + + // Send a new result to a waiting task_id + void send(task_result result) { + std::unique_lock lock(mutex_results); + LOG_VERBOSE("send new result", {}); + for (auto& task_id : waiting_task_ids) { + // LOG_TEE("waiting task id %i \n", task_id); + // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result + if (result.multitask_id == task_id) + { + LOG_VERBOSE("callback_update_multitask", {}); + callback_update_multitask(task_id, result.id, result); + continue; + } + + if (result.id == task_id) + { + LOG_VERBOSE("queue_results.push_back", {}); + queue_results.push_back(result); + condition_results.notify_one(); + return; + } + } + } +}; + +// +// base64 utils (TODO: move to common in the future) +// + +static const std::string base64_chars = + "ABCDEFGHIJKLMNOPQRSTUVWXYZ" + "abcdefghijklmnopqrstuvwxyz" + "0123456789+/"; + +static inline bool is_base64(uint8_t c) +{ + return (isalnum(c) || (c == '+') || (c == '/')); +} + +static inline std::vector base64_decode(const std::string & encoded_string) +{ + int i = 0; + int j = 0; + int in_ = 0; + + int in_len = encoded_string.size(); + + uint8_t char_array_4[4]; + uint8_t char_array_3[3]; + + std::vector ret; + + while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) + { + char_array_4[i++] = encoded_string[in_]; in_++; + if (i == 4) + { + for (i = 0; i <4; i++) + { + char_array_4[i] = base64_chars.find(char_array_4[i]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (i = 0; (i < 3); i++) + { + ret.push_back(char_array_3[i]); + } + i = 0; + } + } + + if (i) + { + for (j = i; j <4; j++) + { + char_array_4[j] = 0; + } + + for (j = 0; j <4; j++) + { + char_array_4[j] = base64_chars.find(char_array_4[j]); + } + + char_array_3[0] = ((char_array_4[0] ) << 2) + ((char_array_4[1] & 0x30) >> 4); + char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2); + char_array_3[2] = ((char_array_4[2] & 0x3) << 6) + char_array_4[3]; + + for (j = 0; (j < i - 1); j++) + { + ret.push_back(char_array_3[j]); + } + } + + return ret; +} + +// +// random string / id +// + +static std::string random_string() +{ + static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); + + std::random_device rd; + std::mt19937 generator(rd()); + + std::string result(32, ' '); + + for (int i = 0; i < 32; ++i) { + result[i] = str[generator() % str.size()]; + } + + return result; +} + +static std::string gen_chatcmplid() +{ + std::stringstream chatcmplid; + chatcmplid << "chatcmpl-" << random_string(); + return chatcmplid.str(); +} From 3b7c914de25c6851396d7f9178249f1ed278120e Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 14:48:15 +0200 Subject: [PATCH 083/101] tests : gitignore test-c.o --- tests/.gitignore | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/.gitignore b/tests/.gitignore index 59be43b99..092dce742 100644 --- a/tests/.gitignore +++ b/tests/.gitignore @@ -1,2 +1,3 @@ * !*.* +test-c.o From 5f1925a8cef81eb9b372faaae34b0dd76d5361d4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 26 Jan 2024 17:09:44 +0200 Subject: [PATCH 084/101] scripts : move run-with-preset.py from root to scripts folder --- run_with_preset.py => scripts/run-with-preset.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) rename run_with_preset.py => scripts/run-with-preset.py (98%) diff --git a/run_with_preset.py b/scripts/run-with-preset.py similarity index 98% rename from run_with_preset.py rename to scripts/run-with-preset.py index 9b4d7ecbe..a18252730 100755 --- a/run_with_preset.py +++ b/scripts/run-with-preset.py @@ -46,7 +46,7 @@ Formatting considerations: - To define multiple "reverse_prompt" properties simultaneously the expected format is a list of strings. - To define a tensor split, pass a list of floats. """ -usage = "run_with_preset.py [-h] [yaml_files ...] [-- ...]" +usage = "run-with-preset.py [-h] [yaml_files ...] [-- ...]" epilog = (" -- specify additional CLI ars to be passed to the binary (override all preset files). " "Unknown args will be ignored.") From 7032f4f6349c17a8352f9f93f7d2122f45469e59 Mon Sep 17 00:00:00 2001 From: snadampal <87143774+snadampal@users.noreply.github.com> Date: Fri, 26 Jan 2024 11:17:59 -0600 Subject: [PATCH 085/101] ggml : update softmax n_task calculation (#5126) updated the n_task calculation to use max number of threads possible. This has improved the prompt eval performance by around 5% for DOT kernels and by around 10% for MMLA kernels on AWS Graviton3. --- ggml.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index ca98fde8a..ef6fd8caf 100644 --- a/ggml.c +++ b/ggml.c @@ -16597,7 +16597,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { } break; case GGML_OP_SOFT_MAX: { - n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0])); + n_tasks = MIN(n_threads, ggml_nrows(node->src[0])); } break; case GGML_OP_CONV_TRANSPOSE_1D: { From 15b4538ff29b280a395a1406d711497d8eaa2564 Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 26 Jan 2024 18:18:26 +0100 Subject: [PATCH 086/101] ggml-alloc : add 10% margin to the buffer sizes (#5149) --- ggml-alloc.c | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/ggml-alloc.c b/ggml-alloc.c index 60141a34d..95a93c99d 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -335,7 +335,9 @@ bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) { } size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) { - return alloc->max_size; + // FIXME: changes in the tensor sizes compared to the measure graph may cause allocations to fail + // to avoid this, we add a 10% margin to the buffer size + return alloc->max_size + alloc->max_size/10; } // graph allocator From 62fead3ea0a30c8d424f4a8373fa14165c7c707f Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 26 Jan 2024 18:59:43 +0100 Subject: [PATCH 087/101] cuda : fix tensor size calculation for non-split buffer (#5145) --- ggml-backend.c | 4 +++- ggml-cuda.cu | 19 +++++-------------- 2 files changed, 8 insertions(+), 15 deletions(-) diff --git a/ggml-backend.c b/ggml-backend.c index 423512def..3fff5fc87 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -30,7 +30,9 @@ size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) { GGML_CALL size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) { // get_alloc_size is optional, defaults to ggml_nbytes if (buft->iface.get_alloc_size) { - return buft->iface.get_alloc_size(buft, tensor); + size_t size = buft->iface.get_alloc_size(buft, tensor); + assert(size >= ggml_nbytes(tensor)); + return size; } return ggml_nbytes(tensor); } diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 05e5d18ab..0d599e20a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -9790,8 +9790,8 @@ static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * s // TODO: mmq/mmv support #endif - const int64_t nb11 = src1->nb[1]; - const int64_t nb1 = dst->nb[1]; + const size_t nb11 = src1->nb[1]; + const size_t nb1 = dst->nb[1]; const struct ggml_tensor * ids = src0; const int32_t id = ((int32_t *) dst->op_params)[0]; @@ -10304,15 +10304,11 @@ GGML_CALL static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t if (ggml_is_quantized(tensor->type)) { // initialize padding to 0 to avoid possible NaN values - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t original_size = ggml_nbytes_split(tensor, nrows_split); + size_t original_size = ggml_nbytes(tensor); size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); if (padded_size > original_size && tensor->view_src == nullptr) { - CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0])); + CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); } } } @@ -10415,12 +10411,7 @@ GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend } GGML_CALL static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { - int64_t row_low = 0; - int64_t row_high = ggml_nrows(tensor); - int64_t nrows_split = row_high - row_low; - - size_t size = ggml_nbytes_split(tensor, nrows_split); - + size_t size = ggml_nbytes(tensor); int64_t ne0 = tensor->ne[0]; if (ggml_is_quantized(tensor->type)) { From bbe7c56c9993af86aa2d84cbe1fd69e1b4300cea Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Fri, 26 Jan 2024 15:34:06 -0500 Subject: [PATCH 088/101] cmake : pass CPU architecture flags to nvcc (#5146) --- CMakeLists.txt | 74 ++++++++++++++++++++++++++------------------------ 1 file changed, 39 insertions(+), 35 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index af3665129..2b2ae532e 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -466,17 +466,17 @@ function(get_flags CCID CCVER) (CCID STREQUAL "Clang" AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0) ) - set(C_FLAGS ${C_FLAGS} -Wdouble-promotion) + list(APPEND C_FLAGS -Wdouble-promotion) endif() elseif (CCID STREQUAL "GNU") set(C_FLAGS -Wdouble-promotion) set(CXX_FLAGS -Wno-array-bounds) if (CCVER VERSION_GREATER_EQUAL 7.1.0) - set(CXX_FLAGS ${CXX_FLAGS} -Wno-format-truncation) + list(APPEND CXX_FLAGS -Wno-format-truncation) endif() if (CCVER VERSION_GREATER_EQUAL 8.1.0) - set(CXX_FLAGS ${CXX_FLAGS} -Wextra-semi) + list(APPEND CXX_FLAGS -Wextra-semi) endif() elseif (CCID MATCHES "Intel") # enable max optimization level when using Intel compiler @@ -510,16 +510,18 @@ if (LLAMA_ALL_WARNINGS) endif() endif() +set(CUDA_CXX_FLAGS "") + if (LLAMA_CUBLAS) set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math) if (NOT MSVC) - set(CUDA_FLAGS ${CUDA_FLAGS} -Wno-pedantic) + list(APPEND CUDA_FLAGS -Wno-pedantic) endif() if (LLAMA_ALL_WARNINGS AND NOT MSVC) set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c) if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "") - set(NVCC_CMD ${NVCC_CMD} -ccbin ${CMAKE_CUDA_HOST_COMPILER}) + list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER}) endif() execute_process( @@ -547,13 +549,8 @@ if (LLAMA_CUBLAS) message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}") get_flags(${CUDA_CCID} ${CUDA_CCVER}) - list(JOIN GF_CXX_FLAGS " " CUDA_CXX_FLAGS) # pass host compiler flags as a single argument - if (NOT CUDA_CXX_FLAGS STREQUAL "") - set(CUDA_FLAGS ${CUDA_FLAGS} -Xcompiler ${CUDA_CXX_FLAGS}) - endif() + list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS}) # This is passed to -Xcompiler later endif() - - add_compile_options("$<$:${CUDA_FLAGS}>") endif() if (WIN32) @@ -618,12 +615,7 @@ if (NOT MSVC) endif() endif() -function(add_compile_option_cpp ARG) - # Adds a compile option to C/C++ only, but not for Cuda. - # Use, e.g., for CPU-architecture flags. - add_compile_options($<$:${ARG}>) - add_compile_options($<$:${ARG}>) -endfunction() +set(ARCH_FLAGS "") if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64")) message(STATUS "ARM detected") @@ -636,19 +628,19 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC else() check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E) if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "") - add_compile_options(-mfp16-format=ieee) + list(APPEND ARCH_FLAGS -mfp16-format=ieee) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6") # Raspberry Pi 1, Zero - add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access) + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7") # Raspberry Pi 2 - add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) + list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations) endif() if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8") # Raspberry Pi 3, 4, Zero 2 (32-bit) - add_compile_options(-mno-unaligned-access) + list(APPEND ARCH_FLAGS -mno-unaligned-access) endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" ) @@ -659,7 +651,7 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE include(cmake/FindSIMD.cmake) endif () if (LLAMA_AVX512) - add_compile_option_cpp(/arch:AVX512) + list(APPEND ARCH_FLAGS /arch:AVX512) # MSVC has no compile-time flags enabling specific # AVX512 extensions, neither it defines the # macros corresponding to the extensions. @@ -673,49 +665,61 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE add_compile_definitions($<$:__AVX512VNNI__>) endif() elseif (LLAMA_AVX2) - add_compile_option_cpp(/arch:AVX2) + list(APPEND ARCH_FLAGS /arch:AVX2) elseif (LLAMA_AVX) - add_compile_option_cpp(/arch:AVX) + list(APPEND ARCH_FLAGS /arch:AVX) endif() else() if (LLAMA_NATIVE) - add_compile_option_cpp(-march=native) + list(APPEND ARCH_FLAGS -march=native) endif() if (LLAMA_F16C) - add_compile_option_cpp(-mf16c) + list(APPEND ARCH_FLAGS -mf16c) endif() if (LLAMA_FMA) - add_compile_option_cpp(-mfma) + list(APPEND ARCH_FLAGS -mfma) endif() if (LLAMA_AVX) - add_compile_option_cpp(-mavx) + list(APPEND ARCH_FLAGS -mavx) endif() if (LLAMA_AVX2) - add_compile_option_cpp(-mavx2) + list(APPEND ARCH_FLAGS -mavx2) endif() if (LLAMA_AVX512) - add_compile_option_cpp(-mavx512f) - add_compile_option_cpp(-mavx512bw) + list(APPEND ARCH_FLAGS -mavx512f) + list(APPEND ARCH_FLAGS -mavx512bw) endif() if (LLAMA_AVX512_VBMI) - add_compile_option_cpp(-mavx512vbmi) + list(APPEND ARCH_FLAGS -mavx512vbmi) endif() if (LLAMA_AVX512_VNNI) - add_compile_option_cpp(-mavx512vnni) + list(APPEND ARCH_FLAGS -mavx512vnni) endif() endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") message(STATUS "PowerPC detected") if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") - add_compile_options(-mcpu=powerpc64le) + list(APPEND ARCH_FLAGS -mcpu=powerpc64le) else() - add_compile_options(-mcpu=native -mtune=native) + list(APPEND ARCH_FLAGS -mcpu=native -mtune=native) #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) endif() else() message(STATUS "Unknown architecture") endif() +add_compile_options("$<$:${ARCH_FLAGS}>") +add_compile_options("$<$:${ARCH_FLAGS}>") + +if (LLAMA_CUBLAS) + list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS}) + list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED) # pass host compiler flags as a single argument + if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "") + list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED}) + endif() + add_compile_options("$<$:${CUDA_FLAGS}>") +endif() + if (MINGW) # Target Windows 8 for PrefetchVirtualMemory add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER}) From a1d6df129bcd3d42cda38c09217d8d4ec4ea3bdd Mon Sep 17 00:00:00 2001 From: 0cc4m Date: Fri, 26 Jan 2024 23:07:32 +0100 Subject: [PATCH 089/101] Add OpenCL add kernel (#5151) * Add OpenCL add kernel * Put add kernel into different string to stay within MSVC string length limit, disable float16 support due to bad results --- ggml-opencl.cpp | 87 +++++++++++++++++++++++++++++++++++++++++++++++-- ggml-opencl.h | 1 + ggml.c | 11 +++++++ 3 files changed, 96 insertions(+), 3 deletions(-) diff --git a/ggml-opencl.cpp b/ggml-opencl.cpp index 2bb93638f..bf9ad964f 100644 --- a/ggml-opencl.cpp +++ b/ggml-opencl.cpp @@ -714,7 +714,6 @@ __kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, dst[row] = tmp[0]; } } - ); @@ -784,6 +783,7 @@ __kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float dst[row] = tmp[0]; } } + ); @@ -799,6 +799,18 @@ __kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y } ); +std::string add_template = MULTILINE_QUOTE( +__kernel void add_f32(__global float * x, const int x_offset, __global float * y, const int y_offset, __global float * dst, const int dst_offset, const int ky) { + const int i = get_group_id(0)*get_local_size(0) + get_local_id(0); + + if (i >= get_global_size(0)) { + return; + } + + dst[dst_offset + i] = x[x_offset + i] + y[y_offset + i%ky]; +} +); + #define CL_CHECK(err) \ do { \ cl_int err_ = (err); \ @@ -878,6 +890,7 @@ static std::string generate_kernels() { } src << mul_kernel << '\n'; } + src << add_template << '\n'; return src.str(); } @@ -893,6 +906,7 @@ static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl; static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl; static cl_kernel mul_f32_cl; +static cl_kernel add_f32_cl; static bool fp16_support; static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) { @@ -1100,9 +1114,10 @@ void ggml_cl_init(void) { char *ext_buffer = (char *)alloca(ext_str_size + 1); clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL); ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated + // Disabled due to faulty outputs // Check if ext_buffer contains cl_khr_fp16 - fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL; - fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false"); + fp16_support = false; // strstr(ext_buffer, "cl_khr_fp16") != NULL; + // fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false"); cl_context_properties properties[] = { (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0 @@ -1150,6 +1165,8 @@ void ggml_cl_init(void) { // mul kernel CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err)); + + CL_CHECK((add_f32_cl = clCreateKernel(program, "add_f32", &err), err)); } static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) { @@ -1458,6 +1475,70 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src ggml_cl_mul_f32(src0, src1, dst); } +static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(src1->backend == GGML_BACKEND_GPU); + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + size_t x_size; + size_t d_size; + + cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0 + cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted. + cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst + + + for (int64_t i03 = 0; i03 < ne03; i03++) { + for (int64_t i02 = 0; i02 < ne02; i02++) { + cl_event ev; + + // copy src0 to device + CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev)); + + const int64_t i13 = i03%ne13; + const int64_t i12 = i02%ne12; + const int i1 = i13*ne12*ne11 + i12*ne11; + + cl_int x_offset = 0; + cl_int y_offset = i1*ne10; + cl_int d_offset = 0; + + size_t global = ne00 * ne01; + cl_int ky = ne10 * ne11; + + CL_CHECK(clSetKernelArg(add_f32_cl, 0, sizeof(cl_mem), &d_X)); + CL_CHECK(clSetKernelArg(add_f32_cl, 1, sizeof(cl_int), &x_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 2, sizeof(cl_mem), &d_Y)); + CL_CHECK(clSetKernelArg(add_f32_cl, 3, sizeof(cl_int), &y_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 4, sizeof(cl_mem), &d_D)); + CL_CHECK(clSetKernelArg(add_f32_cl, 5, sizeof(cl_int), &d_offset)); + CL_CHECK(clSetKernelArg(add_f32_cl, 6, sizeof(cl_int), &ky)); + CL_CHECK(clEnqueueNDRangeKernel(queue, add_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL)); + + CL_CHECK(clReleaseEvent(ev)); + CL_CHECK(clFinish(queue)); + + // copy dst to host + float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3); + CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL)); + } + } + ggml_cl_pool_free(d_X, x_size); + ggml_cl_pool_free(d_D, d_size); +} + +void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32); + ggml_cl_add_f32(src0, src1, dst); +} + static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const int64_t ne00 = src0->ne[0]; const int64_t ne01 = src0->ne[1]; diff --git a/ggml-opencl.h b/ggml-opencl.h index 919b00d63..257a6be6a 100644 --- a/ggml-opencl.h +++ b/ggml-opencl.h @@ -10,6 +10,7 @@ extern "C" { GGML_API void ggml_cl_init(void); GGML_API void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +GGML_API void ggml_cl_add(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); GGML_API bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, const struct ggml_tensor * dst); GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); GGML_API void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize); diff --git a/ggml.c b/ggml.c index ef6fd8caf..8f57003e0 100644 --- a/ggml.c +++ b/ggml.c @@ -7207,6 +7207,17 @@ static void ggml_compute_forward_add_f32( const int ith = params->ith; const int nth = params->nth; +#ifdef GGML_USE_CLBLAST + if (src1->backend == GGML_BACKEND_GPU) { + // TODO: OpenCL kernel support full broadcast + GGML_ASSERT(ggml_can_repeat_rows(src1, src0)); + if (ith == 0) { + ggml_cl_add(src0, src1, dst); + } + return; + } +#endif + const int nr = ggml_nrows(src0); GGML_TENSOR_BINARY_OP_LOCALS From ec903c034131848da9222536ff18da07ec0882a0 Mon Sep 17 00:00:00 2001 From: Maximilian Winter Date: Sat, 27 Jan 2024 14:38:05 +0100 Subject: [PATCH 090/101] server : add self-extend support (#5104) * Ported self extension to server example * Update server.cpp * Fixed prompt caching without self extend * Update server.cpp * Added description to server readme. * Update server.cpp * Update server.cpp * Update server.cpp * Update server.cpp * Update README.md * Changed descriptions * server : formatting * Update examples/server/server.cpp Co-authored-by: Georgi Gerganov * Update examples/server/server.cpp Co-authored-by: Georgi Gerganov * Update server.cpp * Update server.cpp --------- Co-authored-by: Georgi Gerganov --- examples/server/README.md | 3 +- examples/server/server.cpp | 172 +++++++++++++++++++++++++++++++------ 2 files changed, 147 insertions(+), 28 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index fd3034b99..1c92a2041 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -30,7 +30,8 @@ Command line options: - `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled) - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load "a system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime) - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA. - +- `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w` +- `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n` ## Build server is build alongside everything else from the root of the project diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 392836132..af63f2f6f 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -184,6 +184,12 @@ struct llama_client_slot struct llama_sampling_params sparams; llama_sampling_context *ctx_sampling = nullptr; + int32_t ga_i = 0; // group-attention state + int32_t ga_n = 1;// group-attention factor + int32_t ga_w = 512; // group-attention width + + int32_t n_past_se = 0; // self-extend + // multimodal std::vector images; @@ -212,7 +218,8 @@ struct llama_client_slot sent_count = 0; sent_token_probs_index = 0; infill = false; - + ga_i = 0; + n_past_se = 0; generated_token_probs.clear(); for (slot_image & img : images) @@ -399,9 +406,26 @@ struct llama_server_context slot.id = i; slot.n_ctx = n_ctx_slot; - slot.reset(); LOG_TEE(" -> Slot %i - max context: %i\n", slot.id, n_ctx_slot); + + const int ga_n = params.grp_attn_n; + const int ga_w = params.grp_attn_w; + + if (ga_n != 1) { + GGML_ASSERT(ga_n > 0 && "ga_n must be positive"); // NOLINT + GGML_ASSERT(ga_w % ga_n == 0 && "ga_w must be a multiple of ga_n"); // NOLINT + //GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of ga_w"); // NOLINT + //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT + LOG_TEE(" -> Slot %i - self-extend: ga_n = %d, ga_w = %d\n", slot.id, ga_n, ga_w); + } + + slot.ga_i = 0; + slot.ga_n = ga_n; + slot.ga_w = ga_w; + + slot.reset(); + slots.push_back(slot); } @@ -1349,32 +1373,35 @@ struct llama_server_context for (llama_client_slot &slot : slots) { - if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) + if (slot.ga_n == 1) { - // Shift context - const int n_left = slot.n_past - slot.params.n_keep - 1; - const int n_discard = n_left / 2; - - LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); - llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); - llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); - - for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + if (slot.is_processing() && slot.cache_tokens.size() >= (size_t) slot.n_ctx) { - slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + // Shift context + const int n_left = slot.n_past - slot.params.n_keep - 1; + const int n_discard = n_left / 2; + + LOG_TEE("slot %d: context shift - n_keep = %d, n_left = %d, n_discard = %d\n", slot.id, slot.params.n_keep, n_left, n_discard); + llama_kv_cache_seq_rm (ctx, slot.id, slot.params.n_keep + 1 , slot.params.n_keep + n_discard + 1); + llama_kv_cache_seq_shift(ctx, slot.id, slot.params.n_keep + 1 + n_discard, slot.n_past, -n_discard); + + for (size_t i = slot.params.n_keep + 1 + n_discard; i < slot.cache_tokens.size(); i++) + { + slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; + } + + slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + + slot.n_past -= n_discard; + + slot.truncated = true; + + LOG_VERBOSE("context shift", { + { "n_ctx", n_ctx }, + { "n_keep", params.n_keep }, + { "n_left", n_left }, + }); } - - slot.cache_tokens.resize(slot.cache_tokens.size() - n_discard); - - slot.n_past -= n_discard; - - slot.truncated = true; - - LOG_VERBOSE("context shift", { - {"n_ctx", n_ctx}, - {"n_keep", params.n_keep}, - {"n_left", n_left}, - }); } } @@ -1401,7 +1428,8 @@ struct llama_server_context slot.i_batch = batch.n_tokens; - llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true); + const int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past; + llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true); slot.n_past += 1; } @@ -1499,6 +1527,8 @@ struct llama_server_context llama_sampling_reset(slot.ctx_sampling); slot.n_past = 0; + slot.n_past_se = 0; + slot.ga_i = 0; slot.num_prompt_tokens_processed = slot.num_prompt_tokens; } else @@ -1512,6 +1542,25 @@ struct llama_server_context slot.n_past = common_part(slot.cache_tokens, prompt_tokens); slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past; + if (slot.ga_n != 1) + { + int ga_i = 0; + int32_t ga_n = slot.ga_n; + int32_t ga_w = slot.ga_w; + int32_t slot_npast = 0; + for (int k = 0; k < slot.n_past; ++k) + { + while (slot_npast >= ga_i + ga_w) { + const int bd = (ga_w/ga_n)*(ga_n - 1); + slot_npast -= bd; + ga_i += ga_w/ga_n; + } + slot_npast++; + } + slot.n_past_se = slot_npast; + slot.ga_i = ga_i; + } + LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed); } @@ -1526,6 +1575,10 @@ struct llama_server_context // we have to evaluate at least 1 token to generate logits. LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id); slot.n_past--; + if (slot.ga_i > 0) + { + slot.n_past_se--; + } } LOG_VERBOSE("prompt ingested", { @@ -1538,9 +1591,22 @@ struct llama_server_context // process the prefix of first image std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens; + int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past; + int ga_i = slot.ga_i; + int32_t ga_n = slot.ga_n; + int32_t ga_w = slot.ga_w; for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) { - llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); + if (slot.ga_n != 1) + { + while (slot_npast >= ga_i + ga_w) { + const int bd = (ga_w/ga_n)*(ga_n - 1); + slot_npast -= bd; + ga_i += ga_w/ga_n; + } + } + llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot_npast, {slot.id }, false); + slot_npast += 1; } if (has_images && !ingest_images(slot, n_batch)) @@ -1570,6 +1636,36 @@ struct llama_server_context for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i)); + + for (auto & slot : slots) + { + if (slot.ga_n != 1) + { + // context extension via Self-Extend + while (slot.n_past_se >= slot.ga_i + slot.ga_w) + { + const int ib = (slot.ga_n * slot.ga_i) / slot.ga_w; + const int bd = (slot.ga_w / slot.ga_n) * (slot.ga_n - 1); + const int dd = (slot.ga_w / slot.ga_n) - ib * bd - slot.ga_w; + + LOG_TEE("\n"); + LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i, slot.n_past_se, ib * bd, slot.ga_i + ib * bd, slot.n_past_se + ib * bd); + LOG_TEE("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n); + LOG_TEE("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd); + + llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd); + llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n); + llama_kv_cache_seq_shift(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd); + + slot.n_past_se -= bd; + + slot.ga_i += slot.ga_w / slot.ga_n; + + LOG_TEE("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", slot.n_past_se + bd, slot.n_past_se, slot.ga_i); + } + slot.n_past_se += n_tokens; + } + } llama_batch batch_view = { n_tokens, @@ -1583,6 +1679,7 @@ struct llama_server_context }; const int ret = llama_decode(ctx, batch_view); + if (ret != 0) { if (n_batch == 1 || ret < 0) @@ -1728,6 +1825,8 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --override-kv KEY=TYPE:VALUE\n"); printf(" advanced option to override model metadata by key. may be specified multiple times.\n"); printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n"); + printf(" -gan N, --grp-attn-n N Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`"); + printf(" -gaw N, --grp-attn-w N Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`"); printf("\n"); } @@ -1913,6 +2012,25 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } params.n_threads = std::stoi(argv[i]); } + else if (arg == "--grp-attn-n" || arg == "-gan") + { + if (++i >= argc) { + invalid_param = true; + break; + } + + params.grp_attn_n = std::stoi(argv[i]); + } + else if (arg == "--grp-attn-w" || arg == "-gaw") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + + params.grp_attn_w = std::stoi(argv[i]); + } else if (arg == "--threads-batch" || arg == "-tb") { if (++i >= argc) From 35a2ee914308c85ab5cb576467381443ad23f0ac Mon Sep 17 00:00:00 2001 From: Michael Klimenko Date: Sat, 27 Jan 2024 15:25:55 +0100 Subject: [PATCH 091/101] Remove unused data and add fixes (#5154) * Remove unused data and add fixes * Add missing file * Address review comments * Replace the scope of vq allocation --- common/sampling.cpp | 1 + examples/infill/infill.cpp | 2 +- examples/llava/clip.cpp | 1 - examples/server/server.cpp | 2 +- pocs/vdot/vdot.cpp | 1 - tests/test-backend-ops.cpp | 1 - tests/test-llama-grammar.cpp | 1 - 7 files changed, 3 insertions(+), 6 deletions(-) diff --git a/common/sampling.cpp b/common/sampling.cpp index efd7eab6e..e8675a8c0 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -13,6 +13,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_ // will be empty (default) if there are parse errors if (result->parsed_grammar.rules.empty()) { fprintf(stderr, "%s: failed to parse grammar\n", __func__); + delete result; return nullptr; } diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 4a7827876..72fb133b4 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -241,7 +241,7 @@ int main(int argc, char ** argv) { LOG("add_bos: %d\n", add_bos); bool suff_rm_leading_spc = params.escape; - if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) { + if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { params.input_suffix.erase(0, 1); suff_rm_leading_spc = false; } diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 4a0338a37..f2cd86afe 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1277,7 +1277,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i ".*weight", }; - std::vector read_data(512); std::vector work(512); std::vector conv_buf(512); std::vector hist_all(1 << 4, 0); diff --git a/examples/server/server.cpp b/examples/server/server.cpp index af63f2f6f..f58a2acaa 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -681,7 +681,7 @@ struct llama_server_context while ((pos = prompt.find(pattern, pos)) != std::string::npos) { size_t end_prefix = pos; pos += pattern.length(); - size_t end_pos = prompt.find("]", pos); + size_t end_pos = prompt.find(']', pos); if (end_pos != std::string::npos) { std::string image_id = prompt.substr(pos, end_pos - pos); diff --git a/pocs/vdot/vdot.cpp b/pocs/vdot/vdot.cpp index e96372c4b..73ffcd1ca 100644 --- a/pocs/vdot/vdot.cpp +++ b/pocs/vdot/vdot.cpp @@ -243,7 +243,6 @@ int main(int argc, char** argv) { if (useQ4_1) q41.resize(n4); else q40.resize(n4); std::vector q8(n8); - std::vector H(16, 0); double sumt = 0, sumt2 = 0, maxt = 0; double sumqt = 0, sumqt2 = 0, maxqt = 0; double sum = 0, sumq = 0, exactSum = 0; diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 55ce14e0d..e3c656f56 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -102,7 +102,6 @@ static std::vector tensor_to_float(const ggml_tensor * t) { } else if (t->type == GGML_TYPE_I8) { tv.push_back((float)*(int8_t *) &buf[i]); } else if (quantized) { - std::vector vq(ggml_blck_size(t->type)); tt.to_float(&buf[i], vq.data(), ggml_blck_size(t->type)); tv.insert(tv.end(), vq.begin(), vq.end()); } else { diff --git a/tests/test-llama-grammar.cpp b/tests/test-llama-grammar.cpp index 73dd33dd2..78fc41117 100644 --- a/tests/test-llama-grammar.cpp +++ b/tests/test-llama-grammar.cpp @@ -190,7 +190,6 @@ int main() index++; } - std::vector> next_stacks; std::vector next_candidates; next_candidates.resize(24); From e9764230054e01553bdead6f2bfd8e001869599d Mon Sep 17 00:00:00 2001 From: Judd Date: Fri, 26 Jan 2024 21:04:01 +0800 Subject: [PATCH 092/101] ggml : check ggml_add src1 type (ggml/708) Co-authored-by: Judd --- ggml.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index 8f57003e0..11a3114e7 100644 --- a/ggml.c +++ b/ggml.c @@ -7498,7 +7498,12 @@ static void ggml_compute_forward_add( switch (src0->type) { case GGML_TYPE_F32: { - ggml_compute_forward_add_f32(params, src0, src1, dst); + if (src1->type == GGML_TYPE_F32) { + ggml_compute_forward_add_f32(params, src0, src1, dst); + } + else { + GGML_ASSERT(false); + } } break; case GGML_TYPE_F16: { From 753eafed0ebd07af6903771327a1786a7c02cf98 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 27 Jan 2024 16:59:20 +0200 Subject: [PATCH 093/101] sync : ggml --- examples/server/utils.hpp | 1 + scripts/sync-ggml.last | 2 +- 2 files changed, 2 insertions(+), 1 deletion(-) diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index e2b6065f7..70cce0721 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -249,6 +249,7 @@ struct llama_server_queue { } // Start the main loop. This call is blocking + [[noreturn]] void start_loop() { while (true) { // new task arrived diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last index 4d52d946b..efde0069f 100644 --- a/scripts/sync-ggml.last +++ b/scripts/sync-ggml.last @@ -1 +1 @@ -6c1ce0bd591a430c1d3f6797d905194581c878c1 +c2448f88d17395452a587d0176d19ed87e0f7ce1 From 6db2b41a76ee78d5efdd5c3cddd5d7ad3f646855 Mon Sep 17 00:00:00 2001 From: John <78893154+cmp-nct@users.noreply.github.com> Date: Sat, 27 Jan 2024 16:09:18 +0100 Subject: [PATCH 094/101] llava : support for Yi-VL and fix for mobileVLM (#5093) * Support for Yi-VL, templating fix for mobileVLM * ws * Update examples/llava/clip.cpp Co-authored-by: Georgi Gerganov * Update llava-cli.cpp * Update clip.cpp bugfix for new conversions --------- Co-authored-by: Georgi Gerganov --- examples/llava/clip.cpp | 70 ++++++++++++++++++++++++++++++++---- examples/llava/llava-cli.cpp | 32 +++++++++++++++-- 2 files changed, 92 insertions(+), 10 deletions(-) diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index f2cd86afe..9129052a2 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -98,6 +98,7 @@ static std::string format(const char * fmt, ...) { enum projector_type { PROJECTOR_TYPE_MLP, + PROJECTOR_TYPE_MLP_NORM, PROJECTOR_TYPE_LDP, PROJECTOR_TYPE_UNKNOWN, }; @@ -304,10 +305,18 @@ struct clip_vision_model { struct ggml_tensor * projection; // LLaVA projection - struct ggml_tensor * mm_0_w; - struct ggml_tensor * mm_0_b; - struct ggml_tensor * mm_2_w; - struct ggml_tensor * mm_2_b; + struct ggml_tensor * mm_0_w = NULL; + struct ggml_tensor * mm_0_b = NULL; + struct ggml_tensor * mm_2_w = NULL; + struct ggml_tensor * mm_2_b = NULL; + + // Yi type models with mlp+normalization projection + struct ggml_tensor * mm_1_w = NULL; // Yi type models have 0, 1, 3, 4 + struct ggml_tensor * mm_1_b = NULL; + struct ggml_tensor * mm_3_w = NULL; + struct ggml_tensor * mm_3_b = NULL; + struct ggml_tensor * mm_4_w = NULL; + struct ggml_tensor * mm_4_b = NULL; // MobileVLM projection struct ggml_tensor * mm_model_mlp_1_w; @@ -460,6 +469,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 // pre-layernorm { embeddings = ggml_norm(ctx0, embeddings, eps); + ggml_set_name(embeddings, "pre_ln"); embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b); } @@ -575,6 +585,27 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32 embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings); embeddings = ggml_add(ctx0, embeddings, model.mm_2_b); + + } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_0_b); + // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false); + // First LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_1_w), + model.mm_1_b); + + // GELU activation + embeddings = ggml_gelu(ctx0, embeddings); + + // Second linear layer + embeddings = ggml_mul_mat(ctx0, model.mm_3_w, embeddings); + embeddings = ggml_add(ctx0, embeddings, model.mm_3_b); + + // Second LayerNorm + embeddings = ggml_norm(ctx0, embeddings, eps); + embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_4_w), + model.mm_4_b); } else if (ctx->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projector @@ -808,6 +839,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { else { new_clip->proj_type = PROJECTOR_TYPE_MLP; } + if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + if (gguf_find_tensor(ctx, format(TN_LLAVA_PROJ, 3, "weight").c_str()) != -1) { + new_clip->proj_type = PROJECTOR_TYPE_MLP_NORM; + } + } } #ifdef GGML_USE_CUBLAS @@ -956,11 +992,29 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { vision_model.pre_ln_b = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias")); // LLaVA projection - if (new_clip->proj_type == PROJECTOR_TYPE_MLP) { + if (new_clip->proj_type == PROJECTOR_TYPE_MLP || new_clip->proj_type == PROJECTOR_TYPE_MLP_NORM) { vision_model.mm_0_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight")); vision_model.mm_0_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias")); - vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); - vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + try { + // Yi-type llava + vision_model.mm_1_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "weight")); + vision_model.mm_1_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 1, "bias")); + } catch (std::runtime_error & e) { } + try { + // missing in Yi-type llava + vision_model.mm_2_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight")); + vision_model.mm_2_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias")); + } catch (std::runtime_error & e) { } + try { + // Yi-type llava + vision_model.mm_3_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "weight")); + vision_model.mm_3_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 3, "bias")); + } catch (std::runtime_error & e) { } + try { + // Yi-type llava + vision_model.mm_4_w = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "weight")); + vision_model.mm_4_b = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 4, "bias")); + } catch (std::runtime_error & e) { } } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) { // MobileVLM projection @@ -1432,6 +1486,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) { } else if (ctx->proj_type == PROJECTOR_TYPE_MLP) { return ctx->vision_model.mm_2_b->ne[0]; + } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) { + return ctx->vision_model.mm_3_b->ne[0]; } else { std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type]; diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp index d94795fe3..6ac70ba69 100644 --- a/examples/llava/llava-cli.cpp +++ b/examples/llava/llava-cli.cpp @@ -148,10 +148,35 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict; const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama)); - // llava chat format is "\nUSER:\n\nASSISTANT:" - eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos); + std::string system_prompt, user_prompt; + size_t image_pos = prompt.find(""); + if (image_pos != std::string::npos) { + // new templating mode: Provide the full prompt including system message and use as a placeholder for the image + + system_prompt = prompt.substr(0, image_pos); + user_prompt = prompt.substr(image_pos + std::string("").length()); + // We replace \n with actual newlines in user_prompt, just in case -e was not used in templating string + size_t pos = 0; + while ((pos = user_prompt.find("\\n", pos)) != std::string::npos) { + user_prompt.replace(pos, 2, "\n"); + pos += 1; // Advance past the replaced newline + } + while ((pos = system_prompt.find("\\n", pos)) != std::string::npos) { + system_prompt.replace(pos, 2, "\n"); + pos += 1; // Advance past the replaced newline + } + + printf("system_prompt: %s\n", system_prompt.c_str()); + printf("user_prompt: %s\n", user_prompt.c_str()); + } else { + // llava-1.5 native mode + system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:"; + user_prompt = prompt + "\nASSISTANT:"; + } + + eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos); llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past); - eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false); + eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false); // generate the response @@ -162,6 +187,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_ for (int i = 0; i < max_tgt_len; i++) { const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past); if (strcmp(tmp, "") == 0) break; + if (strstr(tmp, "###")) break; // Yi-VL behavior printf("%s", tmp); fflush(stdout); From 39baaf55a160909bb9428bd981014218761a20cb Mon Sep 17 00:00:00 2001 From: Kyle Mistele Date: Sun, 28 Jan 2024 01:55:31 -0600 Subject: [PATCH 095/101] docker : add server-first container images (#5157) * feat: add Dockerfiles for each platform that user ./server instead of ./main * feat: update .github/workflows/docker.yml to build server-first docker containers * doc: add information about running the server with Docker to README.md * doc: add information about running with docker to the server README * doc: update n-gpu-layers to show correct GPU usage * fix(doc): update container tag from `server` to `server-cuda` for README example on running server container with CUDA --- .devops/server-cuda.Dockerfile | 32 +++++++++++++++++++++++ .devops/server-intel.Dockerfile | 25 ++++++++++++++++++ .devops/server-rocm.Dockerfile | 45 +++++++++++++++++++++++++++++++++ .devops/server.Dockerfile | 20 +++++++++++++++ .github/workflows/docker.yml | 4 +++ README.md | 14 +++++++++- examples/server/README.md | 8 ++++++ 7 files changed, 147 insertions(+), 1 deletion(-) create mode 100644 .devops/server-cuda.Dockerfile create mode 100644 .devops/server-intel.Dockerfile create mode 100644 .devops/server-rocm.Dockerfile create mode 100644 .devops/server.Dockerfile diff --git a/.devops/server-cuda.Dockerfile b/.devops/server-cuda.Dockerfile new file mode 100644 index 000000000..4f83904bc --- /dev/null +++ b/.devops/server-cuda.Dockerfile @@ -0,0 +1,32 @@ +ARG UBUNTU_VERSION=22.04 +# This needs to generally match the container host's environment. +ARG CUDA_VERSION=11.7.1 +# Target the CUDA build image +ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION} +# Target the CUDA runtime image +ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION} + +FROM ${BASE_CUDA_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +ARG CUDA_DOCKER_ARCH=all + +RUN apt-get update && \ + apt-get install -y build-essential git + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH} +# Enable cuBLAS +ENV LLAMA_CUBLAS=1 + +RUN make + +FROM ${BASE_CUDA_RUN_CONTAINER} as runtime + +COPY --from=build /app/server /server + +ENTRYPOINT [ "/server" ] diff --git a/.devops/server-intel.Dockerfile b/.devops/server-intel.Dockerfile new file mode 100644 index 000000000..e343d278c --- /dev/null +++ b/.devops/server-intel.Dockerfile @@ -0,0 +1,25 @@ +ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04 +ARG UBUNTU_VERSION=22.04 + +FROM intel/hpckit:$ONEAPI_VERSION as build + +RUN apt-get update && \ + apt-get install -y git + +WORKDIR /app + +COPY . . + +# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance +RUN mkdir build && \ + cd build && \ + cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \ + cmake --build . --config Release --target main server + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/build/bin/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/server" ] diff --git a/.devops/server-rocm.Dockerfile b/.devops/server-rocm.Dockerfile new file mode 100644 index 000000000..e9a31647c --- /dev/null +++ b/.devops/server-rocm.Dockerfile @@ -0,0 +1,45 @@ +ARG UBUNTU_VERSION=22.04 + +# This needs to generally match the container host's environment. +ARG ROCM_VERSION=5.6 + +# Target the CUDA build image +ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete + +FROM ${BASE_ROCM_DEV_CONTAINER} as build + +# Unless otherwise specified, we make a fat build. +# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878 +# This is mostly tied to rocBLAS supported archs. +ARG ROCM_DOCKER_ARCH=\ + gfx803 \ + gfx900 \ + gfx906 \ + gfx908 \ + gfx90a \ + gfx1010 \ + gfx1030 \ + gfx1100 \ + gfx1101 \ + gfx1102 + +COPY requirements.txt requirements.txt +COPY requirements requirements + +RUN pip install --upgrade pip setuptools wheel \ + && pip install -r requirements.txt + +WORKDIR /app + +COPY . . + +# Set nvcc architecture +ENV GPU_TARGETS=${ROCM_DOCKER_ARCH} +# Enable ROCm +ENV LLAMA_HIPBLAS=1 +ENV CC=/opt/rocm/llvm/bin/clang +ENV CXX=/opt/rocm/llvm/bin/clang++ + +RUN make + +ENTRYPOINT [ "/app/server" ] diff --git a/.devops/server.Dockerfile b/.devops/server.Dockerfile new file mode 100644 index 000000000..134588fe2 --- /dev/null +++ b/.devops/server.Dockerfile @@ -0,0 +1,20 @@ +ARG UBUNTU_VERSION=22.04 + +FROM ubuntu:$UBUNTU_VERSION as build + +RUN apt-get update && \ + apt-get install -y build-essential git + +WORKDIR /app + +COPY . . + +RUN make + +FROM ubuntu:$UBUNTU_VERSION as runtime + +COPY --from=build /app/server /server + +ENV LC_ALL=C.utf8 + +ENTRYPOINT [ "/server" ] diff --git a/.github/workflows/docker.yml b/.github/workflows/docker.yml index 825b8f503..94f9161fc 100644 --- a/.github/workflows/docker.yml +++ b/.github/workflows/docker.yml @@ -28,14 +28,18 @@ jobs: config: - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" } # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I # have disabled them for now until the reason why # is understood. - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" } + - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" } - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } + - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" } - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" } + - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" } steps: - name: Check out the repo uses: actions/checkout@v3 diff --git a/README.md b/README.md index 76e48ce8a..cd95f8144 100644 --- a/README.md +++ b/README.md @@ -931,17 +931,20 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th * Create a folder to store big models & intermediate files (ex. /llama/models) #### Images -We have two Docker images available for this project: +We have three Docker images available for this project: 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`) 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`) +3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`) Additionally, there the following images, similar to the above: - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`) - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`) +- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`) - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) +- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`) The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now). @@ -967,6 +970,12 @@ or with a light image: docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 ``` +or with a server image: + +```bash +docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 +``` + ### Docker With CUDA Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container. @@ -976,6 +985,7 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia ```bash docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile . docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile . +docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile . ``` You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture. @@ -989,6 +999,7 @@ The resulting images, are essentially the same as the non-CUDA images: 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file. +3. `local/llama.cpp:server-cuda`: This image only includes the server executable file. #### Usage @@ -997,6 +1008,7 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne ```bash docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1 +docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1 ``` ### Contributing diff --git a/examples/server/README.md b/examples/server/README.md index 1c92a2041..dce4ec47c 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -66,6 +66,14 @@ server.exe -m models\7B\ggml-model.gguf -c 2048 The above command will start a server that by default listens on `127.0.0.1:8080`. You can consume the endpoints with Postman or NodeJS with axios library. You can visit the web front end at the same url. +### Docker: +```bash +docker run -p 8080:8080 -v /path/to/models:/models ggerganov/llama.cpp:server -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 + +# or, with CUDA: +docker run -p 8080:8080 -v /path/to/models:/models --gpus all ggerganov/llama.cpp:server-cuda -m models/7B/ggml-model.gguf -c 512 --host 0.0.0.0 --port 8080 --n-gpu-layers 99 +``` + ## Testing with CURL Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the base OS. From f2e69d28c01303ca9dc79907f89ef120a6ac4a92 Mon Sep 17 00:00:00 2001 From: sharpHL <132747147+sharpHL@users.noreply.github.com> Date: Sun, 28 Jan 2024 16:00:30 +0800 Subject: [PATCH 096/101] llama : add support for Orion-14B (#5118) * add support for Orion-14B(https://huggingface.co/OrionStarAI/Orion-14B-Chat) * flake8 support * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: Georgi Gerganov * Update llama.cpp Co-authored-by: slaren * Update llama.cpp * Update llama.cpp --------- Co-authored-by: lixiaopu Co-authored-by: Georgi Gerganov Co-authored-by: slaren --- convert-hf-to-gguf.py | 81 +++++++++++++++++ gguf-py/gguf/constants.py | 24 ++++- llama.cpp | 187 ++++++++++++++++++++++++++++++++++++++ 3 files changed, 291 insertions(+), 1 deletion(-) diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 7a0a8c3db..6ab7f486e 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -201,6 +201,8 @@ class Model: return PlamoModel if model_architecture == "CodeShellForCausalLM": return CodeShellModel + if model_architecture == "OrionForCausalLM": + return OrionModel return Model def _is_model_safetensors(self) -> bool: @@ -250,6 +252,8 @@ class Model: return gguf.MODEL_ARCH.PLAMO if arch == "CodeShellForCausalLM": return gguf.MODEL_ARCH.CODESHELL + if arch == "OrionForCausalLM": + return gguf.MODEL_ARCH.ORION raise NotImplementedError(f'Architecture "{arch}" not supported!') @@ -572,6 +576,83 @@ class MPTModel(Model): self.gguf_writer.add_tensor("output.weight", data) +class OrionModel(Model): + def set_vocab(self): + self._set_vocab_sentencepiece() + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + hf_repo = self.hparams.get("_name_or_path", "") + + ctx_length = 0 + if "max_sequence_length" in self.hparams: + ctx_length = self.hparams["max_sequence_length"] + elif "max_position_embeddings" in self.hparams: + ctx_length = self.hparams["max_position_embeddings"] + elif "model_max_length" in self.hparams: + ctx_length = self.hparams["model_max_length"] + else: + print("gguf: can not find ctx length parameter.") + sys.exit() + + self.gguf_writer.add_file_type(self.ftype) + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_source_hf_repo(hf_repo) + self.gguf_writer.add_tensor_data_layout("Meta AI original pth") + self.gguf_writer.add_context_length(ctx_length) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"]) + + def write_tensors(self): + # Collect tensors from generator object + model_kv = dict(self.get_tensors()) + block_count = self.hparams["num_hidden_layers"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in model_kv.items(): + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + class BaichuanModel(Model): def set_vocab(self): self._set_vocab_sentencepiece() diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 2d9c33c7d..f5c933a41 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -101,6 +101,7 @@ class MODEL_ARCH(IntEnum): PHI2 = auto() PLAMO = auto() CODESHELL = auto() + ORION = auto() class MODEL_TENSOR(IntEnum): @@ -151,6 +152,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.PHI2: "phi2", MODEL_ARCH.PLAMO: "plamo", MODEL_ARCH.CODESHELL: "codeshell", + MODEL_ARCH.ORION: "orion", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -427,7 +429,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_NORM, MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, - ] + ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], # TODO } @@ -452,6 +470,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.ORION: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } # diff --git a/llama.cpp b/llama.cpp index b03b67e16..4cd0f16eb 100644 --- a/llama.cpp +++ b/llama.cpp @@ -196,6 +196,7 @@ enum llm_arch { LLM_ARCH_PHI2, LLM_ARCH_PLAMO, LLM_ARCH_CODESHELL, + LLM_ARCH_ORION, LLM_ARCH_UNKNOWN, }; @@ -217,6 +218,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PLAMO, "plamo" }, { LLM_ARCH_CODESHELL, "codeshell" }, + { LLM_ARCH_ORION, "orion" }, }; enum llm_kv { @@ -641,6 +643,25 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_ORION, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_ROT_EMBD, "blk.%d.attn_rot_embd" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, { LLM_ARCH_UNKNOWN, @@ -1332,6 +1353,7 @@ enum e_model { MODEL_7B, MODEL_8B, MODEL_13B, + MODEL_14B, MODEL_15B, MODEL_30B, MODEL_34B, @@ -2683,6 +2705,7 @@ static const char * llama_model_type_name(e_model type) { case MODEL_7B: return "7B"; case MODEL_8B: return "8B"; case MODEL_13B: return "13B"; + case MODEL_14B: return "14B"; case MODEL_15B: return "15B"; case MODEL_30B: return "30B"; case MODEL_34B: return "34B"; @@ -2950,7 +2973,15 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_ORION: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps); + switch (hparams.n_layer) { + case 40: model.type = e_model::MODEL_14B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3933,6 +3964,38 @@ static bool llm_load_tensors( layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}); } } break; + case LLM_ARCH_ORION: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + { + model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); + model.output_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + } + for (int i = 0; i < n_layer; ++i) { + ggml_context * ctx_layer = ctx_for_layer(i); + ggml_context * ctx_split = ctx_for_layer_split(i); + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}); + + layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}); + layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}); + layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}); + layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}); + + layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); + layer.ffn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}); + + layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); + } + } break; + + default: throw std::runtime_error("unknown architecture"); } @@ -4563,6 +4626,126 @@ struct llm_build_context { ctx0 = nullptr; } } + struct ggml_cgraph * build_orion() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_view_1d(ctx0, lctx.inp_pos, n_tokens, 0); + cb(inp_pos, "inp_pos", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_view_2d(ctx0, lctx.inp_KQ_mask, n_kv, n_tokens, n_kv*ggml_type_size(lctx.inp_KQ_mask->type), 0); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, lctx.inp_K_shift, LLM_ROPE, n_ctx, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + cur = llm_build_kv(ctx0, model, hparams, kv_self, gf, + model.layers[il].wo, NULL, + Kcur, Vcur, Qcur, KQ_mask, n_ctx, n_tokens, kv_head, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_llama() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -6520,6 +6703,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_codeshell(); } break; + case LLM_ARCH_ORION: + { + result = llm.build_orion(); + } break; default: GGML_ASSERT(false); } From af4980bfedfd8df43b9e4cd1442895e85fee37bc Mon Sep 17 00:00:00 2001 From: Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com> Date: Sun, 28 Jan 2024 00:30:44 -0800 Subject: [PATCH 097/101] readme : add link to rust bindings (#5148) * added link to another set of rust bindings with brief note on differences. * fixed link name --- README.md | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index cd95f8144..44898d2f2 100644 --- a/README.md +++ b/README.md @@ -122,7 +122,8 @@ as the main playground for developing new features for the [ggml](https://github - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp) - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp) - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb) -- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) +- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp) +- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs) - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp) - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s) - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj) From b2b2bf988c098851b4f3831f0cf38394bff75121 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sun, 28 Jan 2024 09:35:14 +0100 Subject: [PATCH 098/101] Tests for min_p, sampling queue (#5147) --- llama.cpp | 5 ++ tests/test-sampling.cpp | 169 ++++++++++++++++++++++++++++++++++++---- 2 files changed, 159 insertions(+), 15 deletions(-) diff --git a/llama.cpp b/llama.cpp index 4cd0f16eb..391c956ec 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8133,6 +8133,11 @@ void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * c } void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int32_t k, size_t min_keep) { + // TODO: move bucket sort to separate function so that top_p/tail_free/typical/softmax first is equally fast + // if (k >= (int32_t)candidates->size) { + // return; + // } + const int64_t t_start_sample_us = ggml_time_us(); k = std::max(k, (int) min_keep); diff --git a/tests/test-sampling.cpp b/tests/test-sampling.cpp index 32e58941c..c3b3d6629 100644 --- a/tests/test-sampling.cpp +++ b/tests/test-sampling.cpp @@ -5,11 +5,10 @@ #undef NDEBUG #endif -#include -#include -#include -#include #include +#include +#include +#include static void dump(const llama_token_data_array * candidates) { for (size_t i = 0; i < candidates->size; i++) { @@ -20,11 +19,11 @@ static void dump(const llama_token_data_array * candidates) { #define DUMP(__candidates) do { printf("%s:%d (%s)\n", __FILE__, __LINE__, __func__); dump((__candidates)); printf("-\n"); } while(0) static void test_top_k(const std::vector & probs, const std::vector & expected_probs, int k) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -41,11 +40,11 @@ static void test_top_k(const std::vector & probs, const std::vector & probs, const std::vector & expected_probs, float p) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -62,11 +61,11 @@ static void test_top_p(const std::vector & probs, const std::vector & probs, const std::vector & expected_probs, float z) { - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -81,12 +80,33 @@ static void test_tfs(const std::vector & probs, const std::vector } } -static void test_typical(const std::vector & probs, const std::vector & expected_probs, float p) { - size_t n_vocab = probs.size(); +static void test_min_p(const std::vector & probs, const std::vector & expected_probs, float p) { + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + DUMP(&candidates_p); + llama_sample_min_p(nullptr, &candidates_p, p, 1); + DUMP(&candidates_p); + llama_sample_softmax(nullptr, &candidates_p); + + GGML_ASSERT(candidates_p.size == expected_probs.size()); + for (size_t i = 0; i < candidates_p.size; i++) { + GGML_ASSERT(fabs(candidates_p.data[i].p - expected_probs[i]) < 1e-3); + } +} + +static void test_typical(const std::vector & probs, const std::vector & expected_probs, float p) { + const size_t n_vocab = probs.size(); + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -107,11 +127,11 @@ static void test_repetition_penalties( ) { GGML_ASSERT(probs.size() == expected_probs.size()); - size_t n_vocab = probs.size(); + const size_t n_vocab = probs.size(); std::vector candidates; candidates.reserve(n_vocab); for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { - float logit = log(probs[token_id]); + const float logit = logf(probs[token_id]); candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); } @@ -128,6 +148,88 @@ static void test_repetition_penalties( } } +static void test_sampler_queue( + const size_t n_vocab, const std::string samplers_sequence, const int top_k, const float top_p, const float min_p +) { + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < (llama_token)n_vocab; token_id++) { + const float logit = logf(token_id); + candidates.emplace_back(llama_token_data{token_id, logit, 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + + llama_token min_token_id = 0; + const llama_token max_token_id = n_vocab-1; + + for (auto s : samplers_sequence) { + switch (s){ + case 'k': llama_sample_top_k (nullptr, &candidates_p, top_k, 1); break; + case 'f': GGML_ASSERT(false && "tail_free test not implemented"); break; + case 'y': GGML_ASSERT(false && "typical test not implemented"); break; + case 'p': llama_sample_top_p (nullptr, &candidates_p, top_p, 1); break; + case 'm': llama_sample_min_p (nullptr, &candidates_p, min_p, 1); break; + case 't': GGML_ASSERT(false && "temperature test not implemented"); break; + default : GGML_ASSERT(false && "Unknown sampler"); break; + } + + llama_sample_softmax(nullptr, &candidates_p); // make sure tokens are sorted for tests + + const int size = candidates_p.size; + + if (s == 'k') { + const int expected_size = std::min(size, top_k); + min_token_id = std::max(min_token_id, (llama_token)(n_vocab - top_k)); + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else if (s == 'p') { + const int softmax_divisor = n_vocab * (n_vocab-1) / 2 - min_token_id * (min_token_id-1) / 2; + const int softmax_numerator_target = ceilf(top_p * softmax_divisor); + + min_token_id = n_vocab; + int expected_size = 0; + int cumsum = 0; + do { // do-while because always at least one token is sampled + min_token_id--; + expected_size++; + + cumsum += min_token_id; + } while (cumsum < softmax_numerator_target); + + // token 0 has p == 0, need special consideration for cumsum because top_p immediately returns + if (min_token_id == 1) { + min_token_id--; + expected_size += 1; + } + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else if (s == 'm') { + int expected_size = ceilf((1.0f-min_p) * n_vocab); + expected_size = std::max(expected_size, 1); + expected_size = std::min(expected_size, size); + + min_token_id = floorf(min_p * n_vocab); + min_token_id = std::max(min_token_id, 1); + min_token_id = std::max(min_token_id, (llama_token)(n_vocab - size)); + min_token_id = std::min(min_token_id, (llama_token)(n_vocab - 1)); + + GGML_ASSERT(size == expected_size); + GGML_ASSERT(candidates_p.data[0].id == max_token_id); + GGML_ASSERT(candidates_p.data[expected_size-1].id == min_token_id); + } else { + GGML_ASSERT(false); + } + } + + printf("Sampler queue %3s OK with n_vocab=%05ld top_k=%05d top_p=%f min_p=%f\n", + samplers_sequence.c_str(), n_vocab, top_k, top_p, min_p); +} + int main(void) { ggml_time_init(); @@ -139,6 +241,15 @@ int main(void) { test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f}, 0.8f); test_top_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 1); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.00f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/1.0f, 0.3f/1.0f, 0.2f/1.0f, 0.1f/1.0f}, 0.24f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.26f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.9f, 0.3f/0.9f, 0.2f/0.9f}, 0.49f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.51f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.7f, 0.3f/0.7f}, 0.74f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 0.76f); + test_min_p({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f/0.4f}, 1.00f); + test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f}, 0.25f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.75f); test_tfs({0.1f, 0.15f, 0.2f, 0.25f, 0.3f}, {0.3f, 0.25f}, 0.99f); @@ -154,6 +265,34 @@ int main(void) { test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2}, {0.499966f, 0.499966f, 0.000023f, 0.000023f, 0.000023f}, 1.0f, 5.0f, 5.0f); test_repetition_penalties({0.2f, 0.2f, 0.2f, 0.2f, 0.2f}, {0, 1, 2, 0, 0}, {0.499977f, 0.499977f, 0.000023f, 0.000023f, 0.000000f}, 1.0f, 5.0f, 5.0f); + test_sampler_queue(10000, "k", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "k", 1, 1.0f, 1.0f); + test_sampler_queue(10000, "p", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.0f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0f, 1e-12); + + test_sampler_queue(10000, "k", 100, 1.0000f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.0002f, 1.0f); + test_sampler_queue(10000, "p", 10000, 0.8000f, 1.0f); + test_sampler_queue(10000, "m", 10000, 1.0000f, 9997.9f/9999.0f); + test_sampler_queue(10000, "m", 10000, 1.0000f, 0.1f); + + test_sampler_queue(10000, "kp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "km", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mp", 100, 0.8f, 9997.9f/9999.0f); + test_sampler_queue(10000, "mp", 100, 0.8f, 0.1f); + + test_sampler_queue(10000, "kpm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "kmp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pkm", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "pmk", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mkp", 100, 0.8f, 0.1f); + test_sampler_queue(10000, "mpk", 100, 0.8f, 0.1f); + printf("OK\n"); return 0; From 9241c3a2ace544aef708334e54bbdddb90208ee8 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Johannes=20G=C3=A4=C3=9Fler?= Date: Sun, 28 Jan 2024 09:59:49 +0100 Subject: [PATCH 099/101] Apply min_p to unsorted tokens (#5115) --- llama.cpp | 54 +++++++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 45 insertions(+), 9 deletions(-) diff --git a/llama.cpp b/llama.cpp index 391c956ec..45eeadbe8 100644 --- a/llama.cpp +++ b/llama.cpp @@ -52,6 +52,7 @@ #include #include #include +#include #include #include #include @@ -8246,21 +8247,56 @@ void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * can return; } - llama_sample_softmax(ctx, candidates); - const int64_t t_start_sample_us = ggml_time_us(); - float scale = candidates->data[0].p; // scale by max prob - size_t i = 1; // first token always matches + bool min_p_applied = false; - for (; i < candidates->size; ++i) { - if (candidates->data[i].p < p * scale && i >= min_keep) { - break; // prob too small + // if the candidates aren't sorted, try the unsorted implementation first + if (!candidates->sorted) { + std::vector filtered_tokens; + + float max_logit = -FLT_MAX; + for (size_t i = 0; i < candidates->size; ++i) { + max_logit = std::max(max_logit, candidates->data[i].logit); + } + const float min_logit = max_logit + logf(p); // min logit for p_i >= p * p_max + + for (size_t i = 0; i < candidates->size; ++i) { + if (candidates->data[i].logit >= min_logit) { + filtered_tokens.push_back(candidates->data[i]); + } + } + + // if we have enough values the operation was a success + if (filtered_tokens.size() >= min_keep) { + memcpy(candidates->data, filtered_tokens.data(), filtered_tokens.size()*sizeof(llama_token_data)); + candidates->size = filtered_tokens.size(); + min_p_applied = true; } } - // Resize the output vector to keep only the matching tokens - candidates->size = i; + // if the candidates are sorted or the unsorted implementation failed, use this implementation + if (!min_p_applied) { + // Sort the logits in descending order + if (!candidates->sorted) { + std::sort(candidates->data, candidates->data + candidates->size, [](const llama_token_data & a, const llama_token_data & b) { + return a.logit > b.logit; + }); + candidates->sorted = true; + } + + const float min_logit = candidates->data[0].logit + logf(p); // min logit for p_i >= p * p_max + size_t i = 1; // first token always matches + + for (; i < candidates->size; ++i) { + if (candidates->data[i].logit < min_logit && i >= min_keep) { + break; // prob too small + } + } + + // Resize the output vector to keep only the matching tokens + candidates->size = i; + } if (ctx) { ctx->t_sample_us += ggml_time_us() - t_start_sample_us; From b764b8f1d079ba44d912801ce6d29bd0d94d51cf Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sun, 28 Jan 2024 16:54:54 +0200 Subject: [PATCH 100/101] flake.lock: Update (#5162) --- flake.lock | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/flake.lock b/flake.lock index 1b253cb44..95e41f333 100644 --- a/flake.lock +++ b/flake.lock @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1705677747, - "narHash": "sha256-eyM3okYtMgYDgmYukoUzrmuoY4xl4FUujnsv/P6I/zI=", + "lastModified": 1706191920, + "narHash": "sha256-eLihrZAPZX0R6RyM5fYAWeKVNuQPYjAkCUBr+JNvtdE=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "bbe7d8f876fbbe7c959c90ba2ae2852220573261", + "rev": "ae5c332cbb5827f6b1f02572496b141021de335f", "type": "github" }, "original": { From 0f648573dde61c510560f68244f70ece7e60d8c1 Mon Sep 17 00:00:00 2001 From: Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com> Date: Sun, 28 Jan 2024 21:26:23 +0530 Subject: [PATCH 101/101] ggml : add unified SYCL backend for Intel GPUs (#2690) * first update for migration * update init_cublas * add debug functio, commit all help code * step 1 * step 2 * step3 add fp16, slower 31->28 * add GGML_LIST_DEVICE function * step 5 format device and print * step6, enhance error check, remove CUDA macro, enhance device id to fix none-zero id issue * support main device is non-zero * step7 add debug for code path, rm log * step 8, rename all macro & func from cuda by sycl * fix error of select non-zero device, format device list * ren ggml-sycl.hpp -> ggml-sycl.h * clear CMAKE to rm unused lib and options * correct queue: rm dtct:get_queue * add print tensor function to debug * fix error: wrong result in 658746bb26702e50f2c59c0e4ada8e9da6010481 * summary dpct definition in one header file to replace folder:dpct * refactor device log * mv dpct definition from folder dpct to ggml-sycl.h * update readme, refactor build script * fix build with sycl * set nthread=1 when sycl, increase performance * add run script, comment debug code * add ls-sycl-device tool * add ls-sycl-device, rm unused files * rm rear space * dos2unix * Update README_sycl.md * fix return type * remove sycl version from include path * restore rm code to fix hang issue * add syc and link for sycl readme * rm original sycl code before refactor * fix code err * add know issue for pvc hang issue * enable SYCL_F16 support * align pr4766 * check for sycl blas, better performance * cleanup 1 * remove extra endif * add build&run script, clean CMakefile, update guide by review comments * rename macro to intel hardware * editor config format * format fixes * format fixes * editor format fix * Remove unused headers * skip build sycl tool for other code path * replace tab by space * fix blas matmul function * fix mac build * restore hip dependency * fix conflict * ren as review comments * mv internal function to .cpp file * export funciton print_sycl_devices(), mv class dpct definition to source file * update CI/action for sycl code, fix CI error of repeat/dup * fix action ID format issue * rm unused strategy * enable llama_f16 in ci * fix conflict * fix build break on MacOS, due to CI of MacOS depend on external ggml, instead of internal ggml * fix ci cases for unsupported data type * revert unrelated changed in cuda cmake remove useless nommq fix typo of GGML_USE_CLBLAS_SYCL * revert hip cmake changes * fix indent * add prefix in func name * revert no mmq * rm cpu blas duplicate * fix no_new_line * fix src1->type==F16 bug. * pass batch offset for F16 src1 * fix batch error * fix wrong code * revert sycl checking in test-sampling * pass void as arguments of ggml_backend_sycl_print_sycl_devices * remove extra blank line in test-sampling * revert setting n_threads in sycl * implement std::isinf for icpx with fast math. * Update ci/run.sh Co-authored-by: Georgi Gerganov * Update examples/sycl/run-llama2.sh Co-authored-by: Georgi Gerganov * Update examples/sycl/run-llama2.sh Co-authored-by: Georgi Gerganov * Update CMakeLists.txt Co-authored-by: Georgi Gerganov * Update CMakeLists.txt Co-authored-by: Georgi Gerganov * Update CMakeLists.txt Co-authored-by: Georgi Gerganov * Update CMakeLists.txt Co-authored-by: Georgi Gerganov * add copyright and MIT license declare * update the cmd example --------- Co-authored-by: jianyuzh Co-authored-by: luoyu-intel Co-authored-by: Meng, Hengyu Co-authored-by: Georgi Gerganov --- .github/workflows/build.yml | 41 + CMakeLists.txt | 46 +- README.md | 11 +- README_sycl.md | 252 + ci/README.md | 4 + ci/run.sh | 11 + common/common.cpp | 26 +- examples/CMakeLists.txt | 3 + examples/server/server.cpp | 6 +- examples/sycl/CMakeLists.txt | 9 + examples/sycl/README.md | 47 + examples/sycl/build.sh | 20 + examples/sycl/ls-sycl-device.cpp | 11 + examples/sycl/run-llama2.sh | 19 + ggml-backend.c | 5 + ggml-sycl.cpp | 15197 +++++++++++++++++++++++++++++ ggml-sycl.h | 27 + ggml.c | 22 +- ggml.h | 1 + llama.cpp | 16 + llama.h | 5 +- tests/test-backend-ops.cpp | 9 +- 22 files changed, 15764 insertions(+), 24 deletions(-) create mode 100644 README_sycl.md create mode 100644 examples/sycl/CMakeLists.txt create mode 100644 examples/sycl/README.md create mode 100755 examples/sycl/build.sh create mode 100644 examples/sycl/ls-sycl-device.cpp create mode 100755 examples/sycl/run-llama2.sh create mode 100644 ggml-sycl.cpp create mode 100644 ggml-sycl.h diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index d22a041a6..e5e435a70 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -143,6 +143,47 @@ jobs: cd build ctest -L main --verbose + ubuntu-22-cmake-sycl: + runs-on: ubuntu-22.04 + + continue-on-error: true + + steps: + - uses: actions/checkout@v2 + + - name: add oneAPI to apt + shell: bash + run: | + cd /tmp + wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB + sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB + rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB + sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main" + + - name: install oneAPI dpcpp compiler + shell: bash + run: | + sudo apt update + sudo apt install intel-oneapi-compiler-dpcpp-cpp + + - name: install oneAPI MKL library + shell: bash + run: | + sudo apt install intel-oneapi-mkl-devel + + - name: Clone + id: checkout + uses: actions/checkout@v3 + + - name: Build + id: cmake_build + run: | + source /opt/intel/oneapi/setvars.sh + mkdir build + cd build + cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx .. + cmake --build . --config Release -j $(nproc) + # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know # how to debug it. # ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124 diff --git a/CMakeLists.txt b/CMakeLists.txt index 2b2ae532e..8b530e3ed 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -1,5 +1,6 @@ cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories. project("llama.cpp" C CXX) +include(CheckIncludeFileCXX) set(CMAKE_EXPORT_COMPILE_COMMANDS ON) @@ -103,6 +104,8 @@ option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" option(LLAMA_METAL_SHADER_DEBUG "llama: compile Metal with -fno-fast-math" OFF) option(LLAMA_MPI "llama: use MPI" OFF) option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF) +option(LLAMA_SYCL "llama: use SYCL" OFF) +option(LLAMA_SYCL_F16 "llama: use 16 bit floats for sycl calculations" OFF) option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) @@ -121,8 +124,12 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake) # # Compile flags # +if (LLAMA_SYCL) + set(CMAKE_CXX_STANDARD 17) +else() + set(CMAKE_CXX_STANDARD 11) +endif() -set(CMAKE_CXX_STANDARD 11) set(CMAKE_CXX_STANDARD_REQUIRED true) set(CMAKE_C_STANDARD 11) set(CMAKE_C_STANDARD_REQUIRED true) @@ -454,6 +461,32 @@ if (LLAMA_HIPBLAS) endif() endif() + +if (LLAMA_SYCL) + if ( NOT DEFINED ENV{ONEAPI_ROOT}) + message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh") + endif() + #todo: AOT + + find_package(IntelSYCL REQUIRED) + if (LLAMA_SYCL_F16) + add_compile_definitions(GGML_SYCL_F16) + endif() + add_compile_definitions(GGML_USE_SYCL) + + add_compile_options(-I./) #include DPCT + add_compile_options(-I/${SYCL_INCLUDE_DIR}) + + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing") + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3") + set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib") + + set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h) + set(GGML_SOURCES_SYCL ggml-sycl.cpp) + + set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread) +endif() + function(get_flags CCID CCVER) set(C_FLAGS "") set(CXX_FLAGS "") @@ -479,10 +512,12 @@ function(get_flags CCID CCVER) list(APPEND CXX_FLAGS -Wextra-semi) endif() elseif (CCID MATCHES "Intel") - # enable max optimization level when using Intel compiler - set(C_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) - set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) - add_link_options(-fuse-ld=lld -static-intel) + if (NOT LLAMA_SYCL) + # enable max optimization level when using Intel compiler + set(C_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector) + add_link_options(-fuse-ld=lld -static-intel) + endif() endif() set(GF_C_FLAGS ${C_FLAGS} PARENT_SCOPE) @@ -799,6 +834,7 @@ add_library(ggml OBJECT ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL} ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI} ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA} + ${GGML_SOURCES_SYCL} ${GGML_HEADERS_SYCL} ) target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES}) diff --git a/README.md b/README.md index 44898d2f2..ecad95e58 100644 --- a/README.md +++ b/README.md @@ -63,7 +63,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant - AVX, AVX2 and AVX512 support for x86 architectures - Mixed F16 / F32 precision - 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support -- CUDA, Metal and OpenCL GPU backend support +- CUDA, Metal, OpenCL, SYCL GPU backend support The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022). Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves @@ -599,6 +599,15 @@ Building the program with BLAS support may lead to some performance improvements You can get a list of platforms and devices from the `clinfo -l` command, etc. +- #### SYCL + + SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators. + + llama.cpp based on SYCL is used to support Intel GPU (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU). + + For detailed info, please refer to [llama.cpp for SYCL](README_sycl.md). + + ### Prepare Data & Run ```bash diff --git a/README_sycl.md b/README_sycl.md new file mode 100644 index 000000000..d5a1818f5 --- /dev/null +++ b/README_sycl.md @@ -0,0 +1,252 @@ +# llama.cpp for SYCL + +[Background](#background) + +[OS](#os) + +[Intel GPU](#intel-gpu) + +[Linux](#linux) + +[Environment Variable](#environment-variable) + +[Known Issue](#known-issue) + +[Todo](#todo) + +## Background + +SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17. + +oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms. + +Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs. + +To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL. + +The llama.cpp for SYCL is used to support Intel GPUs. + +For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building). + +## OS + +|OS|Status|Verified| +|-|-|-| +|Linux|Support|Ubuntu 22.04| +|Windows|Ongoing| | + + +## Intel GPU + +|Intel GPU| Status | Verified Model| +|-|-|-| +|Intel Data Center Max Series| Support| Max 1550| +|Intel Data Center Flex Series| Support| Flex 170| +|Intel Arc Series| Support| Arc 770| +|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake| +|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7| + + +## Linux + +### Setup Environment + +1. Install Intel GPU driver. + +a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html). + +Note: for iGPU, please install the client GPU driver. + +b. Add user to group: video, render. + +``` +sudo usermod -aG render username +sudo usermod -aG video username +``` + +Note: re-login to enable it. + +c. Check + +``` +sudo apt install clinfo +sudo clinfo -l +``` + +Output (example): + +``` +Platform #0: Intel(R) OpenCL Graphics + `-- Device #0: Intel(R) Arc(TM) A770 Graphics + + +Platform #0: Intel(R) OpenCL HD Graphics + `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49] +``` + +2. Install Intel® oneAPI Base toolkit. + + +a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html). + +Recommend to install to default folder: **/opt/intel/oneapi**. + +Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder. + +b. Check + +``` +source /opt/intel/oneapi/setvars.sh + +sycl-ls +``` + +There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**. + +Output (example): +``` +[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2 [2023.16.10.0.17_160000] +[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000] +[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO [23.30.26918.50] +[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918] + +``` + +2. Build locally: + +``` +mkdir -p build +cd build +source /opt/intel/oneapi/setvars.sh + +#for FP16 +#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference + +#for FP32 +cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx + +#build example/main only +#cmake --build . --config Release --target main + +#build all binary +cmake --build . --config Release -v + +``` + +or + +``` +./examples/sycl/build.sh +``` + +Note: + +- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only. + +### Run + +1. Put model file to folder **models** + +2. Enable oneAPI running environment + +``` +source /opt/intel/oneapi/setvars.sh +``` + +3. List device ID + +Run without parameter: + +``` +./build/bin/ls-sycl-device + +or + +./build/bin/main +``` + +Check the ID in startup log, like: + +``` +found 4 SYCL devices: + Device 0: Intel(R) Arc(TM) A770 Graphics, compute capability 1.3, + max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136 + Device 1: Intel(R) FPGA Emulation Device, compute capability 1.2, + max compute_units 24, max work group size 67108864, max sub group size 64, global mem size 67065057280 + Device 2: 13th Gen Intel(R) Core(TM) i7-13700K, compute capability 3.0, + max compute_units 24, max work group size 8192, max sub group size 64, global mem size 67065057280 + Device 3: Intel(R) Arc(TM) A770 Graphics, compute capability 3.0, + max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136 + +``` + +|Attribute|Note| +|-|-| +|compute capability 1.3|Level-zero running time, recommended | +|compute capability 3.0|OpenCL running time, slower than level-zero in most cases| + +4. Set device ID and execute llama.cpp + +Set device ID = 0 by **GGML_SYCL_DEVICE=0** + +``` +GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 +``` +or run by script: + +``` +./examples/sycl/run_llama2.sh +``` + +Note: + +- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue. + + +5. Check the device ID in output + +Like： +``` +Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device +``` + + +## Environment Variable + +#### Build + +|Name|Value|Function| +|-|-|-| +|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path.
For FP32/FP16, LLAMA_SYCL=ON is mandatory.| +|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference.
For FP32, not set it.| +|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path| +|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path| + +#### Running + + +|Name|Value|Function| +|-|-|-| +|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output| +|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG| + +## Known Issue + +- Error: `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`. + + Miss to enable oneAPI running environment. + + Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`. + + +- Hang during startup + + llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block. + + Solution: add **--no-mmap**. + +## Todo + +- Support to build in Windows. + +- Support multiple cards. diff --git a/ci/README.md b/ci/README.md index 65cfe63eb..406470519 100644 --- a/ci/README.md +++ b/ci/README.md @@ -22,4 +22,8 @@ bash ./ci/run.sh ./tmp/results ./tmp/mnt # with CUDA support GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt + +# with SYCL support +source /opt/intel/oneapi/setvars.sh +GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt ``` diff --git a/ci/run.sh b/ci/run.sh index 2427e55a2..82fe247a5 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -10,6 +10,9 @@ # # with CUDA support # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt # +# # with SYCL support +# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt +# if [ -z "$2" ]; then echo "usage: $0 " @@ -40,6 +43,14 @@ if [ ! -z ${GG_BUILD_CUDA} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1" fi +if [ ! -z ${GG_BUILD_SYCL} ]; then + if [ -z ${ONEAPI_ROOT} ]; then + echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh" + exit 1 + fi + + CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON" +fi ## helpers # download a file if it does not exist or if it is outdated diff --git a/common/common.cpp b/common/common.cpp index 6b07f1197..288013676 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -42,6 +42,10 @@ #pragma warning(disable: 4244 4267) // possible loss of data #endif +#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)) +#define GGML_USE_CUBLAS_SYCL +#endif + int32_t get_num_physical_cores() { #ifdef __linux__ // enumerate the set of thread siblings, num entries is num cores @@ -599,9 +603,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { break; } params.main_gpu = std::stoi(argv[i]); -#ifndef GGML_USE_CUBLAS - fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the main GPU has no effect.\n"); -#endif // GGML_USE_CUBLAS +#ifndef GGML_USE_CUBLAS_SYCL + fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the main GPU has no effect.\n"); +#endif // GGML_USE_CUBLAS_SYCL } else if (arg == "--split-mode" || arg == "-sm") { if (++i >= argc) { invalid_param = true; @@ -618,9 +622,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { invalid_param = true; break; } -#ifndef GGML_USE_CUBLAS - fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting the split mode has no effect.\n"); -#endif // GGML_USE_CUBLAS +#ifndef GGML_USE_CUBLAS_SYCL + fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the split mode has no effect.\n"); +#endif // GGML_USE_CUBLAS_SYCL + } else if (arg == "--tensor-split" || arg == "-ts") { if (++i >= argc) { invalid_param = true; @@ -643,9 +648,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { params.tensor_split[i] = 0.0f; } } -#ifndef GGML_USE_CUBLAS - fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Setting a tensor split has no effect.\n"); -#endif // GGML_USE_CUBLAS +#ifndef GGML_USE_CUBLAS_SYCL + fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting a tensor split has no effect.\n"); +#endif // GGML_USE_CUBLAS_SYCL } else if (arg == "--no-mmap") { params.use_mmap = false; } else if (arg == "--numa") { @@ -1007,7 +1012,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n"); printf(" -mg i, --main-gpu i the GPU to use for the model (with split-mode = none),\n"); printf(" or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu); -#endif +#endif // LLAMA_SUPPORTS_GPU_OFFLOAD printf(" --verbose-prompt print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false"); printf(" --no-display-prompt don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false"); printf(" -gan N, --grp-attn-n N\n"); @@ -1514,7 +1519,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false"); fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false"); fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false"); - fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false"); fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false"); fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false"); fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false"); diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index f67d74c55..68ad89964 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -23,6 +23,9 @@ else() add_subdirectory(infill) add_subdirectory(llama-bench) add_subdirectory(llava) + if (LLAMA_SYCL) + add_subdirectory(sycl) + endif() add_subdirectory(main) add_subdirectory(tokenize) add_subdirectory(parallel) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index f58a2acaa..a48582ad9 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2099,7 +2099,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, invalid_param = true; break; } -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL) std::string arg_next = argv[i]; // split string by , and / @@ -2125,7 +2125,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } else if (arg == "--no-mul-mat-q" || arg == "-nommq") { -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL) params.mul_mat_q = false; #else LOG_WARNING("warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n", {}); @@ -2138,7 +2138,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, invalid_param = true; break; } -#ifdef GGML_USE_CUBLAS +#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL) params.main_gpu = std::stoi(argv[i]); #else LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {}); diff --git a/examples/sycl/CMakeLists.txt b/examples/sycl/CMakeLists.txt new file mode 100644 index 000000000..69cf8932e --- /dev/null +++ b/examples/sycl/CMakeLists.txt @@ -0,0 +1,9 @@ +# MIT license +# Copyright (C) 2024 Intel Corporation +# SPDX-License-Identifier: MIT + +set(TARGET ls-sycl-device) +add_executable(${TARGET} ls-sycl-device.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_17) diff --git a/examples/sycl/README.md b/examples/sycl/README.md new file mode 100644 index 000000000..b46f17f39 --- /dev/null +++ b/examples/sycl/README.md @@ -0,0 +1,47 @@ +# llama.cpp/example/sycl + +This example program provide the tools for llama.cpp for SYCL on Intel GPU. + +## Tool + +|Tool Name| Function|Status| +|-|-|-| +|ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support| + +### ls-sycl-device + +List all SYCL devices with ID, compute capability, max work group size, ect. + +1. Build the llama.cpp for SYCL for all targets. + +2. Enable oneAPI running environment + +``` +source /opt/intel/oneapi/setvars.sh +``` + +3. Execute + +``` +./build/bin/ls-sycl-device +``` + +Check the ID in startup log, like: + +``` +found 4 SYCL devices: + Device 0: Intel(R) Arc(TM) A770 Graphics, compute capability 1.3, + max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136 + Device 1: Intel(R) FPGA Emulation Device, compute capability 1.2, + max compute_units 24, max work group size 67108864, max sub group size 64, global mem size 67065057280 + Device 2: 13th Gen Intel(R) Core(TM) i7-13700K, compute capability 3.0, + max compute_units 24, max work group size 8192, max sub group size 64, global mem size 67065057280 + Device 3: Intel(R) Arc(TM) A770 Graphics, compute capability 3.0, + max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136 + +``` + +|Attribute|Note| +|-|-| +|compute capability 1.3|Level-zero running time, recommended | +|compute capability 3.0|OpenCL running time, slower than level-zero in most cases| diff --git a/examples/sycl/build.sh b/examples/sycl/build.sh new file mode 100755 index 000000000..26ad2f7da --- /dev/null +++ b/examples/sycl/build.sh @@ -0,0 +1,20 @@ + +# MIT license +# Copyright (C) 2024 Intel Corporation +# SPDX-License-Identifier: MIT + +mkdir -p build +cd build +source /opt/intel/oneapi/setvars.sh + +#for FP16 +#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference + +#for FP32 +cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx + +#build example/main only +#cmake --build . --config Release --target main + +#build all binary +cmake --build . --config Release -v diff --git a/examples/sycl/ls-sycl-device.cpp b/examples/sycl/ls-sycl-device.cpp new file mode 100644 index 000000000..42847154a --- /dev/null +++ b/examples/sycl/ls-sycl-device.cpp @@ -0,0 +1,11 @@ +/*MIT license + Copyright (C) 2024 Intel Corporation + SPDX-License-Identifier: MIT +*/ + +#include "ggml-sycl.h" + +int main(int argc, char ** argv) { + ggml_backend_sycl_print_sycl_devices(); + return 0; +} diff --git a/examples/sycl/run-llama2.sh b/examples/sycl/run-llama2.sh new file mode 100755 index 000000000..f5f4c1e98 --- /dev/null +++ b/examples/sycl/run-llama2.sh @@ -0,0 +1,19 @@ +#!/bin/bash + +# MIT license +# Copyright (C) 2024 Intel Corporation +# SPDX-License-Identifier: MIT + +INPUT2="Building a website can be done in 10 simple steps:\nStep 1:" +source /opt/intel/oneapi/setvars.sh + +if [ $# -gt 0 ]; then + export GGML_SYCL_DEVICE=$1 +else + export GGML_SYCL_DEVICE=0 +fi +echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE +#export GGML_SYCL_DEBUG=1 +./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 +#./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 5 -e -ngl 33 -t 1 -s 0 + diff --git a/ggml-backend.c b/ggml-backend.c index 3fff5fc87..897a4cb5c 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -339,6 +339,11 @@ GGML_CALL static void ggml_backend_registry_init(void) { ggml_backend_cuda_reg_devices(); #endif +#ifdef GGML_USE_SYCL + extern void ggml_backend_sycl_reg_devices(void); + ggml_backend_sycl_reg_devices(); +#endif + #ifdef GGML_USE_METAL extern GGML_CALL ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); extern GGML_CALL ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void); diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp new file mode 100644 index 000000000..9764d9c35 --- /dev/null +++ b/ggml-sycl.cpp @@ -0,0 +1,15197 @@ +/*MIT license + Copyright (C) 2024 Intel Corporation + SPDX-License-Identifier: MIT +*/ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include + + +#include +#include + +#include "ggml-sycl.h" +#include "ggml.h" +#include "ggml-backend-impl.h" + +/* +Following definition copied from DPCT head files, which are used by ggml-sycl.cpp +*/ +// COPY from DPCT head files +#include +#include +#include + +#if defined(__linux__) +#include +#elif defined(_WIN64) +#ifndef NOMINMAX +#define NOMINMAX +#endif +#include +#else +#error "Only support Windows and Linux." +#endif + +#if defined(__linux__) +#include +#include +#endif +#if defined(_WIN64) +#ifndef NOMINMAX +#define NOMINMAX +#endif +#include +#endif + +#define DPCT_COMPATIBILITY_TEMP (900) + +#if defined(_MSC_VER) +#define __dpct_align__(n) __declspec(align(n)) +#define __dpct_inline__ __forceinline +#else +#define __dpct_align__(n) __attribute__((aligned(n))) +#define __dpct_inline__ __inline__ __attribute__((always_inline)) +#endif + +#if defined(_MSC_VER) +#define __dpct_noinline__ __declspec(noinline) +#else +#define __dpct_noinline__ __attribute__((noinline)) +#endif + +namespace dpct +{ + typedef sycl::queue *queue_ptr; + typedef sycl::event *event_ptr; + typedef char *device_ptr; + typedef uint8_t byte_t; + typedef sycl::buffer buffer_t; + + /// SYCL default exception handler + inline auto exception_handler = [](sycl::exception_list exceptions) + { + for (std::exception_ptr const &e : exceptions) + { + try + { + std::rethrow_exception(e); + } + catch (sycl::exception const &e) + { + std::cerr << "Caught asynchronous SYCL exception:" << std::endl + << e.what() << std::endl + << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + } + } + }; + + enum error_code + { + success = 0, + default_error = 999 + }; + + enum memcpy_direction + { + host_to_host, + host_to_device, + device_to_host, + device_to_device, + automatic + }; + + enum memory_region + { + global = 0, // device global memory + constant, // device constant memory + local, // device local memory + shared, // memory which can be accessed by host and device + }; + + enum class library_data_t : unsigned char + { + real_float = 0, + complex_float, + real_double, + complex_double, + real_half, + complex_half, + real_bfloat16, + complex_bfloat16, + real_int4, + complex_int4, + real_uint4, + complex_uint4, + real_int8, + complex_int8, + real_uint8, + complex_uint8, + real_int16, + complex_int16, + real_uint16, + complex_uint16, + real_int32, + complex_int32, + real_uint32, + complex_uint32, + real_int64, + complex_int64, + real_uint64, + complex_uint64, + real_int8_4, + real_int8_32, + real_uint8_4, + library_data_t_size + }; + + template + struct DataType + { + using T2 = T; + }; + template + struct DataType> + { + using T2 = std::complex; + }; + + static void destroy_event(event_ptr event) + { + delete event; + } + + static inline unsigned int get_tid() + { +#if defined(__linux__) + return syscall(SYS_gettid); +#elif defined(_WIN64) + return GetCurrentThreadId(); +#else +#error "Only support Windows and Linux." +#endif + } + + namespace detail + { + static void get_version(const sycl::device &dev, int &major, int &minor) + { + // Version string has the following format: + // a. OpenCL + // b. + std::string ver; + ver = dev.get_info(); + std::string::size_type i = 0; + while (i < ver.size()) + { + if (isdigit(ver[i])) + break; + i++; + } + major = std::stoi(&(ver[i])); + while (i < ver.size()) + { + if (ver[i] == '.') + break; + i++; + } + i++; + minor = std::stoi(&(ver[i])); + } + + template + class generic_error_type + { + public: + generic_error_type() = default; + generic_error_type(T value) : value{value} {} + operator T() const { return value; } + + private: + T value; + }; + + } // namespace detail + + /// Pitched 2D/3D memory data. + class pitched_data + { + public: + pitched_data() : pitched_data(nullptr, 0, 0, 0) {} + pitched_data(void *data, size_t pitch, size_t x, size_t y) + : _data(data), _pitch(pitch), _x(x), _y(y) {} + + void *get_data_ptr() { return _data; } + void set_data_ptr(void *data) { _data = data; } + + size_t get_pitch() { return _pitch; } + void set_pitch(size_t pitch) { _pitch = pitch; } + + size_t get_x() { return _x; } + void set_x(size_t x) { _x = x; }; + + size_t get_y() { return _y; } + void set_y(size_t y) { _y = y; } + + private: + void *_data; + size_t _pitch, _x, _y; + }; + + class device_info + { + public: + // get interface + const char *get_name() const { return _name; } + char *get_name() { return _name; } + template , + std::enable_if_t> || + std::is_same_v, + int> = 0> + auto get_max_work_item_sizes() const + { + if constexpr (std::is_same_v>) + return sycl::range<3>(_max_work_item_sizes_i[0], + _max_work_item_sizes_i[1], + _max_work_item_sizes_i[2]); + else + { + return _max_work_item_sizes_i; + } + } + template , + std::enable_if_t> || + std::is_same_v, + int> = 0> + auto get_max_work_item_sizes() + { + if constexpr (std::is_same_v>) + return sycl::range<3>(_max_work_item_sizes_i[0], + _max_work_item_sizes_i[1], + _max_work_item_sizes_i[2]); + else + { + return _max_work_item_sizes_i; + } + } + bool get_host_unified_memory() const { return _host_unified_memory; } + int get_major_version() const { return _major; } + int get_minor_version() const { return _minor; } + int get_integrated() const { return _integrated; } + int get_max_clock_frequency() const { return _frequency; } + int get_max_compute_units() const { return _max_compute_units; } + int get_max_work_group_size() const { return _max_work_group_size; } + int get_max_sub_group_size() const { return _max_sub_group_size; } + int get_max_work_items_per_compute_unit() const + { + return _max_work_items_per_compute_unit; + } + int get_max_register_size_per_work_group() const + { + return _max_register_size_per_work_group; + } + template || + std::is_same_v, + int> = 0> + auto get_max_nd_range_size() const + { + if constexpr (std::is_same_v) + return _max_nd_range_size; + else + return _max_nd_range_size_i; + } + template || + std::is_same_v, + int> = 0> + auto get_max_nd_range_size() + { + if constexpr (std::is_same_v) + return _max_nd_range_size; + else + return _max_nd_range_size_i; + } + size_t get_global_mem_size() const { return _global_mem_size; } + size_t get_local_mem_size() const { return _local_mem_size; } + /// Returns the maximum clock rate of device's global memory in kHz. If + /// compiler does not support this API then returns default value 3200000 kHz. + unsigned int get_memory_clock_rate() const { return _memory_clock_rate; } + /// Returns the maximum bus width between device and memory in bits. If + /// compiler does not support this API then returns default value 64 bits. + unsigned int get_memory_bus_width() const { return _memory_bus_width; } + uint32_t get_device_id() const { return _device_id; } + std::array get_uuid() const { return _uuid; } + /// Returns global memory cache size in bytes. + unsigned int get_global_mem_cache_size() const + { + return _global_mem_cache_size; + } + + // set interface + void set_name(const char *name) + { + size_t length = strlen(name); + if (length < 256) + { + std::memcpy(_name, name, length + 1); + } + else + { + std::memcpy(_name, name, 255); + _name[255] = '\0'; + } + } + void set_max_work_item_sizes(const sycl::range<3> max_work_item_sizes) + { + for (int i = 0; i < 3; ++i) + _max_work_item_sizes_i[i] = max_work_item_sizes[i]; + } + [[deprecated]] void + set_max_work_item_sizes(const sycl::id<3> max_work_item_sizes) + { + for (int i = 0; i < 3; ++i) + { + _max_work_item_sizes_i[i] = max_work_item_sizes[i]; + } + } + void set_host_unified_memory(bool host_unified_memory) + { + _host_unified_memory = host_unified_memory; + } + void set_major_version(int major) { _major = major; } + void set_minor_version(int minor) { _minor = minor; } + void set_integrated(int integrated) { _integrated = integrated; } + void set_max_clock_frequency(int frequency) { _frequency = frequency; } + void set_max_compute_units(int max_compute_units) + { + _max_compute_units = max_compute_units; + } + void set_global_mem_size(size_t global_mem_size) + { + _global_mem_size = global_mem_size; + } + void set_local_mem_size(size_t local_mem_size) + { + _local_mem_size = local_mem_size; + } + void set_max_work_group_size(int max_work_group_size) + { + _max_work_group_size = max_work_group_size; + } + void set_max_sub_group_size(int max_sub_group_size) + { + _max_sub_group_size = max_sub_group_size; + } + void + set_max_work_items_per_compute_unit(int max_work_items_per_compute_unit) + { + _max_work_items_per_compute_unit = max_work_items_per_compute_unit; + } + void set_max_nd_range_size(int max_nd_range_size[]) + { + for (int i = 0; i < 3; i++) + { + _max_nd_range_size[i] = max_nd_range_size[i]; + _max_nd_range_size_i[i] = max_nd_range_size[i]; + } + } + void set_memory_clock_rate(unsigned int memory_clock_rate) + { + _memory_clock_rate = memory_clock_rate; + } + void set_memory_bus_width(unsigned int memory_bus_width) + { + _memory_bus_width = memory_bus_width; + } + void + set_max_register_size_per_work_group(int max_register_size_per_work_group) + { + _max_register_size_per_work_group = max_register_size_per_work_group; + } + void set_device_id(uint32_t device_id) + { + _device_id = device_id; + } + void set_uuid(std::array uuid) + { + _uuid = std::move(uuid); + } + void set_global_mem_cache_size(unsigned int global_mem_cache_size) + { + _global_mem_cache_size = global_mem_cache_size; + } + + private: + char _name[256]; + int _max_work_item_sizes_i[3]; + bool _host_unified_memory = false; + int _major; + int _minor; + int _integrated = 0; + int _frequency; + // Set estimated value 3200000 kHz as default value. + unsigned int _memory_clock_rate = 3200000; + // Set estimated value 64 bits as default value. + unsigned int _memory_bus_width = 64; + unsigned int _global_mem_cache_size; + int _max_compute_units; + int _max_work_group_size; + int _max_sub_group_size; + int _max_work_items_per_compute_unit; + int _max_register_size_per_work_group; + size_t _global_mem_size; + size_t _local_mem_size; + size_t _max_nd_range_size[3]; + int _max_nd_range_size_i[3]; + uint32_t _device_id; + std::array _uuid; + }; + + static int get_major_version(const sycl::device &dev) + { + int major, minor; + detail::get_version(dev, major, minor); + return major; + } + + static int get_minor_version(const sycl::device &dev) + { + int major, minor; + detail::get_version(dev, major, minor); + return minor; + } + + static void get_device_info(device_info &out, const sycl::device &dev) + { + device_info prop; + prop.set_name(dev.get_info().c_str()); + + int major, minor; + detail::get_version(dev, major, minor); + prop.set_major_version(major); + prop.set_minor_version(minor); + + prop.set_max_work_item_sizes( +#if (__SYCL_COMPILER_VERSION && __SYCL_COMPILER_VERSION < 20220902) + // oneAPI DPC++ compiler older than 2022/09/02, where max_work_item_sizes + // is an enum class element + dev.get_info()); +#else + // SYCL 2020-conformant code, max_work_item_sizes is a struct templated by + // an int + dev.get_info>()); +#endif + prop.set_host_unified_memory(dev.has(sycl::aspect::usm_host_allocations)); + + prop.set_max_clock_frequency( + dev.get_info() * 1000); + + prop.set_max_compute_units( + dev.get_info()); + prop.set_max_work_group_size( + dev.get_info()); + prop.set_global_mem_size(dev.get_info()); + prop.set_local_mem_size(dev.get_info()); + +#if (defined(SYCL_EXT_INTEL_DEVICE_INFO) && SYCL_EXT_INTEL_DEVICE_INFO >= 6) + if (dev.has(sycl::aspect::ext_intel_memory_clock_rate)) + { + unsigned int tmp = + dev.get_info(); + if (tmp != 0) + prop.set_memory_clock_rate(1000 * tmp); + } + if (dev.has(sycl::aspect::ext_intel_memory_bus_width)) + { + prop.set_memory_bus_width( + dev.get_info()); + } + if (dev.has(sycl::aspect::ext_intel_device_id)) + { + prop.set_device_id( + dev.get_info()); + } + if (dev.has(sycl::aspect::ext_intel_device_info_uuid)) + { + prop.set_uuid(dev.get_info()); + } +#elif defined(_MSC_VER) && !defined(__clang__) +#pragma message("get_device_info: querying memory_clock_rate and \ + memory_bus_width are not supported by the compiler used. \ + Use 3200000 kHz as memory_clock_rate default value. \ + Use 64 bits as memory_bus_width default value.") +#else +#warning "get_device_info: querying memory_clock_rate and \ + memory_bus_width are not supported by the compiler used. \ + Use 3200000 kHz as memory_clock_rate default value. \ + Use 64 bits as memory_bus_width default value." +#endif + + size_t max_sub_group_size = 1; + std::vector sub_group_sizes = + dev.get_info(); + + for (const auto &sub_group_size : sub_group_sizes) + { + if (max_sub_group_size < sub_group_size) + max_sub_group_size = sub_group_size; + } + + prop.set_max_sub_group_size(max_sub_group_size); + + prop.set_max_work_items_per_compute_unit( + dev.get_info()); + int max_nd_range_size[] = {0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF}; + prop.set_max_nd_range_size(max_nd_range_size); + + // Estimates max register size per work group, feel free to update the value + // according to device properties. + prop.set_max_register_size_per_work_group(65536); + + prop.set_global_mem_cache_size( + dev.get_info()); + out = prop; + } + + /// dpct device extension + class device_ext : public sycl::device + { + typedef std::mutex mutex_type; + + public: + device_ext() : sycl::device(), _ctx(*this) {} + ~device_ext() + { + std::lock_guard lock(m_mutex); + clear_queues(); + } + device_ext(const sycl::device &base) : sycl::device(base), _ctx(*this) + { + std::lock_guard lock(m_mutex); + init_queues(); + } + + int is_native_atomic_supported() { return 0; } + int get_major_version() const + { + return dpct::get_major_version(*this); + } + + int get_minor_version() const + { + return dpct::get_minor_version(*this); + } + + int get_max_compute_units() const + { + return get_device_info().get_max_compute_units(); + } + + /// Return the maximum clock frequency of this device in KHz. + int get_max_clock_frequency() const + { + return get_device_info().get_max_clock_frequency(); + } + + int get_integrated() const { return get_device_info().get_integrated(); } + + int get_max_sub_group_size() const + { + return get_device_info().get_max_sub_group_size(); + } + + int get_max_register_size_per_work_group() const + { + return get_device_info().get_max_register_size_per_work_group(); + } + + int get_max_work_group_size() const + { + return get_device_info().get_max_work_group_size(); + } + + int get_mem_base_addr_align() const + { + return get_info(); + } + + size_t get_global_mem_size() const + { + return get_device_info().get_global_mem_size(); + } + + /// Get the number of bytes of free and total memory on the SYCL device. + /// \param [out] free_memory The number of bytes of free memory on the SYCL device. + /// \param [out] total_memory The number of bytes of total memory on the SYCL device. + void get_memory_info(size_t &free_memory, size_t &total_memory) + { +#if (defined(__SYCL_COMPILER_VERSION) && __SYCL_COMPILER_VERSION >= 20221105) + if (!has(sycl::aspect::ext_intel_free_memory)) + { + std::cerr << "get_memory_info: ext_intel_free_memory is not supported." << std::endl; + free_memory = 0; + } + else + { + free_memory = get_info(); + } +#else + std::cerr << "get_memory_info: ext_intel_free_memory is not supported." << std::endl; + free_memory = 0; +#if defined(_MSC_VER) && !defined(__clang__) +#pragma message("Querying the number of bytes of free memory is not supported") +#else +#warning "Querying the number of bytes of free memory is not supported" +#endif +#endif + total_memory = get_device_info().get_global_mem_size(); + } + + void get_device_info(device_info &out) const + { + dpct::get_device_info(out, *this); + } + + device_info get_device_info() const + { + device_info prop; + dpct::get_device_info(prop, *this); + return prop; + } + + void reset() + { + std::lock_guard lock(m_mutex); + clear_queues(); + init_queues(); + } + + sycl::queue &in_order_queue() { return *_q_in_order; } + + sycl::queue &out_of_order_queue() { return *_q_out_of_order; } + + sycl::queue &default_queue() + { +#ifdef DPCT_USM_LEVEL_NONE + return out_of_order_queue(); +#else + return in_order_queue(); +#endif // DPCT_USM_LEVEL_NONE + } + + void queues_wait_and_throw() + { + std::unique_lock lock(m_mutex); + std::vector> current_queues( + _queues); + lock.unlock(); + for (const auto &q : current_queues) + { + q->wait_and_throw(); + } + // Guard the destruct of current_queues to make sure the ref count is safe. + lock.lock(); + } + + sycl::queue *create_queue(bool enable_exception_handler = false) + { +#ifdef DPCT_USM_LEVEL_NONE + return create_out_of_order_queue(enable_exception_handler); +#else + return create_in_order_queue(enable_exception_handler); +#endif // DPCT_USM_LEVEL_NONE + } + + sycl::queue *create_in_order_queue(bool enable_exception_handler = false) + { + std::lock_guard lock(m_mutex); + return create_queue_impl(enable_exception_handler, + sycl::property::queue::in_order()); + } + + sycl::queue *create_out_of_order_queue(bool enable_exception_handler = false) + { + std::lock_guard lock(m_mutex); + return create_queue_impl(enable_exception_handler); + } + + void destroy_queue(sycl::queue *&queue) + { + std::lock_guard lock(m_mutex); + _queues.erase(std::remove_if(_queues.begin(), _queues.end(), + [=](const std::shared_ptr &q) -> bool + { + return q.get() == queue; + }), + _queues.end()); + queue = nullptr; + } + void set_saved_queue(sycl::queue *q) + { + std::lock_guard lock(m_mutex); + _saved_queue = q; + } + sycl::queue *get_saved_queue() const + { + std::lock_guard lock(m_mutex); + return _saved_queue; + } + sycl::context get_context() const { return _ctx; } + + private: + void clear_queues() + { + _queues.clear(); + _q_in_order = _q_out_of_order = _saved_queue = nullptr; + } + + void init_queues() + { + _q_in_order = create_queue_impl(true, sycl::property::queue::in_order()); + _q_out_of_order = create_queue_impl(true); + _saved_queue = &default_queue(); + } + + /// Caller should acquire resource \p m_mutex before calling this function. + template + sycl::queue *create_queue_impl(bool enable_exception_handler, + Properties... properties) + { + sycl::async_handler eh = {}; + if (enable_exception_handler) + { + eh = exception_handler; + } + _queues.push_back(std::make_shared( + _ctx, *this, eh, + sycl::property_list( +#ifdef DPCT_PROFILING_ENABLED + sycl::property::queue::enable_profiling(), +#endif + properties...))); + + return _queues.back().get(); + } + + void get_version(int &major, int &minor) const + { + detail::get_version(*this, major, minor); + } + sycl::queue *_q_in_order, *_q_out_of_order; + sycl::queue *_saved_queue; + sycl::context _ctx; + std::vector> _queues; + mutable mutex_type m_mutex; + }; + + /// device manager + class dev_mgr + { + public: + device_ext ¤t_device() + { + unsigned int dev_id = current_device_id(); + check_id(dev_id); + return *_devs[dev_id]; + } + device_ext &cpu_device() const + { + std::lock_guard lock(m_mutex); + if (_cpu_device == -1) + { + throw std::runtime_error("no valid cpu device"); + } + else + { + return *_devs[_cpu_device]; + } + } + device_ext &get_device(unsigned int id) const + { + std::lock_guard lock(m_mutex); + check_id(id); + return *_devs[id]; + } + unsigned int current_device_id() const + { + std::lock_guard lock(m_mutex); + auto it = _thread2dev_map.find(get_tid()); + if (it != _thread2dev_map.end()) + return it->second; + return DEFAULT_DEVICE_ID; + } + + /// Select device with a device ID. + /// \param [in] id The id of the device which can + /// be obtained through get_device_id(const sycl::device). + void select_device(unsigned int id) + { + std::lock_guard lock(m_mutex); + check_id(id); + _thread2dev_map[get_tid()] = id; + } + unsigned int device_count() { return _devs.size(); } + + unsigned int get_device_id(const sycl::device &dev) + { + unsigned int id = 0; + for (auto dev_item : _devs) + { + if (*dev_item == dev) + { + break; + } + id++; + } + return id; + } + + template + std::enable_if_t< + std::is_invocable_r_v> + select_device(const DeviceSelector &selector = sycl::gpu_selector_v) + { + sycl::device selected_device = sycl::device(selector); + unsigned int selected_device_id = get_device_id(selected_device); + select_device(selected_device_id); + } + + /// Returns the instance of device manager singleton. + static dev_mgr &instance() + { + static dev_mgr d_m; + return d_m; + } + dev_mgr(const dev_mgr &) = delete; + dev_mgr &operator=(const dev_mgr &) = delete; + dev_mgr(dev_mgr &&) = delete; + dev_mgr &operator=(dev_mgr &&) = delete; + + private: + mutable std::recursive_mutex m_mutex; + dev_mgr() + { + sycl::device default_device = + sycl::device(sycl::default_selector_v); + _devs.push_back(std::make_shared(default_device)); + + std::vector sycl_all_devs = + sycl::device::get_devices(sycl::info::device_type::all); + // Collect other devices except for the default device. + if (default_device.is_cpu()) + _cpu_device = 0; + for (auto &dev : sycl_all_devs) + { + if (dev == default_device) + { + continue; + } + _devs.push_back(std::make_shared(dev)); + if (_cpu_device == -1 && dev.is_cpu()) + { + _cpu_device = _devs.size() - 1; + } + } + } + void check_id(unsigned int id) const + { + if (id >= _devs.size()) + { + throw std::runtime_error("invalid device id"); + } + } + std::vector> _devs; + /// DEFAULT_DEVICE_ID is used, if current_device_id() can not find current + /// thread id in _thread2dev_map, which means default device should be used + /// for the current thread. + const unsigned int DEFAULT_DEVICE_ID = 0; + /// thread-id to device-id map. + std::map _thread2dev_map; + int _cpu_device = -1; + }; + + static inline sycl::queue &get_default_queue() + { + return dev_mgr::instance().current_device().default_queue(); + } + + namespace detail + { + enum class pointer_access_attribute + { + host_only = 0, + device_only, + host_device, + end + }; + + static pointer_access_attribute get_pointer_attribute(sycl::queue &q, + const void *ptr) + { +#ifdef DPCT_USM_LEVEL_NONE + return mem_mgr::instance().is_device_ptr(ptr) + ? pointer_access_attribute::device_only + : pointer_access_attribute::host_only; +#else + switch (sycl::get_pointer_type(ptr, q.get_context())) + { + case sycl::usm::alloc::unknown: + return pointer_access_attribute::host_only; + case sycl::usm::alloc::device: + return pointer_access_attribute::device_only; + case sycl::usm::alloc::shared: + case sycl::usm::alloc::host: + return pointer_access_attribute::host_device; + } +#endif + } + + template + inline constexpr std::uint64_t get_type_combination_id(ArgT Val) + { + static_assert((unsigned char)library_data_t::library_data_t_size <= + std::numeric_limits::max() && + "library_data_t size exceeds limit."); + static_assert(std::is_same_v, "Unsupported ArgT"); + return (std::uint64_t)Val; + } + + template + inline constexpr std::uint64_t get_type_combination_id(FirstT FirstVal, + RestT... RestVal) + { + static_assert((std::uint8_t)library_data_t::library_data_t_size <= + std::numeric_limits::max() && + "library_data_t size exceeds limit."); + static_assert(sizeof...(RestT) <= 8 && "Too many parameters"); + static_assert(std::is_same_v, "Unsupported FirstT"); + return get_type_combination_id(RestVal...) << 8 | ((std::uint64_t)FirstVal); + } + + class mem_mgr + { + mem_mgr() + { + // Reserved address space, no real memory allocation happens here. +#if defined(__linux__) + mapped_address_space = + (byte_t *)mmap(nullptr, mapped_region_size, PROT_NONE, + MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); +#elif defined(_WIN64) + mapped_address_space = (byte_t *)VirtualAlloc( + NULL, // NULL specified as the base address parameter + mapped_region_size, // Size of allocation + MEM_RESERVE, // Allocate reserved pages + PAGE_NOACCESS); // Protection = no access +#else +#error "Only support Windows and Linux." +#endif + next_free = mapped_address_space; + }; + + public: + using buffer_id_t = int; + + struct allocation + { + buffer_t buffer; + byte_t *alloc_ptr; + size_t size; + }; + + ~mem_mgr() + { +#if defined(__linux__) + munmap(mapped_address_space, mapped_region_size); +#elif defined(_WIN64) + VirtualFree(mapped_address_space, 0, MEM_RELEASE); +#else +#error "Only support Windows and Linux." +#endif + }; + + mem_mgr(const mem_mgr &) = delete; + mem_mgr &operator=(const mem_mgr &) = delete; + mem_mgr(mem_mgr &&) = delete; + mem_mgr &operator=(mem_mgr &&) = delete; + + /// Allocate + void *mem_alloc(size_t size) + { + if (!size) + return nullptr; + std::lock_guard lock(m_mutex); + if (next_free + size > mapped_address_space + mapped_region_size) + { + throw std::runtime_error("dpct_malloc: out of memory for virtual memory pool"); + } + // Allocation + sycl::range<1> r(size); + buffer_t buf(r); + allocation A{buf, next_free, size}; + // Map allocation to device pointer + void *result = next_free; + m_map.emplace(next_free + size, A); + // Update pointer to the next free space. + next_free += (size + extra_padding + alignment - 1) & ~(alignment - 1); + + return result; + } + + /// Deallocate + void mem_free(const void *ptr) + { + if (!ptr) + return; + std::lock_guard lock(m_mutex); + auto it = get_map_iterator(ptr); + m_map.erase(it); + } + + /// map: device pointer -> allocation(buffer, alloc_ptr, size) + allocation translate_ptr(const void *ptr) + { + std::lock_guard lock(m_mutex); + auto it = get_map_iterator(ptr); + return it->second; + } + + /// Check if the pointer represents device pointer or not. + bool is_device_ptr(const void *ptr) const + { + std::lock_guard lock(m_mutex); + return (mapped_address_space <= ptr) && + (ptr < mapped_address_space + mapped_region_size); + } + + /// Returns the instance of memory manager singleton. + static mem_mgr &instance() + { + static mem_mgr m; + return m; + } + + private: + std::map m_map; + mutable std::mutex m_mutex; + byte_t *mapped_address_space; + byte_t *next_free; + const size_t mapped_region_size = 128ull * 1024 * 1024 * 1024; + const size_t alignment = 256; + /// This padding may be defined to some positive value to debug + /// out of bound accesses. + const size_t extra_padding = 0; + + std::map::iterator get_map_iterator(const void *ptr) + { + auto it = m_map.upper_bound((byte_t *)ptr); + if (it == m_map.end()) + { + // Not a virtual pointer. + throw std::runtime_error("can not get buffer from non-virtual pointer"); + } + const allocation &alloc = it->second; + if (ptr < alloc.alloc_ptr) + { + // Out of bound. + // This may happen if there's a gap between allocations due to alignment + // or extra padding and pointer points to this gap. + throw std::runtime_error("invalid virtual pointer"); + } + return it; + } + }; + + template + class accessor; + template + class memory_traits + { + public: + static constexpr sycl::access::target target = + sycl::access::target::device; + static constexpr sycl::access_mode mode = + (Memory == constant) ? sycl::access_mode::read + : sycl::access_mode::read_write; + static constexpr size_t type_size = sizeof(T); + using element_t = + typename std::conditional::type; + using value_t = typename std::remove_cv::type; + template + using accessor_t = typename std::conditional< + Memory == local, sycl::local_accessor, + sycl::accessor>::type; + using pointer_t = T *; + }; + + static inline void *dpct_malloc(size_t size, sycl::queue &q) + { +#ifdef DPCT_USM_LEVEL_NONE + return mem_mgr::instance().mem_alloc(size * sizeof(byte_t)); +#else + return sycl::malloc_device(size, q.get_device(), q.get_context()); +#endif // DPCT_USM_LEVEL_NONE + } + +#define PITCH_DEFAULT_ALIGN(x) (((x) + 31) & ~(0x1F)) + static inline void *dpct_malloc(size_t &pitch, size_t x, size_t y, size_t z, + sycl::queue &q) + { + pitch = PITCH_DEFAULT_ALIGN(x); + return dpct_malloc(pitch * y * z, q); + } + + /** + * @brief Sets \p value to the first \p size elements starting from \p dev_ptr in \p q. + * @tparam valueT The type of the element to be set. + * @param [in] q The queue in which the operation is done. + * @param [in] dev_ptr Pointer to the virtual device memory address. + * @param [in] value The value to be set. + * @param [in] size Number of elements to be set to the value. + * @return An event representing the memset operation. + */ + template + static inline sycl::event dpct_memset(sycl::queue &q, void *dev_ptr, + valueT value, size_t size) + { +#ifdef DPCT_USM_LEVEL_NONE + auto &mm = mem_mgr::instance(); + assert(mm.is_device_ptr(dev_ptr)); + auto alloc = mm.translate_ptr(dev_ptr); + size_t offset = (valueT *)dev_ptr - (valueT *)alloc.alloc_ptr; + + return q.submit([&](sycl::handler &cgh) + { + auto r = sycl::range<1>(size); + auto o = sycl::id<1>(offset); + auto new_buffer = alloc.buffer.reinterpret( + sycl::range<1>(alloc.size / sizeof(valueT))); + sycl::accessor + acc(new_buffer, cgh, r, o); + cgh.fill(acc, value); }); +#else + return q.fill(dev_ptr, value, size); +#endif // DPCT_USM_LEVEL_NONE + } + + /** + * @brief Sets \p value to the 3D memory region pointed by \p data in \p q. + * @tparam valueT The type of the element to be set. + * @param [in] q The queue in which the operation is done. + * @param [in] data Pointer to the pitched device memory region. + * @param [in] value The value to be set. + * @param [in] size 3D memory region by number of elements. + * @return An event list representing the memset operations. + */ + template + static inline std::vector + dpct_memset(sycl::queue &q, pitched_data data, valueT value, + sycl::range<3> size) + { + std::vector event_list; + size_t slice = data.get_pitch() * data.get_y(); + unsigned char *data_surface = (unsigned char *)data.get_data_ptr(); + for (size_t z = 0; z < size.get(2); ++z) + { + unsigned char *data_ptr = data_surface; + for (size_t y = 0; y < size.get(1); ++y) + { + event_list.push_back(dpct_memset(q, data_ptr, value, size.get(0))); + data_ptr += data.get_pitch(); + } + data_surface += slice; + } + return event_list; + } + + /** + * @brief Sets \p val to the pitched 2D memory region pointed by \p ptr in \p q. + * @tparam valueT The type of the element to be set. + * @param [in] q The queue in which the operation is done. + * @param [in] ptr Pointer to the virtual device memory. + * @param [in] pitch The pitch size by number of elements, including padding. + * @param [in] val The value to be set. + * @param [in] x The width of memory region by number of elements. + * @param [in] y The height of memory region by number of elements. + * @return An event list representing the memset operations. + */ + template + static inline std::vector + dpct_memset(sycl::queue &q, void *ptr, size_t pitch, valueT val, size_t x, + size_t y) + { + return dpct_memset(q, pitched_data(ptr, pitch, x, 1), val, + sycl::range<3>(x, y, 1)); + } + + static memcpy_direction deduce_memcpy_direction(sycl::queue &q, void *to_ptr, + const void *from_ptr, + memcpy_direction dir) + { + switch (dir) + { + case memcpy_direction::host_to_host: + case memcpy_direction::host_to_device: + case memcpy_direction::device_to_host: + case memcpy_direction::device_to_device: + return dir; + case memcpy_direction::automatic: + { + // table[to_attribute][from_attribute] + static const memcpy_direction + direction_table[static_cast(pointer_access_attribute::end)] + [static_cast(pointer_access_attribute::end)] = + {{memcpy_direction::host_to_host, + memcpy_direction::device_to_host, + memcpy_direction::host_to_host}, + {memcpy_direction::host_to_device, + memcpy_direction::device_to_device, + memcpy_direction::device_to_device}, + {memcpy_direction::host_to_host, + memcpy_direction::device_to_device, + memcpy_direction::device_to_device}}; + return direction_table[static_cast(get_pointer_attribute( + q, to_ptr))][static_cast(get_pointer_attribute(q, from_ptr))]; + } + default: + throw std::runtime_error("dpct_memcpy: invalid direction value"); + } + } + + static sycl::event + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, size_t size, + memcpy_direction direction, + const std::vector &dep_events = {}) + { + if (!size) + return sycl::event{}; +#ifdef DPCT_USM_LEVEL_NONE + auto &mm = mem_mgr::instance(); + auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction); + + switch (real_direction) + { + case host_to_host: + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); }); + case host_to_device: + { + auto alloc = mm.translate_ptr(to_ptr); + size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto o = sycl::id<1>(offset); + sycl::accessor + acc(alloc.buffer, cgh, r, o); + cgh.copy(from_ptr, acc); }); + } + case device_to_host: + { + auto alloc = mm.translate_ptr(from_ptr); + size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto o = sycl::id<1>(offset); + sycl::accessor + acc(alloc.buffer, cgh, r, o); + cgh.copy(acc, to_ptr); }); + } + case device_to_device: + { + auto to_alloc = mm.translate_ptr(to_ptr); + auto from_alloc = mm.translate_ptr(from_ptr); + size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr; + size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto to_o = sycl::id<1>(to_offset); + auto from_o = sycl::id<1>(from_offset); + sycl::accessor + to_acc(to_alloc.buffer, cgh, r, to_o); + sycl::accessor + from_acc(from_alloc.buffer, cgh, r, from_o); + cgh.copy(from_acc, to_acc); }); + } + default: + throw std::runtime_error("dpct_memcpy: invalid direction value"); + } +#else + return q.memcpy(to_ptr, from_ptr, size, dep_events); +#endif // DPCT_USM_LEVEL_NONE + } + + // Get actual copy range and make sure it will not exceed range. + static inline size_t get_copy_range(sycl::range<3> size, size_t slice, + size_t pitch) + { + return slice * (size.get(2) - 1) + pitch * (size.get(1) - 1) + size.get(0); + } + + static inline size_t get_offset(sycl::id<3> id, size_t slice, + size_t pitch) + { + return slice * id.get(2) + pitch * id.get(1) + id.get(0); + } + + /// copy 3D matrix specified by \p size from 3D matrix specified by \p from_ptr + /// and \p from_range to another specified by \p to_ptr and \p to_range. + static inline std::vector + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, + sycl::range<3> to_range, sycl::range<3> from_range, + sycl::id<3> to_id, sycl::id<3> from_id, + sycl::range<3> size, memcpy_direction direction, + const std::vector &dep_events = {}) + { + // RAII for host pointer + class host_buffer + { + void *_buf; + size_t _size; + sycl::queue &_q; + const std::vector &_deps; // free operation depends + + public: + host_buffer(size_t size, sycl::queue &q, + const std::vector &deps) + : _buf(std::malloc(size)), _size(size), _q(q), _deps(deps) {} + void *get_ptr() const { return _buf; } + size_t get_size() const { return _size; } + ~host_buffer() + { + if (_buf) + { + _q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(_deps); + cgh.host_task([buf = _buf] { std::free(buf); }); }); + } + } + }; + std::vector event_list; + + size_t to_slice = to_range.get(1) * to_range.get(0), + from_slice = from_range.get(1) * from_range.get(0); + unsigned char *to_surface = + (unsigned char *)to_ptr + get_offset(to_id, to_slice, to_range.get(0)); + const unsigned char *from_surface = + (const unsigned char *)from_ptr + + get_offset(from_id, from_slice, from_range.get(0)); + + if (to_slice == from_slice && to_slice == size.get(1) * size.get(0)) + { + return {dpct_memcpy(q, to_surface, from_surface, to_slice * size.get(2), + direction, dep_events)}; + } + direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction); + size_t size_slice = size.get(1) * size.get(0); + switch (direction) + { + case host_to_host: + for (size_t z = 0; z < size.get(2); ++z) + { + unsigned char *to_ptr = to_surface; + const unsigned char *from_ptr = from_surface; + if (to_range.get(0) == from_range.get(0) && + to_range.get(0) == size.get(0)) + { + event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size_slice, + direction, dep_events)); + } + else + { + for (size_t y = 0; y < size.get(1); ++y) + { + event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size.get(0), + direction, dep_events)); + to_ptr += to_range.get(0); + from_ptr += from_range.get(0); + } + } + to_surface += to_slice; + from_surface += from_slice; + } + break; + case host_to_device: + { + host_buffer buf(get_copy_range(size, to_slice, to_range.get(0)), q, + event_list); + std::vector host_events; + if (to_slice == size_slice) + { + // Copy host data to a temp host buffer with the shape of target. + host_events = + dpct_memcpy(q, buf.get_ptr(), from_surface, to_range, from_range, + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, + host_to_host, dep_events); + } + else + { + // Copy host data to a temp host buffer with the shape of target. + host_events = dpct_memcpy( + q, buf.get_ptr(), from_surface, to_range, from_range, + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, host_to_host, + // If has padding data, not sure whether it is useless. So fill temp + // buffer with it. + std::vector{ + dpct_memcpy(q, buf.get_ptr(), to_surface, buf.get_size(), + device_to_host, dep_events)}); + } + // Copy from temp host buffer to device with only one submit. + event_list.push_back(dpct_memcpy(q, to_surface, buf.get_ptr(), + buf.get_size(), host_to_device, + host_events)); + break; + } + case device_to_host: + { + host_buffer buf(get_copy_range(size, from_slice, from_range.get(0)), q, + event_list); + // Copy from host temp buffer to host target with reshaping. + event_list = dpct_memcpy( + q, to_surface, buf.get_ptr(), to_range, from_range, sycl::id<3>(0, 0, 0), + sycl::id<3>(0, 0, 0), size, host_to_host, + // Copy from device to temp host buffer with only one submit. + std::vector{dpct_memcpy(q, buf.get_ptr(), from_surface, + buf.get_size(), + device_to_host, dep_events)}); + break; + } + case device_to_device: +#ifdef DPCT_USM_LEVEL_NONE + { + auto &mm = mem_mgr::instance(); + auto to_alloc = mm.translate_ptr(to_surface); + auto from_alloc = mm.translate_ptr(from_surface); + size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr; + size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr; + event_list.push_back(q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto to_o = sycl::id<1>(to_offset); + auto from_o = sycl::id<1>(from_offset); + sycl::accessor + to_acc(to_alloc.buffer, cgh, + get_copy_range(size, to_slice, to_range.get(0)), to_o); + sycl::accessor + from_acc(from_alloc.buffer, cgh, + get_copy_range(size, from_slice, from_range.get(0)), from_o); + cgh.parallel_for( + size, + [=](sycl::id<3> id) { + to_acc[get_offset(id, to_slice, to_range.get(0))] = + from_acc[get_offset(id, from_slice, from_range.get(0))]; + }); })); + } +#else + event_list.push_back(q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + cgh.parallel_for( + size, + [=](sycl::id<3> id) { + to_surface[get_offset(id, to_slice, to_range.get(0))] = + from_surface[get_offset(id, from_slice, from_range.get(0))]; + }); })); +#endif + break; + default: + throw std::runtime_error("dpct_memcpy: invalid direction value"); + } + return event_list; + } + + /// memcpy 2D/3D matrix specified by pitched_data. + static inline std::vector + dpct_memcpy(sycl::queue &q, pitched_data to, sycl::id<3> to_id, + pitched_data from, sycl::id<3> from_id, sycl::range<3> size, + memcpy_direction direction = automatic) + { + return dpct_memcpy(q, to.get_data_ptr(), from.get_data_ptr(), + sycl::range<3>(to.get_pitch(), to.get_y(), 1), + sycl::range<3>(from.get_pitch(), from.get_y(), 1), to_id, from_id, + size, direction); + } + + /// memcpy 2D matrix with pitch. + static inline std::vector + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, + size_t to_pitch, size_t from_pitch, size_t x, size_t y, + memcpy_direction direction = automatic) + { + return dpct_memcpy(q, to_ptr, from_ptr, sycl::range<3>(to_pitch, y, 1), + sycl::range<3>(from_pitch, y, 1), + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), + sycl::range<3>(x, y, 1), direction); + } + + namespace deprecated + { + + template + class usm_allocator + { + private: + using Alloc = sycl::usm_allocator; + Alloc _impl; + + public: + using value_type = typename std::allocator_traits::value_type; + using pointer = typename std::allocator_traits::pointer; + using const_pointer = typename std::allocator_traits::const_pointer; + using void_pointer = typename std::allocator_traits::void_pointer; + using const_void_pointer = + typename std::allocator_traits::const_void_pointer; + using reference = typename std::allocator_traits::value_type &; + using const_reference = + const typename std::allocator_traits::value_type &; + using difference_type = + typename std::allocator_traits::difference_type; + using size_type = typename std::allocator_traits::size_type; + using propagate_on_container_copy_assignment = typename std::allocator_traits< + Alloc>::propagate_on_container_copy_assignment; + using propagate_on_container_move_assignment = typename std::allocator_traits< + Alloc>::propagate_on_container_move_assignment; + using propagate_on_container_swap = + typename std::allocator_traits::propagate_on_container_swap; + using is_always_equal = + typename std::allocator_traits::is_always_equal; + + template + struct rebind + { + typedef usm_allocator other; + }; + + usm_allocator() : _impl(dpct::get_default_queue()) {} + ~usm_allocator() {} + usm_allocator(const usm_allocator &other) : _impl(other._impl) {} + usm_allocator(usm_allocator &&other) : _impl(std::move(other._impl)) {} + pointer address(reference r) { return &r; } + const_pointer address(const_reference r) { return &r; } + pointer allocate(size_type cnt, const_void_pointer hint = nullptr) + { + return std::allocator_traits::allocate(_impl, cnt, hint); + } + void deallocate(pointer p, size_type cnt) + { + std::allocator_traits::deallocate(_impl, p, cnt); + } + size_type max_size() const + { + return std::allocator_traits::max_size(_impl); + } + bool operator==(const usm_allocator &other) const { return _impl == other._impl; } + bool operator!=(const usm_allocator &other) const { return _impl != other._impl; } + }; + + } // namespace deprecated + + inline void dpct_free(void *ptr, + const sycl::queue &q) + { + if (ptr) + { +#ifdef DPCT_USM_LEVEL_NONE + detail::mem_mgr::instance().mem_free(ptr); +#else + sycl::free(ptr, q.get_context()); +#endif // DPCT_USM_LEVEL_NONE + } + } + + template + inline auto get_memory(const void *x) + { + T *new_x = reinterpret_cast(const_cast(x)); +#ifdef DPCT_USM_LEVEL_NONE + return dpct::get_buffer>(new_x); +#else + return new_x; +#endif + } + + template + inline typename DataType::T2 get_value(const T *s, sycl::queue &q) + { + using Ty = typename DataType::T2; + Ty s_h; + if (get_pointer_attribute(q, s) == pointer_access_attribute::device_only) + detail::dpct_memcpy(q, (void *)&s_h, (void *)s, sizeof(T), device_to_host) + .wait(); + else + s_h = *reinterpret_cast(s); + return s_h; + } + + } // namespace detail + + template + inline auto get_value(const T *s, sycl::queue &q) + { + return detail::get_value(s, q); + } + + namespace detail + { + template + inline void gemm_impl(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, int k, + const void *alpha, const void *a, int lda, const void *b, + int ldb, const void *beta, void *c, int ldc) + { +#ifndef __INTEL_MKL__ + throw std::runtime_error("The oneAPI Math Kernel Library (oneMKL) Interfaces " + "Project does not support this API."); +#else + Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); + Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); + auto data_a = get_memory(a); + auto data_b = get_memory(b); + auto data_c = get_memory(c); + oneapi::mkl::blas::column_major::gemm( + q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, + data_b, ldb, beta_value, data_c, ldc); +#endif + } + + template + class vectorized_binary + { + public: + inline VecT operator()(VecT a, VecT b, const BinaryOperation binary_op) + { + VecT v4; + for (size_t i = 0; i < v4.size(); ++i) + { + v4[i] = binary_op(a[i], b[i]); + } + return v4; + } + }; + + template + class vectorized_binary< + VecT, BinaryOperation, + std::void_t>> + { + public: + inline VecT operator()(VecT a, VecT b, const BinaryOperation binary_op) + { + return binary_op(a, b).template as(); + } + }; + + template + inline void gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, int k, + const void *alpha, const void **a, int lda, + const void **b, int ldb, const void *beta, void **c, + int ldc, int batch_size) + { + struct matrix_info_t + { + oneapi::mkl::transpose transpose_info[2]; + Ts value_info[2]; + std::int64_t size_info[3]; + std::int64_t ld_info[3]; + std::int64_t groupsize_info; + }; + + Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); + Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); + + matrix_info_t *matrix_info = + (matrix_info_t *)std::malloc(sizeof(matrix_info_t)); + matrix_info->transpose_info[0] = a_trans; + matrix_info->transpose_info[1] = b_trans; + matrix_info->value_info[0] = alpha_value; + matrix_info->value_info[1] = beta_value; + matrix_info->size_info[0] = m; + matrix_info->size_info[1] = n; + matrix_info->size_info[2] = k; + matrix_info->ld_info[0] = lda; + matrix_info->ld_info[1] = ldb; + matrix_info->ld_info[2] = ldc; + matrix_info->groupsize_info = batch_size; + + sycl::event e = oneapi::mkl::blas::column_major::gemm_batch( + q, matrix_info->transpose_info, matrix_info->transpose_info + 1, + matrix_info->size_info, matrix_info->size_info + 1, + matrix_info->size_info + 2, matrix_info->value_info, + reinterpret_cast(a), matrix_info->ld_info, + reinterpret_cast(b), matrix_info->ld_info + 1, + matrix_info->value_info + 1, reinterpret_cast(c), + matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info)); + + q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(e); + cgh.host_task([=] { std::free(matrix_info); }); }); + } + + template + inline void + gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, + int k, const void *alpha, const void *a, int lda, + long long int stride_a, const void *b, int ldb, + long long int stride_b, const void *beta, void *c, + int ldc, long long int stride_c, int batch_size) + { + Ts alpha_value = dpct::get_value(reinterpret_cast(alpha), q); + Ts beta_value = dpct::get_value(reinterpret_cast(beta), q); + auto data_a = get_memory(a); + auto data_b = get_memory(b); + auto data_c = get_memory(c); + oneapi::mkl::blas::column_major::gemm_batch( + q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, + stride_a, data_b, ldb, stride_b, beta_value, + data_c, ldc, stride_c, batch_size); + } + + } // namespace detail + + template + inline unsigned vectorized_binary(unsigned a, unsigned b, + const BinaryOperation binary_op) + { + sycl::vec v0{a}, v1{b}; + auto v2 = v0.as(); + auto v3 = v1.as(); + auto v4 = + detail::vectorized_binary()(v2, v3, binary_op); + v0 = v4.template as>(); + return v0; + } + + static void async_dpct_memcpy(void *to_ptr, const void *from_ptr, size_t size, + memcpy_direction direction = automatic, + sycl::queue &q = dpct::get_default_queue()) + { + detail::dpct_memcpy(q, to_ptr, from_ptr, size, direction); + } + + static inline unsigned int select_device(unsigned int id) + { + dev_mgr::instance().select_device(id); + return id; + } + + template + T permute_sub_group_by_xor(sycl::sub_group g, T x, unsigned int mask, + int logical_sub_group_size = 32) + { + unsigned int id = g.get_local_linear_id(); + unsigned int start_index = + id / logical_sub_group_size * logical_sub_group_size; + unsigned int target_offset = (id % logical_sub_group_size) ^ mask; + return sycl::select_from_group(g, x, + target_offset < logical_sub_group_size + ? start_index + target_offset + : id); + } + + template + sycl::vec extract_and_sign_or_zero_extend4(T val) + { + return sycl::vec(val) + .template as, int8_t, uint8_t>, 4>>() + .template convert(); + } + + template + using dot_product_acc_t = + std::conditional_t && std::is_unsigned_v, + uint32_t, int32_t>; + + template + inline auto dp4a(T1 a, T2 b, T3 c) + { + dot_product_acc_t res = c; + auto va = extract_and_sign_or_zero_extend4(a); + auto vb = extract_and_sign_or_zero_extend4(b); + res += va[0] * vb[0]; + res += va[1] * vb[1]; + res += va[2] * vb[2]; + res += va[3] * vb[3]; + return res; + } + + struct sub_sat + { + template + auto operator()(const T x, const T y) const + { + return sycl::sub_sat(x, y); + } + }; + + template + inline T vectorized_min(T a, T b) + { + sycl::vec v0{a}, v1{b}; + auto v2 = v0.template as(); + auto v3 = v1.template as(); + auto v4 = sycl::min(v2, v3); + v0 = v4.template as>(); + return v0; + } + + inline float pow(const float a, const int b) { return sycl::pown(a, b); } + inline double pow(const double a, const int b) { return sycl::pown(a, b); } + inline float pow(const float a, const float b) { return sycl::pow(a, b); } + inline double pow(const double a, const double b) { return sycl::pow(a, b); } + template + inline typename std::enable_if_t, T> + pow(const T a, const U b) + { + return sycl::pow(a, static_cast(b)); + } + template + inline typename std::enable_if_t, double> + pow(const T a, const U b) + { + return sycl::pow(static_cast(a), static_cast(b)); + } + + inline double min(const double a, const float b) + { + return sycl::fmin(a, static_cast(b)); + } + inline double min(const float a, const double b) + { + return sycl::fmin(static_cast(a), b); + } + inline float min(const float a, const float b) { return sycl::fmin(a, b); } + inline double min(const double a, const double b) { return sycl::fmin(a, b); } + inline std::uint32_t min(const std::uint32_t a, const std::int32_t b) + { + return sycl::min(a, static_cast(b)); + } + inline std::uint32_t min(const std::int32_t a, const std::uint32_t b) + { + return sycl::min(static_cast(a), b); + } + inline std::int32_t min(const std::int32_t a, const std::int32_t b) + { + return sycl::min(a, b); + } + inline std::uint32_t min(const std::uint32_t a, const std::uint32_t b) + { + return sycl::min(a, b); + } + inline std::uint64_t min(const std::uint64_t a, const std::int64_t b) + { + return sycl::min(a, static_cast(b)); + } + inline std::uint64_t min(const std::int64_t a, const std::uint64_t b) + { + return sycl::min(static_cast(a), b); + } + inline std::int64_t min(const std::int64_t a, const std::int64_t b) + { + return sycl::min(a, b); + } + inline std::uint64_t min(const std::uint64_t a, const std::uint64_t b) + { + return sycl::min(a, b); + } + inline std::uint64_t min(const std::uint64_t a, const std::int32_t b) + { + return sycl::min(a, static_cast(b)); + } + inline std::uint64_t min(const std::int32_t a, const std::uint64_t b) + { + return sycl::min(static_cast(a), b); + } + inline std::uint64_t min(const std::uint64_t a, const std::uint32_t b) + { + return sycl::min(a, static_cast(b)); + } + inline std::uint64_t min(const std::uint32_t a, const std::uint64_t b) + { + return sycl::min(static_cast(a), b); + } + // max function overloads. + // For floating-point types, `float` or `double` arguments are acceptable. + // For integer types, `std::uint32_t`, `std::int32_t`, `std::uint64_t` or + // `std::int64_t` type arguments are acceptable. + inline double max(const double a, const float b) + { + return sycl::fmax(a, static_cast(b)); + } + inline double max(const float a, const double b) + { + return sycl::fmax(static_cast(a), b); + } + inline float max(const float a, const float b) { return sycl::fmax(a, b); } + inline double max(const double a, const double b) { return sycl::fmax(a, b); } + inline std::uint32_t max(const std::uint32_t a, const std::int32_t b) + { + return sycl::max(a, static_cast(b)); + } + inline std::uint32_t max(const std::int32_t a, const std::uint32_t b) + { + return sycl::max(static_cast(a), b); + } + inline std::int32_t max(const std::int32_t a, const std::int32_t b) + { + return sycl::max(a, b); + } + inline std::uint32_t max(const std::uint32_t a, const std::uint32_t b) + { + return sycl::max(a, b); + } + inline std::uint64_t max(const std::uint64_t a, const std::int64_t b) + { + return sycl::max(a, static_cast(b)); + } + inline std::uint64_t max(const std::int64_t a, const std::uint64_t b) + { + return sycl::max(static_cast(a), b); + } + inline std::int64_t max(const std::int64_t a, const std::int64_t b) + { + return sycl::max(a, b); + } + inline std::uint64_t max(const std::uint64_t a, const std::uint64_t b) + { + return sycl::max(a, b); + } + inline std::uint64_t max(const std::uint64_t a, const std::int32_t b) + { + return sycl::max(a, static_cast(b)); + } + inline std::uint64_t max(const std::int32_t a, const std::uint64_t b) + { + return sycl::max(static_cast(a), b); + } + inline std::uint64_t max(const std::uint64_t a, const std::uint32_t b) + { + return sycl::max(a, static_cast(b)); + } + inline std::uint64_t max(const std::uint32_t a, const std::uint64_t b) + { + return sycl::max(static_cast(a), b); + } + + inline void + has_capability_or_fail(const sycl::device &dev, + const std::initializer_list &props) + { + for (const auto &it : props) + { + if (dev.has(it)) + continue; + switch (it) + { + case sycl::aspect::fp64: + throw std::runtime_error("'double' is not supported in '" + + dev.get_info() + + "' device"); + break; + case sycl::aspect::fp16: + throw std::runtime_error("'half' is not supported in '" + + dev.get_info() + + "' device"); + break; + default: +#define __SYCL_ASPECT(ASPECT, ID) \ + case sycl::aspect::ASPECT: \ + return #ASPECT; +#define __SYCL_ASPECT_DEPRECATED(ASPECT, ID, MESSAGE) __SYCL_ASPECT(ASPECT, ID) +#define __SYCL_ASPECT_DEPRECATED_ALIAS(ASPECT, ID, MESSAGE) + auto getAspectNameStr = [](sycl::aspect AspectNum) -> std::string + { + switch (AspectNum) + { +#include +#include + default: + return "unknown aspect"; + } + }; +#undef __SYCL_ASPECT_DEPRECATED_ALIAS +#undef __SYCL_ASPECT_DEPRECATED +#undef __SYCL_ASPECT + throw std::runtime_error( + "'" + getAspectNameStr(it) + "' is not supported in '" + + dev.get_info() + "' device"); + } + break; + } + } + + static inline unsigned int get_current_device_id() + { + return dev_mgr::instance().current_device_id(); + } + + static inline device_ext &get_current_device() + { + return dev_mgr::instance().current_device(); + } + + static inline sycl::queue &get_in_order_queue() + { + return dev_mgr::instance().current_device().in_order_queue(); + } + + static sycl::event + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, size_t size, + memcpy_direction direction, + const std::vector &dep_events = {}) + { + if (!size) + return sycl::event{}; +#ifdef DPCT_USM_LEVEL_NONE + auto &mm = mem_mgr::instance(); + auto real_direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction); + + switch (real_direction) + { + case host_to_host: + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + cgh.host_task([=] { std::memcpy(to_ptr, from_ptr, size); }); }); + case host_to_device: + { + auto alloc = mm.translate_ptr(to_ptr); + size_t offset = (byte_t *)to_ptr - alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto o = sycl::id<1>(offset); + sycl::accessor + acc(alloc.buffer, cgh, r, o); + cgh.copy(from_ptr, acc); }); + } + case device_to_host: + { + auto alloc = mm.translate_ptr(from_ptr); + size_t offset = (byte_t *)from_ptr - alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto o = sycl::id<1>(offset); + sycl::accessor + acc(alloc.buffer, cgh, r, o); + cgh.copy(acc, to_ptr); }); + } + case device_to_device: + { + auto to_alloc = mm.translate_ptr(to_ptr); + auto from_alloc = mm.translate_ptr(from_ptr); + size_t to_offset = (byte_t *)to_ptr - to_alloc.alloc_ptr; + size_t from_offset = (byte_t *)from_ptr - from_alloc.alloc_ptr; + return q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto r = sycl::range<1>(size); + auto to_o = sycl::id<1>(to_offset); + auto from_o = sycl::id<1>(from_offset); + sycl::accessor + to_acc(to_alloc.buffer, cgh, r, to_o); + sycl::accessor + from_acc(from_alloc.buffer, cgh, r, from_o); + cgh.copy(from_acc, to_acc); }); + } + default: + throw std::runtime_error("dpct_memcpy: invalid direction value"); + } +#else + return q.memcpy(to_ptr, from_ptr, size, dep_events); +#endif // DPCT_USM_LEVEL_NONE + } + + // Get actual copy range and make sure it will not exceed range. + static inline size_t get_copy_range(sycl::range<3> size, size_t slice, + size_t pitch) + { + return slice * (size.get(2) - 1) + pitch * (size.get(1) - 1) + size.get(0); + } + + static inline size_t get_offset(sycl::id<3> id, size_t slice, + size_t pitch) + { + return slice * id.get(2) + pitch * id.get(1) + id.get(0); + } + + /// copy 3D matrix specified by \p size from 3D matrix specified by \p from_ptr + /// and \p from_range to another specified by \p to_ptr and \p to_range. + static inline std::vector + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, + sycl::range<3> to_range, sycl::range<3> from_range, + sycl::id<3> to_id, sycl::id<3> from_id, + sycl::range<3> size, memcpy_direction direction, + const std::vector &dep_events = {}) + { + // RAII for host pointer + class host_buffer + { + void *_buf; + size_t _size; + sycl::queue &_q; + const std::vector &_deps; // free operation depends + + public: + host_buffer(size_t size, sycl::queue &q, + const std::vector &deps) + : _buf(std::malloc(size)), _size(size), _q(q), _deps(deps) {} + void *get_ptr() const { return _buf; } + size_t get_size() const { return _size; } + ~host_buffer() + { + if (_buf) + { + _q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(_deps); + cgh.host_task([buf = _buf] { std::free(buf); }); }); + } + } + }; + std::vector event_list; + + size_t to_slice = to_range.get(1) * to_range.get(0), + from_slice = from_range.get(1) * from_range.get(0); + unsigned char *to_surface = + (unsigned char *)to_ptr + get_offset(to_id, to_slice, to_range.get(0)); + const unsigned char *from_surface = + (const unsigned char *)from_ptr + + get_offset(from_id, from_slice, from_range.get(0)); + + if (to_slice == from_slice && to_slice == size.get(1) * size.get(0)) + { + return {dpct_memcpy(q, to_surface, from_surface, to_slice * size.get(2), + direction, dep_events)}; + } + direction = detail::deduce_memcpy_direction(q, to_ptr, from_ptr, direction); + size_t size_slice = size.get(1) * size.get(0); + switch (direction) + { + case host_to_host: + for (size_t z = 0; z < size.get(2); ++z) + { + unsigned char *to_ptr = to_surface; + const unsigned char *from_ptr = from_surface; + if (to_range.get(0) == from_range.get(0) && + to_range.get(0) == size.get(0)) + { + event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size_slice, + direction, dep_events)); + } + else + { + for (size_t y = 0; y < size.get(1); ++y) + { + event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size.get(0), + direction, dep_events)); + to_ptr += to_range.get(0); + from_ptr += from_range.get(0); + } + } + to_surface += to_slice; + from_surface += from_slice; + } + break; + case host_to_device: + { + host_buffer buf(get_copy_range(size, to_slice, to_range.get(0)), q, + event_list); + std::vector host_events; + if (to_slice == size_slice) + { + // Copy host data to a temp host buffer with the shape of target. + host_events = + dpct_memcpy(q, buf.get_ptr(), from_surface, to_range, from_range, + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, + host_to_host, dep_events); + } + else + { + // Copy host data to a temp host buffer with the shape of target. + host_events = dpct_memcpy( + q, buf.get_ptr(), from_surface, to_range, from_range, + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, host_to_host, + // If has padding data, not sure whether it is useless. So fill temp + // buffer with it. + std::vector{ + dpct_memcpy(q, buf.get_ptr(), to_surface, buf.get_size(), + device_to_host, dep_events)}); + } + // Copy from temp host buffer to device with only one submit. + event_list.push_back(dpct_memcpy(q, to_surface, buf.get_ptr(), + buf.get_size(), host_to_device, + host_events)); + break; + } + case device_to_host: + { + host_buffer buf(get_copy_range(size, from_slice, from_range.get(0)), q, + event_list); + // Copy from host temp buffer to host target with reshaping. + event_list = dpct_memcpy( + q, to_surface, buf.get_ptr(), to_range, from_range, sycl::id<3>(0, 0, 0), + sycl::id<3>(0, 0, 0), size, host_to_host, + // Copy from device to temp host buffer with only one submit. + std::vector{dpct_memcpy(q, buf.get_ptr(), from_surface, + buf.get_size(), + device_to_host, dep_events)}); + break; + } + case device_to_device: +#ifdef DPCT_USM_LEVEL_NONE + { + auto &mm = mem_mgr::instance(); + auto to_alloc = mm.translate_ptr(to_surface); + auto from_alloc = mm.translate_ptr(from_surface); + size_t to_offset = (byte_t *)to_surface - to_alloc.alloc_ptr; + size_t from_offset = (byte_t *)from_surface - from_alloc.alloc_ptr; + event_list.push_back(q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + auto to_o = sycl::id<1>(to_offset); + auto from_o = sycl::id<1>(from_offset); + sycl::accessor + to_acc(to_alloc.buffer, cgh, + get_copy_range(size, to_slice, to_range.get(0)), to_o); + sycl::accessor + from_acc(from_alloc.buffer, cgh, + get_copy_range(size, from_slice, from_range.get(0)), from_o); + cgh.parallel_for( + size, + [=](sycl::id<3> id) { + to_acc[get_offset(id, to_slice, to_range.get(0))] = + from_acc[get_offset(id, from_slice, from_range.get(0))]; + }); })); + } +#else + event_list.push_back(q.submit([&](sycl::handler &cgh) + { + cgh.depends_on(dep_events); + cgh.parallel_for( + size, + [=](sycl::id<3> id) { + to_surface[get_offset(id, to_slice, to_range.get(0))] = + from_surface[get_offset(id, from_slice, from_range.get(0))]; + }); })); +#endif + break; + default: + throw std::runtime_error("dpct_memcpy: invalid direction value"); + } + return event_list; + } + + /// memcpy 2D/3D matrix specified by pitched_data. + static inline std::vector + dpct_memcpy(sycl::queue &q, pitched_data to, sycl::id<3> to_id, + pitched_data from, sycl::id<3> from_id, sycl::range<3> size, + memcpy_direction direction = automatic) + { + return dpct_memcpy(q, to.get_data_ptr(), from.get_data_ptr(), + sycl::range<3>(to.get_pitch(), to.get_y(), 1), + sycl::range<3>(from.get_pitch(), from.get_y(), 1), to_id, from_id, + size, direction); + } + + /// memcpy 2D matrix with pitch. + static inline std::vector + dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, + size_t to_pitch, size_t from_pitch, size_t x, size_t y, + memcpy_direction direction = automatic) + { + return dpct_memcpy(q, to_ptr, from_ptr, sycl::range<3>(to_pitch, y, 1), + sycl::range<3>(from_pitch, y, 1), + sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), + sycl::range<3>(x, y, 1), direction); + } + + inline void gemm(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, int k, + const void *alpha, const void *a, library_data_t a_type, + int lda, const void *b, library_data_t b_type, int ldb, + const void *beta, void *c, library_data_t c_type, int ldc, + library_data_t scaling_type) + { + if (scaling_type == library_data_t::real_float && + c_type == library_data_t::complex_float) + { + scaling_type = library_data_t::complex_float; + } + else if (scaling_type == library_data_t::real_double && + c_type == library_data_t::complex_double) + { + scaling_type = library_data_t::complex_double; + } + + std::uint64_t key = + detail::get_type_combination_id(a_type, b_type, c_type, scaling_type); + switch (key) + { + case detail::get_type_combination_id( + library_data_t::real_float, library_data_t::real_float, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_double, library_data_t::real_double, + library_data_t::real_double, library_data_t::real_double): + { + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_float, library_data_t::complex_float, + library_data_t::complex_float, library_data_t::complex_float): + { + detail::gemm_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_double, library_data_t::complex_double, + library_data_t::complex_double, library_data_t::complex_double): + { + detail::gemm_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_half): + { + detail::gemm_impl(q, a_trans, b_trans, m, n, k, alpha, a, + lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_impl(q, a_trans, b_trans, m, n, k, alpha, a, lda, b, + ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_float): + { + float alpha_value = + dpct::get_value(reinterpret_cast(alpha), q); + float beta_value = + dpct::get_value(reinterpret_cast(beta), q); + sycl::half alpha_half(alpha_value); + sycl::half beta_half(beta_value); + detail::gemm_impl(q, a_trans, b_trans, m, n, k, &alpha_half, + a, lda, b, ldb, &beta_half, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_bfloat16, library_data_t::real_float): + { + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_int32, library_data_t::real_int32): + { + float alpha_float = + dpct::get_value(reinterpret_cast(alpha), q); + float beta_float = + dpct::get_value(reinterpret_cast(beta), q); + detail::gemm_impl( + q, a_trans, b_trans, m, n, k, &alpha_float, a, lda, b, ldb, &beta_float, c, ldc); + break; + } + default: + throw std::runtime_error("the combination of data type is unsupported"); + } + } // gemm() + + /// Computes a batch of matrix-matrix product with general matrices. + /// \param [in] q The queue where the routine should be executed. + /// \param [in] a_trans Specifies the operation applied to A. + /// \param [in] b_trans Specifies the operation applied to B. + /// \param [in] m Specifies the number of rows of the matrix op(A) and of the matrix C. + /// \param [in] n Specifies the number of columns of the matrix op(B) and of the matrix C. + /// \param [in] k Specifies the number of columns of the matrix op(A) and the number of rows of the matrix op(B). + /// \param [in] alpha Scaling factor for the matrix-matrix product. + /// \param [in] a Input matrix A. + /// \param [in] a_type Data type of the matrix A. + /// \param [in] lda Leading dimension of A. + /// \param [in] b Input matrix B. + /// \param [in] b_type Data type of the matrix B. + /// \param [in] ldb Leading dimension of B. + /// \param [in] beta Scaling factor for matrix C. + /// \param [in, out] c Input/Output matrix C. + /// \param [in] c_type Data type of the matrix C. + /// \param [in] ldc Leading dimension of C. + /// \param [in] batch_size Specifies the number of matrix multiply operations to perform. + /// \param [in] scaling_type Data type of the scaling factors. + inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, int k, + const void *alpha, const void *a[], + library_data_t a_type, int lda, const void *b[], + library_data_t b_type, int ldb, const void *beta, + void *c[], library_data_t c_type, int ldc, + int batch_size, library_data_t scaling_type) + { +#ifdef DPCT_USM_LEVEL_NONE + throw std::runtime_error("this API is unsupported when USM level is none"); +#else + if (scaling_type == library_data_t::real_float && + c_type == library_data_t::complex_float) + { + scaling_type = library_data_t::complex_float; + } + else if (scaling_type == library_data_t::real_double && + c_type == library_data_t::complex_double) + { + scaling_type = library_data_t::complex_double; + } + + std::uint64_t key = + detail::get_type_combination_id(a_type, b_type, c_type, scaling_type); + switch (key) + { + case detail::get_type_combination_id( + library_data_t::real_float, library_data_t::real_float, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_double, library_data_t::real_double, + library_data_t::real_double, library_data_t::real_double): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_float, library_data_t::complex_float, + library_data_t::complex_float, library_data_t::complex_float): + { + detail::gemm_batch_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_double, library_data_t::complex_double, + library_data_t::complex_double, library_data_t::complex_double): + { + detail::gemm_batch_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_half): + { + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, + a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } +#ifdef __INTEL_MKL__ + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_bfloat16, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, a, lda, + b, ldb, beta, c, ldc, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_int32, library_data_t::real_int32): + { + float alpha_float = + dpct::get_value(reinterpret_cast(alpha), q); + float beta_float = + dpct::get_value(reinterpret_cast(beta), q); + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, &alpha_float, + a, lda, b, ldb, &beta_float, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, + batch_size); + break; + } +#endif + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_float): + { + float alpha_value = + dpct::get_value(reinterpret_cast(alpha), q); + float beta_value = + dpct::get_value(reinterpret_cast(beta), q); + sycl::half alpha_half(alpha_value); + sycl::half beta_half(beta_value); + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, b, ldb, &beta_half, c, ldc, + batch_size); + break; + } + default: + throw std::runtime_error("the combination of data type is unsupported"); + } +#endif + } + + /// Computes a batch of matrix-matrix product with general matrices. + /// \param [in] q The queue where the routine should be executed. + /// \param [in] a_trans Specifies the operation applied to A. + /// \param [in] b_trans Specifies the operation applied to B. + /// \param [in] m Specifies the number of rows of the matrix op(A) and of the matrix C. + /// \param [in] n Specifies the number of columns of the matrix op(B) and of the matrix C. + /// \param [in] k Specifies the number of columns of the matrix op(A) and the number of rows of the matrix op(B). + /// \param [in] alpha Scaling factor for the matrix-matrix product. + /// \param [in] a Input matrix A. + /// \param [in] a_type Data type of the matrix A. + /// \param [in] lda Leading dimension of A. + /// \param [in] stride_a Stride between the different A matrices. + /// \param [in] b Input matrix B. + /// \param [in] b_type Data type of the matrix B. + /// \param [in] ldb Leading dimension of B. + /// \param [in] stride_b Stride between the different B matrices. + /// \param [in] beta Scaling factor for matrix C. + /// \param [in, out] c Input/Output matrix C. + /// \param [in] c_type Data type of the matrix C. + /// \param [in] ldc Leading dimension of C. + /// \param [in] stride_c Stride between the different C matrices. + /// \param [in] batch_size Specifies the number of matrix multiply operations to perform. + /// \param [in] scaling_type Data type of the scaling factors. + inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans, + oneapi::mkl::transpose b_trans, int m, int n, int k, + const void *alpha, const void *a, library_data_t a_type, + int lda, long long int stride_a, const void *b, + library_data_t b_type, int ldb, long long int stride_b, + const void *beta, void *c, library_data_t c_type, + int ldc, long long int stride_c, int batch_size, + library_data_t scaling_type) + { + if (scaling_type == library_data_t::real_float && + c_type == library_data_t::complex_float) + { + scaling_type = library_data_t::complex_float; + } + else if (scaling_type == library_data_t::real_double && + c_type == library_data_t::complex_double) + { + scaling_type = library_data_t::complex_double; + } + + std::uint64_t key = + detail::get_type_combination_id(a_type, b_type, c_type, scaling_type); + switch (key) + { + case detail::get_type_combination_id( + library_data_t::real_float, library_data_t::real_float, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_double, library_data_t::real_double, + library_data_t::real_double, library_data_t::real_double): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_float, library_data_t::complex_float, + library_data_t::complex_float, library_data_t::complex_float): + { + detail::gemm_batch_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::complex_double, library_data_t::complex_double, + library_data_t::complex_double, library_data_t::complex_double): + { + detail::gemm_batch_impl, std::complex, + std::complex, std::complex>( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_half): + { + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, + a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } +#ifdef __INTEL_MKL__ + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_bfloat16, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_bfloat16, library_data_t::real_bfloat16, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, a, lda, + stride_a, b, ldb, stride_b, beta, c, ldc, + stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_int32, library_data_t::real_int32): + { + detail::gemm_batch_impl(q, a_trans, b_trans, m, n, k, alpha, + a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_int8, library_data_t::real_int8, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_float, library_data_t::real_float): + { + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, + beta, c, ldc, stride_c, batch_size); + break; + } +#endif + case detail::get_type_combination_id( + library_data_t::real_half, library_data_t::real_half, + library_data_t::real_half, library_data_t::real_float): + { + float alpha_value = + dpct::get_value(reinterpret_cast(alpha), q); + float beta_value = + dpct::get_value(reinterpret_cast(beta), q); + sycl::half alpha_half(alpha_value); + sycl::half beta_half(beta_value); + detail::gemm_batch_impl( + q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, stride_a, b, ldb, stride_b, + &beta_half, c, ldc, stride_c, batch_size); + break; + } + default: + throw std::runtime_error("the combination of data type is unsupported"); + } + } + + static inline void + async_dpct_memcpy(void *to_ptr, size_t to_pitch, const void *from_ptr, + size_t from_pitch, size_t x, size_t y, + memcpy_direction direction = automatic, + sycl::queue &q = get_default_queue()) + { + detail::dpct_memcpy(q, to_ptr, from_ptr, to_pitch, from_pitch, x, y, + direction); + } + + using err0 = detail::generic_error_type; + using err1 = detail::generic_error_type; + +} // COPY from DPCT head files + + +static int g_ggml_sycl_debug=0; +#define GGML_SYCL_DEBUG(...) do{if(g_ggml_sycl_debug) printf(__VA_ARGS__);}while(0) + +#define CHECK_TRY_ERROR(expr) \ + [&]() { \ + try { \ + expr; \ + return dpct::success; \ + } catch (std::exception const &e) { \ + std::cerr << e.what()<< "\nException caught at file:" << __FILE__ \ + << ", line:" << __LINE__ <<", func:"<<__func__<< std::endl; \ + return dpct::default_error; \ + } \ + }() + +// #define DEBUG_SYCL_MALLOC + +static int g_work_group_size = 0; +// typedef sycl::half ggml_fp16_t; + +#define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP +#define VER_4VEC 610 //todo for hardward optimize. +#define VER_GEN9 700 //todo for hardward optimize. +#define VER_GEN12 1000000 //todo for hardward optimize. +#define VER_GEN13 (VER_GEN12 + 1030) //todo for hardward optimize. + +#define GGML_SYCL_MAX_NODES 8192 //TODO: adapt to hardwares + + +//define for XMX in Intel GPU +//TODO: currently, it's not used for XMX really. +#define SYCL_USE_XMX + +// max batch size to use MMQ kernels when tensor cores are available +#define XMX_MAX_BATCH_SIZE 32 + + +#if defined(_MSC_VER) +#pragma warning(disable: 4244 4267) // possible loss of data +#endif + +static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size"); + +static void crash(){ + int *ptr = NULL; + *ptr = 0; +} + +static void ggml_sycl_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) { + fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg); + fprintf(stderr, " in function %s at %s:%d\n", func, file, line); + GGML_ASSERT(!"SYCL error"); +} + +#define SYCL_CHECK(err) do { \ + auto err_ = (err); if (err_ != 0) ggml_sycl_error( \ + #err, __func__, __FILE__, __LINE__, \ + "Meet error in this line code!"); \ +} while (0) + +#if DPCT_COMPAT_RT_VERSION >= 11100 +#define GGML_SYCL_ASSUME(x) __builtin_assume(x) +#else +#define GGML_SYCL_ASSUME(x) +#endif // DPCT_COMPAT_RT_VERSION >= 11100 + +#ifdef GGML_SYCL_F16 +typedef sycl::half dfloat; // dequantize float +typedef sycl::half2 dfloat2; +#else +typedef float dfloat; // dequantize float +typedef sycl::float2 dfloat2; +#endif //GGML_SYCL_F16 + +bool ggml_sycl_loaded(void); +void * ggml_sycl_host_malloc(size_t size); +void ggml_sycl_host_free(void * ptr); +bool ggml_sycl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst); +void ggml_sycl_set_tensor_split(const float * tensor_split); +void ggml_sycl_transform_tensor(void * data, struct ggml_tensor * tensor); +void ggml_sycl_free_data(struct ggml_tensor * tensor); +void ggml_sycl_assign_buffers(struct ggml_tensor * tensor); +void ggml_sycl_assign_buffers_no_scratch(struct ggml_tensor * tensor); +void ggml_sycl_assign_buffers_force_inplace(struct ggml_tensor * tensor); +void ggml_sycl_assign_buffers_no_alloc(struct ggml_tensor * tensor); +void ggml_sycl_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset); +void ggml_sycl_copy_to_device(struct ggml_tensor * tensor); +void ggml_sycl_set_main_device(int main_device); +void ggml_sycl_set_mul_mat_q(bool mul_mat_q); +void ggml_sycl_set_scratch_size(size_t scratch_size); +void ggml_sycl_free_scratch(void); +int ggml_sycl_get_device_count(void); +void ggml_sycl_get_device_description(int device, char * description, size_t description_size); +bool ggml_backend_is_sycl(ggml_backend_t backend); +int ggml_backend_sycl_get_device(ggml_backend_t backend); +int get_main_device(); +void print_ggml_tensor(const char*name, struct ggml_tensor *src); +void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt); + +static __dpct_inline__ int get_int_from_int8(const int8_t *x8, const int &i32) { + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __dpct_inline__ int get_int_from_uint8(const uint8_t *x8, + const int &i32) { + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + + int x32 = 0; + x32 |= x16[0] << 0; + x32 |= x16[1] << 16; + + return x32; +} + +static __dpct_inline__ int get_int_from_int8_aligned(const int8_t *x8, + const int &i32) { + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + +static __dpct_inline__ int get_int_from_uint8_aligned(const uint8_t *x8, + const int &i32) { + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment +} + +template +using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y, + int k, dpct::queue_ptr stream); +typedef to_t_sycl_t to_fp32_sycl_t; +typedef to_t_sycl_t to_fp16_sycl_t; + +typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v); +typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v); +typedef void (*cpy_kernel_t)(const char * cx, char * cdst); +typedef void (*ggml_sycl_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst); +typedef void (*ggml_sycl_op_mul_mat_t)( + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream); +typedef void (*ggml_sycl_op_flatten_t)(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream); + +// QK = number of values after dequantization +// QR = QK / number of values before dequantization +// QI = number of 32 bit integers before dequantization + +#define QK4_0 32 +#define QR4_0 2 +#define QI4_0 (QK4_0 / (4 * QR4_0)) +typedef struct dpct_type_471834 { + sycl::half d; // delta + uint8_t qs[QK4_0 / 2]; // nibbles / quants +} block_q4_0; +static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding"); + +#define QK4_1 32 +#define QR4_1 2 +#define QI4_1 (QK4_1 / (4 * QR4_1)) +typedef struct dpct_type_143705 { + sycl::half2 dm; // dm.x = delta, dm.y = min + uint8_t qs[QK4_1 / 2]; // nibbles / quants +} block_q4_1; +static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding"); + +#define QK5_0 32 +#define QR5_0 2 +#define QI5_0 (QK5_0 / (4 * QR5_0)) +typedef struct dpct_type_673649 { + sycl::half d; // delta + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_0 / 2]; // nibbles / quants +} block_q5_0; +static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding"); + +#define QK5_1 32 +#define QR5_1 2 +#define QI5_1 (QK5_1 / (4 * QR5_1)) +typedef struct dpct_type_135589 { + sycl::half2 dm; // dm.x = delta, dm.y = min + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[QK5_1 / 2]; // nibbles / quants +} block_q5_1; +static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding"); + +#define QK8_0 32 +#define QR8_0 1 +#define QI8_0 (QK8_0 / (4 * QR8_0)) +typedef struct dpct_type_122878 { + sycl::half d; // delta + int8_t qs[QK8_0]; // quants +} block_q8_0; +static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding"); + +#define QK8_1 32 +#define QR8_1 1 +#define QI8_1 (QK8_1 / (4 * QR8_1)) +typedef struct dpct_type_143721 { + sycl::half2 ds; // ds.x = delta, ds.y = sum + int8_t qs[QK8_0]; // quants +} block_q8_1; +static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding"); + +typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs); +typedef void (*allocate_tiles_sycl_t)(int **x_ql, sycl::half2 **x_dm, + int **x_qh, int **x_sc); +typedef void (*load_tiles_sycl_t)(const void *__restrict__ vx, + int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, + int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, + const int &i_max, const int &k, + const int &blocks_per_row); +typedef float (*vec_dot_q_mul_mat_sycl_t)( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ms, + const int &i, const int &j, const int &k); + +//================================= k-quants + +#ifdef GGML_QKK_64 +#define QK_K 64 +#define K_SCALE_SIZE 4 +#else +#define QK_K 256 +#define K_SCALE_SIZE 12 +#endif + +#define QR2_K 4 +#define QI2_K (QK_K / (4*QR2_K)) +typedef struct dpct_type_619598 { + uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits + uint8_t qs[QK_K/4]; // quants + sycl::half2 dm; // super-block scale for quantized scales/mins +} block_q2_K; +static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding"); + +#define QR3_K 4 +#define QI3_K (QK_K / (4*QR3_K)) +typedef struct dpct_type_138576 { + uint8_t hmask[QK_K/8]; // quants - high bit + uint8_t qs[QK_K/4]; // quants - low 2 bits +#ifdef GGML_QKK_64 + uint8_t scales[2]; // scales, quantized with 8 bits +#else + uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits +#endif + sycl::half d; // super-block scale +} block_q3_K; +//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding"); + +#define QR4_K 2 +#define QI4_K (QK_K / (4*QR4_K)) +#ifdef GGML_QKK_64 +typedef struct { + half dm[2]; // super-block scales/mins + uint8_t scales[2]; // 4-bit block scales/mins + uint8_t qs[QK_K/2]; // 4--bit quants +} block_q4_K; +static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding"); +#else +typedef struct dpct_type_154943 { + sycl::half2 dm; // super-block scale for quantized scales/mins + uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits + uint8_t qs[QK_K/2]; // 4--bit quants +} block_q4_K; +static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding"); +#endif + +#define QR5_K 2 +#define QI5_K (QK_K / (4*QR5_K)) +#ifdef GGML_QKK_64 +typedef struct { + half d; // super-block scale + int8_t scales[QK_K/16]; // block scales + uint8_t qh[QK_K/8]; // quants, high bit + uint8_t qs[QK_K/2]; // quants, low 4 bits +} block_q5_K; +static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding"); +#else +typedef struct dpct_type_866817 { + sycl::half2 dm; // super-block scale for quantized scales/mins + uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits + uint8_t qh[QK_K/8]; // quants, high bit + uint8_t qs[QK_K/2]; // quants, low 4 bits +} block_q5_K; +static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding"); +#endif + +#define QR6_K 2 +#define QI6_K (QK_K / (4*QR6_K)) +typedef struct dpct_type_107281 { + uint8_t ql[QK_K/2]; // quants, lower 4 bits + uint8_t qh[QK_K/4]; // quants, upper 2 bits + int8_t scales[QK_K/16]; // scales + sycl::half d; // delta +} block_q6_K; +static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding"); + +#define WARP_SIZE 32 +#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses + +#define SYCL_GELU_BLOCK_SIZE 256 +#define SYCL_SILU_BLOCK_SIZE 256 +#define SYCL_TANH_BLOCK_SIZE 256 +#define SYCL_RELU_BLOCK_SIZE 256 +#define SYCL_SQR_BLOCK_SIZE 256 +#define SYCL_CPY_BLOCK_SIZE 32 +#define SYCL_SCALE_BLOCK_SIZE 256 +#define SYCL_CLAMP_BLOCK_SIZE 256 +#define SYCL_ROPE_BLOCK_SIZE 256 +#define SYCL_SOFT_MAX_BLOCK_SIZE 1024 +#define SYCL_ALIBI_BLOCK_SIZE 32 +#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32 +#define SYCL_QUANTIZE_BLOCK_SIZE 256 +#define SYCL_DEQUANTIZE_BLOCK_SIZE 256 +#define SYCL_GET_ROWS_BLOCK_SIZE 256 +#define SYCL_UPSCALE_BLOCK_SIZE 256 +#define SYCL_CONCAT_BLOCK_SIZE 256 +#define SYCL_PAD_BLOCK_SIZE 256 +#define SYCL_ACC_BLOCK_SIZE 256 +#define SYCL_IM2COL_BLOCK_SIZE 256 + +// dmmv = dequantize_mul_mat_vec +#ifndef GGML_SYCL_DMMV_X +#define GGML_SYCL_DMMV_X 32 +#endif +#ifndef GGML_SYCL_MMV_Y +#define GGML_SYCL_MMV_Y 1 +#endif + +#ifndef K_QUANTS_PER_ITERATION +#define K_QUANTS_PER_ITERATION 2 +#else +static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2"); +#endif + +#ifndef GGML_SYCL_PEER_MAX_BATCH_SIZE +#define GGML_SYCL_PEER_MAX_BATCH_SIZE 128 +#endif // GGML_SYCL_PEER_MAX_BATCH_SIZE + +#define MUL_MAT_SRC1_COL_STRIDE 128 + +#define MAX_STREAMS 8 +static dpct::queue_ptr g_syclStreams[GGML_SYCL_MAX_DEVICES][MAX_STREAMS] = { + {0}}; + +struct ggml_tensor_extra_gpu { + void * data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split tensors + dpct::event_ptr + events[GGML_SYCL_MAX_DEVICES] + [MAX_STREAMS]; // events for synchronizing multiple GPUs +}; + +inline dpct::err0 ggml_sycl_set_device(const int device) try { + int current_device; + + SYCL_CHECK(CHECK_TRY_ERROR( + current_device = dpct::dev_mgr::instance().current_device_id())); + + // GGML_SYCL_DEBUG("ggml_sycl_set_device device=%d, current_device=%d\n", device, current_device); + if (device == current_device) { + return 0; + } + + return CHECK_TRY_ERROR(dpct::select_device(device)); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + crash(); + std::exit(1); +} + +static int g_device_count = -1; +static int g_all_sycl_device_count = -1; +static int g_main_device = -1; +static int g_main_device_index = -1; + +static float g_tensor_split[GGML_SYCL_MAX_DEVICES] = {0}; + +struct sycl_device_capabilities { + int cc; // compute capability + bool vmm; // virtual memory support + size_t vmm_granularity; // granularity of virtual memory + int device_id; +}; + +static sycl_device_capabilities g_device_caps[GGML_SYCL_MAX_DEVICES] = { {0, false, 0, -1} }; + +struct sycl_device_id2index { + int index; +}; + +static sycl_device_id2index g_sycl_device_id2index[GGML_SYCL_MAX_DEVICES] = { {-1} }; + +static void * g_scratch_buffer = nullptr; +static size_t g_scratch_size = 0; // disabled by default +static size_t g_scratch_offset = 0; + +static dpct::queue_ptr g_sycl_handles[GGML_SYCL_MAX_DEVICES] = {nullptr}; + +int get_main_device(){ + return g_main_device; +} + +[[noreturn]] +static void bad_arch(const sycl::stream &stream_ct1) { + stream_ct1 << "ERROR: ggml-sycl was compiled without support for the " + "current GPU architecture.\n"; + // __trap(); + std::exit(1); + + (void) bad_arch; // suppress unused function warning +} + +void log_ggml_var_device(const char*name, float *src, size_t total_elements, bool src_on_device){ + if(!g_ggml_sycl_debug) return; + if(!src){ + printf("GGML Tensor:%s skip to save for NULL pointer\n", name); + return; + } + char filename[1024]; + sprintf(filename, "%s.txt", name); + printf("GGML Tensor:%s save to %s\n", name, filename); + + size_t total_size = total_elements*sizeof(float); + float *local_buf = NULL; + // printf("total_size %d2, src_on_device %d\n", total_size, src_on_device); + if(src_on_device) { + local_buf = (float *) ggml_sycl_host_malloc(total_size); + // printf("local buf %p size %d bytes\n", local_buf, total_size); + ggml_sycl_set_device(g_main_device); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + main_stream->memcpy(local_buf, src, total_size); + } + else { + local_buf = (float *)src; + // printf("local buf from src-> data %p\n", local_buf); + } + + std::ofstream logfile; + logfile.open(filename); + // printf("local buf element %d\n", total_elements); + for(int i=0; ibackend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT; + float *src_data =NULL; + if(src_on_device) { + ggml_tensor_extra_gpu * src_extra = (ggml_tensor_extra_gpu *) src->extra; + src_data = (float*)src_extra->data_device[g_main_device_index]; + } + else { + src_data = (float *)src->data; + } + + log_ggml_var_device(name, src_data, total_elements, src_on_device); +} + +static int log_file_name_idx=0; +void log_tensor_with_cnt(const char* name, struct ggml_tensor * src, int stop_cnt) { + stop_cnt = 4; + if(log_file_name_idx>=stop_cnt) return; + char filename[1280]; + sprintf(filename, "%s_%07d", name, log_file_name_idx); + log_file_name_idx++; + print_ggml_tensor(filename, src); + // print_ggml_tensor("ggml_sycl_rms_norm_src0", (ggml_tensor *)src0); + // print_ggml_tensor("ggml_sycl_rms_norm_src1", (ggml_tensor *)src1); + // int *ptr = NULL; + // *ptr = 0; +} + +static __dpct_inline__ float warp_reduce_sum(float x, + const sycl::nd_item<3> &item_ct1) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + /* + DPCT1096:98: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask); + } + return x; +} + +static __dpct_inline__ sycl::float2 +warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3> &item_ct1) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(), + mask); + a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(), + mask); + } + return a; +} + +static __dpct_inline__ float warp_reduce_max(float x, + const sycl::nd_item<3> &item_ct1) { +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + /* + DPCT1096:97: The right-most dimension of the work-group used in the SYCL + kernel that calls this function may be less than "32". The function + "dpct::permute_sub_group_by_xor" may return an unexpected result on the + CPU device. Modify the size of the work-group to ensure that the value + of the right-most dimension is a multiple of "32". + */ + x = sycl::fmax(x, dpct::permute_sub_group_by_xor( + item_ct1.get_sub_group(), x, mask)); + } + return x; +} + +static __dpct_inline__ float op_repeat(const float a, const float b) { + return b; +} + +static __dpct_inline__ float op_add(const float a, const float b) { + return a + b; +} + +static __dpct_inline__ float op_mul(const float a, const float b) { + return a * b; +} + +static __dpct_inline__ float op_div(const float a, const float b) { + return a / b; +} + +template +static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst, + int ne0, int ne1, int ne2, int ne3, + int ne10, int ne11, int ne12, int ne13, + /*int s0, */ int s1, int s2, int s3, + /*int s10,*/ int s11, int s12, int s13, + const sycl::nd_item<3> &item_ct1) { + const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0)) / + ne3; + const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0)) % + ne3; + + if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { + return; + } + + const int i11 = i1 % ne11; + const int i12 = i2 % ne12; + const int i13 = i3 % ne13; + + const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; + const size_t i_dst = i_src0; + + const src0_t * src0_row = src0 + i_src0; + const src1_t * src1_row = src1 + i_src1; + dst_t * dst_row = dst + i_dst; + + for (int i0 = i0s; i0 < ne0; + i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) { + const int i10 = i0 % ne10; + dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); + } +} + +template +static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst, + int ne0, int ne1, int ne2, int ne3, + int ne10, int ne11, int ne12, int ne13, + /*int s0, */ int s1, int s2, int s3, + /*int s10,*/ int s11, int s12, int s13, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + const int i3 = i/(ne2*ne1*ne0); + const int i2 = (i/(ne1*ne0)) % ne2; + const int i1 = (i/ne0) % ne1; + const int i0 = i % ne0; + + if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) { + return; + } + + const int i11 = i1 % ne11; + const int i12 = i2 % ne12; + const int i13 = i3 % ne13; + + const size_t i_src0 = i3*s3 + i2*s2 + i1*s1; + const size_t i_src1 = i13*s13 + i12*s12 + i11*s11; + const size_t i_dst = i_src0; + + const src0_t * src0_row = src0 + i_src0; + const src1_t * src1_row = src1 + i_src1; + dst_t * dst_row = dst + i_dst; + + const int i10 = i0 % ne10; + dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]); +} + +static void acc_f32(const float * x, const float * y, float * dst, const int ne, + const int ne10, const int ne11, const int ne12, + const int nb1, const int nb2, int offset, const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + if (i >= ne) { + return; + } + int src1_idx = i - offset; + int oz = src1_idx / nb2; + int oy = (src1_idx - (oz * nb2)) / nb1; + int ox = src1_idx % nb1; + if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) { + dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11]; + } else { + dst[i] = x[i]; + } +} + +static void gelu_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const float GELU_COEF_A = 0.044715f; + const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + float xi = x[i]; + dst[i] = 0.5f * xi * + (1.0f + + sycl::tanh(SQRT_2_OVER_PI * xi * (1.0f + GELU_COEF_A * xi * xi))); +} + +static void silu_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + dst[i] = x[i] / (1.0f + sycl::native::exp(-x[i])); +} + +static void gelu_quick_f32(const float *x, float *dst, int k, + const sycl::nd_item<3> &item_ct1) { + const float GELU_QUICK_COEF = -1.702f; + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + if (i >= k) { + return; + } + dst[i] = x[i] * (1.0f / (1.0f + sycl::native::exp(GELU_QUICK_COEF * x[i]))); +} + +static void tanh_f32(const float *x, float *dst, int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + if (i >= k) { + return; + } + dst[i] = sycl::tanh((float)(x[i])); +} + +static void relu_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + dst[i] = sycl::fmax((float)(x[i]), (float)0); +} + +static void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + if (i >= k) { + return; + } + dst[i] = sycl::fmax((float)(x[i]), (float)0) + + sycl::fmin((float)(x[i]), 0.0f) * negative_slope; +} + +static void sqr_f32(const float * x, float * dst, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + dst[i] = x[i] * x[i]; +} + +static void norm_f32(const float * x, float * dst, const int ncols, const float eps, + const sycl::nd_item<3> &item_ct1, sycl::float2 *s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + + sycl::float2 mean_var = sycl::float2(0.f, 0.f); + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row*ncols + col]; + mean_var.x() += xi; + mean_var.y() += xi * xi; + } + + // sum up partial sums + mean_var = warp_reduce_sum(mean_var, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = mean_var; + } + /* + DPCT1118:0: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + mean_var = s_sum[lane_id]; + mean_var = warp_reduce_sum(mean_var, item_ct1); + } + + const float mean = mean_var.x() / ncols; + const float var = mean_var.y() / ncols - mean * mean; + const float inv_std = sycl::rsqrt(var + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std; + } +} + +static void concat_f32(const float *x,const float *y, float *dst, const int ne0, const int ne02, + const sycl::nd_item<3> &item_ct1) { + int nidx = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (nidx >= ne0) { + return; + } + // operation + int offset_dst = nidx + item_ct1.get_group(1) * ne0 + + item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1); + if (item_ct1.get_group(0) < ne02) { // src0 + int offset_src = + nidx + item_ct1.get_group(1) * ne0 + + item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1); + dst[offset_dst] = x[offset_src]; + } else { + int offset_src = + nidx + item_ct1.get_group(1) * ne0 + + (item_ct1.get_group(0) - ne02) * ne0 * item_ct1.get_group_range(1); + dst[offset_dst] = y[offset_src]; + } +} + +static void upscale_f32(const float *x, float *dst, const int ne00, const int nb02, const int scale_factor, + const sycl::nd_item<3> &item_ct1) { + int ne0 = ne00 * scale_factor; + int nidx = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (nidx >= ne0) { + return; + } + // operation + int i00 = nidx / scale_factor; + int i01 = item_ct1.get_group(1) / scale_factor; + int offset_src = i00 + i01 * ne00 + item_ct1.get_group(0) * nb02; + int offset_dst = nidx + item_ct1.get_group(1) * ne0 + + item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1); + dst[offset_dst] = x[offset_src]; +} + +static void pad_f32(const float *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02, + const sycl::nd_item<3> &item_ct1) { + int nidx = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (nidx >= ne0) { + return; + } + + // operation + int offset_dst = nidx + item_ct1.get_group(1) * ne0 + + item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1); + if (nidx < ne00 && item_ct1.get_group(1) < ne01 && + item_ct1.get_group(0) < ne02) { + int offset_src = nidx + item_ct1.get_group(1) * ne00 + + item_ct1.get_group(0) * ne00 * ne01; + dst[offset_dst] = x[offset_src]; + } else { + dst[offset_dst] = 0.0f; + } +} + +static void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps, + const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { + int start = item_ct1.get_group(2) * group_size; + int end = start + group_size; + + start += item_ct1.get_local_id(2); + + if (end >= ne_elements) { + end = ne_elements; + } + + float tmp = 0.0f; // partial sum for thread in warp + + for (int j = start; j < end; j += block_size) { + tmp += x[j]; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:1: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:54: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float mean = tmp / group_size; + tmp = 0.0f; + + for (int j = start; j < end; j += block_size) { + float xi = x[j] - mean; + dst[j] = xi; + tmp += xi * xi; + } + + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:2: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:55: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + float variance = tmp / group_size; + float scale = sycl::rsqrt(variance + eps); + for (int j = start; j < end; j += block_size) { + dst[j] *= scale; + } +} + +static void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps, + const sycl::nd_item<3> &item_ct1, float *s_sum, int block_size) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + const int tid = item_ct1.get_local_id(2); + + float tmp = 0.0f; // partial sum for thread in warp + + for (int col = tid; col < ncols; col += block_size) { + const float xi = x[row*ncols + col]; + tmp += xi * xi; + } + + // sum up partial sums + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + + int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + if (lane_id == 0) { + s_sum[warp_id] = tmp; + } + /* + DPCT1118:3: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + item_ct1.barrier(sycl::access::fence_space::local_space); + tmp = s_sum[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float mean = tmp / ncols; + const float scale = sycl::rsqrt(mean + eps); + + for (int col = tid; col < ncols; col += block_size) { + dst[row*ncols + col] = scale * x[row*ncols + col]; + } +} + +static __dpct_inline__ void dequantize_q4_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q4_0 * x = (const block_q4_0 *) vx; + + const dfloat d = x[ib].d; + + const int vui = x[ib].qs[iqs]; + + v.x() = vui & 0xF; + v.y() = vui >> 4; + +#ifdef GGML_SYCL_F16 + // v = v - {8.0f, 8.0f}; + // v = v * {d, d}; + v.s0() = (v.s0() - 8.0f) * d; + v.s1() = (v.s1() - 8.0f) * d; + +#else + v.x() = (v.x() - 8.0f) * d; + v.y() = (v.y() - 8.0f) * d; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q4_1(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q4_1 * x = (const block_q4_1 *) vx; + + const dfloat d = x[ib].dm[0]; + const dfloat m = x[ib].dm[1]; + + const int vui = x[ib].qs[iqs]; + + v.x() = vui & 0xF; + v.y() = vui >> 4; + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + // v = v + {m, m}; + v.s0() = (v.s0() * d) + m; + v.s1() = (v.s1() * d) + m; + +#else + v.x() = (v.x() * d) + m; + v.y() = (v.y() * d) + m; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q5_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q5_0 * x = (const block_q5_0 *) vx; + + const dfloat d = x[ib].d; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_SYCL_F16 + // v = v - {16.0f, 16.0f}; + // v = v * {d, d}; + v.s0() = (v.s0() - 16.0f) * d; + v.s1() = (v.s1() - 16.0f) * d; + +#else + v.x() = (v.x() - 16.0f) * d; + v.y() = (v.y() - 16.0f) * d; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q5_1(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q5_1 * x = (const block_q5_1 *) vx; + + const dfloat d = x[ib].dm[0]; + const dfloat m = x[ib].dm[1]; + + uint32_t qh; + memcpy(&qh, x[ib].qh, sizeof(qh)); + + const int xh_0 = ((qh >> (iqs + 0)) << 4) & 0x10; + const int xh_1 = ((qh >> (iqs + 12)) ) & 0x10; + + v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0); + v.y() = ((x[ib].qs[iqs] >> 4) | xh_1); + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + // v = v + {m, m}; + v.s0() = (v.s0() * d) + m; + v.s1() = (v.s1() * d) + m; +#else + v.x() = (v.x() * d) + m; + v.y() = (v.y() * d) + m; +#endif // GGML_SYCL_F16 +} + +static __dpct_inline__ void dequantize_q8_0(const void *vx, const int ib, + const int iqs, dfloat2 &v) { + const block_q8_0 * x = (const block_q8_0 *) vx; + + const dfloat d = x[ib].d; + + v.x() = x[ib].qs[iqs + 0]; + v.y() = x[ib].qs[iqs + 1]; + +#ifdef GGML_SYCL_F16 + // v = v * {d, d}; + v.s0() *= d; + v.s1() *= d; +#else + v.x() *= d; + v.y() *= d; +#endif // GGML_SYCL_F16 +} + +//================================== k-quants + +template +static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_q2_K * x = (const block_q2_K *) vx; + + const int tid = item_ct1.get_local_id(2); +#if QK_K == 256 + const int n = tid/32; + const int l = tid - 32*n; + const int is = 8*n + l/16; + + const uint8_t q = x[i].qs[32*n + l]; + dst_t * y = yy + i*QK_K + 128*n; + + float dall = x[i].dm[0]; + float dmin = x[i].dm[1]; + y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4); + y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4); + y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4); +#else + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const uint8_t q = x[i].qs[il] >> (2*is); + dst_t * y = yy + i*QK_K + 16*is + il; + float dall = __low2half(x[i].dm); + float dmin = __high2half(x[i].dm); + y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4); + y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4); +#endif + +} + +template +static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + + const int i = item_ct1.get_group(2); + const block_q3_K * x = (const block_q3_K *) vx; + +#if QK_K == 256 + const int r = item_ct1.get_local_id(2) / 4; + const int tid = r/2; + const int is0 = r%2; + const int l0 = 16 * is0 + 4 * (item_ct1.get_local_id(2) % 4); + const int n = tid / 4; + const int j = tid - 4*n; + + uint8_t m = 1 << (4*n + j); + int is = 8*n + 2*j + is0; + int shift = 2*j; + + int8_t us = is < 4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) : + is < 8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) : + is < 12 ? (x[i].scales[is-8] >> 4) | (((x[i].scales[is+0] >> 4) & 3) << 4) : + (x[i].scales[is-8] >> 4) | (((x[i].scales[is-4] >> 6) & 3) << 4); + float d_all = x[i].d; + float dl = d_all * (us - 32); + + dst_t * y = yy + i*QK_K + 128*n + 32*j; + const uint8_t * q = x[i].qs + 32*n; + const uint8_t * hm = x[i].hmask; + + for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4)); +#else + const int tid = threadIdx.x; + const int is = tid/16; // 0 or 1 + const int il = tid%16; // 0...15 + const int im = il/8; // 0...1 + const int in = il%8; // 0...7 + + dst_t * y = yy + i*QK_K + 16*is + il; + + const uint8_t q = x[i].qs[il] >> (2*is); + const uint8_t h = x[i].hmask[in] >> (2*is + im); + const float d = (float)x[i].d; + + if (is == 0) { + y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } else { + y[ 0] = d * ((x[i].scales[0] >> 4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4)); + y[32] = d * ((x[i].scales[1] >> 4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4)); + } +#endif + +} + +#if QK_K == 256 +static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) { + if (j < 4) { + d = q[j] & 63; m = q[j + 4] & 63; + } else { + d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4); + m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4); + } +} +#endif + +template +static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q4_K * x = (const block_q4_K *) vx; + + const int i = item_ct1.get_group(2); + +#if QK_K == 256 + // assume 32 threads + const int tid = item_ct1.get_local_id(2); + const int il = tid/8; + const int ir = tid%8; + const int is = 2*il; + const int n = 4; + + dst_t * y = yy + i*QK_K + 64*il + n*ir; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint8_t * q = x[i].qs + 32*il + n*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + for (int l = 0; l < n; ++l) { + y[l + 0] = d1 * (q[l] & 0xF) - m1; + y[l +32] = d2 * (q[l] >> 4) - m2; + } +#else + const int tid = threadIdx.x; + const uint8_t * q = x[i].qs; + dst_t * y = yy + i*QK_K; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4); + y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >> 4) - m * (x[i].scales[1] >> 4); +#endif +} + +template +static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q5_K * x = (const block_q5_K *) vx; + + const int i = item_ct1.get_group(2); + +#if QK_K == 256 + // assume 64 threads - this is very slightly better than the one below + const int tid = item_ct1.get_local_id(2); + const int il = tid/16; // il is in 0...3 + const int ir = tid%16; // ir is in 0...15 + const int is = 2*il; // is is in 0...6 + + dst_t * y = yy + i*QK_K + 64*il + 2*ir; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint8_t * ql = x[i].qs + 32*il + 2*ir; + const uint8_t * qh = x[i].qh + 2*ir; + + uint8_t sc, m; + get_scale_min_k4(is + 0, x[i].scales, sc, m); + const float d1 = dall * sc; const float m1 = dmin * m; + get_scale_min_k4(is + 1, x[i].scales, sc, m); + const float d2 = dall * sc; const float m2 = dmin * m; + + uint8_t hm = 1 << (2*il); + y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1; + y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1; + hm <<= 1; + y[32] = d2 * ((ql[ 0] >> 4) + (qh[ 0] & hm ? 16 : 0)) - m2; + y[33] = d2 * ((ql[ 1] >> 4) + (qh[ 1] & hm ? 16 : 0)) - m2; +#else + const int tid = threadIdx.x; + const uint8_t q = x[i].qs[tid]; + const int im = tid/8; // 0...3 + const int in = tid%8; // 0...7 + const int is = tid/16; // 0 or 1 + const uint8_t h = x[i].qh[in] >> im; + const float d = x[i].d; + dst_t * y = yy + i*QK_K + tid; + y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16)); + y[32] = d * x[i].scales[is+2] * ((q >> 4) - ((h >> 4) & 1 ? 0 : 16)); +#endif +} + +template +static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy, + const sycl::nd_item<3> &item_ct1) { + const block_q6_K * x = (const block_q6_K *) vx; + + const int i = item_ct1.get_group(2); +#if QK_K == 256 + + // assume 64 threads - this is very slightly better than the one below + const int tid = item_ct1.get_local_id(2); + const int ip = tid/32; // ip is 0 or 1 + const int il = tid - 32*ip; // 0...32 + const int is = 8*ip + il/16; + + dst_t * y = yy + i*QK_K + 128*ip + il; + + const float d = x[i].d; + + const uint8_t * ql = x[i].ql + 64*ip + il; + const uint8_t qh = x[i].qh[32*ip + il]; + const int8_t * sc = x[i].scales + is; + + y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32); + y[64] = d * sc[4] * ((int8_t)((ql[ 0] >> 4) | (((qh >> 4) & 3) << 4)) - 32); + y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32); +#else + + // assume 32 threads + const int tid = threadIdx.x; + const int ip = tid/16; // 0 or 1 + const int il = tid - 16*ip; // 0...15 + + dst_t * y = yy + i*QK_K + 16*ip + il; + + const float d = x[i].d; + + const uint8_t ql = x[i].ql[16*ip + il]; + const uint8_t qh = x[i].qh[il] >> (2*ip); + const int8_t * sc = x[i].scales; + + y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32); + y[32] = d * sc[ip+2] * ((int8_t)((ql >> 4) | (((qh >> 4) & 3) << 4)) - 32); +#endif +} + +/* +DPCT1110:4: The total declared local variable size in device function +dequantize_mul_mat_vec_q2_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q2_K * x = (const block_q2_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 16/K_QUANTS_PER_ITERATION; + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int s_offset = 8*im; + const int y_offset = 128*im + l0; + + uint32_t aux[4]; + const uint8_t * d = (const uint8_t *)aux; + const uint8_t * m = (const uint8_t *)(aux + 2); + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset); + aux[0] = a[0] & 0x0f0f0f0f; + aux[1] = a[1] & 0x0f0f0f0f; + aux[2] = (a[0] >> 4) & 0x0f0f0f0f; + aux[3] = (a[1] >> 4) & 0x0f0f0f0f; + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3) + + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3) + + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3) + + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3) + + y[l+16] * d[1] * ((q[l+16] >> 0) & 3) + + y[l+48] * d[3] * ((q[l+16] >> 2) & 3) + + y[l+80] * d[5] * ((q[l+16] >> 4) & 3) + +y[l+112] * d[7] * ((q[l+16] >> 6) & 3); + sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6] + + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7]; + + } + tmp += dall * sum1 - dmin * sum2; + + } +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; + + uint32_t uaux[2]; + const uint8_t * d = (const uint8_t *)uaux; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint32_t * s = (const uint32_t *)x[i].scales; + + uaux[0] = s[0] & 0x0f0f0f0f; + uaux[1] = (s[0] >> 4) & 0x0f0f0f0f; + + const float2 dall = __half22float2(x[i].dm); + + float sum1 = 0, sum2 = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t ql = q[l]; + sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3) + + y[l+16] * d[1] * ((ql >> 2) & 3) + + y[l+32] * d[2] * ((ql >> 4) & 3) + + y[l+48] * d[3] * ((ql >> 6) & 3); + sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7]; + } + tmp += dall.x * sum1 - dall.y * sum2; + } +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:5: The total declared local variable size in device function +dequantize_mul_mat_vec_q3_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q3_K * x = (const block_q3_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + + const uint16_t kmask1 = 0x0303; + const uint16_t kmask2 = 0x0f0f; + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int n = K_QUANTS_PER_ITERATION; // iterations in the inner loop + const int step = 16/K_QUANTS_PER_ITERATION; + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0....15 or 0...7 + + const uint8_t m = 1 << (4*im); + + const int l0 = n*in; // 0...15 or 0...14 in steps of 2 + const int q_offset = 32*im + l0; + const int y_offset = 128*im + l0; + + uint16_t utmp[4]; + const int8_t * s = (const int8_t *)utmp; + + const uint16_t s_shift = 4*im; + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * q = x[i].qs + q_offset; + const uint8_t * h = x[i].hmask + l0; + + const uint16_t * a = (const uint16_t *)x[i].scales; + utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4); + utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4); + utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4); + utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4); + + const float d = x[i].d; + + float sum = 0; + for (int l = 0; l < n; ++l) { + sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4)) + + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4)) + + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4)) + + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4)); + sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4)) + + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4)) + + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4)) + + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4)); + } + tmp += d * sum; + + } +#else + + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 or 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0....1 or 0...3 + const int offset = tid * K_QUANTS_PER_ITERATION; // 0...15 or 0...14 + const int in = offset/8; // 0 or 1 + const int im = offset%8; // 0...7 + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + offset; + const uint8_t * q = x[i].qs + offset; + const uint8_t * s = x[i].scales; + + const float dall = (float)x[i].d; + + float sum = 0; + for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) { + const uint8_t hl = x[i].hmask[im+l] >> in; + const uint8_t ql = q[l]; + sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4)) + + y[l+16] * dall * ((s[0] >> 4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4)) + + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4)) + + y[l+48] * dall * ((s[1] >> 4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:6: The total declared local variable size in device function +dequantize_mul_mat_vec_q4_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q4_K * x = (const block_q4_K *)vx + ib0; + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1 + + const int step = 8/K_QUANTS_PER_ITERATION; // 8 or 4 + + const int il = tid/step; // 0...3 + const int ir = tid - step*il; // 0...7 or 0...3 + const int n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + +#if K_QUANTS_PER_ITERATION == 2 + uint32_t q32[4]; + const uint8_t * q4 = (const uint8_t *)q32; +#else + uint16_t q16[4]; + const uint8_t * q4 = (const uint8_t *)q16; +#endif + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + +#if K_QUANTS_PER_ITERATION == 2 + const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset); + const uint32_t * q2 = q1 + 16; + + q32[0] = q1[0] & 0x0f0f0f0f; + q32[1] = q1[0] & 0xf0f0f0f0; + q32[2] = q2[0] & 0x0f0f0f0f; + q32[3] = q2[0] & 0xf0f0f0f0; + + sycl::float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 4; ++l) { + s.x() += y1[l] * q4[l + 0]; s.y() += y1[l + 32] * q4[l + 4]; + s.z() += y2[l] * q4[l + 8]; s.w() += y2[l + 32] * q4[l + 12]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f / 16.f + + s.z() * sc[4] + s.w() * sc[5] * 1.f / 16.f) - + dmin * smin; +#else + const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset); + const uint16_t * q2 = q1 + 32; + + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[0] & 0xf0f0; + q16[2] = q2[0] & 0x0f0f; + q16[3] = q2[0] & 0xf0f0; + + float4 s = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + for (int l = 0; l < 2; ++l) { + s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2]; + s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6]; + smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7]; + } + tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin; +#endif + + } +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + + const int step = tid * K_QUANTS_PER_ITERATION; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + float tmp = 0; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const float * y = yy + i*QK_K + step; + const uint16_t * a = (const uint16_t *)x[i].scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + const float d = (float)x[i].dm[0]; + const float m = (float)x[i].dm[1]; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2]) + + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2]) + + y[j+32] * (d * s[1] * (q[j+ 0] >> 4) - m * s[3]) + + y[j+48] * (d * s[1] * (q[j+16] >> 4) - m * s[3]); + } + tmp += sum; + } + +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { + dst[row] = tmp; + } +} + +/* +DPCT1110:7: The total declared local variable size in device function +dequantize_mul_mat_vec_q5_k exceeds 128 bytes and may cause high register +pressure. Consult with your hardware vendor to find the total register size +available and adjust the code, or use smaller sub-group size to avoid high +register pressure. +*/ +static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx, + const float *__restrict__ yy, + float *__restrict__ dst, + const int ncols, + const sycl::nd_item<3> &item_ct1) { + + const int row = item_ct1.get_group(2); + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q5_K * x = (const block_q5_K *)vx + ib0; + + float tmp = 0; // partial sum for thread in warp + +#if QK_K == 256 + const uint16_t kmask1 = 0x3f3f; + const uint16_t kmask2 = 0x0f0f; + const uint16_t kmask3 = 0xc0c0; + + const int tid = item_ct1.get_local_id(2) / 2; // 0...15 + const int ix = item_ct1.get_local_id(2) % 2; + + const int il = tid/4; // 0...3 + const int ir = tid - 4*il;// 0...3 + const int n = 2; + + const int im = il/2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 + const int in = il%2; + + const int l0 = n*(2*ir + in); + const int q_offset = 32*im + l0; + const int y_offset = 64*im + l0; + + const uint8_t hm1 = 1 << (2*im); + const uint8_t hm2 = hm1 << 4; + + uint16_t aux[4]; + const uint8_t * sc = (const uint8_t *)aux; + + uint16_t q16[8]; + const uint8_t * q4 = (const uint8_t *)q16; + + for (int i = ix; i < num_blocks_per_row; i += 2) { + + const uint8_t * ql1 = x[i].qs + q_offset; + const uint8_t * qh = x[i].qh + l0; + const float * y1 = yy + i*QK_K + y_offset; + const float * y2 = y1 + 128; + + const float dall = x[i].dm[0]; + const float dmin = x[i].dm[1]; + + const uint16_t * a = (const uint16_t *)x[i].scales; + aux[0] = a[im+0] & kmask1; + aux[1] = a[im+2] & kmask1; + aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2); + aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2); + + sycl::float4 sum = {0.f, 0.f, 0.f, 0.f}; + float smin = 0; + const uint16_t * q1 = (const uint16_t *)ql1; + const uint16_t * q2 = q1 + 32; + q16[0] = q1[0] & 0x0f0f; + q16[1] = q1[8] & 0x0f0f; + q16[2] = (q1[0] >> 4) & 0x0f0f; + q16[3] = (q1[8] >> 4) & 0x0f0f; + q16[4] = q2[0] & 0x0f0f; + q16[5] = q2[8] & 0x0f0f; + q16[6] = (q2[0] >> 4) & 0x0f0f; + q16[7] = (q2[8] >> 4) & 0x0f0f; + for (int l = 0; l < n; ++l) { + sum.x() += + y1[l + 0] * (q4[l + 0] + (qh[l + 0] & (hm1 << 0) ? 16 : 0)) + + y1[l + 16] * (q4[l + 2] + (qh[l + 16] & (hm1 << 0) ? 16 : 0)); + sum.y() += + y1[l + 32] * (q4[l + 4] + (qh[l + 0] & (hm1 << 1) ? 16 : 0)) + + y1[l + 48] * (q4[l + 6] + (qh[l + 16] & (hm1 << 1) ? 16 : 0)); + sum.z() += + y2[l + 0] * (q4[l + 8] + (qh[l + 0] & (hm2 << 0) ? 16 : 0)) + + y2[l + 16] * (q4[l + 10] + (qh[l + 16] & (hm2 << 0) ? 16 : 0)); + sum.w() += + y2[l + 32] * (q4[l + 12] + (qh[l + 0] & (hm2 << 1) ? 16 : 0)) + + y2[l + 48] * (q4[l + 14] + (qh[l + 16] & (hm2 << 1) ? 16 : 0)); + smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3] + + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7]; + } + tmp += dall * (sum.x() * sc[0] + sum.y() * sc[1] + sum.z() * sc[4] + + sum.w() * sc[5]) - + dmin * smin; + } + +#else + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...15 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); + const int step = tid * K_QUANTS_PER_ITERATION; + const int im = step/8; + const int in = step%8; + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + const uint8_t * q = x[i].qs + step; + const int8_t * s = x[i].scales; + const float * y = yy + i*QK_K + step; + const float d = x[i].d; + float sum = 0.f; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + const uint8_t h = x[i].qh[in+j] >> im; + sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16)) + + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16)) + + y[j+32] * d * s[2] * ((q[j+ 0] >> 4) - ((h >> 4) & 1 ? 0 : 16)) + + y[j+48] * d * s[3] * ((q[j+16] >> 4) - ((h >> 6) & 1 ? 0 : 16)); + } + tmp += sum; + } +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows, + const sycl::nd_item<3> &item_ct1) { + + static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); + + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + if (row > nrows) return; + + const int num_blocks_per_row = ncols / QK_K; + const int ib0 = row*num_blocks_per_row; + + const block_q6_K * x = (const block_q6_K *)vx + ib0; + +#if QK_K == 256 + + const int tid = + item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16 + const int ix = + item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0, 1 + + const int step = 16/K_QUANTS_PER_ITERATION; // 16 or 8 + + const int im = tid/step; // 0 or 1. 0 computes 0..., 1 computes 128... + const int in = tid - step*im; // 0...15 or 0...7 + +#if K_QUANTS_PER_ITERATION == 1 + const int l0 = K_QUANTS_PER_ITERATION*in; // 0...15 + const int is = 0; +#else + const int l0 = 4 * in; // 0, 4, 8, ..., 28 + const int is = in / 4; +#endif + const int ql_offset = 64*im + l0; + const int qh_offset = 32*im + l0; + const int s_offset = 8*im + is; + const int y_offset = 128*im + l0; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + y_offset; + const uint8_t * ql = x[i].ql + ql_offset; + const uint8_t * qh = x[i].qh + qh_offset; + const int8_t * s = x[i].scales + s_offset; + + const float d = x[i].d; + +#if K_QUANTS_PER_ITERATION == 1 + float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32) + + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32) + + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32) + + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32) + + y[64] * s[4] * d * ((int8_t)((ql[ 0] >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32) + + y[80] * s[5] * d * ((int8_t)((ql[16] >> 4) | ((qh[16] & 0x30) >> 0)) - 32) + + y[96] * s[6] * d * ((int8_t)((ql[32] >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32) + +y[112] * s[7] * d * ((int8_t)((ql[48] >> 4) | ((qh[16] & 0xc0) >> 2)) - 32); + tmp += sum; +#else + float sum = 0; + for (int l = 0; l < 4; ++l) { + sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32) + + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32) + + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32) + + y[l+96] * s[6] * d * ((int8_t)((ql[l+32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32); + } + tmp += sum; +#endif + + } + +#else + + const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION); // 0...7 + const int ix = threadIdx.x%(2*K_QUANTS_PER_ITERATION); // 0...3 + + const int step = tid * K_QUANTS_PER_ITERATION; + + float tmp = 0; // partial sum for thread in warp + + for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) { + + const float * y = yy + i * QK_K + step; + const uint8_t * ql = x[i].ql + step; + const uint8_t * qh = x[i].qh + step; + const int8_t * s = x[i].scales; + + const float d = x[i+0].d; + + float sum = 0; + for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) { + sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32) + + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32) + + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >> 4) | ((qh[j] & 0x30) >> 0)) - 32) + + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >> 4) | ((qh[j] & 0xc0) >> 2)) - 32); + } + tmp += sum; + + } + +#endif + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { + dst[row] = tmp; + } +} + +static void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const sycl::half *x = (const sycl::half *)vx; + + // automatic half -> float type cast if dfloat == float + v.x() = x[ib + iqs + 0]; + v.y() = x[ib + iqs + 1]; +} + +static void convert_f32(const void * vx, const int ib, const int iqs, dfloat2 & v){ + const float * x = (const float *) vx; + + // automatic half -> float type cast if dfloat == float + v.x() = x[ib + iqs + 0]; + v.y() = x[ib + iqs + 1]; +} + +static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded, + const sycl::nd_item<3> &item_ct1) { + const int ix = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (ix >= kx_padded) { + return; + } + + const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + + const int i_padded = iy*kx_padded + ix; + + block_q8_1 * y = (block_q8_1 *) vy; + + const int ib = i_padded / QK8_1; // block index + const int iqs = i_padded % QK8_1; // quant index + + const float xi = ix < kx ? x[iy*kx + ix] : 0.0f; + float amax = sycl::fabs((float)xi); + float sum = xi; + +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + amax = sycl::fmax(amax, dpct::permute_sub_group_by_xor( + item_ct1.get_sub_group(), amax, mask)); + sum += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), sum, mask); + } + + const float d = amax / 127; + const int8_t q = amax == 0.0f ? 0 : sycl::round(xi / d); + + y[ib].qs[iqs] = q; + + if (iqs > 0) { + return; + } + + reinterpret_cast(y[ib].ds.x()) = d; + reinterpret_cast(y[ib].ds.y()) = sum; +} + +template +static void k_get_rows( + const void * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { + + const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2)) * + 2; + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03; + + const int ib = i00/qk; // block index + const int iqs = (i00%qk)/qr; // quant index + const int iybs = i00 - i00%qk; // dst block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(src0_row, ib, iqs, v); + + dst_row[iybs + iqs + 0] = v.x(); + dst_row[iybs + iqs + y_offset] = v.y(); +} + +template +static void k_get_rows_float( + const src0_t * src0, const int32_t * src1, dst_t * dst, + int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/ + /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/ + /*size_t s0,*/ size_t s1, size_t s2, size_t s3, + /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03, + size_t s10, size_t s11, size_t s12, + const sycl::nd_item<3> &item_ct1/*, size_t s13*/) { + + const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2); + const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) / + ne12; + const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) + + item_ct1.get_local_id(0)) % + ne12; + + if (i00 >= ne00) { + return; + } + + const int i01 = src1[i10*s10 + i11*s11 + i12*s12]; + + dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3; + const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03); + + dst_row[i00] = src0_row[i00]; +} + +template +static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + 2 * item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + const int ib = i/qk; // block index + const int iqs = (i%qk)/qr; // quant index + const int iybs = i - i%qk; // y block start index + const int y_offset = qr == 1 ? 1 : qk/2; + + // dequantize + dfloat2 v; + dequantize_kernel(vx, ib, iqs, v); + + y[iybs + iqs + 0] = v.x(); + y[iybs + iqs + y_offset] = v.y(); +} + +// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called +// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q + +#define VDR_Q4_0_Q8_1_MMVQ 2 +#define VDR_Q4_0_Q8_1_MMQ 4 + +template +static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u, + const float &d4, + const sycl::half2 &ds8) { + int sumi = 0; +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); + sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + } + + const sycl::float2 ds8f = + ds8.convert(); + + // second part effectively subtracts 8 from each quant value + return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y()); +} + +#define VDR_Q4_1_Q8_1_MMVQ 2 +#define VDR_Q4_1_Q8_1_MMQ 4 + +template +static __dpct_inline__ float vec_dot_q4_1_q8_1_impl(const int *v, const int *u, + const sycl::half2 &dm4, + const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + const int vi0 = (v[i] >> 0) & 0x0F0F0F0F; + const int vi1 = (v[i] >> 4) & 0x0F0F0F0F; + + // SIMD dot product of quantized values + sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi); + sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi); + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm4 * ds8).convert(); + const float d4d8 = tmp.x(); + const float m4s8 = tmp.y(); +#else + const sycl::float2 dm4f = + dm4.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d4d8 = dm4f.x() * ds8f.x(); + const float m4s8 = dm4f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it + return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1)); +} + +#define VDR_Q5_0_Q8_1_MMVQ 2 +#define VDR_Q5_0_Q8_1_MMQ 4 + +template +static __dpct_inline__ float +vec_dot_q5_0_q8_1_impl(const int *vl, const int *vh, const int *u, + const float &d5, const sycl::half2 &ds8) { + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = dpct::dp4a(vi0, u[2 * i + 0], + sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = dpct::dp4a(vi1, u[2 * i + 1], + sumi); // SIMD dot product of quantized values + } + + const sycl::float2 ds8f = + ds8.convert(); + + // second part effectively subtracts 16 from each quant value + return d5 * (sumi * ds8f.x() - (16 * vdr / QI5_0) * ds8f.y()); +} + +#define VDR_Q5_1_Q8_1_MMVQ 2 +#define VDR_Q5_1_Q8_1_MMQ 4 + +template +static __dpct_inline__ float +vec_dot_q5_1_q8_1_impl(const int *vl, const int *vh, const int *u, + const sycl::half2 &dm5, const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + int vi0 = (vl[i] >> 0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits + vi0 |= (vh[i] << 4) & 0x00000010; // 0 -> 4 + vi0 |= (vh[i] << 11) & 0x00001000; // 1 -> 12 + vi0 |= (vh[i] << 18) & 0x00100000; // 2 -> 20 + vi0 |= (vh[i] << 25) & 0x10000000; // 3 -> 28 + sumi = dpct::dp4a(vi0, u[2 * i + 0], + sumi); // SIMD dot product of quantized values + + int vi1 = (vl[i] >> 4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits + vi1 |= (vh[i] >> 12) & 0x00000010; // 16 -> 4 + vi1 |= (vh[i] >> 5) & 0x00001000; // 17 -> 12 + vi1 |= (vh[i] << 2) & 0x00100000; // 18 -> 20 + vi1 |= (vh[i] << 9) & 0x10000000; // 19 -> 28 + sumi = dpct::dp4a(vi1, u[2 * i + 1], + sumi); // SIMD dot product of quantized values + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm5 * ds8).convert(); + const float d5d8 = tmp.x(); + const float m5s8 = tmp.y(); + + +#else + const sycl::float2 dm5f = + dm5.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d5d8 = dm5f.x() * ds8f.x(); + const float m5s8 = dm5f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it + return sumi*d5d8 + m5s8 / (QI5_1 / vdr); +} + +#define VDR_Q8_0_Q8_1_MMVQ 2 +#define VDR_Q8_0_Q8_1_MMQ 8 + +template +static __dpct_inline__ float vec_dot_q8_0_q8_1_impl(const int *v, const int *u, + const float &d8_0, + const float &d8_1) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = dpct::dp4a(v[i], u[i], sumi); + } + + return d8_0*d8_1 * sumi; +} + +template +static __dpct_inline__ float vec_dot_q8_1_q8_1_impl(const int *v, const int *u, + const sycl::half2 &dm8, + const sycl::half2 &ds8) { + + int sumi = 0; + +#pragma unroll + for (int i = 0; i < vdr; ++i) { + // SIMD dot product of quantized values + sumi = dpct::dp4a(v[i], u[i], sumi); + } + +#ifdef GGML_SYCL_F16 + const sycl::float2 tmp = + (dm8 * ds8).convert(); + const float d8d8 = tmp.x(); + const float m8s8 = tmp.y(); +#else + const sycl::float2 dm8f = + dm8.convert(); + const sycl::float2 ds8f = + ds8.convert(); + const float d8d8 = dm8f.x() * ds8f.x(); + const float m8s8 = dm8f.y() * ds8f.y(); +#endif // GGML_SYCL_F16 + + // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it + return sumi*d8d8 + m8s8 / (QI8_1 / vdr); +} + +#define VDR_Q2_K_Q8_1_MMVQ 1 +#define VDR_Q2_K_Q8_1_MMQ 2 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q2_K_q8_1_impl_mmvq( + const int &v, const int *__restrict__ u, const uint8_t *__restrict__ scales, + const sycl::half2 &dm2, const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR2_K; ++i) { + const int sc = scales[2*i]; + + const int vi = (v >> (2*i)) & 0x03030303; + + sumf_d += + d8[i] * (dpct::dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + sumf_m += d8[i] * + dpct::dp4a( + m, u[i], + 0); // multiply constant q2_K part with sum of q8_1 values + } + + const sycl::float2 dm2f = + dm2.convert(); + + return dm2f.x() * sumf_d - dm2f.y() * sumf_m; +} + +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q2_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ scales, + const sycl::half2 &dm2, const float &d8) { + + int sumi_d = 0; + int sumi_m = 0; + +#pragma unroll + for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) { + int sumi_d_sc = 0; + + const int sc = scales[i0 / (QI8_1/2)]; + + // fill int with 4x m + int m = sc >> 4; + m |= m << 8; + m |= m << 16; + +#pragma unroll + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_d_sc = dpct::dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product + sumi_m = dpct::dp4a(m, u[i], + sumi_m); // multiply sum of q8_1 values with m + } + + sumi_d += sumi_d_sc * (sc & 0xF); + } + + const sycl::float2 dm2f = + dm2.convert(); + + return d8 * (dm2f.x() * sumi_d - dm2f.y() * sumi_m); +} + +#define VDR_Q3_K_Q8_1_MMVQ 1 +#define VDR_Q3_K_Q8_1_MMQ 2 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q3_K_q8_1_impl_mmvq( + const int &vl, const int &vh, const int *__restrict__ u, + const uint8_t *__restrict__ scales, const int &scale_offset, + const float &d3, const float *__restrict__ d8) { + + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + const int isc = scale_offset + 2*i; + + const int isc_low = isc % (QK_K/32); + const int sc_shift_low = 4 * (isc / (QK_K/32)); + const int sc_low = (scales[isc_low] >> sc_shift_low) & 0xF; + + const int isc_high = isc % (QK_K/64); + const int sc_shift_high = 2 * (isc / (QK_K/64)); + const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4; + + const int sc = (sc_low | sc_high) - 32; + + const int vil = (vl >> (2*i)) & 0x03030303; + + const int vih = ((vh >> i) << 2) & 0x04040404; + + const int vi = + dpct::vectorized_binary(vil, vih, dpct::sub_sat()); + + sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d3 * sumf; +} + +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q3_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const int8_t *__restrict__ scales, const float &d3, + const float &d8) { + + int sumi = 0; + +#pragma unroll + for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) { + int sumi_sc = 0; + + for (int i = i0; i < i0 + QI8_1/2; ++i) { + sumi_sc = dpct::dp4a(v[i], u[i], sumi_sc); // SIMD dot product + } + + sumi += sumi_sc * scales[i0 / (QI8_1/2)]; + } + + return d3*d8 * sumi; +} + +#define VDR_Q4_K_Q8_1_MMVQ 2 +#define VDR_Q4_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_vmmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K; ++i) { + const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F; + const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F; + + const int dot1 = + dpct::dp4a(v1i, u[2 * i + 1], + dpct::dp4a(v0i, u[2 * i + 0], 0)); // SIMD dot product + const int dot2 = + dpct::dp4a(0x01010101, u[2 * i + 1], + dpct::dp4a(0x01010101, u[2 * i + 0], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); // multiply constant part of q4_K with sum of q8_1 values + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + +// contiguous u/y values +static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_mmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = dpct::dp4a((v[j] >> (4 * i)) & 0x0F0F0F0F, + u[i * QI8_1 + j], sumi_d); // SIMD dot product + } + + const sycl::float2 ds8f = + ds8[i].convert(); + + sumf_d += ds8f.x() * (sc[i] * sumi_d); + sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + +#define VDR_Q5_K_Q8_1_MMVQ 2 +#define VDR_Q5_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_vmmq( + const int *__restrict__ vl, const int *__restrict__ vh, + const int *__restrict__ u, const uint8_t *__restrict__ sc, + const uint8_t *__restrict__ m, const sycl::half2 &dm5, + const float *__restrict__ d8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F; + const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F; + + const int vh0i = ((vh[0] >> i) << 4) & 0x10101010; + const int vh1i = ((vh[1] >> i) << 4) & 0x10101010; + + const int v0i = vl0i | vh0i; + const int v1i = vl1i | vh1i; + + const int dot1 = + dpct::dp4a(v0i, u[2 * i + 0], + dpct::dp4a(v1i, u[2 * i + 1], 0)); // SIMD dot product + const int dot2 = + dpct::dp4a(0x01010101, u[2 * i + 0], + dpct::dp4a(0x01010101, u[2 * i + 1], 0)); // sum of u + + sumf_d += d8[i] * (dot1 * sc[i]); + sumf_m += d8[i] * (dot2 * m[i]); + + } + + const sycl::float2 dm5f = + dm5.convert(); + + return dm5f.x() * sumf_d - dm5f.y() * sumf_m; +} + +// contiguous u/y values +static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_mmq( + const int *__restrict__ v, const int *__restrict__ u, + const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m, + const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) { + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + +#pragma unroll + for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) { + int sumi_d = 0; + +#pragma unroll + for (int j = 0; j < QI8_1; ++j) { + sumi_d = dpct::dp4a(v[i * QI8_1 + j], u[i * QI8_1 + j], + sumi_d); // SIMD dot product + } + + const sycl::float2 ds8f = + ds8[i].convert(); + + sumf_d += ds8f.x() * (sc[i] * sumi_d); + sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val + } + + const sycl::float2 dm4f = + dm4.convert(); + + return dm4f.x() * sumf_d - dm4f.y() * sumf_m; +} + +#define VDR_Q6_K_Q8_1_MMVQ 1 +#define VDR_Q6_K_Q8_1_MMQ 8 + +// contiguous v/x values +static __dpct_inline__ float +vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh, + const int *__restrict__ u, + const int8_t *__restrict__ scales, const float &d, + const float *__restrict__ d8) { + + float sumf = 0.0f; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + const int sc = scales[4*i]; + + const int vil = (vl >> (4*i)) & 0x0F0F0F0F; + + const int vih = ((vh >> (4*i)) << 4) & 0x30303030; + + const int vi = dpct::vectorized_binary( + (vil | vih), 0x20202020, dpct::sub_sat()); // vi = (vil | vih) - 32 + + sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product + } + + return d*sumf; +} + +// contiguous u/y values +static __dpct_inline__ float +vec_dot_q6_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u, + const int8_t *__restrict__ sc, const float &d6, + const float *__restrict__ d8) { + + float sumf_d = 0.0f; + +#pragma unroll + for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) { + sycl::int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale + +#pragma unroll + for (int i = i0; i < i0 + 2; ++i) { + sumi_d.x() = dpct::dp4a(v[2 * i + 0], u[2 * i + 0], + sumi_d.x()); // SIMD dot product + sumi_d.x() = dpct::dp4a(v[2 * i + 1], u[2 * i + 1], + sumi_d.x()); // SIMD dot product + + sumi_d.y() = dpct::dp4a(v[2 * i + 4], u[2 * i + 4], + sumi_d.y()); // SIMD dot product + sumi_d.y() = dpct::dp4a(v[2 * i + 5], u[2 * i + 5], + sumi_d.y()); // SIMD dot product + } + + sumf_d += d8[i0 / 4] * + (sc[i0 / 2 + 0] * sumi_d.x() + sc[i0 / 2 + 1] * sumi_d.y()); + } + + return d6 * sumf_d; +} + +static __dpct_inline__ float +vec_dot_q4_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq; + + int v[VDR_Q4_0_Q8_1_MMVQ]; + int u[2*VDR_Q4_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8(bq4_0->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0); + } + + return vec_dot_q4_0_q8_1_impl(v, u, bq4_0->d, bq8_1->ds); +} + +template +static __dpct_inline__ void +allocate_tiles_q4_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q4_0, float *tile_x_d_q4_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q4_0; + *x_dm = (sycl::half2 *)tile_x_d_q4_0; +} + +template +static __dpct_inline__ void +load_tiles_q4_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_0; + const int kqsx = k % QI4_0; + + const block_q4_0 * bx0 = (const block_q4_0 *) vx; + + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) { + int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q4_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const float * x_dmf = (const float *) x_dm; + + int u[2*VDR_Q4_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE]; + } + + return vec_dot_q4_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +static __dpct_inline__ float +vec_dot_q4_1_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq; + + int v[VDR_Q4_1_Q8_1_MMVQ]; + int u[2*VDR_Q4_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_uint8_aligned(bq4_1->qs, iqs + i); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1); + } + + return vec_dot_q4_1_q8_1_impl(v, u, bq4_1->dm, bq8_1->ds); +} + +template +static __dpct_inline__ void +allocate_tiles_q4_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q4_1, sycl::half2 *tile_x_dm_q4_1) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q4_1; + *x_dm = tile_x_dm_q4_1; +} + +template +static __dpct_inline__ void +load_tiles_q4_1(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_1; + const int kqsx = k % QI4_1; + + const block_q4_1 * bx0 = (const block_q4_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) { + int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm; + } +} + +static __dpct_inline__ float vec_dot_q4_1_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + + int u[2*VDR_Q4_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE]; + } + + return vec_dot_q4_1_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1], + y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +static __dpct_inline__ float +vec_dot_q5_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq; + + int vl[VDR_Q5_0_Q8_1_MMVQ]; + int vh[VDR_Q5_0_Q8_1_MMVQ]; + int u[2*VDR_Q5_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8(bq5_0->qs, iqs + i); + vh[i] = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0); + } + + return vec_dot_q5_0_q8_1_impl(vl, vh, u, bq5_0->d, bq8_1->ds); +} + +template +static __dpct_inline__ void +allocate_tiles_q5_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_0, float *tile_x_d_q5_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_ql_q5_0; + *x_dm = (sycl::half2 *)tile_x_d_q5_0; +} + +template +static __dpct_inline__ void +load_tiles_q5_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_0; + const int kqsx = k % QI5_0; + + const block_q5_0 * bx0 = (const block_q5_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8(bxi->qs, kqsx); + const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + qs0 = dpct::vectorized_binary( + qs0, 0x10101010, dpct::sub_sat()); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + qs1 = dpct::vectorized_binary( + qs1, 0x10101010, dpct::sub_sat()); // subtract 16 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_0; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) { + int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q5_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + int u[2*VDR_Q5_0_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE]; + } + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +static __dpct_inline__ float +vec_dot_q5_1_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq; + + int vl[VDR_Q5_1_Q8_1_MMVQ]; + int vh[VDR_Q5_1_Q8_1_MMVQ]; + int u[2*VDR_Q5_1_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) { + vl[i] = get_int_from_uint8_aligned(bq5_1->qs, iqs + i); + vh[i] = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i)); + u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1); + } + + return vec_dot_q5_1_q8_1_impl(vl, vh, u, bq5_1->dm, bq8_1->ds); +} + +template +static __dpct_inline__ void +allocate_tiles_q5_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_1, sycl::half2 *tile_x_dm_q5_1) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_ql_q5_1; + *x_dm = tile_x_dm_q5_1; +} + +template +static __dpct_inline__ void +load_tiles_q5_1(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_1; + const int kqsx = k % QI5_1; + + const block_q5_1 * bx0 = (const block_q5_1 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1)); + + int qs0 = (ql >> 0) & 0x0F0F0F0F; + qs0 |= (qh << 4) & 0x00000010; // 0 -> 4 + qs0 |= (qh << 11) & 0x00001000; // 1 -> 12 + qs0 |= (qh << 18) & 0x00100000; // 2 -> 20 + qs0 |= (qh << 25) & 0x10000000; // 3 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0; + + int qs1 = (ql >> 4) & 0x0F0F0F0F; + qs1 |= (qh >> 12) & 0x00000010; // 16 -> 4 + qs1 |= (qh >> 5) & 0x00001000; // 17 -> 12 + qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 + qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 + + x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_1; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) { + int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm; + } +} + +static __dpct_inline__ float vec_dot_q5_1_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); + const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; + + int u[2*VDR_Q5_1_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) { + u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l) % WARP_SIZE]; + u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE]; + } + + return vec_dot_q8_1_q8_1_impl + (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]); +} + +static __dpct_inline__ float +vec_dot_q8_0_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq; + + int v[VDR_Q8_0_Q8_1_MMVQ]; + int u[VDR_Q8_0_Q8_1_MMVQ]; + +#pragma unroll + for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) { + v[i] = get_int_from_int8(bq8_0->qs, iqs + i); + u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i); + } + + return vec_dot_q8_0_q8_1_impl(v, u, bq8_0->d, + bq8_1->ds[0]); +} + +template +static __dpct_inline__ void +allocate_tiles_q8_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_qs_q8_0, float *tile_x_d_q8_0) { + (void)x_qh; (void)x_sc; + + *x_ql = tile_x_qs_q8_0; + *x_dm = (sycl::half2 *)tile_x_d_q8_0; +} + +template +static __dpct_inline__ void +load_tiles_q8_0(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; (void)x_sc; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI8_0; + const int kqsx = k % QI8_0; + float * x_dmf = (float *) x_dm; + + const block_q8_0 * bx0 = (const block_q8_0 *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI8_0; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) { + int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d; + } +} + +static __dpct_inline__ float vec_dot_q8_0_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; (void)x_sc; + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + return vec_dot_q8_0_q8_1_impl + (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0], + y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]); +} + +static __dpct_inline__ float +vec_dot_q2_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q2_K * bq2_K = (const block_q2_K *) vbq; + + const int bq8_offset = QR2_K * (iqs / QI8_1); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const uint8_t * scales = bq2_K->scales + scale_offset; + + const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs); + int u[QR2_K]; + float d8[QR2_K]; + +#pragma unroll + for (int i = 0; i < QR2_K; ++ i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + i].ds[0]; + } + + return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8); +} + +template +static __dpct_inline__ void +allocate_tiles_q2_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q2_K, sycl::half2 *tile_x_dm_q2_K, + int *tile_x_sc_q2_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q2_K; + *x_dm = tile_x_dm_q2_K; + *x_sc = tile_x_sc_q2_K; +} + +template +static __dpct_inline__ void +load_tiles_q2_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI2_K; + const int kqsx = k % QI2_K; + + const block_q2_K * bx0 = (const block_q2_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI2_K; + const int kbxd = k % blocks_per_tile_x_row; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) { + int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4)); + } +} + +static __dpct_inline__ float vec_dot_q2_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const int kbx = k / QI2_K; + const int ky = (k % QI2_K) * QR2_K; + const float * y_df = (const float *) y_ds; + + int v[QR2_K*VDR_Q2_K_Q8_1_MMQ]; + + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2); + const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2)); + +#pragma unroll + for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) { + v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303; + } + + const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4; + + const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE; + return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]); +} + +static __dpct_inline__ float +vec_dot_q3_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q3_K * bq3_K = (const block_q3_K *) vbq; + + const int bq8_offset = QR3_K * (iqs / (QI3_K/2)); + const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2); + + const float d = bq3_K->d; + + const int vl = get_int_from_uint8(bq3_K->qs, iqs); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset; + + int u[QR3_K]; + float d8[QR3_K]; + +#pragma unroll + for (int i = 0; i < QR3_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + i].ds[0]; + } + + return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8); +} + +template +static __dpct_inline__ void +allocate_tiles_q3_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q3_K, sycl::half2 *tile_x_dm_q3_K, + int *tile_x_qh_q3_K, int *tile_x_sc_q3_K) { + + *x_ql = tile_x_ql_q3_K; + *x_dm = tile_x_dm_q3_K; + *x_qh = tile_x_qh_q3_K; + *x_sc = tile_x_sc_q3_K; +} + +template +static __dpct_inline__ void +load_tiles_q3_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI3_K; + const int kqsx = k % QI3_K; + + const block_q3_K * bx0 = (const block_q3_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI3_K; + const int kbxd = k % blocks_per_tile_x_row; + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) { + int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) { + int i = i0 + i_offset * 2 + k / (WARP_SIZE/2); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2); + + // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted + x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2)); + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) { + int i = i0 + i_offset * 4 + k / (WARP_SIZE/4); + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4); + + const int ksc = k % (QI3_K/4); + + const int ksc_low = ksc % (QI3_K/8); + const int shift_low = 4 * (ksc / (QI3_K/8)); + const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F; + + const int ksc_high = QI3_K/8; + const int shift_high = 2 * ksc; + const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030; + + const int sc = dpct::vectorized_binary( + sc_low | sc_high, 0x20202020, dpct::sub_sat()); + + x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc; + } +} + +static __dpct_inline__ float vec_dot_q3_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + + const int kbx = k / QI3_K; + const int ky = (k % QI3_K) * QR3_K; + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; + + int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; + +#pragma unroll + for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) { + const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2); + const int shift = 2 * ((ky % 32) / 8); + const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303; + + const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8); + const int vlh = (vh << 2) & 0x04040404; + + v[l] = dpct::vectorized_binary(vll, vlh, dpct::sub_sat()); + } + + const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE; + return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]); +} + +static __dpct_inline__ float +vec_dot_q4_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + +#ifndef GGML_QKK_64 + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + int v[2]; + int u[2*QR4_K]; + float d8[QR4_K]; + + // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6 + const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2)); + + // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12 + // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44 + // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76 + // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108 + + const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + v[0] = q4[0]; + v[1] = q4[4]; + + const uint16_t * scales = (const uint16_t *)bq4_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + + for (int i = 0; i < QR4_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = bq8i->ds[0]; + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8); + +#else + +#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics + const block_q4_K * bq4_K = (const block_q4_K *) vbq; + + float sumf_d = 0.0f; + float sumf_m = 0.0f; + + uint16_t aux16[2]; + const uint8_t * s = (const uint8_t *)aux16; + + const uint16_t * a = (const uint16_t *)bq4_K->scales; + aux16[0] = a[0] & 0x0f0f; + aux16[1] = (a[0] >> 4) & 0x0f0f; + + const float dall = bq4_K->dm[0]; + const float dmin = bq4_K->dm[1]; + + const float d8_1 = __low2float(bq8_1[0].ds); + const float d8_2 = __low2float(bq8_1[1].ds); + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * q4 = (const int *)bq4_K->qs + (iqs/2); + const int v1 = q4[0]; + const int v2 = q4[4]; + + const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0)); + const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0)); + const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0)); + const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0)); + + sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]); + sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]); + + return dall * sumf_d - dmin * sumf_m; + +#else + bad_arch(); +#endif // __SYCL_ARCH__ >= VER_4VEC + +#endif +} + +template +static __dpct_inline__ void +allocate_tiles_q4_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q4_K, sycl::half2 *tile_x_dm_q4_K, + int *tile_x_sc_q4_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q4_K; + *x_dm = tile_x_dm_q4_K; + *x_sc = tile_x_sc_q4_K; +} + +template +static __dpct_inline__ void +load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI4_K; // == 0 if QK_K == 256 + const int kqsx = k % QI4_K; // == k if QK_K == 256 + + const block_q4_K * bx0 = (const block_q4_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx; + + x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) { + int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm; +#else + x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]}; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + +static __dpct_inline__ float vec_dot_q4_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); + + const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE; + return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]); +} + +static __dpct_inline__ float +vec_dot_q5_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + +#ifndef GGML_QKK_64 + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + int vl[2]; + int vh[2]; + int u[2*QR5_K]; + float d8[QR5_K]; + + const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2)); + const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4)); + const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4)); + + vl[0] = ql[0]; + vl[1] = ql[4]; + + vh[0] = qh[0] >> bq8_offset; + vh[1] = qh[4] >> bq8_offset; + + const uint16_t * scales = (const uint16_t *)bq5_K->scales; + uint16_t aux[2]; + const int j = bq8_offset/2; + if (j < 2) { + aux[0] = scales[j+0] & 0x3f3f; + aux[1] = scales[j+2] & 0x3f3f; + } else { + aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2); + aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2); + } + const uint8_t * sc = (const uint8_t *)aux; + const uint8_t * m = sc + 2; + +#pragma unroll + for (int i = 0; i < QR5_K; ++i) { + const block_q8_1 * bq8i = bq8_1 + bq8_offset + i; + d8[i] = bq8i->ds[0]; + + const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4); + u[2*i+0] = q8[0]; + u[2*i+1] = q8[4]; + } + + return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8); + +#else + +#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics + const block_q5_K * bq5_K = (const block_q5_K *) vbq; + + const int8_t * s = bq5_K->scales; + + const float d = bq5_K->d; + + const float d8_1 = __low2half(bq8_1[0].ds); + const float d8_2 = __low2half(bq8_1[1].ds); + + const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2)); + const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4); + const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2)); + const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4); + + const int * ql = (const int *)bq5_K->qs + (iqs/2); + const int vl1 = ql[0]; + const int vl2 = ql[4]; + + const int step = 4 * (iqs/2); // 0, 4, 8, 12 + const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6 + const int in = step%8; // 0, 4, 0, 4 + const int vh = (*((const int *)(bq5_K->qh + in))) >> im; + + const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f); + const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f); + const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f); + const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f); + + const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1]) + + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]); + + return d * sumf_d; + +#else + bad_arch(); +#endif // __SYCL_ARCH__ >= VER_4VEC + +#endif +} + +template +static __dpct_inline__ void +allocate_tiles_q5_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql_q5_K, sycl::half2 *tile_x_dm_q5_K, + int *tile_x_sc_q5_K) { + (void)x_qh; + + *x_ql = tile_x_ql_q5_K; + *x_dm = tile_x_dm_q5_K; + *x_sc = tile_x_sc_q5_K; +} + +template +static __dpct_inline__ void +load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI5_K; // == 0 if QK_K == 256 + const int kqsx = k % QI5_K; // == k if QK_K == 256 + + const block_q5_K * bx0 = (const block_q5_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR5_K*kqsx; + + const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4)); + const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010; + const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010; + + const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0; + const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4); + + x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0; + x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1; + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) { + int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd; + +#if QK_K == 256 + x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm; +#endif + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); + + const int * scales = (const int *) bxi->scales; + + const int ksc = k % (WARP_SIZE/8); + + // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8 + int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits + scales8 |= (scales[ksc/2] >> (2 * (ksc % 2))) & 0x30303030; // upper 2 bits + + x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8; + } +} + +static __dpct_inline__ float vec_dot_q5_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); + + const int index_x = i * (QR5_K*WARP_SIZE + 1) + QR5_K*k; + const int index_y = j * WARP_SIZE + (QR5_K*k) % WARP_SIZE; + return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8, + x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]); +} + +static __dpct_inline__ float +vec_dot_q6_K_q8_1(const void *__restrict__ vbq, + const block_q8_1 *__restrict__ bq8_1, const int &iqs) { + + const block_q6_K * bq6_K = (const block_q6_K *) vbq; + + const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4); + const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8); + const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4)); + + const int vl = get_int_from_uint8(bq6_K->ql, iqs); + const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift; + + const int8_t * scales = bq6_K->scales + scale_offset; + + int u[QR6_K]; + float d8[QR6_K]; + +#pragma unroll + for (int i = 0; i < QR6_K; ++i) { + u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1); + d8[i] = bq8_1[bq8_offset + 2 * i].ds[0]; + } + + return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8); +} + +template +static __dpct_inline__ void +allocate_tiles_q6_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc, + int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_sc) { + (void)x_qh; + + *x_ql = tile_x_ql; + *x_dm = tile_x_dm; + *x_sc = tile_x_sc; +} + +template +static __dpct_inline__ void +load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql, + sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh, + int *__restrict__ x_sc, const int &i_offset, const int &i_max, + const int &k, const int &blocks_per_row) { + (void)x_qh; + + GGML_SYCL_ASSUME(i_offset >= 0); + GGML_SYCL_ASSUME(i_offset < nwarps); + GGML_SYCL_ASSUME(k >= 0); + GGML_SYCL_ASSUME(k < WARP_SIZE); + + const int kbx = k / QI6_K; // == 0 if QK_K == 256 + const int kqsx = k % QI6_K; // == k if QK_K == 256 + + const block_q6_K * bx0 = (const block_q6_K *) vx; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { + int i = i0 + i_offset; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx; + const int ky = QR6_K*kqsx; + + const int ql = get_int_from_uint8(bxi->ql, kqsx); + const int ql0 = (ql >> 0) & 0x0F0F0F0F; + const int ql1 = (ql >> 4) & 0x0F0F0F0F; + + const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4)); + const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030; + const int qh1 = (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) & 0x30303030; + + const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0; + const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2); + + x_ql[i * (2 * WARP_SIZE + 1) + kq0] = + dpct::vectorized_binary(ql0 | qh0, 0x20202020, + dpct::sub_sat()); + x_ql[i * (2 * WARP_SIZE + 1) + kq1] = + dpct::vectorized_binary(ql1 | qh1, 0x20202020, + dpct::sub_sat()); + } + + const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256 + const int kbxd = k % blocks_per_tile_x_row; // == 0 if QK_K == 256 + float * x_dmf = (float *) x_dm; + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) { + int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd; + + x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d; + } + +#pragma unroll + for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) { + int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y; + + if (need_check) { + i = sycl::min(i, i_max); + } + + const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4; + + x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8)); + } +} + +static __dpct_inline__ float vec_dot_q6_K_q8_1_mul_mat( + const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm, + const int *__restrict__ x_qh, const int *__restrict__ x_sc, + const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds, + const int &i, const int &j, const int &k) { + (void)x_qh; + + const float * x_dmf = (const float *) x_dm; + const float * y_df = (const float *) y_ds; + + const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]); + + const int index_x = i * (QR6_K*WARP_SIZE + 1) + QR6_K*k; + const int index_y = j * WARP_SIZE + (QR6_K*k) % WARP_SIZE; + return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]); +} + +template +/* +DPCT1110:8: The total declared local variable size in device function mul_mat_q +exceeds 128 bytes and may cause high register pressure. Consult with your +hardware vendor to find the total register size available and adjust the code, +or use smaller sub-group size to avoid high register pressure. +*/ +static __dpct_inline__ void +mul_mat_q(const void *__restrict__ vx, const void *__restrict__ vy, + float *__restrict__ dst, const int ncols_x, const int nrows_x, + const int ncols_y, const int nrows_y, const int nrows_dst, + int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_qh, + int *tile_x_sc, const sycl::nd_item<3> &item_ct1, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + const int blocks_per_row_x = ncols_x / qk; + const int blocks_per_col_y = nrows_y / QK8_1; + const int blocks_per_warp = WARP_SIZE / qi; + + const int & ncols_dst = ncols_y; + + const int row_dst_0 = item_ct1.get_group(2) * mmq_y; + const int & row_x_0 = row_dst_0; + + const int col_dst_0 = item_ct1.get_group(1) * mmq_x; + const int & col_y_0 = col_dst_0; + + float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; + + for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { + + load_tiles(x + row_x_0 * blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, + tile_x_qh, tile_x_sc, item_ct1.get_local_id(1), + nrows_x - row_x_0 - 1, item_ct1.get_local_id(2), + blocks_per_row_x); + +#pragma unroll + for (int ir = 0; ir < qr; ++ir) { + const int kqs = ir * WARP_SIZE + item_ct1.get_local_id(2); + const int kbxd = kqs / QI8_1; + +#pragma unroll + for (int i = 0; i < mmq_x; i += nwarps) { + const int col_y_eff = dpct::min( + (unsigned int)(col_y_0 + item_ct1.get_local_id(1) + i), + ncols_y - 1); // to prevent out-of-bounds memory accesses + + const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd]; + + const int index_y = (item_ct1.get_local_id(1) + i) * WARP_SIZE + + kqs % WARP_SIZE; + tile_y_qs[index_y] = get_int_from_int8_aligned( + by0->qs, item_ct1.get_local_id(2) % QI8_1); + } + +#pragma unroll + for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) { + const int ids = + (ids0 + item_ct1.get_local_id(1) * QI8_1 + + item_ct1.get_local_id(2) / (WARP_SIZE / QI8_1)) % + mmq_x; + const int kby = item_ct1.get_local_id(2) % (WARP_SIZE / QI8_1); + const int col_y_eff = sycl::min(col_y_0 + ids, ncols_y - 1); + + // if the sum is not needed it's faster to transform the scale to f32 ahead of time + const sycl::half2 *dsi_src = + &y[col_y_eff * blocks_per_col_y + ib0 * (qk / QK8_1) + + ir * (WARP_SIZE / QI8_1) + kby] + .ds; + sycl::half2 *dsi_dst = + &tile_y_ds[ids * (WARP_SIZE / QI8_1) + kby]; + if (need_sum) { + *dsi_dst = *dsi_src; + } else { + float * dfi_dst = (float *) dsi_dst; + *dfi_dst = (*dsi_src)[0]; + } + } + + /* + DPCT1118:9: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:56: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + +// #pragma unroll // unrolling this loop causes too much register pressure + for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) { +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + sum[i / WARP_SIZE][j / nwarps] += vec_dot( + tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, + tile_y_qs, tile_y_ds, item_ct1.get_local_id(2) + i, + item_ct1.get_local_id(1) + j, k); + } + } + } + + /* + DPCT1118:10: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:57: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + } + } + +#pragma unroll + for (int j = 0; j < mmq_x; j += nwarps) { + const int col_dst = col_dst_0 + j + item_ct1.get_local_id(1); + + if (col_dst >= ncols_dst) { + return; + } + +#pragma unroll + for (int i = 0; i < mmq_y; i += WARP_SIZE) { + const int row_dst = row_dst_0 + item_ct1.get_local_id(2) + i; + + if (row_dst >= nrows_dst) { + continue; + } + + dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps]; + } + } +} + +#define MMQ_X_Q4_0_RDNA2 64 +#define MMQ_Y_Q4_0_RDNA2 128 +#define NWARPS_Q4_0_RDNA2 8 +#define MMQ_X_Q4_0_RDNA1 64 +#define MMQ_Y_Q4_0_RDNA1 64 +#define NWARPS_Q4_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_0_AMPERE 4 +#define MMQ_Y_Q4_0_AMPERE 32 +#define NWARPS_Q4_0_AMPERE 4 +#else +#define MMQ_X_Q4_0_AMPERE 64 +#define MMQ_Y_Q4_0_AMPERE 128 +#define NWARPS_Q4_0_AMPERE 4 +#endif +#define MMQ_X_Q4_0_PASCAL 64 +#define MMQ_Y_Q4_0_PASCAL 64 +#define NWARPS_Q4_0_PASCAL 8 + +template static void + mul_mat_q4_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_0, float *tile_x_d_q4_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + + const int mmq_x = MMQ_X_Q4_0_AMPERE; + const int mmq_y = MMQ_Y_Q4_0_AMPERE; + const int nwarps = NWARPS_Q4_0_AMPERE; + allocate_tiles_q4_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q4_0, tile_x_d_q4_0); + mul_mat_q, VDR_Q4_0_Q8_1_MMQ, + vec_dot_q4_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q4_1_RDNA2 64 +#define MMQ_Y_Q4_1_RDNA2 128 +#define NWARPS_Q4_1_RDNA2 8 +#define MMQ_X_Q4_1_RDNA1 64 +#define MMQ_Y_Q4_1_RDNA1 64 +#define NWARPS_Q4_1_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_1_AMPERE 4 +#define MMQ_Y_Q4_1_AMPERE 32 +#define NWARPS_Q4_1_AMPERE 4 +#else +#define MMQ_X_Q4_1_AMPERE 64 +#define MMQ_Y_Q4_1_AMPERE 128 +#define NWARPS_Q4_1_AMPERE 4 +#endif +#define MMQ_X_Q4_1_PASCAL 64 +#define MMQ_Y_Q4_1_PASCAL 64 +#define NWARPS_Q4_1_PASCAL 8 + +template static void + mul_mat_q4_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_1, + sycl::half2 *tile_x_dm_q4_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q4_1_AMPERE; + const int mmq_y = MMQ_Y_Q4_1_AMPERE; + const int nwarps = NWARPS_Q4_1_AMPERE; + allocate_tiles_q4_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q4_1, tile_x_dm_q4_1); + mul_mat_q, VDR_Q4_1_Q8_1_MMQ, + vec_dot_q4_1_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_0_RDNA2 64 +#define MMQ_Y_Q5_0_RDNA2 128 +#define NWARPS_Q5_0_RDNA2 8 +#define MMQ_X_Q5_0_RDNA1 64 +#define MMQ_Y_Q5_0_RDNA1 64 +#define NWARPS_Q5_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_0_AMPERE 4 +#define MMQ_Y_Q5_0_AMPERE 32 +#define NWARPS_Q5_0_AMPERE 4 +#else +#define MMQ_X_Q5_0_AMPERE 128 +#define MMQ_Y_Q5_0_AMPERE 64 +#define NWARPS_Q5_0_AMPERE 4 +#endif +#define MMQ_X_Q5_0_PASCAL 64 +#define MMQ_Y_Q5_0_PASCAL 64 +#define NWARPS_Q5_0_PASCAL 8 + +template static void + mul_mat_q5_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_0, float *tile_x_d_q5_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_0_AMPERE; + const int mmq_y = MMQ_Y_Q5_0_AMPERE; + const int nwarps = NWARPS_Q5_0_AMPERE; + allocate_tiles_q5_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_0, tile_x_d_q5_0); + mul_mat_q, VDR_Q5_0_Q8_1_MMQ, + vec_dot_q5_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_1_RDNA2 64 +#define MMQ_Y_Q5_1_RDNA2 128 +#define NWARPS_Q5_1_RDNA2 8 +#define MMQ_X_Q5_1_RDNA1 64 +#define MMQ_Y_Q5_1_RDNA1 64 +#define NWARPS_Q5_1_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_1_AMPERE 4 +#define MMQ_Y_Q5_1_AMPERE 32 +#define NWARPS_Q5_1_AMPERE 4 +#else +#define MMQ_X_Q5_1_AMPERE 128 +#define MMQ_Y_Q5_1_AMPERE 64 +#define NWARPS_Q5_1_AMPERE 4 +#endif +#define MMQ_X_Q5_1_PASCAL 64 +#define MMQ_Y_Q5_1_PASCAL 64 +#define NWARPS_Q5_1_PASCAL 8 + +template static void +mul_mat_q5_1( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_1, + sycl::half2 *tile_x_dm_q5_1, int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_1_AMPERE; + const int mmq_y = MMQ_Y_Q5_1_AMPERE; + const int nwarps = NWARPS_Q5_1_AMPERE; + allocate_tiles_q5_1(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_1, tile_x_dm_q5_1); + mul_mat_q, VDR_Q5_1_Q8_1_MMQ, + vec_dot_q5_1_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q8_0_RDNA2 64 +#define MMQ_Y_Q8_0_RDNA2 128 +#define NWARPS_Q8_0_RDNA2 8 +#define MMQ_X_Q8_0_RDNA1 64 +#define MMQ_Y_Q8_0_RDNA1 64 +#define NWARPS_Q8_0_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q8_0_AMPERE 4 +#define MMQ_Y_Q8_0_AMPERE 32 +#define NWARPS_Q8_0_AMPERE 4 +#else +#define MMQ_X_Q8_0_AMPERE 128 +#define MMQ_Y_Q8_0_AMPERE 64 +#define NWARPS_Q8_0_AMPERE 4 +#endif +#define MMQ_X_Q8_0_PASCAL 64 +#define MMQ_Y_Q8_0_PASCAL 64 +#define NWARPS_Q8_0_PASCAL 8 + +template static void + mul_mat_q8_0( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q8_0, float *tile_x_d_q8_0, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q8_0_AMPERE; + const int mmq_y = MMQ_Y_Q8_0_AMPERE; + const int nwarps = NWARPS_Q8_0_AMPERE; + allocate_tiles_q8_0(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_qs_q8_0, tile_x_d_q8_0); + mul_mat_q, VDR_Q8_0_Q8_1_MMQ, + vec_dot_q8_0_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q2_K_RDNA2 64 +#define MMQ_Y_Q2_K_RDNA2 128 +#define NWARPS_Q2_K_RDNA2 8 +#define MMQ_X_Q2_K_RDNA1 128 +#define MMQ_Y_Q2_K_RDNA1 32 +#define NWARPS_Q2_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q2_K_AMPERE 4 +#define MMQ_Y_Q2_K_AMPERE 32 +#define NWARPS_Q2_K_AMPERE 4 +#else +#define MMQ_X_Q2_K_AMPERE 64 +#define MMQ_Y_Q2_K_AMPERE 128 +#define NWARPS_Q2_K_AMPERE 4 +#endif +#define MMQ_X_Q2_K_PASCAL 64 +#define MMQ_Y_Q2_K_PASCAL 64 +#define NWARPS_Q2_K_PASCAL 8 + +template static void +mul_mat_q2_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q2_K, + sycl::half2 *tile_x_dm_q2_K, int *tile_x_sc_q2_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q2_K_AMPERE; + const int mmq_y = MMQ_Y_Q2_K_AMPERE; + const int nwarps = NWARPS_Q2_K_AMPERE; + allocate_tiles_q2_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q2_K, tile_x_dm_q2_K, tile_x_sc_q2_K); + mul_mat_q, VDR_Q2_K_Q8_1_MMQ, + vec_dot_q2_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q3_K_RDNA2 128 +#define MMQ_Y_Q3_K_RDNA2 64 +#define NWARPS_Q3_K_RDNA2 8 +#define MMQ_X_Q3_K_RDNA1 32 +#define MMQ_Y_Q3_K_RDNA1 128 +#define NWARPS_Q3_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q3_K_AMPERE 4 +#define MMQ_Y_Q3_K_AMPERE 32 +#define NWARPS_Q3_K_AMPERE 4 +#else +#define MMQ_X_Q3_K_AMPERE 128 +#define MMQ_Y_Q3_K_AMPERE 128 +#define NWARPS_Q3_K_AMPERE 4 +#endif +#define MMQ_X_Q3_K_PASCAL 64 +#define MMQ_Y_Q3_K_PASCAL 64 +#define NWARPS_Q3_K_PASCAL 8 + +template static void +mul_mat_q3_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q3_K, + sycl::half2 *tile_x_dm_q3_K, int *tile_x_qh_q3_K, int *tile_x_sc_q3_K, + int *tile_y_qs, sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q3_K_AMPERE; + const int mmq_y = MMQ_Y_Q3_K_AMPERE; + const int nwarps = NWARPS_Q3_K_AMPERE; + allocate_tiles_q3_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q3_K, tile_x_dm_q3_K, tile_x_qh_q3_K, + tile_x_sc_q3_K); + mul_mat_q, VDR_Q3_K_Q8_1_MMQ, + vec_dot_q3_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q4_K_RDNA2 64 +#define MMQ_Y_Q4_K_RDNA2 128 +#define NWARPS_Q4_K_RDNA2 8 +#define MMQ_X_Q4_K_RDNA1 32 +#define MMQ_Y_Q4_K_RDNA1 64 +#define NWARPS_Q4_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q4_K_AMPERE 4 +#define MMQ_Y_Q4_K_AMPERE 32 +#define NWARPS_Q4_K_AMPERE 4 +#else +#define MMQ_X_Q4_K_AMPERE 64 +#define MMQ_Y_Q4_K_AMPERE 128 +#define NWARPS_Q4_K_AMPERE 4 +#endif +#define MMQ_X_Q4_K_PASCAL 64 +#define MMQ_Y_Q4_K_PASCAL 64 +#define NWARPS_Q4_K_PASCAL 8 + +template static void + mul_mat_q4_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q4_K, + sycl::half2 *tile_x_dm_q4_K, int *tile_x_sc_q4_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q4_K_AMPERE; + const int mmq_y = MMQ_Y_Q4_K_AMPERE; + const int nwarps = NWARPS_Q4_K_AMPERE; + allocate_tiles_q4_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q4_K, tile_x_dm_q4_K, tile_x_sc_q4_K); + mul_mat_q, VDR_Q4_K_Q8_1_MMQ, + vec_dot_q4_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q5_K_RDNA2 64 +#define MMQ_Y_Q5_K_RDNA2 128 +#define NWARPS_Q5_K_RDNA2 8 +#define MMQ_X_Q5_K_RDNA1 32 +#define MMQ_Y_Q5_K_RDNA1 64 +#define NWARPS_Q5_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q5_K_AMPERE 4 +#define MMQ_Y_Q5_K_AMPERE 32 +#define NWARPS_Q5_K_AMPERE 4 +#else +#define MMQ_X_Q5_K_AMPERE 64 +#define MMQ_Y_Q5_K_AMPERE 128 +#define NWARPS_Q5_K_AMPERE 4 +#endif +#define MMQ_X_Q5_K_PASCAL 64 +#define MMQ_Y_Q5_K_PASCAL 64 +#define NWARPS_Q5_K_PASCAL 8 + +template static void +mul_mat_q5_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_K, + sycl::half2 *tile_x_dm_q5_K, int *tile_x_sc_q5_K, int *tile_y_qs, + sycl::half2 *tile_y_ds) { + int * tile_x_ql = nullptr; + sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q5_K_AMPERE; + const int mmq_y = MMQ_Y_Q5_K_AMPERE; + const int nwarps = NWARPS_Q5_K_AMPERE; + allocate_tiles_q5_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql_q5_K, tile_x_dm_q5_K, tile_x_sc_q5_K); + mul_mat_q, VDR_Q5_K_Q8_1_MMQ, + vec_dot_q5_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +#define MMQ_X_Q6_K_RDNA2 64 +#define MMQ_Y_Q6_K_RDNA2 128 +#define NWARPS_Q6_K_RDNA2 8 +#define MMQ_X_Q6_K_RDNA1 32 +#define MMQ_Y_Q6_K_RDNA1 64 +#define NWARPS_Q6_K_RDNA1 8 +#if defined(SYCL_USE_XMX) +#define MMQ_X_Q6_K_AMPERE 4 +#define MMQ_Y_Q6_K_AMPERE 32 +#define NWARPS_Q6_K_AMPERE 4 +#else +#define MMQ_X_Q6_K_AMPERE 64 +#define MMQ_Y_Q6_K_AMPERE 64 +#define NWARPS_Q6_K_AMPERE 4 +#endif +#define MMQ_X_Q6_K_PASCAL 64 +#define MMQ_Y_Q6_K_PASCAL 64 +#define NWARPS_Q6_K_PASCAL 8 + +template static void + mul_mat_q6_K( + const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst, + const sycl::nd_item<3> &item_ct1, int *tile_x_ql, sycl::half2 *tile_x_dm, + int *tile_x_sc, int *tile_y_qs, sycl::half2 *tile_y_ds) { + // int * tile_x_ql = nullptr; + // sycl::half2 *tile_x_dm = nullptr; + int * tile_x_qh = nullptr; + // int * tile_x_sc = nullptr; + +//sycl_todo: change according to hardware + const int mmq_x = MMQ_X_Q6_K_AMPERE; + const int mmq_y = MMQ_Y_Q6_K_AMPERE; + const int nwarps = NWARPS_Q6_K_AMPERE; + allocate_tiles_q6_K(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc, + tile_x_ql, tile_x_dm, tile_x_sc); + mul_mat_q, VDR_Q6_K_Q8_1_MMQ, + vec_dot_q6_K_q8_1_mul_mat>( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql, + tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds); +} + +template +static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int blocks_per_row = ncols / qk; + const int blocks_per_warp = vdr * WARP_SIZE / qi; + +// partial sum for each thread + float tmp = 0.0f; + + const block_q_t * x = (const block_q_t *) vx; + const block_q8_1 * y = (const block_q8_1 *) vy; + + for (int i = 0; i < blocks_per_row; i += blocks_per_warp) { + const int ibx = row * blocks_per_row + i + + item_ct1.get_local_id(2) / (qi / vdr); // x block index + + const int iby = (i + item_ct1.get_local_id(2) / (qi / vdr)) * + (qk / QK8_1); // y block index that aligns with ibx + + const int iqs = + vdr * + (item_ct1.get_local_id(2) % + (qi / vdr)); // x block quant index when casting the quants to int + + tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs); + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[row] = tmp; + } +} + +template +static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows, + const sycl::nd_item<3> &item_ct1) { + // qk = quantized weights per x block + // qr = number of quantized weights per data value in x block + const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (row >= nrows) { + return; + } + + const int tid = item_ct1.get_local_id(2); + + const int iter_stride = 2*GGML_SYCL_DMMV_X; + const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter + const int y_offset = qr == 1 ? 1 : qk/2; + +// partial sum for each thread +#ifdef GGML_SYCL_F16 + sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics +#else + float tmp = 0.0f; +#endif // GGML_SYCL_F16 + + for (int i = 0; i < ncols; i += iter_stride) { + const int col = i + vals_per_iter*tid; + const int ib = (row*ncols + col)/qk; // x block index + const int iqs = (col%qk)/qr; // x quant index + const int iybs = col - col%qk; // y block start index + +// processing >2 values per i iter is faster for fast GPUs +#pragma unroll + for (int j = 0; j < vals_per_iter; j += 2) { + // process 2 vals per j iter + + // dequantize + // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val + dfloat2 v; + dequantize_kernel(vx, ib, iqs + j/qr, v); + + // matrix multiplication + // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2 +#ifdef GGML_SYCL_F16 + dfloat2 t1{y[iybs + iqs + j / qr + 0], + y[iybs + iqs + j / qr + y_offset]}; + + tmp += v * t1; +#else + tmp += v.x() * y[iybs + iqs + j / qr + 0]; + tmp += v.y() * y[iybs + iqs + j / qr + y_offset]; +#endif // GGML_SYCL_F16 + } + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (tid == 0) { +#ifdef GGML_SYCL_F16 + dst[row] = tmp.x() + tmp.y(); +#else + dst[row] = tmp; +#endif // GGML_SYCL_F16 + } +} + +static void mul_mat_p021_f16_f32( + const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, + const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y, + const sycl::nd_item<3> &item_ct1) { + + const sycl::half *x = (const sycl::half *)vx; + + const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0); + const int channel_x = channel / (nchannels_y / nchannels_x); + + const int nrows_y = ncols_x; + const int nrows_dst = nrows_x; + const int row_dst = row_x; + + float tmp = 0.0f; + + for (int col_x0 = 0; col_x0 < ncols_x; + col_x0 += item_ct1.get_local_range(2)) { + const int col_x = col_x0 + item_ct1.get_local_id(2); + + if (col_x >= ncols_x) { + break; + } + + // x is transposed and permuted + const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x; + const float xi = + sycl::vec(x[ix]) + .convert()[0]; + + const int row_y = col_x; + + + // y is not transposed but permuted + const int iy = channel*nrows_y + row_y; + + tmp += xi * y[iy]; + } + + // dst is not transposed and not permuted + const int idst = channel*nrows_dst + row_dst; + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[idst] = tmp; + } +} + +static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous + const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, + const int row_stride_x, const int channel_stride_x, const int channel_x_divisor, + const sycl::nd_item<3> &item_ct1) { + + const sycl::half *x = (const sycl::half *)vx; + + const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) + + item_ct1.get_local_id(0); + const int channel_x = channel / channel_x_divisor; + + const int nrows_y = ncols_x; + const int nrows_dst = nrows_x; + const int row_dst = row_x; + + const int idst = channel*nrows_dst + row_dst; + + float tmp = 0.0f; + + for (int col_x0 = 0; col_x0 < ncols_x; + col_x0 += item_ct1.get_local_range(2)) { + const int col_x = col_x0 + item_ct1.get_local_id(2); + + if (col_x >= ncols_x) { + break; + } + + const int row_y = col_x; + + const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x; + const int iy = channel*nrows_y + row_y; + + const float xi = + sycl::vec(x[ix]) + .convert()[0]; + + tmp += xi * y[iy]; + } + + // sum up partial sums and write back result +#pragma unroll + for (int mask = 16; mask > 0; mask >>= 1) { + tmp += + dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask); + } + + if (item_ct1.get_local_id(2) == 0) { + dst[idst] = tmp; + } +} + +static void cpy_1_f32_f32(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + float * dsti = (float *) cdsti; + + *dsti = *xi; +} + +static void cpy_1_f32_f16(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + sycl::half *dsti = (sycl::half *)cdsti; + + *dsti = sycl::vec(*xi) + .convert()[0]; +} + +static void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const sycl::half *xi = (const sycl::half *)cxi; + sycl::half *dsti = (sycl::half *)cdsti; + + *dsti = *xi; +} + +static void cpy_1_i16_i16(const char * cxi, char * cdsti) { + const int16_t *xi = (const int16_t *)cxi; + int16_t *dsti = (int16_t *)cdsti; + + *dsti = *xi; +} + +static void cpy_1_i32_i32(const char * cxi, char * cdsti) { + const int32_t *xi = (const int32_t *)cxi; + int32_t *dsti = (int32_t *)cdsti; + + *dsti = *xi; +} + +template +static void cpy_f32_f16(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, + const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= ne) { + return; + } + + // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor + // then combine those indices with the corresponding byte offsets to get the total offsets + const int i02 = i / (ne00*ne01); + const int i01 = (i - i02*ne01*ne00) / ne00; + const int i00 = i - i02*ne01*ne00 - i01*ne00; + const int x_offset = i00*nb00 + i01*nb01 + i02*nb02; + + const int i12 = i / (ne10*ne11); + const int i11 = (i - i12*ne10*ne11) / ne10; + const int i10 = i - i12*ne10*ne11 - i11*ne10; + const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12; + + cpy_1(cx + x_offset, cdst + dst_offset); +} + +static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q8_0 * dsti = (block_q8_0 *) cdsti; + + float amax = 0.0f; // absolute max + + for (int j = 0; j < QK8_0; j++) { + const float v = xi[j]; + amax = sycl::fmax(amax, sycl::fabs((float)v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK8_0; ++j) { + const float x0 = xi[j]*id; + + dsti->qs[j] = sycl::round((float)x0); + } +} + +static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_0 * dsti = (block_q4_0 *) cdsti; + + float amax = 0.0f; + float vmax = 0.0f; + + for (int j = 0; j < QK4_0; ++j) { + const float v = xi[j]; + if (amax < sycl::fabs((float)v)) { + amax = sycl::fabs((float)v); + vmax = v; + } + } + + const float d = vmax / -8; + const float id = d ? 1.0f/d : 0.0f; + + dsti->d = d; + + for (int j = 0; j < QK4_0/2; ++j) { + const float x0 = xi[0 + j]*id; + const float x1 = xi[QK4_0/2 + j]*id; + + const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) { + const float * xi = (const float *) cxi; + block_q4_1 * dsti = (block_q4_1 *) cdsti; + + float vmin = FLT_MAX; + float vmax = -FLT_MAX; + + for (int j = 0; j < QK4_1; ++j) { + const float v = xi[j]; + + if (v < vmin) vmin = v; + if (v > vmax) vmax = v; + } + + const float d = (vmax - vmin) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + dsti->dm.x() = d; + dsti->dm.y() = vmin; + + for (int j = 0; j < QK4_1/2; ++j) { + const float x0 = (xi[0 + j] - vmin)*id; + const float x1 = (xi[QK4_1/2 + j] - vmin)*id; + + const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f)); + + dsti->qs[j] = xi0; + dsti->qs[j] |= xi1 << 4; + } +} + +template +static void cpy_f32_q(const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, + const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, + const sycl::nd_item<3> &item_ct1) { + const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2)) * + qk; + + if (i >= ne) { + return; + } + + const int i02 = i / (ne00*ne01); + const int i01 = (i - i02*ne01*ne00) / ne00; + const int i00 = (i - i02*ne01*ne00 - i01*ne00); + const int x_offset = i00*nb00 + i01*nb01 + i02*nb02; + + const int i12 = i / (ne10*ne11); + const int i11 = (i - i12*ne10*ne11) / ne10; + const int i10 = (i - i12*ne10*ne11 - i11*ne10)/qk; + const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12; + + cpy_blck(cx + x_offset, cdst + dst_offset); +} + +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); + return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); +} + +struct rope_corr_dims { + float v[4]; +}; + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale); + } + *cos_theta = sycl::cos(theta) * mscale; + *sin_theta = sycl::sin(theta) * mscale; +} + +// rope == RoPE == rotary positional embedding +template +static void rope( + const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + float ext_factor, float attn_factor, rope_corr_dims corr_dims +, + const sycl::nd_item<3> &item_ct1) { + const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (col >= ncols) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int i = row*ncols + col; + const int i2 = row/p_delta_rows; + + const int p = has_pos ? pos[i2] : 0; + const float theta_base = p * dpct::pow(freq_base, -float(col) / ncols); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + 1]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + 1] = x0*sin_theta + x1*cos_theta; +} + +template +static void rope_neox( + const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims +, + const sycl::nd_item<3> &item_ct1) { + const int col = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1)); + + if (col >= ncols) { + return; + } + + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int ib = col / n_dims; + const int ic = col % n_dims; + + if (ib > 0) { + const int i = row*ncols + ib*n_dims + ic; + + dst[i + 0] = x[i + 0]; + dst[i + 1] = x[i + 1]; + + return; + } + + const int i = row*ncols + ib*n_dims + ic/2; + const int i2 = row/p_delta_rows; + + float cur_rot = inv_ndims * ic - ib; + + const int p = has_pos ? pos[i2] : 0; + const float theta_base = + p * freq_scale * dpct::pow(theta_scale, col / 2.0f); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + n_dims/2]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; +} + +static void rope_glm_f32( + const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + int n_ctx +, const sycl::nd_item<3> &item_ct1) { + const int col = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + const int half_n_dims = ncols/4; + + if (col >= half_n_dims) { + return; + } + + const int row = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i = row*ncols + col; + const int i2 = row/p_delta_rows; + + const float col_theta_scale = dpct::pow(freq_base, -2.0f * col / ncols); + // FIXME: this is likely wrong + const int p = pos != nullptr ? pos[i2] : 0; + + const float theta = sycl::min(p, n_ctx - 2) * freq_scale * col_theta_scale; + const float sin_theta = sycl::sin((float)theta); + const float cos_theta = sycl::cos((float)theta); + + const float x0 = x[i + 0]; + const float x1 = x[i + half_n_dims]; + + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta; + + const float block_theta = + ((float)sycl::max(p - n_ctx - 2, 0)) * col_theta_scale; + const float sin_block_theta = sycl::sin((float)block_theta); + const float cos_block_theta = sycl::cos((float)block_theta); + + const float x2 = x[i + half_n_dims * 2]; + const float x3 = x[i + half_n_dims * 3]; + + dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta; + dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta; +} + +static void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows, + const int n_heads_log2_floor, const float m0, const float m1, + const sycl::nd_item<3> &item_ct1) { + const int col = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (col >= ncols) { + return; + } + + const int row = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int i = row*ncols + col; + + const int k = row/k_rows; + + float m_k; + if (k < n_heads_log2_floor) { + m_k = dpct::pow(m0, k + 1); + } else { + m_k = dpct::pow(m1, 2 * (k - n_heads_log2_floor) + 1); + } + + dst[i] = col * m_k + x[i]; +} + +static void k_sum_rows_f32(const float * x, float * dst, const int ncols, + const sycl::nd_item<3> &item_ct1) { + const int row = item_ct1.get_group(1); + const int col = item_ct1.get_local_id(2); + + float sum = 0.0f; + for (int i = col; i < ncols; i += item_ct1.get_local_range(2)) { + sum += x[row * ncols + i]; + } + + sum = warp_reduce_sum(sum, item_ct1); + + if (col == 0) { + dst[row] = sum; + } +} + +template +static inline void swap(T & a, T & b) { + T tmp = a; + a = b; + b = tmp; +} + +template +static void k_argsort_f32_i32(const float * x, int * dst, const int ncols, + const sycl::nd_item<3> &item_ct1) { + // bitonic sort + int col = item_ct1.get_local_id(2); + int row = item_ct1.get_group(1); + + if (col >= ncols) return; + + const float * x_row = x + row * ncols; + int * dst_row = dst + row * ncols; + + // initialize indices + if (col < ncols) { + dst_row[col] = col; + } + /* + DPCT1065:58: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for better + performance if there is no access to global memory. + */ + item_ct1.barrier(); + + for (int k = 2; k <= ncols; k *= 2) { + for (int j = k / 2; j > 0; j /= 2) { + int ixj = col ^ j; + if (ixj > col) { + if ((col & k) == 0) { + if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) { + swap(dst_row[col], dst_row[ixj]); + } + } else { + if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) { + swap(dst_row[col], dst_row[ixj]); + } + } + } + /* + DPCT1118:11: SYCL group functions and algorithms must be encountered + in converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:59: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + } + } +} + +static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past, + const sycl::nd_item<3> &item_ct1) { + const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) + + item_ct1.get_local_id(1); + const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (col >= ncols) { + return; + } + + const int i = row*ncols + col; + //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i]; + //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU + dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX; +} + +static void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale, + const sycl::nd_item<3> &item_ct1, float *buf) { + const int tid = item_ct1.get_local_id(2); + const int rowx = item_ct1.get_group(2); + const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension + + const int block_size = item_ct1.get_local_range(2); + + const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE; + const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE; + + float max_val = -INFINITY; + + for (int col = tid; col < ncols; col += block_size) { + const int ix = rowx*ncols + col; + const int iy = rowy*ncols + col; + max_val = sycl::max(max_val, x[ix] * scale + (y ? y[iy] : 0.0f)); + } + + // find the max value in the block + max_val = warp_reduce_max(max_val, item_ct1); + if (block_size > WARP_SIZE) { + if (warp_id == 0) { + buf[lane_id] = -INFINITY; + } + /* + DPCT1118:12: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:60: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + + if (lane_id == 0) { + buf[warp_id] = max_val; + } + /* + DPCT1118:13: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:61: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + + max_val = buf[lane_id]; + max_val = warp_reduce_max(max_val, item_ct1); + } + + float tmp = 0.f; + + for (int col = tid; col < ncols; col += block_size) { + const int ix = rowx*ncols + col; + const int iy = rowy*ncols + col; + const float val = + sycl::native::exp((x[ix] * scale + (y ? y[iy] : 0.0f)) - max_val); + tmp += val; + dst[ix] = val; + } + + // find the sum of exps in the block + tmp = warp_reduce_sum(tmp, item_ct1); + if (block_size > WARP_SIZE) { + if (warp_id == 0) { + buf[lane_id] = 0.f; + } + /* + DPCT1118:14: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:62: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + + if (lane_id == 0) { + buf[warp_id] = tmp; + } + /* + DPCT1118:15: SYCL group functions and algorithms must be encountered in + converged control flow. You may need to adjust the code. + */ + /* + DPCT1065:63: Consider replacing sycl::nd_item::barrier() with + sycl::nd_item::barrier(sycl::access::fence_space::local_space) for + better performance if there is no access to global memory. + */ + item_ct1.barrier(); + + tmp = buf[lane_id]; + tmp = warp_reduce_sum(tmp, item_ct1); + } + + const float inv_tmp = 1.f / tmp; + + for (int col = tid; col < ncols; col += block_size) { + const int i = rowx*ncols + col; + dst[i] *= inv_tmp; + } +} + +static void scale_f32(const float * x, float * dst, const float scale, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + dst[i] = scale * x[i]; +} + +static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) + + item_ct1.get_local_id(2); + + if (i >= k) { + return; + } + + dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); +} + +static void im2col_f32_f16(const float *x, sycl::half *dst, int offset_delta, + int IW, int IH, int OW, int KW, int KH, + int pelements, int CHW, int s0, int s1, int p0, + int p1, int d0, int d1, + const sycl::nd_item<3> &item_ct1) { + const int i = item_ct1.get_local_id(2) + + item_ct1.get_group(2) * item_ct1.get_local_range(2); + if (i >= pelements) { + return; + } + + const int ksize = OW * (KH > 1 ? KW : 1); + const int kx = i / ksize; + const int kd = kx * ksize; + const int ky = (i - kd) / OW; + const int ix = i % OW; + + const int64_t iiw = ix * s0 + kx * d0 - p0; + const int64_t iih = item_ct1.get_group(1) * s1 + ky * d1 - p1; + + const int64_t offset_dst = + (item_ct1.get_group(1) * OW + ix) * CHW + + (item_ct1.get_group(0) * (KW * KH) + ky * KW + kx); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = + sycl::vec(0.0f) + .convert()[0]; + } else { + const int64_t offset_src = item_ct1.get_group(0) * offset_delta; + dst[offset_dst] = + sycl::vec(x[offset_src + iih * IW + iiw]) + .convert()[0]; + } +} + +template +static void get_rows_sycl(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const void *src0_dd, + const int32_t *src1_dd, float *dst_dd, + dpct::queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE); + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + GGML_ASSERT(ne00 % 2 == 0); + + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_get_rows( + src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, + s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); + }); + + (void) dst; +} + +template +static void get_rows_sycl_float(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const src0_t *src0_dd, const int32_t *src1_dd, + float *dst_dd, dpct::queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE); + const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE; + const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x); + + // strides in elements + //const size_t s0 = nb0 / ggml_element_size(dst); + const size_t s1 = nb1 / ggml_element_size(dst); + const size_t s2 = nb2 / ggml_element_size(dst); + const size_t s3 = nb3 / ggml_element_size(dst); + + const size_t s10 = nb10 / ggml_element_size(src1); + const size_t s11 = nb11 / ggml_element_size(src1); + const size_t s12 = nb12 / ggml_element_size(src1); + //const size_t s13 = nb13 / ggml_element_size(src1); + + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, + s3, nb01, nb02, nb03, s10, s11, s12, item_ct1); + }); + } + + (void) dst; +} + +template +struct bin_bcast_sycl { + template + void operator()(const struct ggml_tensor *src0, + const struct ggml_tensor *src1, struct ggml_tensor *dst, + const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd, + dpct::queue_ptr stream) { + + GGML_TENSOR_BINARY_OP_LOCALS + + int nr0 = ne10/ne0; + int nr1 = ne11/ne1; + int nr2 = ne12/ne2; + int nr3 = ne13/ne3; + + int nr[4] = { nr0, nr1, nr2, nr3 }; + + // collapse dimensions until first broadcast dimension + int64_t cne0[] = {ne0, ne1, ne2, ne3}; + int64_t cne1[] = {ne10, ne11, ne12, ne13}; + size_t cnb0[] = {nb0, nb1, nb2, nb3}; + size_t cnb1[] = {nb10, nb11, nb12, nb13}; + auto collapse = [](int64_t cne[]) { + cne[0] *= cne[1]; + cne[1] = cne[2]; + cne[2] = cne[3]; + cne[3] = 1; + }; + + auto collapse_nb = [](size_t cnb[], int64_t cne[]) { + cnb[1] *= cne[1]; + cnb[2] *= cne[2]; + cnb[3] *= cne[3]; + }; + + for (int i = 0; i < 4; i++) { + if (nr[i] != 1) { + break; + } + if (i > 0) { + collapse_nb(cnb0, cne0); + collapse_nb(cnb1, cne1); + collapse(cne0); + collapse(cne1); + } + } + { + int64_t ne0 = cne0[0]; + int64_t ne1 = cne0[1]; + int64_t ne2 = cne0[2]; + int64_t ne3 = cne0[3]; + + int64_t ne10 = cne1[0]; + int64_t ne11 = cne1[1]; + int64_t ne12 = cne1[2]; + int64_t ne13 = cne1[3]; + + size_t nb0 = cnb0[0]; + size_t nb1 = cnb0[1]; + size_t nb2 = cnb0[2]; + size_t nb3 = cnb0[3]; + + size_t nb10 = cnb1[0]; + size_t nb11 = cnb1[1]; + size_t nb12 = cnb1[2]; + size_t nb13 = cnb1[3]; + + size_t s0 = nb0 / sizeof(dst_t); + size_t s1 = nb1 / sizeof(dst_t); + size_t s2 = nb2 / sizeof(dst_t); + size_t s3 = nb3 / sizeof(dst_t); + + size_t s10 = nb10 / sizeof(src1_t); + size_t s11 = nb11 / sizeof(src1_t); + size_t s12 = nb12 / sizeof(src1_t); + size_t s13 = nb13 / sizeof(src1_t); + + GGML_ASSERT(s0 == 1); + GGML_ASSERT(s10 == 1); + + const int block_size = 128; + + int64_t hne0 = std::max(ne0/2LL, 1LL); + + sycl::range<3> block_dims(1, 1, 1); + block_dims[2] = std::min(hne0, block_size); + block_dims[1] = std::min( + ne1, block_size / (unsigned int)block_dims[2]); + block_dims[0] = std::min( + std::min( + ne2 * ne3, block_size / (unsigned int)block_dims[2] / + (unsigned int)block_dims[1]), + 64U); + + sycl::range<3> block_nums( + (ne2 * ne3 + block_dims[0] - 1) / block_dims[0], + (ne1 + block_dims[1] - 1) / block_dims[1], + (hne0 + block_dims[2] - 1) / block_dims[2]); + + if (block_nums[0] > 65535) { + // this is the maximum number of blocks in z direction, fallback to 1D grid kernel + int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * + sycl::range<3>(1, 1, block_size), + sycl::range<3>(1, 1, block_size)), + [=](sycl::nd_item<3> item_ct1) { + k_bin_bcast_unravel( + src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, + ne10, ne11, ne12, ne13, s1, s2, s3, s11, s12, + s13, item_ct1); + }); + } + } else { + /* + DPCT1049:16: The work-group size passed to the SYCL kernel may + exceed the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if + needed. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_bin_bcast(src0_dd, src1_dd, dst_dd, ne0, ne1, + ne2, ne3, ne10, ne11, ne12, ne13, + s1, s2, s3, s11, s12, s13, + item_ct1); + }); + } + } + } +}; + +static void acc_f32_sycl(const float *x, const float *y, float *dst, + const int n_elements, const int ne10, const int ne11, + const int ne12, const int nb1, const int nb2, + const int offset, dpct::queue_ptr stream) { + int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, + item_ct1); + }); +} + +static void gelu_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + gelu_f32(x, dst, k, item_ct1); + }); +} + +static void silu_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + silu_f32(x, dst, k, item_ct1); + }); +} + +static void gelu_quick_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + gelu_quick_f32(x, dst, k, item_ct1); + }); +} + +static void tanh_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + tanh_f32(x, dst, k, item_ct1); + }); +} + +static void relu_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + relu_f32(x, dst, k, item_ct1); + }); +} + +static void leaky_relu_f32_sycl(const float *x, float *dst, const int k, + const float negative_slope, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + leaky_relu_f32(x, dst, k, negative_slope, item_ct1); + }); +} + +static void sqr_f32_sycl(const float *x, float *dst, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + sqr_f32(x, dst, k, item_ct1); + }); +} + +static void norm_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, const float eps, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(32), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); + }); + }); + } else { + const int work_group_size = g_work_group_size; + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:17: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1( + sycl::range<1>(32), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void group_norm_f32_sycl(const float *x, float *dst, + const int num_groups, const int group_size, + const int ne_elements, dpct::queue_ptr stream) { + static const float eps = 1e-6f; + if (group_size < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); + + const float eps_ct4 = eps; + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + group_norm_f32( + x, dst, group_size, ne_elements, eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); + }); + }); + } else { + const int work_group_size = g_work_group_size; + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:18: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); + + const float eps_ct4 = eps; + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + group_norm_f32(x, dst, group_size, ne_elements, + eps_ct4, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void concat_f32_sycl(const float *x, const float *y, float *dst, + const int ne0, int ne1, int ne2, int ne02, + dpct::queue_ptr stream) { + int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE; + sycl::range<3> gridDim(ne2, ne1, num_blocks); + stream->parallel_for( + sycl::nd_range<3>(gridDim * + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + concat_f32(x, y, dst, ne0, ne02, item_ct1); + }); +} + +static void upscale_f32_sycl(const float *x, float *dst, const int ne00, + const int ne01, const int ne02, + const int scale_factor, dpct::queue_ptr stream) { + int ne0 = (ne00 * scale_factor); + int num_blocks = (ne0 + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE; + sycl::range<3> gridDim(ne02, (ne01 * scale_factor), num_blocks); + stream->parallel_for( + sycl::nd_range<3>(gridDim * + sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_UPSCALE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + upscale_f32(x, dst, ne00, ne00 * ne01, scale_factor, item_ct1); + }); +} + +static void pad_f32_sycl(const float *x, float *dst, const int ne00, + const int ne01, const int ne02, const int ne0, + const int ne1, const int ne2, dpct::queue_ptr stream) { + int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE; + sycl::range<3> gridDim(ne2, ne1, num_blocks); + stream->parallel_for( + sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + pad_f32(x, dst, ne0, ne00, ne01, ne02, item_ct1); + }); +} + +static void rms_norm_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, const float eps, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % WARP_SIZE == 0); + // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE); + if (ncols < 1024) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), WARP_SIZE); + }); + }); + } else { + const int work_group_size = g_work_group_size; + const sycl::range<3> block_dims(1, 1, work_group_size); + /* + DPCT1049:19: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor s_sum_acc_ct1(sycl::range<1>(32), + cgh); + + cgh.parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, + block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + rms_norm_f32(x, dst, ncols, eps, item_ct1, + s_sum_acc_ct1.get_pointer(), work_group_size); + }); + }); + } +} + +static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx, + const int ky, const int kx_padded, + dpct::queue_ptr stream) { + const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE; + const sycl::range<3> num_blocks(1, ky, block_num_x); + const sycl::range<3> block_size(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(num_blocks * block_size, block_size), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + quantize_q8_1(x, vy, kx, kx_padded, item_ct1); + }); + } +} + +template +static void dequantize_block_sycl(const void *__restrict__ vx, + dst_t *__restrict__ y, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>( + sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block(vx, y, k, item_ct1); + }); + } +} + +template +static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 64), + sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q2_K(vx, y, item_ct1); + }); + } +#else + dequantize_block_q2_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 64), + sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q3_K(vx, y, item_ct1); + }); + } +#else + dequantize_block_q3_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 32), + sycl::range<3>(1, 1, 32)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q4_K(vx, y, item_ct1); + }); + } +} + +template +static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 64), + sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q5_K(vx, y, item_ct1); + }); + } +#else + dequantize_block_q5_K<<>>(vx, y); +#endif +} + +template +static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k, + dpct::queue_ptr stream) { + const int nb = k / QK_K; +#if QK_K == 256 + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * + sycl::range<3>(1, 1, 64), + sycl::range<3>(1, 1, 64)), + [=](sycl::nd_item<3> item_ct1) { + dequantize_block_q6_K(vx, y, item_ct1); + }); + } +#else + dequantize_block_q6_K<<>>(vx, y); +#endif +} + +static to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_block_sycl; + case GGML_TYPE_Q4_1: + return dequantize_block_sycl; + case GGML_TYPE_Q5_0: + return dequantize_block_sycl; + case GGML_TYPE_Q5_1: + return dequantize_block_sycl; + case GGML_TYPE_Q8_0: + return dequantize_block_sycl; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_sycl; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_sycl; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_sycl; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_sycl; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_sycl; + case GGML_TYPE_F32: + return dequantize_block_sycl<1, 1, convert_f32>; + default: + return nullptr; + } +} + +static to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) { + switch (type) { + case GGML_TYPE_Q4_0: + return dequantize_block_sycl; + case GGML_TYPE_Q4_1: + return dequantize_block_sycl; + case GGML_TYPE_Q5_0: + return dequantize_block_sycl; + case GGML_TYPE_Q5_1: + return dequantize_block_sycl; + case GGML_TYPE_Q8_0: + return dequantize_block_sycl; + case GGML_TYPE_Q2_K: + return dequantize_row_q2_K_sycl; + case GGML_TYPE_Q3_K: + return dequantize_row_q3_K_sycl; + case GGML_TYPE_Q4_K: + return dequantize_row_q4_K_sycl; + case GGML_TYPE_Q5_K: + return dequantize_row_q5_K_sycl; + case GGML_TYPE_Q6_K: + return dequantize_row_q6_K_sycl; + case GGML_TYPE_F16: + return dequantize_block_sycl<1, 1, convert_f16>; + default: + return nullptr; + } +} + +static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec( + vx, y, dst, ncols, nrows, item_ct1); + }); + } +} + +static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 + const int block_num_y = (nrows + ny - 1) / ny; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, ny, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1); + }); +} + +static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, ny, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1); + }); +} + +static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, ny, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1); + }); +} + +static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const sycl::range<3> block_dims(1, 1, 32); + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1); + }); +} + +static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int ny = 2 / K_QUANTS_PER_ITERATION; + const int block_num_y = (nrows + ny - 1) / ny; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, ny, 32); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1); + }); +} + +static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, + nrows, item_ct1); + }); + } +} + +static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK4_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK5_1 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK8_0 == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols, + const int nrows, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % QK_K == 0); + const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y; + const sycl::range<3> block_nums(1, 1, block_num_y); + const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE); + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_q(vx, vy, dst, ncols, nrows, + item_ct1); + }); +} + +int get_device_index_by_id(int id){ + int res = g_sycl_device_id2index[id].index; + // GGML_SYCL_DEBUG("get_device_index_by_id id=%d device_index=%d\n", id, res); + GGML_ASSERT(res>=0); + return res; +} + +int get_device_id_by_index(int index){ + int res = g_device_caps[index].device_id; + GGML_ASSERT(res>=0); + return res; +} + + +int get_current_device_index(){ + return get_device_index_by_id(dpct::dev_mgr::instance().current_device_id()); +} + +static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_0_RDNA2; + mmq_y = MMQ_Y_Q4_0_RDNA2; + nwarps = NWARPS_Q4_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_0_RDNA1; + mmq_y = MMQ_Y_Q4_0_RDNA1; + nwarps = NWARPS_Q4_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_0_AMPERE; + mmq_y = MMQ_Y_Q4_0_AMPERE; + nwarps = NWARPS_Q4_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_0_PASCAL; + mmq_y = MMQ_Y_Q4_0_PASCAL; + nwarps = NWARPS_Q4_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:20: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q4_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q4_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_0_acc_ct1.get_pointer(), + tile_x_d_q4_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:21: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q4_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q4_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_0_acc_ct1.get_pointer(), + tile_x_d_q4_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_1_RDNA2; + mmq_y = MMQ_Y_Q4_1_RDNA2; + nwarps = NWARPS_Q4_1_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_1_RDNA1; + mmq_y = MMQ_Y_Q4_1_RDNA1; + nwarps = NWARPS_Q4_1_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_1_AMPERE; + mmq_y = MMQ_Y_Q4_1_AMPERE; + nwarps = NWARPS_Q4_1_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_1_PASCAL; + mmq_y = MMQ_Y_Q4_1_PASCAL; + nwarps = NWARPS_Q4_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:22: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q4_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); + sycl::local_accessor tile_x_dm_q4_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_1_acc_ct1.get_pointer(), + tile_x_dm_q4_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:23: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q4_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh); + sycl::local_accessor tile_x_dm_q4_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q4_1_acc_ct1.get_pointer(), + tile_x_dm_q4_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_0_RDNA2; + mmq_y = MMQ_Y_Q5_0_RDNA2; + nwarps = NWARPS_Q5_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_0_RDNA1; + mmq_y = MMQ_Y_Q5_0_RDNA1; + nwarps = NWARPS_Q5_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_0_AMPERE; + mmq_y = MMQ_Y_Q5_0_AMPERE; + nwarps = NWARPS_Q5_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_0_PASCAL; + mmq_y = MMQ_Y_Q5_0_PASCAL; + nwarps = NWARPS_Q5_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:24: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_0_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q5_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_0_acc_ct1.get_pointer(), + tile_x_d_q5_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:25: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_0_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q5_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_0_acc_ct1.get_pointer(), + tile_x_d_q5_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_1_RDNA2; + mmq_y = MMQ_Y_Q5_1_RDNA2; + nwarps = NWARPS_Q5_1_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_1_RDNA1; + mmq_y = MMQ_Y_Q5_1_RDNA1; + nwarps = NWARPS_Q5_1_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_1_AMPERE; + mmq_y = MMQ_Y_Q5_1_AMPERE; + nwarps = NWARPS_Q5_1_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_1_PASCAL; + mmq_y = MMQ_Y_Q5_1_PASCAL; + nwarps = NWARPS_Q5_1_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:26: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_1_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q5_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_1_acc_ct1.get_pointer(), + tile_x_dm_q5_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:27: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_1_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q5_1_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_1( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_1_acc_ct1.get_pointer(), + tile_x_dm_q5_1_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q8_0_RDNA2; + mmq_y = MMQ_Y_Q8_0_RDNA2; + nwarps = NWARPS_Q8_0_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q8_0_RDNA1; + mmq_y = MMQ_Y_Q8_0_RDNA1; + nwarps = NWARPS_Q8_0_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q8_0_AMPERE; + mmq_y = MMQ_Y_Q8_0_AMPERE; + nwarps = NWARPS_Q8_0_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q8_0_PASCAL; + mmq_y = MMQ_Y_Q8_0_PASCAL; + nwarps = NWARPS_Q8_0_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:28: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q8_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q8_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q8_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q8_0_acc_ct1.get_pointer(), + tile_x_d_q8_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:29: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_qs_q8_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_d_q8_0_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0), + cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q8_0( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_qs_q8_0_acc_ct1.get_pointer(), + tile_x_d_q8_0_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q2_K_RDNA2; + mmq_y = MMQ_Y_Q2_K_RDNA2; + nwarps = NWARPS_Q2_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q2_K_RDNA1; + mmq_y = MMQ_Y_Q2_K_RDNA1; + nwarps = NWARPS_Q2_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q2_K_AMPERE; + mmq_y = MMQ_Y_Q2_K_AMPERE; + nwarps = NWARPS_Q2_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q2_K_PASCAL; + mmq_y = MMQ_Y_Q2_K_PASCAL; + nwarps = NWARPS_Q2_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:30: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), + cgh); + sycl::local_accessor tile_x_sc_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q2_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q2_K_acc_ct1.get_pointer(), + tile_x_dm_q2_K_acc_ct1.get_pointer(), + tile_x_sc_q2_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:31: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K), + cgh); + sycl::local_accessor tile_x_sc_q2_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q2_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q2_K_acc_ct1.get_pointer(), + tile_x_dm_q2_K_acc_ct1.get_pointer(), + tile_x_sc_q2_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + +#if QK_K == 256 + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q3_K_RDNA2; + mmq_y = MMQ_Y_Q3_K_RDNA2; + nwarps = NWARPS_Q3_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q3_K_RDNA1; + mmq_y = MMQ_Y_Q3_K_RDNA1; + nwarps = NWARPS_Q3_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q3_K_AMPERE; + mmq_y = MMQ_Y_Q3_K_AMPERE; + nwarps = NWARPS_Q3_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q3_K_PASCAL; + mmq_y = MMQ_Y_Q3_K_PASCAL; + nwarps = NWARPS_Q3_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:32: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), + cgh); + sycl::local_accessor tile_x_qh_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); + sycl::local_accessor tile_x_sc_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q3_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q3_K_acc_ct1.get_pointer(), + tile_x_dm_q3_K_acc_ct1.get_pointer(), + tile_x_qh_q3_K_acc_ct1.get_pointer(), + tile_x_sc_q3_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:33: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K), + cgh); + sycl::local_accessor tile_x_qh_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh); + sycl::local_accessor tile_x_sc_q3_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q3_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q3_K_acc_ct1.get_pointer(), + tile_x_dm_q3_K_acc_ct1.get_pointer(), + tile_x_qh_q3_K_acc_ct1.get_pointer(), + tile_x_sc_q3_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +#endif +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q4_K_RDNA2; + mmq_y = MMQ_Y_Q4_K_RDNA2; + nwarps = NWARPS_Q4_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q4_K_RDNA1; + mmq_y = MMQ_Y_Q4_K_RDNA1; + nwarps = NWARPS_Q4_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q4_K_AMPERE; + mmq_y = MMQ_Y_Q4_K_AMPERE; + nwarps = NWARPS_Q4_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q4_K_PASCAL; + mmq_y = MMQ_Y_Q4_K_PASCAL; + nwarps = NWARPS_Q4_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:34: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), + cgh); + sycl::local_accessor tile_x_sc_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q4_K_acc_ct1.get_pointer(), + tile_x_dm_q4_K_acc_ct1.get_pointer(), + tile_x_sc_q4_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:35: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K), + cgh); + sycl::local_accessor tile_x_sc_q4_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q4_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q4_K_acc_ct1.get_pointer(), + tile_x_dm_q4_K_acc_ct1.get_pointer(), + tile_x_sc_q4_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q5_K_RDNA2; + mmq_y = MMQ_Y_Q5_K_RDNA2; + nwarps = NWARPS_Q5_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q5_K_RDNA1; + mmq_y = MMQ_Y_Q5_K_RDNA1; + nwarps = NWARPS_Q5_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q5_K_AMPERE; + mmq_y = MMQ_Y_Q5_K_AMPERE; + nwarps = NWARPS_Q5_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q5_K_PASCAL; + mmq_y = MMQ_Y_Q5_K_PASCAL; + nwarps = NWARPS_Q5_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:36: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), + cgh); + sycl::local_accessor tile_x_sc_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_K_acc_ct1.get_pointer(), + tile_x_dm_q5_K_acc_ct1.get_pointer(), + tile_x_sc_q5_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:37: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K), + cgh); + sycl::local_accessor tile_x_sc_q5_K_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q5_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_q5_K_acc_ct1.get_pointer(), + tile_x_dm_q5_K_acc_ct1.get_pointer(), + tile_x_sc_q5_K_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy, + float *dst, const int ncols_x, + const int nrows_x, const int ncols_y, + const int nrows_y, const int nrows_dst, + dpct::queue_ptr stream) try { + + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + const int compute_capability = g_device_caps[id].cc; + + int mmq_x, mmq_y, nwarps; + if (compute_capability >= VER_GEN13) { + mmq_x = MMQ_X_Q6_K_RDNA2; + mmq_y = MMQ_Y_Q6_K_RDNA2; + nwarps = NWARPS_Q6_K_RDNA2; + } else if (compute_capability >= VER_GEN12) { + mmq_x = MMQ_X_Q6_K_RDNA1; + mmq_y = MMQ_Y_Q6_K_RDNA1; + nwarps = NWARPS_Q6_K_RDNA1; + } else if (compute_capability >= VER_GEN9) { + mmq_x = MMQ_X_Q6_K_AMPERE; + mmq_y = MMQ_Y_Q6_K_AMPERE; + nwarps = NWARPS_Q6_K_AMPERE; + } else if (compute_capability >= VER_4VEC) { + mmq_x = MMQ_X_Q6_K_PASCAL; + mmq_y = MMQ_Y_Q6_K_PASCAL; + nwarps = NWARPS_Q6_K_PASCAL; + } else { + GGML_ASSERT(false); + } + + const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y; + const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x; + const sycl::range<3> block_nums(1, block_num_y, block_num_x); + const sycl::range<3> block_dims(1, nwarps, WARP_SIZE); + + if (nrows_x % mmq_y == 0) { + const bool need_check = false; + /* + DPCT1049:38: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), + cgh); + sycl::local_accessor tile_x_sc_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q6_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_acc_ct1.get_pointer(), + tile_x_dm_acc_ct1.get_pointer(), + tile_x_sc_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } else { + const bool need_check = true; + /* + DPCT1049:39: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->submit([&](sycl::handler &cgh) { + sycl::local_accessor tile_x_ql_acc_ct1( + sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh); + sycl::local_accessor tile_x_dm_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K), + cgh); + sycl::local_accessor tile_x_sc_acc_ct1( + sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh); + sycl::local_accessor tile_y_qs_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE), cgh); + sycl::local_accessor tile_y_ds_acc_ct1( + sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + mul_mat_q6_K( + vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, + nrows_dst, item_ct1, + tile_x_ql_acc_ct1.get_pointer(), + tile_x_dm_acc_ct1.get_pointer(), + tile_x_sc_acc_ct1.get_pointer(), + tile_y_qs_acc_ct1.get_pointer(), + tile_y_ds_acc_ct1.get_pointer()); + }); + }); + } + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y, + float *dst, const int ncols_x, + const int nrows_x, + const int nchannels_x, + const int nchannels_y, + dpct::queue_ptr stream) { + + const sycl::range<3> block_nums(nchannels_y, nrows_x, 1); + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x, + nchannels_y, item_ct1); + }); + } +} + +static void ggml_mul_mat_vec_nc_f16_f32_sycl( + const void *vx, const float *y, float *dst, const int ncols_x, + const int nrows_x, const int row_stride_x, const int nchannels_x, + const int nchannels_y, const int channel_stride_x, dpct::queue_ptr stream) { + + const sycl::range<3> block_nums(nchannels_y, nrows_x, 1); + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x, + row_stride_x, channel_stride_x, + nchannels_y / nchannels_x, item_ct1); + }); + } +} + +static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, nb00, nb01, + nb02, ne10, ne11, nb10, nb11, nb12, + item_ct1); + }); + } +} + +static void ggml_cpy_f32_f16_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, nb00, nb01, + nb02, ne10, ne11, nb10, nb11, nb12, + item_ct1); + }); + } +} + +static void ggml_cpy_f32_q8_0_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + GGML_ASSERT(ne % QK8_0 == 0); + const int num_blocks = ne / QK8_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), + sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q( + cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, + ne10, ne11, nb10, nb11, nb12, item_ct1); + }); +} + +static void ggml_cpy_f32_q4_0_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + GGML_ASSERT(ne % QK4_0 == 0); + const int num_blocks = ne / QK4_0; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), + sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q( + cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, + ne10, ne11, nb10, nb11, nb12, item_ct1); + }); +} + +static void ggml_cpy_f32_q4_1_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + GGML_ASSERT(ne % QK4_1 == 0); + const int num_blocks = ne / QK4_1; + stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), + sycl::range<3>(1, 1, 1)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_q( + cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, + ne10, ne11, nb10, nb11, nb12, item_ct1); + }); +} + +static void ggml_cpy_f16_f16_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, nb00, nb01, + nb02, ne10, ne11, nb10, nb11, nb12, + item_ct1); + }); + } +} + +static void ggml_cpy_i16_i16_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, nb00, nb01, + nb02, ne10, ne11, nb10, nb11, nb12, + item_ct1); + }); + } +} + +static void ggml_cpy_i32_i32_sycl(const char *cx, char *cdst, const int ne, + const int ne00, const int ne01, + const int nb00, const int nb01, + const int nb02, const int ne10, + const int ne11, const int nb10, + const int nb11, const int nb12, + dpct::queue_ptr stream) { + + const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE; + { + // dpct::has_capability_or_fail(stream->get_device(), + // {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + cpy_f32_f16(cx, cdst, ne, ne00, ne01, nb00, nb01, + nb02, ne10, ne11, nb10, nb11, nb12, + item_ct1); + }); + } +} + +static void scale_f32_sycl(const float *x, float *dst, const float scale, + const int k, dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_SCALE_BLOCK_SIZE - 1) / SYCL_SCALE_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + scale_f32(x, dst, scale, k, item_ct1); + }); +} + +static void clamp_f32_sycl(const float *x, float *dst, const float min, + const float max, const int k, + dpct::queue_ptr stream) { + const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE; + stream->parallel_for( + sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + clamp_f32(x, dst, min, max, k, item_ct1); + }); +} + +template +static void rope_sycl(const T *x, T *dst, int ncols, int nrows, + const int32_t *pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, + rope_corr_dims corr_dims, dpct::queue_ptr stream) { + GGML_ASSERT(ncols % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nrows); + if (pos == nullptr) { + /* + DPCT1049:40: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope(x, dst, ncols, pos, freq_scale, p_delta_rows, + freq_base, ext_factor, attn_factor, corr_dims, + item_ct1); + }); + } else { + /* + DPCT1049:41: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope(x, dst, ncols, pos, freq_scale, p_delta_rows, + freq_base, ext_factor, attn_factor, corr_dims, + item_ct1); + }); + } +} + +template +static void rope_neox_sycl(const T *x, T *dst, int ncols, int n_dims, int nrows, + const int32_t *pos, float freq_scale, + int p_delta_rows, float freq_base, float ext_factor, + float attn_factor, rope_corr_dims corr_dims, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int num_blocks_x = (ncols + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> block_nums(1, num_blocks_x, nrows); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.0f / n_dims; + + if (pos == nullptr) { + /* + DPCT1049:42: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, + item_ct1); + }); + } else { + /* + DPCT1049:43: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_neox(x, dst, ncols, n_dims, pos, freq_scale, + p_delta_rows, ext_factor, attn_factor, + corr_dims, theta_scale, inv_ndims, item_ct1); + }); + } +} + +static void rope_glm_f32_sycl(const float *x, float *dst, int ncols, int nrows, + const int32_t *pos, float freq_scale, + int p_delta_rows, float freq_base, int n_ctx, + dpct::queue_ptr stream) { + GGML_ASSERT(ncols % 4 == 0); + const sycl::range<3> block_dims(1, 1, SYCL_ROPE_BLOCK_SIZE / 4); + const int num_blocks_x = (ncols + SYCL_ROPE_BLOCK_SIZE - 1) / SYCL_ROPE_BLOCK_SIZE; + const sycl::range<3> block_nums(1, nrows, num_blocks_x); + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + rope_glm_f32(x, dst, ncols, pos, freq_scale, + p_delta_rows, freq_base, n_ctx, + item_ct1); + }); +} + +static void alibi_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, const int k_rows, + const int n_heads_log2_floor, const float m0, + const float m1, dpct::queue_ptr stream) { + const sycl::range<3> block_dims(1, 1, SYCL_ALIBI_BLOCK_SIZE); + const int num_blocks_x = (ncols + SYCL_ALIBI_BLOCK_SIZE - 1) / (SYCL_ALIBI_BLOCK_SIZE); + const sycl::range<3> block_nums(1, nrows, num_blocks_x); + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + alibi_f32(x, dst, ncols, k_rows, + n_heads_log2_floor, m0, m1, item_ct1); + }); +} + +static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols, + const int nrows, dpct::queue_ptr stream) { + const sycl::range<3> block_dims(1, 1, WARP_SIZE); + const sycl::range<3> block_nums(1, nrows, 1); + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) + [[intel::reqd_sub_group_size(32)]] { + k_sum_rows_f32(x, dst, ncols, item_ct1); + }); +} + +static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols, + const int nrows, ggml_sort_order order, + dpct::queue_ptr stream) { + // bitonic sort requires ncols to be power of 2 + GGML_ASSERT((ncols & (ncols - 1)) == 0); + + const sycl::range<3> block_dims(1, 1, ncols); + const sycl::range<3> block_nums(1, nrows, 1); + if (order == GGML_SORT_ASC) { + /* + DPCT1049:44: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_argsort_f32_i32(x, dst, ncols, item_ct1); + }); + } else if (order == GGML_SORT_DESC) { + /* + DPCT1049:45: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + stream->parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_argsort_f32_i32(x, dst, ncols, item_ct1); + }); + } else { + GGML_ASSERT(false); + } +} + +static void diag_mask_inf_f32_sycl(const float *x, float *dst, + const int ncols_x, const int nrows_x, + const int rows_per_channel, const int n_past, + dpct::queue_ptr stream) { + const sycl::range<3> block_dims(1, SYCL_DIAG_MASK_INF_BLOCK_SIZE, 1); + const int block_num_x = (ncols_x + SYCL_DIAG_MASK_INF_BLOCK_SIZE - 1) / SYCL_DIAG_MASK_INF_BLOCK_SIZE; + const sycl::range<3> block_nums(1, block_num_x, nrows_x); + stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + diag_mask_inf_f32(x, dst, ncols_x, + rows_per_channel, n_past, + item_ct1); + }); +} + +static void soft_max_f32_sycl(const float *x, const float *y, float *dst, + const int ncols_x, const int nrows_x, + const int nrows_y, const float scale, + dpct::queue_ptr stream) { + int nth = WARP_SIZE; + while (nth < ncols_x && nth < SYCL_SOFT_MAX_BLOCK_SIZE) nth *= 2; + const sycl::range<3> block_dims(1, 1, nth); + const sycl::range<3> block_nums(1, 1, nrows_x); + /* + DPCT1049:46: The work-group size passed to the SYCL kernel may exceed the + limit. To get the device limit, query info::device::max_work_group_size. + Adjust the work-group size if needed. + */ + stream->submit([&](sycl::handler &cgh) { + /* + DPCT1101:96: 'SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE' expression was + replaced with a value. Modify the code to use the original expression, + provided in comments, if it is correct. + */ + sycl::local_accessor buf_acc_ct1( + sycl::range<1>(32 /*SYCL_SOFT_MAX_BLOCK_SIZE/WARP_SIZE*/), cgh); + + cgh.parallel_for( + sycl::nd_range<3>(block_nums * block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] { + soft_max_f32(x, y, dst, ncols_x, nrows_y, scale, item_ct1, + buf_acc_ct1.get_pointer()); + }); + }); +} + +static void im2col_f32_f16_sycl(const float *x, sycl::half *dst, int IW, int IH, + int OW, int OH, int KW, int KH, int IC, + int offset_delta, int s0, int s1, int p0, + int p1, int d0, int d1, + dpct::queue_ptr stream) { + const int parallel_elements = OW * KW * KH; + const int num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE; + sycl::range<3> block_nums(IC, OH, num_blocks); + { + dpct::has_capability_or_fail(stream->get_device(), + {sycl::aspect::fp16}); + + stream->parallel_for( + sycl::nd_range<3>(block_nums * + sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE), + sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)), + [=](sycl::nd_item<3> item_ct1) { + im2col_f32_f16(x, dst, offset_delta, IW, IH, OW, KW, KH, + parallel_elements, (IC * KH * KW), s0, s1, p0, + p1, d0, d1, item_ct1); + }); + } +} + +// buffer pool for sycl +#define MAX_SYCL_BUFFERS 256 + +struct scoped_spin_lock { + std::atomic_flag& lock; + scoped_spin_lock(std::atomic_flag& lock) : lock(lock) { + while (lock.test_and_set(std::memory_order_acquire)) { + ; // spin + } + } + ~scoped_spin_lock() { + lock.clear(std::memory_order_release); + } + scoped_spin_lock(const scoped_spin_lock&) = delete; + scoped_spin_lock& operator=(const scoped_spin_lock&) = delete; +}; + +static std::atomic_flag g_sycl_pool_lock = ATOMIC_FLAG_INIT; + +// #define DEBUG_SYCL_MALLOC +struct sycl_buffer { + void * ptr = nullptr; + size_t size = 0; +}; + +static sycl_buffer g_sycl_buffer_pool[GGML_SYCL_MAX_DEVICES][MAX_SYCL_BUFFERS]; +static size_t g_sycl_pool_size[GGML_SYCL_MAX_DEVICES] = {0}; + +static void *ggml_sycl_pool_malloc_leg(size_t size, size_t *actual_size) try { + scoped_spin_lock lock(g_sycl_pool_lock); + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg index %d\n", id); +#ifdef DEBUG_SYCL_MALLOC + int nnz = 0; + size_t max_size = 0; +#endif + size_t best_diff = 1ull << 36; + int ibest = -1; + for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) { + sycl_buffer& b = g_sycl_buffer_pool[id][i]; + if (b.ptr != nullptr) { +#ifdef DEBUG_SYCL_MALLOC + ++nnz; + if (b.size > max_size) max_size = b.size; +#endif + if (b.size >= size) { + size_t diff = b.size - size; + if (diff < best_diff) { + best_diff = diff; + ibest = i; + if (!best_diff) { + void * ptr = b.ptr; + *actual_size = b.size; + b.ptr = nullptr; + b.size = 0; + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 1 %p\n", ptr); + return ptr; + } + } + } + } + } + if (ibest >= 0) { + sycl_buffer& b = g_sycl_buffer_pool[id][ibest]; + void * ptr = b.ptr; + *actual_size = b.size; + b.ptr = nullptr; + b.size = 0; + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return 2 %p\n", ptr); + return ptr; + } + void * ptr; + size_t look_ahead_size = (size_t) (1.05 * size); + look_ahead_size = 256 * ((look_ahead_size + 255)/256); + + const dpct::queue_ptr stream = g_syclStreams[id][0]; + SYCL_CHECK( + CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device( + look_ahead_size, *stream))); + *actual_size = look_ahead_size; + g_sycl_pool_size[id] += look_ahead_size; + +#ifdef DEBUG_SYCL_MALLOC + fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz, + (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024)); +#endif + // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg return %p\n", ptr); + return ptr; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_pool_free_leg(void *ptr, size_t size) try { + scoped_spin_lock lock(g_sycl_pool_lock); + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + + const dpct::queue_ptr stream = g_syclStreams[id][0]; + for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) { + sycl_buffer& b = g_sycl_buffer_pool[id][i]; + if (b.ptr == nullptr) { + b.ptr = ptr; + b.size = size; + return; + } + } + fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_SYCL_BUFFERS\n"); + SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *stream))); + g_sycl_pool_size[id] -= size; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +// pool with virtual memory +/* +DPCT1082:64: Migration of CUmemGenericAllocationHandle type is not supported. +*/ +// static std::vector +// g_sycl_pool_handles[GGML_SYCL_MAX_DEVICES]; +static dpct::device_ptr g_sycl_pool_addr[GGML_SYCL_MAX_DEVICES] = {0}; +static size_t g_sycl_pool_used[GGML_SYCL_MAX_DEVICES] = {0}; +static const size_t SYCL_POOL_VMM_MAX_SIZE = 1ull << 36; // 64 GB + +static void *ggml_sycl_pool_malloc_vmm(size_t size, size_t *actual_size) try { + + return NULL; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_pool_free_vmm(void *ptr, size_t size) try { + scoped_spin_lock lock(g_sycl_pool_lock); + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = dpct::dev_mgr::instance().current_device_id())); + +#ifdef DEBUG_SYCL_MALLOC + printf("sycl pool[%d]: freed %llu bytes at %llx\n", id, (unsigned long long) size, ptr); +#endif + + g_sycl_pool_used[id] -= size; + + // all deallocations must be in reverse order of the allocations + GGML_ASSERT(ptr == (void *) (g_sycl_pool_addr[id] + g_sycl_pool_used[id])); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void *ggml_sycl_pool_malloc(size_t size, size_t *actual_size) try { + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + if (g_device_caps[id].vmm) { + return ggml_sycl_pool_malloc_vmm(size, actual_size); + } else { + return ggml_sycl_pool_malloc_leg(size, actual_size); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_pool_free(void *ptr, size_t size) try { + int id; + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + if (g_device_caps[id].vmm) { + ggml_sycl_pool_free_vmm(ptr, size); + } else { + ggml_sycl_pool_free_leg(ptr, size); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + + +template +struct sycl_pool_alloc { + T * ptr = nullptr; + size_t actual_size = 0; + + // size is in number of elements + T * alloc(size_t size) { + GGML_ASSERT(ptr == nullptr); + ptr = (T *) ggml_sycl_pool_malloc(size * sizeof(T), &this->actual_size); + // GGML_SYCL_DEBUG("alloc %lu return %p actual size=%lu\n", size * sizeof(T), ptr, this->actual_size); + return ptr; + } + + sycl_pool_alloc(size_t size) { + alloc(size); + } + + ~sycl_pool_alloc() { + if (ptr != nullptr) { + ggml_sycl_pool_free(ptr, actual_size); + } + } + + T * get() { + return ptr; + } + + sycl_pool_alloc() = default; + sycl_pool_alloc(const sycl_pool_alloc &) = delete; + sycl_pool_alloc(sycl_pool_alloc &&) = delete; + sycl_pool_alloc& operator=(const sycl_pool_alloc &) = delete; + sycl_pool_alloc& operator=(sycl_pool_alloc &&) = delete; +}; + +static bool g_sycl_loaded = false; + +bool ggml_sycl_loaded(void) { + return g_sycl_loaded; +} + +void ggml_backend_sycl_print_sycl_devices(){ + int device_count = dpct::dev_mgr::instance().device_count(); + fprintf(stderr, "found %d SYCL devices:\n", device_count); + for (int id = 0; id < device_count; ++id) { + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(id)))); + sycl::device cur_device = dpct::dev_mgr::instance().get_device(id); + fprintf(stderr, " Device %d: %s,\tcompute capability %d.%d,\n\tmax compute_units %d,\tmax work group size %d,\tmax sub group size %d,\tglobal mem size %lu\n", id, + prop.get_name(), prop.get_major_version(), + prop.get_minor_version(), + prop.get_max_compute_units(), + prop.get_max_work_group_size(), + prop.get_max_sub_group_size(), + prop.get_global_mem_size() + ); + } + // fprintf(stderr, "\n"); +} + +int get_sycl_env(const char* env_name, int default_val){ + char * user_device_string = getenv(env_name); + int user_number = default_val; + + unsigned n; + if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1) { + user_number = (int)n; + } else { + user_number=default_val; + } + return user_number; +} + +int get_work_group_size(int user_device_id){ + dpct::device_info prop; + dpct::get_device_info( + prop, + dpct::dev_mgr::instance().get_device(user_device_id)); + return prop.get_max_work_group_size(); +} + +void ggml_init_sycl() try { + static bool initialized = false; + + if (!initialized) { + g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0); + + printf("GGML_SYCL_DEBUG=%d\n", g_ggml_sycl_debug); + + int user_device_id = get_sycl_env("GGML_SYCL_DEVICE", 0); + + if (CHECK_TRY_ERROR(g_all_sycl_device_count = + dpct::dev_mgr::instance().device_count()) != + 0) { + initialized = true; + g_sycl_loaded = false; + return; + } + GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES); + int64_t total_vram = 0; + +#if defined(GGML_SYCL_FP16) + fprintf(stderr, "%s: GGML_SYCL_FP16: yes\n", __func__); +#else + fprintf(stderr, "%s: GGML_SYCL_FP16: no\n", __func__); +#endif + + +#if defined(SYCL_USE_XMX) + fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__); +#else + fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); +#endif + ggml_backend_sycl_print_sycl_devices(); + for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { + g_sycl_device_id2index[id].index = -1; + g_device_caps[id].vmm = 0; + g_device_caps[id].device_id = -1; + g_device_caps[id].cc = 0; + g_tensor_split[id] = 0; + } + + int device_inx = -1; + for (int id = 0; id < g_all_sycl_device_count; ++id) { + if(id!=user_device_id) continue; + + device_inx++; + int device_vmm = 0; + + g_device_caps[device_inx].vmm = !!device_vmm; + g_device_caps[device_inx].device_id = id; + g_sycl_device_id2index[id].index = device_inx; + + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(id)))); + + // fprintf(stderr, + // " Device %d: %s, compute capability %d.%d, VMM: %s\n", id, + // prop.get_name(), prop.get_major_version(), + // prop.get_minor_version(), device_vmm ? "yes" : "no"); + + g_tensor_split[device_inx] = total_vram; + total_vram += prop.get_global_mem_size(); + + g_device_caps[device_inx].cc = + 100 * prop.get_major_version() + 10 * prop.get_minor_version(); + + // printf("g_device_caps[%d].cc=%d\n", device_inx, g_device_caps[device_inx].cc); + } + device_inx = -1; + for (int id = 0; id < g_all_sycl_device_count; ++id) { + if(id!=user_device_id) continue; + device_inx++; + g_tensor_split[device_inx] /= total_vram; + } + + device_inx = -1; + for (int id = 0; id < g_all_sycl_device_count; ++id) { + if(id!=user_device_id) continue; + device_inx++; + SYCL_CHECK(ggml_sycl_set_device(id)); + + // create sycl streams + for (int is = 0; is < MAX_STREAMS; ++is) { + /* + DPCT1025:88: The SYCL queue is created ignoring the flag and + priority options. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + g_syclStreams[device_inx][is] = + dpct::get_current_device().create_queue())); + } + + const dpct::queue_ptr stream = g_syclStreams[device_inx][0]; + // create sycl handle + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[device_inx] = + stream)); + /* + DPCT1027:89: The call to syclSetMathMode was replaced with 0 + because this functionality is redundant in SYCL. + */ + SYCL_CHECK(0); + } + + // configure logging to stdout + // SYCL_CHECK(syclLoggerConfigure(1, 1, 0, nullptr)); + + //hardcode, force set to 1 device + g_device_count = 1; + ggml_sycl_set_main_device(user_device_id); + ggml_sycl_set_device(user_device_id); + g_work_group_size = get_work_group_size(user_device_id); + // fprintf(stderr, "Using Device %d\n", user_device_id); + + // for (int id = 0; id < g_all_sycl_device_count; ++id) { + // GGML_SYCL_DEBUG("id=%d g_device_caps[%d].device_id=%d g_sycl_device_id2index[%d].index=%d ", id, id, + // g_device_caps[id].device_id, id, g_sycl_device_id2index[id].index); + // } + + initialized = true; + g_sycl_loaded = true; + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + + +void ggml_sycl_set_tensor_split(const float * tensor_split) { + if (tensor_split == nullptr) { + return; + } + bool all_zero = true; + for (int i = 0; i < g_device_count; ++i) { + if (tensor_split[i] != 0.0f) { + all_zero = false; + break; + } + } + if (all_zero) { + return; + } + float split_sum = 0.0f; + for (int i = 0; i < g_device_count; ++i) { + g_tensor_split[i] = split_sum; + split_sum += tensor_split[i]; + } + for (int i = 0; i < g_device_count; ++i) { + g_tensor_split[i] /= split_sum; + } +} + +void *ggml_sycl_host_malloc(size_t size) try { + if (getenv("GGML_SYCL_NO_PINNED") != nullptr) { + return nullptr; + } + + void * ptr = nullptr; + //allow to use dpct::get_in_order_queue() for host malloc + dpct::err0 err = CHECK_TRY_ERROR( + ptr = (void *)sycl::malloc_host(size, dpct::get_in_order_queue())); + /* + DPCT1000:82: Error handling if-stmt was detected but could not be rewritten. + */ + if (err != 0) { + // clear the error + /* + DPCT1026:83: The call to syclGetLastError was removed because this + functionality is redundant in SYCL. + */ + /* + DPCT1001:81: The statement could not be removed. + */ + fprintf( + stderr, + "WARNING: failed to allocate %.2f MB of pinned memory: %s\n", + /* + DPCT1009:84: SYCL uses exceptions to report errors and does not use + the error codes. The original code was commented out and a warning + string was inserted. You need to rewrite this code. + */ + size / 1024.0 / 1024.0, + "syclGetErrorString is not supported" /*syclGetErrorString(err)*/); + return nullptr; + } + + return ptr; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_host_free(void *ptr) try { + //allow to use dpct::get_in_order_queue() for host malloc + SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue()))); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst, + const struct ggml_tensor *src, + int64_t i3, int64_t i2, + int64_t i1_low, int64_t i1_high, + dpct::queue_ptr stream) try { + + dpct::memcpy_direction kind; + char * src_ptr; + if (src->backend == GGML_BACKEND_CPU) { + kind = dpct::host_to_device; + src_ptr = (char *) src->data; + // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d GGML_BACKEND_CPU src_ptr %p\n", src_ptr); + } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) { + GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1])); + kind = dpct::device_to_device; + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra; + int id; + SYCL_CHECK(CHECK_TRY_ERROR( + id = get_current_device_index())); + // GGML_SYCL_DEBUG("current device index %d\n", id); + src_ptr = (char *) extra->data_device[id]; + } else { + // GGML_SYCL_DEBUG("GGML_ASSERT(false)\n"); + GGML_ASSERT(false); + } + char * dst_ptr = (char *) dst; + + const int64_t ne0 = src->ne[0]; + const int64_t nb0 = src->nb[0]; + const int64_t nb1 = src->nb[1]; + const int64_t nb2 = src->nb[2]; + const int64_t nb3 = src->nb[3]; + const enum ggml_type type = src->type; + const int64_t ts = ggml_type_size(type); + const int64_t bs = ggml_blck_size(type); + int64_t i1_diff = i1_high - i1_low; + + const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3; + if (nb0 == ts && nb1 == ts*ne0/bs) { + // GGML_SYCL_DEBUG("stream->memcpy: dst_ptr=%p, x=%p, size=%lu\n", dst_ptr, x, i1_diff * nb1); + // return CHECK_TRY_ERROR(stream->memcpy(dst_ptr, x, i1_diff * nb1)); + return CHECK_TRY_ERROR(dpct::async_dpct_memcpy(dst_ptr, x, i1_diff * nb1, + kind, *stream)); + + } else if (nb0 == ts) { + return CHECK_TRY_ERROR( + dpct::async_dpct_memcpy(dst_ptr, ts * ne0 / bs, x, nb1, + ts * ne0 / bs, i1_diff, kind, *stream)); + } else { + for (int64_t i1 = 0; i1 < i1_diff; i1++) { + const void * rx = (const void *) ((const char *) x + i1*nb1); + void * rd = (void *) (dst_ptr + i1*ts*ne0/bs); + // pretend the row is a matrix with cols=1 + dpct::err0 r = CHECK_TRY_ERROR(dpct::async_dpct_memcpy( + rd, ts / bs, rx, nb0, ts / bs, ne0, kind, *stream)); + /* + DPCT1001:85: The statement could not be removed. + */ + /* + DPCT1000:86: Error handling if-stmt was detected but could not be + rewritten. + */ + if (r != 0) return r; + } + return 0; + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_op_get_rows(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_d, const float *src1_d, + float *dst_d, const dpct::queue_ptr &stream) { + + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type)); + GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type)); + GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type)); + + const int32_t * src1_i32 = (const int32_t *) src1_d; + + switch (src0->type) { + case GGML_TYPE_F16: + get_rows_sycl_float(src0, src1, dst, (const sycl::half *)src0_d, + src1_i32, dst_d, stream); + break; + case GGML_TYPE_F32: + get_rows_sycl_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q4_0: + get_rows_sycl(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q4_1: + get_rows_sycl(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q5_0: + get_rows_sycl(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q5_1: + get_rows_sycl(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + case GGML_TYPE_Q8_0: + get_rows_sycl(src0, src1, dst, src0_d, src1_i32, dst_d, stream); + break; + default: + // TODO: k-quants + fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type)); + GGML_ASSERT(false); + break; + } +} + +template +inline void ggml_sycl_op_bin_bcast(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) { + op()(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { + op()(src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, + (sycl::half *)dst_dd, main_stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { + op()(src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd, + main_stream); + } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) { + op()(src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd, + main_stream); + } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) { + op()(src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd, + main_stream); + } else { + fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__, + ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type)); + GGML_ASSERT(false); + } +} + +static void ggml_sycl_op_repeat(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_d, const float *src1_d, + float *dst_d, + const dpct::queue_ptr &main_stream) { + + ggml_sycl_op_bin_bcast>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream); + + (void) src1; + (void) src1_d; +} + +inline void ggml_sycl_op_add(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + ggml_sycl_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); +} + +inline void ggml_sycl_op_acc(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported + + int nb1 = dst->op_params[0] / 4; // 4 bytes of float32 + int nb2 = dst->op_params[1] / 4; // 4 bytes of float32 + // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused + int offset = dst->op_params[3] / 4; // offset in bytes + + acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream); + + (void) dst; +} + +inline void ggml_sycl_op_mul(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + ggml_sycl_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); +} + +inline void ggml_sycl_op_div(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + ggml_sycl_op_bin_bcast>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream); +} + +inline void ggml_sycl_op_gelu(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_silu(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_gelu_quick(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_tanh(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_relu(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_leaky_relu(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float negative_slope; + memcpy(&negative_slope, dst->op_params, sizeof(float)); + + leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_sqr(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_norm(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_group_norm(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + int num_groups = dst->op_params[0]; + int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups); + group_norm_f32_sycl(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_concat(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + + for (int i3 = 0; i3 < dst->ne[3]; i3++) { + concat_f32_sycl(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream); + } + + (void) src1; + (void) dst; +} + +inline void ggml_sycl_op_upscale(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors + + const int scale_factor = dst->op_params[0]; + + upscale_f32_sycl(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_pad(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT(dst->type == GGML_TYPE_F32); + GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors + + pad_f32_sycl(src0_dd, dst_dd, + src0->ne[0], src0->ne[1], src0->ne[2], + dst->ne[0], dst->ne[1], dst->ne[2], main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_rms_norm(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + float eps; + memcpy(&eps, dst->op_params, sizeof(float)); + + rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_mul_mat_q( + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) try { + + const int64_t ne00 = src0->ne[0]; + + const int64_t ne10 = src1->ne[0]; + GGML_ASSERT(ne10 % QK8_1 == 0); + + const int64_t ne0 = dst->ne[0]; + + const int64_t row_diff = row_high - row_low; + + int device_id; + SYCL_CHECK( + CHECK_TRY_ERROR(device_id = dpct::dev_mgr::instance().current_device_id())); + + // the main device has a larger memory buffer to hold the results from all GPUs + // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into + const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + ggml_mul_mat_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_1: + ggml_mul_mat_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_0: + ggml_mul_mat_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_1: + ggml_mul_mat_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q8_0: + ggml_mul_mat_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q2_K: + ggml_mul_mat_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q3_K: + ggml_mul_mat_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q4_K: + ggml_mul_mat_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q5_K: + ggml_mul_mat_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + case GGML_TYPE_Q6_K: + ggml_mul_mat_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddf_i; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static int64_t get_row_rounding(ggml_type type) { + int64_t min_compute_capability = INT_MAX; + int64_t max_compute_capability = INT_MIN; + for (int64_t id = 0; id < g_device_count; ++id) { + if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + if (min_compute_capability > g_device_caps[id].cc) { + min_compute_capability = g_device_caps[id].cc; + } + if (max_compute_capability < g_device_caps[id].cc) { + max_compute_capability = g_device_caps[id].cc; + } + } + } + + switch(type) { + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + return max_compute_capability >= VER_GEN9 ? 128 : 64; + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + return 64; + case GGML_TYPE_F16: + case GGML_TYPE_F32: + return 1; + case GGML_TYPE_Q2_K: + case GGML_TYPE_Q3_K: + case GGML_TYPE_Q4_K: + case GGML_TYPE_Q5_K: + return max_compute_capability >= VER_GEN9 ? 128 : 64; + case GGML_TYPE_Q6_K: + return 64; + default: + GGML_ASSERT(false); + } +} + +inline void ggml_sycl_op_mul_mat_vec_q( + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) { + + GGML_ASSERT(ggml_nrows(src1) == 1); + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + switch (src0->type) { + case GGML_TYPE_Q4_0: + mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddf_i; + (void) src1_ncols; + (void) src1_padded_row_size; +} + +inline void ggml_sycl_op_dequantize_mul_mat_vec( + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) { + + const int64_t ne00 = src0->ne[0]; + const int64_t row_diff = row_high - row_low; + + // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics +#ifdef GGML_SYCL_F16 + sycl_pool_alloc src1_dfloat_a; + sycl::half *src1_dfloat = nullptr; // dfloat == half + + bool src1_convert_f16 = + src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 || + src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 || + src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16; + + if (src1_convert_f16) { + if (src1->type == GGML_TYPE_F16) { + src1_dfloat = (sycl::half *)src1->data + src1_padded_row_size; + } else { + src1_dfloat = src1_dfloat_a.alloc(ne00); + ggml_cpy_f32_f16_sycl((const char *)src1_ddf_i, (char *)src1_dfloat, + ne00, ne00, 1, sizeof(float), 0, 0, ne00, 1, + sizeof(sycl::half), 0, 0, stream); + } + } +#else + const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion +#endif // GGML_SYCL_F16 + + switch (src0->type) { + case GGML_TYPE_Q4_0: + dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_1: + dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_0: + dequantize_mul_mat_vec_q5_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_1: + dequantize_mul_mat_vec_q5_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q8_0: + dequantize_mul_mat_vec_q8_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q2_K: + dequantize_mul_mat_vec_q2_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q3_K: + dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q4_K: + dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q5_K: + dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_Q6_K: + dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream); + break; + case GGML_TYPE_F16: + convert_mul_mat_vec_f16_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream); + break; + default: + GGML_ASSERT(false); + break; + } + + (void) src1; + (void) dst; + (void) src1_ddq_i; + (void) src1_ncols; + (void) src1_padded_row_size; +} + +inline void ggml_sycl_op_mul_mat_sycl( + const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst, + const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i, + float *dst_dd_i, const int64_t row_low, const int64_t row_high, + const int64_t src1_ncols, const int64_t src1_padded_row_size, + const dpct::queue_ptr &stream) try { + + GGML_ASSERT(src0_dd_i != nullptr); + GGML_ASSERT(src1_ddf_i != nullptr); + GGML_ASSERT(dst_dd_i != nullptr); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne10 = src1->ne[0]; + + const int64_t ne0 = dst->ne[0]; + + const int64_t row_diff = row_high - row_low; + + int id; + int device_id = dpct::dev_mgr::instance().current_device_id(); + SYCL_CHECK( + CHECK_TRY_ERROR(id = get_current_device_index())); + + // the main device has a larger memory buffer to hold the results from all GPUs + // ldc == nrows of the matrix that cuBLAS writes into + int ldc = dst->backend == GGML_BACKEND_GPU && device_id == g_main_device ? ne0 : row_diff; + + const int compute_capability = g_device_caps[id].cc; +#ifdef GGML_SYCL_F16 + bool use_fp16 = true; // TODO(Yu) SYCL capability check +#else + bool use_fp16 = false; +#endif + // if (compute_capability >= VER_GEN9 && (src0->type == GGML_TYPE_F16 || + // ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == + // src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) { + if ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && + use_fp16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && + dst->op_params[0] == GGML_PREC_DEFAULT) { + + // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 + // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp16 path\n"); + sycl_pool_alloc src0_as_f16; + if (src0->type != GGML_TYPE_F16) { + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + size_t ne = row_diff*ne00; + src0_as_f16.alloc(ne); + to_fp16_sycl(src0_dd_i, src0_as_f16.get(), ne, stream); + } + const sycl::half *src0_ptr = src0->type == GGML_TYPE_F16 + ? (const sycl::half *)src0_dd_i + : src0_as_f16.get(); + + sycl_pool_alloc src1_as_f16; + if (src1->type != GGML_TYPE_F16) { + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + size_t ne = src1_ncols*ne10; + src1_as_f16.alloc(ne); + to_fp16_sycl(src1_ddf_i, src1_as_f16.get(), ne, stream); + } + const sycl::half *src1_ptr = src1->type == GGML_TYPE_F16 + ? (const sycl::half *)src1->data + src1_padded_row_size + : src1_as_f16.get(); + sycl_pool_alloc dst_f16(row_diff * src1_ncols); + + const sycl::half alpha_f16 = 1.0f; + const sycl::half beta_f16 = 0.0f; + + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[id] = stream)); + SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm( + *g_sycl_handles[id], oneapi::mkl::transpose::trans, + oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, + &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00, + src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16, + dst_f16.get(), dpct::library_data_t::real_half, ldc, + dpct::library_data_t::real_half))); + + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); + to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream); + } + else { + // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp32 path\n"); + sycl_pool_alloc src0_ddq_as_f32; + + if (src0->type != GGML_TYPE_F32) { + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type); + GGML_ASSERT(to_fp32_sycl != nullptr); + src0_ddq_as_f32.alloc(row_diff*ne00); + to_fp32_sycl(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream); + } + const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get(); + + const float alpha = 1.0f; + const float beta = 0.0f; + + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[id] = stream)); + SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm( + *g_sycl_handles[id], oneapi::mkl::transpose::trans, + oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, + dpct::get_value(&alpha, *g_sycl_handles[id]), src0_ddf_i, ne00, + src1_ddf_i, ne10, dpct::get_value(&beta, *g_sycl_handles[id]), + dst_dd_i, ldc))); + } + + (void) dst; + (void) src1_ddq_i; + (void) src1_padded_row_size; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16); + GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); + GGML_ASSERT(src0->type == dst->type); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne2 = dst->ne[2]; + const int64_t nrows = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + + // RoPE alteration for extended context + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + + const int32_t * pos = nullptr; + if ((mode & 1) == 0) { + GGML_ASSERT(src1->type == GGML_TYPE_I32); + GGML_ASSERT(src1->ne[0] == ne2); + pos = (const int32_t *) src1_dd; + } + + const bool is_neox = mode & 2; + const bool is_glm = mode & 4; + + rope_corr_dims corr_dims; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); + + // compute + if (is_glm) { + GGML_ASSERT(false); + rope_glm_f32_sycl(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream); + } else if (is_neox) { + if (src0->type == GGML_TYPE_F32) { + rope_neox_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_neox_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, + ne00, n_dims, nrows, pos, freq_scale, ne01, + freq_base, ext_factor, attn_factor, corr_dims, + main_stream); + } else { + GGML_ASSERT(false); + } + } else { + if (src0->type == GGML_TYPE_F32) { + rope_sycl( + (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); + } else if (src0->type == GGML_TYPE_F16) { + rope_sycl((const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, + nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream); + } else { + GGML_ASSERT(false); + } + } + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t nrows = ggml_nrows(src0); + + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_head = ((int32_t *) dst->op_params)[1]; + float max_bias; + memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float)); + + //GGML_ASSERT(ne01 + n_past == ne00); + GGML_ASSERT(n_head == ne02); + + const int n_heads_log2_floor = 1 << (int) floor(log2(n_head)); + + const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor); + const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor); + + alibi_f32_sycl(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream); + + (void) src1; + (void) src1_dd; +} + +inline void ggml_sycl_op_im2col(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; + + const int64_t IC = src1->ne[is_2D ? 2 : 1]; + const int64_t IH = is_2D ? src1->ne[1] : 1; + const int64_t IW = src1->ne[0]; + + const int64_t KH = is_2D ? src0->ne[1] : 1; + const int64_t KW = src0->ne[0]; + + const int64_t OH = is_2D ? dst->ne[2] : 1; + const int64_t OW = dst->ne[1]; + + const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 + + im2col_f32_f16_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, + IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream); + + (void) src0; + (void) src0_dd; +} + +inline void ggml_sycl_op_sum_rows(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ncols = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_argsort(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_I32); + + const int64_t ncols = src0->ne[0]; + const int64_t nrows = ggml_nrows(src0); + + enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0]; + + argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_diag_mask_inf(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int nrows0 = ggml_nrows(src0); + + const int n_past = ((int32_t *) dst->op_params)[0]; + + diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_soft_max(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const float *src0_dd, const float *src1_dd, + float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional + + const int64_t ne00 = src0->ne[0]; + const int64_t nrows_x = ggml_nrows(src0); + const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1; + + float scale = 1.0f; + memcpy(&scale, dst->op_params, sizeof(float)); + + soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream); + + (void) dst; +} + +inline void ggml_sycl_op_scale(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float scale; + memcpy(&scale, dst->op_params, sizeof(float)); + + scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream); + /* + DPCT1010:87: SYCL uses exceptions to report errors and does not use the + error codes. The call was replaced with 0. You need to rewrite this code. + */ + SYCL_CHECK(0); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_sycl_op_clamp(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst, const float *src0_dd, + const float *src1_dd, float *dst_dd, + const dpct::queue_ptr &main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + float min; + float max; + memcpy(&min, dst->op_params, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); + + clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream); + /* + DPCT1010:88: SYCL uses exceptions to report errors and does not use the + error codes. The call was replaced with 0. You need to rewrite this code. + */ + SYCL_CHECK(0); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +static void ggml_sycl_op_flatten(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + const ggml_sycl_op_flatten_t op) try { + const int64_t nrows0 = ggml_nrows(src0); + + const bool use_src1 = src1 != nullptr; + const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1; + + GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT( dst->backend != GGML_BACKEND_GPU_SPLIT); + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr; + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + + const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT; + const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU; + + // dd = data device + float * src0_ddf = nullptr; + float * src1_ddf = nullptr; + float * dst_ddf = nullptr; + + sycl_pool_alloc src0_f; + sycl_pool_alloc src1_f; + sycl_pool_alloc dst_f; + + ggml_sycl_set_device(g_main_device); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + // GGML_SYCL_DEBUG("g_main_device_index=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n", + // g_main_device_index, main_stream, src0_on_device, src1_on_device, dst_on_device); + + if (src0_on_device) { + src0_ddf = (float *) src0_extra->data_device[g_main_device_index]; + } else { + src0_ddf = src0_f.alloc(ggml_nelements(src0)); + // GGML_SYCL_DEBUG("before ggml_sycl_cpy_tensor_2d src0_ddf=%p, src0=%p\n", src0_ddf, src0); + SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream)); + } + + if (use_src1) { + if (src1_on_device) { + src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + } else { + src1_ddf = src1_f.alloc(ggml_nelements(src1)); + SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream)); + } + } + if (dst_on_device) { + dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + // printf("zjy dst_ddf=%p main_stream=%p g_main_device_index=%d\n", dst_ddf, main_stream, g_main_device_index); + } else { + dst_ddf = dst_f.alloc(ggml_nelements(dst)); + } + + // GGML_SYCL_DEBUG("op src0=%p, src1=%p, dst=%p, src0_ddf=%p, src1_ddf=%p, dst_ddf=%p, main_stream=%p\n", + // src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); + // do the computation + op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream); + /* + DPCT1010:89: SYCL uses exceptions to report errors and does not use the + error codes. The call was replaced with 0. You need to rewrite this code. + */ + SYCL_CHECK(0); + + // copy dst to host if necessary + if (!dst_on_device) { + SYCL_CHECK(CHECK_TRY_ERROR( + main_stream->memcpy(dst->data, dst_ddf, ggml_nbytes(dst)))); + } + + if (dst->backend == GGML_BACKEND_CPU) { + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::get_current_device().queues_wait_and_throw())); + } + // print_ggml_tensor("tensor", dst); +} +catch (sycl::exception const &exc) { + + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_set_peer_access(const int n_tokens) { + static bool peer_access_enabled = false; + + const bool enable_peer_access = n_tokens <= GGML_SYCL_PEER_MAX_BATCH_SIZE; + + if (peer_access_enabled == enable_peer_access) { + return; + } + +#ifdef NDEBUG + for (int id = 0; id < g_device_count; ++id) { + SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + // SYCL_CHECK(syclDeviceSynchronize()); + } + + for (int id = 0; id < g_device_count; ++id) { + SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + int device_id = g_device_caps[id].device_id; + + for (int id_other = 0; id_other < g_device_count; ++id_other) { + int device_id_other = g_device_caps[id_other].device_id; + if (device_id == id_other) { + continue; + } + if (device_id != g_main_device && device_id_other != g_main_device) { + continue; + } + + int can_access_peer; + // SYCL_CHECK(syclDeviceCanAccessPeer(&can_access_peer, id, id_other)); + // if (can_access_peer) { + // if (enable_peer_access) { + // SYCL_CHECK(syclDeviceEnablePeerAccess(id_other, 0)); + // } else { + // SYCL_CHECK(syclDeviceDisablePeerAccess(id_other)); + // } + // } + } + } +#endif // NDEBUG + + peer_access_enabled = enable_peer_access; +} + +static void ggml_sycl_op_mul_mat(const ggml_tensor *src0, + const ggml_tensor *src1, ggml_tensor *dst, + ggml_sycl_op_mul_mat_t op, + const bool convert_src1_to_q8_1) try { + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + const int64_t nrows0 = ggml_nrows(src0); + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + const int64_t nrows1 = ggml_nrows(src1); + + GGML_ASSERT(ne03 == ne13); + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + + const int nb2 = dst->nb[2]; + const int nb3 = dst->nb[3]; + + GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT); + + GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0); + + const int64_t i02_divisor = ne12 / ne02; + + const size_t src0_ts = ggml_type_size(src0->type); + const size_t src0_bs = ggml_blck_size(src0->type); + const size_t q8_1_ts = sizeof(block_q8_1); + const size_t q8_1_bs = QK8_1; + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + + const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT; + const bool src0_is_contiguous = ggml_is_contiguous(src0); + const bool src1_is_contiguous = ggml_is_contiguous(src1); + + int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING); + + const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT; + GGML_ASSERT(!(split && ne02 > 1)); + GGML_ASSERT(!(split && ne03 > 1)); + GGML_ASSERT(!(split && ne02 < ne12)); + + // dd = data device + char * src0_dd[GGML_SYCL_MAX_DEVICES] = {nullptr}; + float * src1_ddf[GGML_SYCL_MAX_DEVICES] = {nullptr}; // float + char * src1_ddq[GGML_SYCL_MAX_DEVICES] = {nullptr}; // q8_1 + float * dst_dd[GGML_SYCL_MAX_DEVICES] = {nullptr}; + + // as = actual size + size_t src0_as[GGML_SYCL_MAX_DEVICES] = {0}; + size_t src1_asf[GGML_SYCL_MAX_DEVICES] = {0}; + size_t src1_asq[GGML_SYCL_MAX_DEVICES] = {0}; + size_t dst_as[GGML_SYCL_MAX_DEVICES] = {0}; + + int64_t row_low[GGML_SYCL_MAX_DEVICES]; + int64_t row_high[GGML_SYCL_MAX_DEVICES]; + + int used_devices = 0; + + for (int64_t id = 0; id < g_device_count; ++id) { + // by default, use all rows + row_low[id] = 0; + row_high[id] = ne01; + + // for multi GPU, get the row boundaries from tensor split + // and round to mul_mat_q tile sizes + if (split) { + const int64_t rounding = get_row_rounding(src0->type); + + if (id != 0) { + row_low[id] = ne01*g_tensor_split[id]; + if (row_low[id] < ne01) { + row_low[id] -= row_low[id] % rounding; + } + } + + if (id != g_device_count - 1) { + row_high[id] = ne01*g_tensor_split[id + 1]; + if (row_high[id] < ne01) { + row_high[id] -= row_high[id] % rounding; + } + } + } + } + for (int64_t id = 0; id < g_device_count; ++id) { + + if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { + continue; + } + + used_devices++; + + const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device_index; + + ggml_sycl_set_device(get_device_id_by_index(id)); + const dpct::queue_ptr stream = g_syclStreams[id][0]; + + if (src0_on_device && src0_is_contiguous) { + src0_dd[id] = (char *) src0_extra->data_device[id]; + } else { + // const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); + src0_dd[id] = (char *) ggml_sycl_pool_malloc(ggml_nbytes(src0), &src0_as[id]); + } + + if (src1_on_device && src1_is_contiguous) { + src1_ddf[id] = (float *) src1_extra->data_device[id]; + } else { + src1_ddf[id] = (float *) ggml_sycl_pool_malloc(ggml_nbytes(src1), &src1_asf[id]); + } + + if (convert_src1_to_q8_1) { + src1_ddq[id] = (char *) ggml_sycl_pool_malloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs, &src1_asq[id]); + + if (src1_on_device && src1_is_contiguous) { + quantize_row_q8_1_sycl(src1_ddf[id], src1_ddq[id], ne10, nrows1, src1_padded_col_size, stream); + /* + DPCT1010:90: SYCL uses exceptions to report errors and does not + use the error codes. The call was replaced with 0. You need to + rewrite this code. + */ + SYCL_CHECK(0); + } + } + + if (dst_on_device) { + dst_dd[id] = (float *) dst_extra->data_device[id]; + } else { + const size_t size_dst_ddf = split ? (row_high[id]-row_low[id])*ne1*sizeof(float) : ggml_nbytes(dst); + dst_dd[id] = (float *) ggml_sycl_pool_malloc(size_dst_ddf, &dst_as[id]); + } + } + + // if multiple devices are used they need to wait for the main device + // here an event is recorded that signals that the main device has finished calculating the input data + if (split && used_devices > 1) { + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + /* + DPCT1024:91: The original code returned the error code that was further + consumed by the program logic. This original code was replaced with 0. + You may need to rewrite the program logic consuming the error code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + *src0_extra->events[g_main_device_index][0] = + g_syclStreams[g_main_device_index][0]->ext_oneapi_submit_barrier())); + } + + const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; + for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) { + const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; + const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; + + for (int64_t id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { + continue; + } + + const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device_index; + const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device_index; + const int64_t row_diff = row_high[id] - row_low[id]; + + ggml_sycl_set_device(get_device_id_by_index(id)); + const dpct::queue_ptr stream = g_syclStreams[id][is]; + + // wait for main GPU data if necessary + if (split && (id != g_main_device_index || is != 0)) { + SYCL_CHECK(CHECK_TRY_ERROR(stream->ext_oneapi_submit_barrier( + {*src0_extra->events[g_main_device_index][0]}))); + } + + for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) { + const int64_t i03 = i0 / ne12; + const int64_t i02 = i0 % ne12; + + const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs; + + // for split tensors the data begins at i0 == i0_offset_low + char * src0_dd_i = src0_dd[id] + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs; + float * src1_ddf_i = src1_ddf[id] + (i0*ne11 + src1_col_0) * ne10; + char * src1_ddq_i = src1_ddq[id] + src1_ddq_i_offset; + float * dst_dd_i = dst_dd[id] + (i0*ne1 + src1_col_0) * (dst_on_device ? ne0 : row_diff); + + // the main device memory buffer can be on VRAM scratch, with space for all partial results + // in that case an offset on dst_ddf_i is needed + if (dst->backend == GGML_BACKEND_GPU && id == g_main_device_index) { + dst_dd_i += row_low[id]; // offset is 0 if no tensor split + } + + // copy src0, src1 to device if necessary + if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) { + if (id != g_main_device_index) { + if (convert_src1_to_q8_1) { + char * src1_ddq_i_source = src1_ddq[g_main_device_index] + src1_ddq_i_offset; + SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( + src1_ddq_i, src1_ddq_i_source, + src1_ncols * src1_padded_col_size * q8_1_ts / + q8_1_bs))); + } else { + float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device_index]; + src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10; + SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( + src1_ddf_i, src1_ddf_i_source, + src1_ncols * ne10 * sizeof(float)))); + } + } + } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) { + SYCL_CHECK(ggml_sycl_cpy_tensor_2d( + src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream)); + } else { + GGML_ASSERT(false); + } + + if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) { + quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream); + /* + DPCT1010:92: SYCL uses exceptions to report errors and does + not use the error codes. The call was replaced with 0. You + need to rewrite this code. + */ + SYCL_CHECK(0); + } + + if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) { + SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, row_low[id], row_high[id], stream)); + } + if (src1->type == GGML_TYPE_F16) { + src1_padded_col_size = (i0 * ne11 + src1_col_0) * ne10; + } + // do the computation + op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i, + row_low[id], row_high[id], src1_ncols, src1_padded_col_size, stream); + /* + DPCT1010:93: SYCL uses exceptions to report errors and does not + use the error codes. The call was replaced with 0. You need to + rewrite this code. + */ + SYCL_CHECK(0); + + // copy dst to host or other device if necessary + if (!dst_on_device) { + void * dst_off_device; + dpct::memcpy_direction kind; + if (dst->backend == GGML_BACKEND_CPU) { + dst_off_device = dst->data; + kind = dpct::device_to_host; + } else if (dst->backend == GGML_BACKEND_GPU) { + dst_off_device = dst_extra->data_device[g_main_device_index]; + kind = dpct::device_to_device; + } else { + GGML_ASSERT(false); + } + if (split) { + // src0 = weight matrix is saved as a transposed matrix for better memory layout. + // dst is NOT transposed. + // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU. + // Instead they need to be copied to the correct slice in ne0 = dst row index. + // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results. + float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); + GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); + dhf_dst_i += src1_col_0*ne0 + row_low[id]; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy( + dhf_dst_i, ne0 * sizeof(float), dst_dd_i, + row_diff * sizeof(float), row_diff * sizeof(float), + src1_ncols, kind, *stream))); + } else { + float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3); + GGML_ASSERT(dst->nb[1] == ne0*sizeof(float)); + dhf_dst_i += src1_col_0*ne0; + SYCL_CHECK(CHECK_TRY_ERROR( + stream->memcpy(dhf_dst_i, dst_dd_i, + src1_ncols * ne0 * sizeof(float)))); + } + } + + // add event for the main device to wait on until other device is done + if (split && (id != g_main_device_index || is != 0)) { + /* + DPCT1024:94: The original code returned the error code that + was further consumed by the program logic. This original + code was replaced with 0. You may need to rewrite the + program logic consuming the error code. + */ + SYCL_CHECK(CHECK_TRY_ERROR( + *src0_extra->events[id][is] = + stream->ext_oneapi_submit_barrier())); + } + } + } + } + + for (int64_t id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device_index) || row_low[id] == row_high[id]) { + continue; + } + SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + + // free buffers again when done + if (dst_as[id] > 0) { + ggml_sycl_pool_free(dst_dd[id], dst_as[id]); + } + if (src1_asq[id] > 0) { + ggml_sycl_pool_free(src1_ddq[id], src1_asq[id]); + } + if (src1_asf[id] > 0) { + ggml_sycl_pool_free(src1_ddf[id], src1_asf[id]); + } + if (src0_as[id] > 0) { + ggml_sycl_pool_free(src0_dd[id], src0_as[id]); + } + } + + // main device waits for all other devices to be finished + if (split && g_device_count > 1) { + int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE; + is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS; + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + for (int64_t id = 0; id < g_device_count; ++id) { + if (row_low[id] == row_high[id]) { + continue; + } + for (int64_t is = 0; is < is_max; ++is) { + SYCL_CHECK(CHECK_TRY_ERROR( + g_syclStreams[g_main_device_index][0]->ext_oneapi_submit_barrier( + {*src0_extra->events[id][is]}))); + } + } + } + + if (dst->backend == GGML_BACKEND_CPU) { + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::get_current_device().queues_wait_and_throw())); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_repeat); +} + +static void ggml_sycl_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_get_rows); +} + +static void ggml_sycl_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_add); + // log_tensor_with_cnt("log_ggml_sycl_add_src0", (struct ggml_tensor *) src0, 6); + // log_tensor_with_cnt("log_ggml_sycl_add_src1", (struct ggml_tensor *)src1, 6); + // log_tensor_with_cnt("log_ggml_sycl_add_dst", dst, 6); +} + +static void ggml_sycl_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_acc); +} + +static void ggml_sycl_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_mul); + // log_tensor_with_cnt("log_ggml_sycl_mul_src0", (struct ggml_tensor *)src0, 6); + // log_tensor_with_cnt("log_ggml_sycl_mul_src1", (struct ggml_tensor *)src1, 6); + // log_tensor_with_cnt("log_ggml_sycl_mul_dst", dst, 6); + +} + +static void ggml_sycl_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_div); +} + +static void ggml_sycl_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_gelu); +} + +static void ggml_sycl_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_silu); +} + +static void ggml_sycl_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_gelu_quick); +} + +static void ggml_sycl_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_tanh); +} + +static void ggml_sycl_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_relu); +} + +static void ggml_sycl_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_leaky_relu); +} + +static void ggml_sycl_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_sqr); +} + +static void ggml_sycl_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_norm); +} + +static void ggml_sycl_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_group_norm); +} + +static void ggml_sycl_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_concat); +} + +static void ggml_sycl_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_upscale); +} + +static void ggml_sycl_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_pad); +} + + +static void ggml_sycl_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_SYCL_DEBUG("call %s\n", __func__); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_rms_norm); + // log_tensor_with_cnt("log_ggml_sycl_rms_norm_src0", (struct ggml_tensor *)src0, 6); + // log_tensor_with_cnt("log_ggml_sycl_rms_norm_src1", (struct ggml_tensor *)src1, 6); + // log_tensor_with_cnt("log_ggml_sycl_rms_norm_dst", dst, 6); +} + +bool ggml_sycl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + if (!g_sycl_loaded) return false; + + const int64_t ne10 = src1->ne[0]; + + const int64_t ne0 = dst->ne[0]; + const int64_t ne1 = dst->ne[1]; + + // TODO: find the optimal values for these + return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && + src1->type == GGML_TYPE_F32 && + dst->type == GGML_TYPE_F32 && + (ne0 >= 32 && ne1 >= 32 && ne10 >= 32); +} + +static void ggml_sycl_mul_mat_vec_p021(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst) try { + GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1)); + GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation + GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + + const int64_t ne12 = src1->ne[2]; + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + void * src0_ddq = src0_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + + ggml_mul_mat_p021_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_mul_mat_vec_nc(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst) try { + GGML_ASSERT(!ggml_is_transposed(src0)); + GGML_ASSERT(!ggml_is_transposed(src1)); + GGML_ASSERT(!ggml_is_permuted(src0)); + GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + + const int64_t nb01 = src0->nb[1]; + const int64_t nb02 = src0->nb[2]; + + const int64_t ne12 = src1->ne[2]; + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + void * src0_ddq = src0_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + + const int64_t row_stride_x = nb01 / sizeof(sycl::half); + const int64_t channel_stride_x = nb02 / sizeof(sycl::half); + + ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void k_compute_batched_ptrs(const sycl::half *src0_as_f16, + const sycl::half *src1_as_f16, char *dst, + const void **ptrs_src, void **ptrs_dst, + int64_t ne12, int64_t ne13, int64_t ne23, + size_t nb02, size_t nb03, size_t nb12, + size_t nb13, size_t nbd2, size_t nbd3, + int64_t r2, int64_t r3, + const sycl::nd_item<3> &item_ct1) { + int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) + + item_ct1.get_local_id(2); + int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) + + item_ct1.get_local_id(1); + + if (i13 >= ne13 || i12 >= ne12) { + return; + } + + int64_t i03 = i13 / r3; + int64_t i02 = i12 / r2; + + ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; + ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2; + ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst + i12*nbd2 + i13*nbd3; +} + +static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst) try { + GGML_ASSERT(!ggml_is_transposed(src0)); + GGML_ASSERT(!ggml_is_transposed(src1)); + + GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00); + const int64_t ne01 = src0->ne[1]; + const int64_t ne02 = src0->ne[2]; + const int64_t ne03 = src0->ne[3]; + + const int64_t nb01 = src0->nb[1]; + const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02); + const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03); + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + const int64_t nb11 = src1->nb[1]; + const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12); + const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13); + + const int64_t ne1 = ggml_nelements(src1); + const int64_t ne = ggml_nelements(dst); + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + + SYCL_CHECK( + CHECK_TRY_ERROR(g_sycl_handles[g_main_device_index] = main_stream)); + + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + void * src0_ddq = src0_extra->data_device[g_main_device_index]; + sycl::half *src0_as_f16 = (sycl::half *)src0_ddq; + + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + + // convert src1 to fp16 + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + + sycl_pool_alloc src1_as_f16(ne1); + to_fp16_sycl(src1_ddf, src1_as_f16.get(), ne1, main_stream); + + sycl_pool_alloc dst_f16; + char * dst_t; + + dpct::library_data_t cu_compute_type = dpct::library_data_t::real_half; + dpct::library_data_t cu_data_type = dpct::library_data_t::real_half; + + // dst strides + size_t nbd2 = dst->nb[2]; + size_t nbd3 = dst->nb[3]; + + const sycl::half alpha_f16 = 1.0f; + const sycl::half beta_f16 = 0.0f; + + const float alpha_f32 = 1.0f; + const float beta_f32 = 0.0f; + + const void * alpha = &alpha_f16; + const void * beta = &beta_f16; + + if (dst->op_params[0] == GGML_PREC_DEFAULT) { + dst_t = (char *) dst_f16.alloc(ne); + + nbd2 /= sizeof(float) / sizeof(sycl::half); + nbd3 /= sizeof(float) / sizeof(sycl::half); + } else { + dst_t = (char *) dst_ddf; + + cu_compute_type = dpct::library_data_t::real_float; + cu_data_type = dpct::library_data_t::real_float; + + alpha = &alpha_f32; + beta = &beta_f32; + } + + GGML_ASSERT(ne12 % ne02 == 0); + GGML_ASSERT(ne13 % ne03 == 0); + + // broadcast factors + const int64_t r2 = ne12/ne02; + const int64_t r3 = ne13/ne03; + +#if 0 + // use syclGemmEx + { + for (int i13 = 0; i13 < ne13; ++i13) { + for (int i12 = 0; i12 < ne12; ++i12) { + int i03 = i13 / r3; + int i02 = i12 / r2; + + SYCL_CHECK( + syclGemmEx(g_sycl_handles[g_main_device_index], CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + alpha, (const char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3] , SYCL_R_16F, nb01/sizeof(half), + (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, SYCL_R_16F, nb11/sizeof(float), + beta, ( char *) dst_t + i12*nbd2 + i13*nbd3, cu_data_type, ne01, + cu_compute_type, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + } + } + } +#else + if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) { + // there is no broadcast and src0, src1 are contiguous across dims 2, 3 + // use syclGemmStridedBatchedEx + SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( + *g_sycl_handles[g_main_device_index], oneapi::mkl::transpose::trans, + oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, + (const char *)src0_as_f16, dpct::library_data_t::real_half, + nb01 / sizeof(sycl::half), src0->nb[2] / sizeof(sycl::half), + (const char *)src1_as_f16.get(), dpct::library_data_t::real_half, + nb11 / sizeof(float), src1->nb[2] / sizeof(float), beta, + (char *)dst_t, cu_data_type, ne01, dst->nb[2] / sizeof(float), + ne12 * ne13, cu_compute_type))); + } else { + // use syclGemmBatchedEx + const int ne23 = ne12*ne13; + + sycl_pool_alloc ptrs_src(2*ne23); + sycl_pool_alloc< void *> ptrs_dst(1*ne23); + + sycl::range<3> block_dims(1, ne12, ne13); + /* + DPCT1049:47: The work-group size passed to the SYCL kernel may exceed + the limit. To get the device limit, query + info::device::max_work_group_size. Adjust the work-group size if needed. + */ + { + dpct::has_capability_or_fail(main_stream->get_device(), + {sycl::aspect::fp16}); + + main_stream->submit([&](sycl::handler &cgh) { + const sycl::half *src1_as_f16_get_ct1 = src1_as_f16.get(); + const void **ptrs_src_get_ct3 = ptrs_src.get(); + void **ptrs_dst_get_ct4 = ptrs_dst.get(); + + cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), + [=](sycl::nd_item<3> item_ct1) { + k_compute_batched_ptrs( + src0_as_f16, src1_as_f16_get_ct1, + dst_t, ptrs_src_get_ct3, + ptrs_dst_get_ct4, ne12, ne13, ne23, + nb02, nb03, nb12, nb13, nbd2, nbd3, r2, + r3, item_ct1); + }); + }); + } + /* + DPCT1010:95: SYCL uses exceptions to report errors and does not use the + error codes. The call was replaced with 0. You need to rewrite this + code. + */ + SYCL_CHECK(0); + + SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( + *g_sycl_handles[g_main_device_index], oneapi::mkl::transpose::trans, + oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha, + (const void **)(ptrs_src.get() + 0 * ne23), + dpct::library_data_t::real_half, nb01 / sizeof(sycl::half), + (const void **)(ptrs_src.get() + 1 * ne23), + dpct::library_data_t::real_half, nb11 / sizeof(float), beta, + (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, + cu_compute_type))); + } +#endif + + if (dst->op_params[0] == GGML_PREC_DEFAULT) { + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); + to_fp32_sycl(dst_f16.get(), dst_ddf, ne, main_stream); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + const bool all_on_device = + (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) && + (src1->backend == GGML_BACKEND_GPU) && + ( dst->backend == GGML_BACKEND_GPU); + + const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT; + + int64_t min_compute_capability = INT_MAX; + for (int64_t id = 0; id < g_device_count; ++id) { + if (min_compute_capability > g_device_caps[id].cc && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { + min_compute_capability = g_device_caps[id].cc; + } + } + +#ifdef SYCL_USE_XMX + const bool use_xmx = true; +#else + const bool use_xmx = false; +#endif + + // debug helpers + //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]); + //printf(" %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]); + //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]); + //printf(" %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]); + //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); + //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); + + if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { + // KQ single-batch + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_p021\n"); + ggml_sycl_mul_mat_vec_p021(src0, src1, dst); + } else if (!split && all_on_device && !use_xmx && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { + // KQV single-batch + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_vec_nc\n"); + ggml_sycl_mul_mat_vec_nc(src0, src1, dst); + } else if (!split && all_on_device && use_xmx && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { + // KQ + KQV multi-batch + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat_mat_batched_sycl\n"); + ggml_sycl_mul_mat_mat_batched_sycl(src0, src1, dst); + } else if (src0->type == GGML_TYPE_F32) { + // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat\n"); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); + } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) { + // GGML_SYCL_DEBUG("ggml_is_quantized or GGML_TYPE_F16\n"); + if (src1->ne[1] == 1 && src0->ne[0] % GGML_SYCL_DMMV_X == 0) { +#ifdef GGML_SYCL_FORCE_DMMV + const bool use_mul_mat_vec_q = false; +#else + const bool use_mul_mat_vec_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type) && ggml_nrows(src1) == 1; +#endif // GGML_SYCL_FORCE_DMMV + + if (use_mul_mat_vec_q) { + // NOTE: this kernel does not support ggml_nrows(src1) > 1 + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_vec_q path\n"); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true); + } else { + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_dequantize_mul_mat_vec path\n"); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false); + } + } else { + bool use_mul_mat_q = min_compute_capability >= VER_4VEC && ggml_is_quantized(src0->type); + + if (use_xmx && min_compute_capability >= VER_GEN9 && src1->ne[1] > XMX_MAX_BATCH_SIZE) { + use_mul_mat_q = false; + } + + if (use_mul_mat_q) { + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_q path\n"); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_q, true); + } else { + // GGML_SYCL_DEBUG("ggml_sycl_mul_mat ggml_sycl_op_mul_mat_sycl path\n"); + ggml_sycl_op_mul_mat(src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false); + } + } + } else { + GGML_ASSERT(false); + } +} + +#if 0 +template +static __global__ void k_compute_batched_ptrs_id( + const void ** ptrs_src, void ** ptrs_dst, + int ne12, int ne13, + int ne23, + int nb02, int nb03, + int nb12, int nb13, + int nb2, int nb3, + int r2, int r3, + ggml_type src0_type, half * src0_as_f16, int64_t src0_ne, + const half * src1_f16, half * dst_f16, + const int32_t * ids, const int id, + Srcs... src0s) { + + int i = ids[id]; + + half * src0_f16; + const void * srcs_ar[] = { (const half *) src0s... }; + if (src0_type == GGML_TYPE_F16) { + src0_f16 = (half *) srcs_ar[i]; + } else { + src0_f16 = src0_as_f16; + if (threadIdx.x == 0 && threadIdx.y == 0) { + const to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(src0_type); + to_fp16(srcs_ar[i], src0_f16, src0_ne, syclStreamFireAndForget); + } + } + + int i13 = blockIdx.x * blockDim.x + threadIdx.x; + int i12 = blockIdx.y * blockDim.y + threadIdx.y; + + if (i13 >= ne13 || i12 >= ne12) { + return; + } + + int i03 = i13 / r3; + int i02 = i12 / r2; + + ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_f16 + i02*nb02 + i03*nb03; + ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_f16 + i12*nb12/2 + i13*nb13/2; + ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst_f16 + i12* nb2/2 + i13* nb3/2; +} + +static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) { + const struct ggml_tensor * ids = dst->src[0]; + const struct ggml_tensor * src1 = dst->src[1]; + const struct ggml_tensor * src00 = dst->src[2]; + + const int id = dst->op_params[0]; + + GGML_ASSERT(!ggml_is_transposed(src00)); + GGML_ASSERT(!ggml_is_transposed(src1)); + + GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + + const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00); + const int64_t ne01 = src00->ne[1]; + const int64_t ne02 = src00->ne[2]; + const int64_t ne03 = src00->ne[3]; + + //const int64_t nb01 = src00->nb[1]; + const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02); + const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03); + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + const int64_t ne12 = src1->ne[2]; + const int64_t ne13 = src1->ne[3]; + + //const int64_t nb11 = src1->nb[1]; + const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12); + const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13); + + const int64_t ne1 = ggml_nelements(src1); + const int64_t ne = ggml_nelements(dst); + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + syclStream_t main_stream = g_syclStreams[g_main_device_index][0]; + + SYCL_CHECK(syclSetStream(g_sycl_handles[g_main_device_index], main_stream)); + + //ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + //void * src0_ddq = src0_extra->data_device[g_main_device_index]; + //half * src0_as_f16 = (half *) src0_ddq; + + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + float * src1_ddf = (float *) src1_extra->data_device[g_main_device_index]; + + ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra; + float * dst_ddf = (float *) dst_extra->data_device[g_main_device_index]; + + // convert src1 to fp16 + const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type); + GGML_ASSERT(to_fp16_sycl != nullptr); + + size_t src1_as = 0; + half * src1_as_f16 = (half *) ggml_sycl_pool_malloc(ne1 * sizeof(half), &src1_as); + to_fp16_sycl(src1_ddf, src1_as_f16, ne1, main_stream); + + size_t dst_as = 0; + half * dst_f16 = (half *) ggml_sycl_pool_malloc(ne * sizeof(half), &dst_as); + + GGML_ASSERT(ne12 % ne02 == 0); + GGML_ASSERT(ne13 % ne03 == 0); + + // broadcast factors + const int64_t r2 = ne12/ne02; + const int64_t r3 = ne13/ne03; + + const half alpha_f16 = 1.0f; + const half beta_f16 = 0.0f; + + // use syclGemmBatchedEx + const int ne23 = ne12*ne13; + + const void ** ptrs_src = nullptr; + void ** ptrs_dst = nullptr; + + size_t ptrs_src_s = 0; + size_t ptrs_dst_s = 0; + + ptrs_src = (const void **) ggml_sycl_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s); + ptrs_dst = ( void **) ggml_sycl_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s); + + int64_t src0_ne = ggml_nelements(src00); + half * src0_as_f16 = nullptr; + size_t src0_as = 0; + if (src00->type != GGML_TYPE_F16) { + src0_as_f16 = (half *) ggml_sycl_pool_malloc(src0_ne * sizeof(half), &src0_as); + } + + static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6"); + dim3 block_dims(ne13, ne12); + k_compute_batched_ptrs_id<<<1, block_dims, 0, main_stream>>>( + ptrs_src, ptrs_dst, + ne12, ne13, + ne23, + ne00*ne01*sizeof(half), ne00*ne01*ne02*sizeof(half), + nb12, nb13, + dst->nb[2], dst->nb[3], + r2, r3, + src00->type, src0_as_f16, src0_ne, + src1_as_f16, dst_f16, + (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index], id, + dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device_index] : nullptr, + dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device_index] : nullptr, + dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device_index] : nullptr, + dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device_index] : nullptr + ); + SYCL_CHECK(syclGetLastError()); + + SYCL_CHECK( + syclGemmBatchedEx(g_sycl_handles[g_main_device_index], CUBLAS_OP_T, CUBLAS_OP_N, + ne01, ne11, ne10, + &alpha_f16, (const void **) (ptrs_src + 0*ne23), SYCL_R_16F, ne00, + (const void **) (ptrs_src + 1*ne23), SYCL_R_16F, ne10, + &beta_f16, ( void **) (ptrs_dst + 0*ne23), SYCL_R_16F, ne01, + ne23, + CUBLAS_COMPUTE_16F, + CUBLAS_GEMM_DEFAULT_TENSOR_OP)); + + if (src0_as != 0) { + ggml_sycl_pool_free(src0_as_f16, src0_as); + } + if (ptrs_src_s != 0) { + ggml_sycl_pool_free(ptrs_src, ptrs_src_s); + } + if (ptrs_dst_s != 0) { + ggml_sycl_pool_free(ptrs_dst, ptrs_dst_s); + } + + const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16); + to_fp32_sycl(dst_f16, dst_ddf, ne, main_stream); + + ggml_sycl_pool_free(src1_as_f16, src1_as); + ggml_sycl_pool_free(dst_f16, dst_as); +} +#endif + +static void ggml_sycl_mul_mat_id(const ggml_tensor *src0, + const ggml_tensor *src1, + ggml_tensor *dst) try { +#if 0 + ggml_sycl_mul_mat_id_sycl(dst); + // TODO: mmq/mmv support +#endif + + const int64_t nb11 = src1->nb[1]; + const int64_t nb1 = dst->nb[1]; + + const struct ggml_tensor * ids = src0; + const int32_t id = ((int32_t *) dst->op_params)[0]; + const int32_t n_as = ((int32_t *) dst->op_params)[1]; + + std::vector ids_host(ggml_nbytes(ids)); + + const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + + if (ids->backend == GGML_BACKEND_GPU) { + const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device_index]; + SYCL_CHECK(CHECK_TRY_ERROR( + stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids)))); + SYCL_CHECK(CHECK_TRY_ERROR(stream->wait())); + } else { + memcpy(ids_host.data(), ids->data, ggml_nbytes(ids)); + } + + const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra; + const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra; + + ggml_tensor_extra_gpu src1_row_extra; + ggml_tensor_extra_gpu dst_row_extra; + + ggml_tensor src1_row = *src1; + ggml_tensor dst_row = *dst; + + src1_row.backend = GGML_BACKEND_GPU; + dst_row.backend = GGML_BACKEND_GPU; + + src1_row.extra = &src1_row_extra; + dst_row.extra = &dst_row_extra; + + char * src1_original = src1->backend == GGML_BACKEND_CPU ? + (char *) src1->data : (char *) src1_extra->data_device[g_main_device_index]; + char * dst_original = dst->backend == GGML_BACKEND_CPU ? + (char *) dst->data : (char *) dst_extra->data_device[g_main_device_index]; + + if (src1->ne[1] == 1) { + GGML_ASSERT(src1->backend == GGML_BACKEND_GPU); + GGML_ASSERT(dst->backend == GGML_BACKEND_GPU); + + for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { + //int32_t row_id; + //SYCL_CHECK(syclMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), syclMemcpyDeviceToHost, g_syclStreams[g_main_device][0])); + //SYCL_CHECK(syclStreamSynchronize(g_syclStreams[g_main_device][0])); + + const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); + + GGML_ASSERT(row_id >= 0 && row_id < n_as); + + const struct ggml_tensor * src0_row = dst->src[row_id + 2]; + + src1_row_extra.data_device[g_main_device_index] = src1_original + i01*src1->nb[1]; + src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set? + + dst_row_extra.data_device[g_main_device_index] = dst_original + i01*dst->nb[1]; + dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set? + + ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row); + } + } else { + sycl_pool_alloc src1_contiguous(sizeof(float)*ggml_nelements(src1)); + sycl_pool_alloc dst_contiguous(sizeof(float)*ggml_nelements(dst)); + + src1_row_extra.data_device[g_main_device_index] = src1_contiguous.get(); + dst_row_extra.data_device[g_main_device_index] = dst_contiguous.get(); + + const dpct::memcpy_direction src1_kind = + src1->backend == GGML_BACKEND_CPU ? dpct::host_to_device + : dpct::device_to_device; + const dpct::memcpy_direction dst_kind = dst->backend == GGML_BACKEND_CPU + ? dpct::device_to_host + : dpct::device_to_device; + + for (int32_t row_id = 0; row_id < n_as; ++row_id) { + const struct ggml_tensor * src0_row = dst->src[row_id + 2]; + + int64_t num_src1_rows = 0; + for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { + const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); + + if (row_id_i != row_id) { + continue; + } + + GGML_ASSERT(row_id >= 0 && row_id < n_as); + + SYCL_CHECK(CHECK_TRY_ERROR( + stream->memcpy(src1_contiguous.get() + num_src1_rows * nb11, + src1_original + i01 * nb11, nb11))); + num_src1_rows++; + } + + if (num_src1_rows == 0) { + continue; + } + + src1_row.ne[1] = num_src1_rows; + dst_row.ne[1] = num_src1_rows; + + src1_row.nb[1] = nb11; + src1_row.nb[2] = num_src1_rows*nb11; + src1_row.nb[3] = num_src1_rows*nb11; + + dst_row.nb[1] = nb1; + dst_row.nb[2] = num_src1_rows*nb1; + dst_row.nb[3] = num_src1_rows*nb1; + + ggml_sycl_mul_mat(src0_row, &src1_row, &dst_row); + + num_src1_rows = 0; + for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) { + const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]); + + if (row_id_i != row_id) { + continue; + } + + GGML_ASSERT(row_id >= 0 && row_id < n_as); + + SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy( + dst_original + i01 * nb1, + dst_contiguous.get() + num_src1_rows * nb1, nb1))); + num_src1_rows++; + } + } + } + + if (dst->backend == GGML_BACKEND_CPU) { + SYCL_CHECK(CHECK_TRY_ERROR(stream->wait())); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_scale); +} + +static void ggml_sycl_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_clamp); +} + +static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1, + ggml_tensor *dst) try { + const int64_t ne = ggml_nelements(src0); + GGML_ASSERT(ne == ggml_nelements(src1)); + + GGML_ASSERT(src0->backend == GGML_BACKEND_GPU); + GGML_ASSERT(src1->backend == GGML_BACKEND_GPU); + + GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX); + GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX); + + const int64_t ne00 = src0->ne[0]; + const int64_t ne01 = src0->ne[1]; + GGML_ASSERT(src0->ne[3] == 1); + + const int64_t nb00 = src0->nb[0]; + const int64_t nb01 = src0->nb[1]; + const int64_t nb02 = src0->nb[2]; + + const int64_t ne10 = src1->ne[0]; + const int64_t ne11 = src1->ne[1]; + GGML_ASSERT(src1->ne[3] == 1); + + const int64_t nb10 = src1->nb[0]; + const int64_t nb11 = src1->nb[1]; + const int64_t nb12 = src1->nb[2]; + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0]; + + const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra; + const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra; + + char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; + char * src1_ddc = (char *) src1_extra->data_device[g_main_device_index]; + + if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) { + ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f32_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) { + ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) { + ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) { + ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) { + ggml_cpy_i16_i16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) { + ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else { + fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, + ggml_type_name(src0->type), ggml_type_name(src1->type)); + GGML_ASSERT(false); + } + + (void) dst; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_sycl_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + // TODO: why do we pass dst as src1 here? + ggml_sycl_cpy(src0, dst, nullptr); + (void) src1; +} + +static void ggml_sycl_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_diag_mask_inf); +} + +static void ggml_sycl_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_soft_max); +} + +static void ggml_sycl_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_rope); +} + +static void ggml_sycl_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_alibi); +} + +static void ggml_sycl_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_im2col); +} + +static void ggml_sycl_sum_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(ggml_is_contiguous(src0)); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_sum_rows); +} + +static void ggml_sycl_argsort(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + GGML_ASSERT(ggml_is_contiguous(src0)); + ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_argsort); +} + +static void ggml_sycl_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + (void) src0; + (void) src1; + (void) dst; +} + +static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) { + static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function"); + + return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]); +} + +void ggml_sycl_transform_tensor(void *data, struct ggml_tensor *tensor) try { + const int64_t nrows = ggml_nrows(tensor); + + const int64_t ne0 = tensor->ne[0]; + + const size_t nb1 = tensor->nb[1]; + + ggml_backend_type backend = tensor->backend; + ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu; + memset(extra, 0, sizeof(*extra)); + + for (int64_t id = 0; id < g_device_count; ++id) { + if (backend == GGML_BACKEND_GPU && id != g_main_device_index) { + continue; + } + ggml_sycl_set_device(get_device_id_by_index(id)); + const dpct::queue_ptr stream = g_syclStreams[id][0]; + + int64_t row_low, row_high; + if (backend == GGML_BACKEND_GPU) { + row_low = 0; + row_high = nrows; + } else if (backend == GGML_BACKEND_GPU_SPLIT) { + const int64_t rounding = get_row_rounding(tensor->type); + + row_low = id == 0 ? 0 : nrows*g_tensor_split[id]; + row_low -= row_low % rounding; + + if (id == g_device_count - 1) { + row_high = nrows; + } else { + row_high = nrows*g_tensor_split[id + 1]; + row_high -= row_high % rounding; + } + } else { + GGML_ASSERT(false); + } + if (row_low == row_high) { + continue; + } + + int64_t nrows_split = row_high - row_low; + + const size_t offset_split = row_low*nb1; + size_t size = ggml_nbytes_split(tensor, nrows_split); + const size_t original_size = size; + + // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + + char * buf; + SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device( + size, *stream))); + char * buf_host = (char *)data + offset_split; + + // set padding to 0 to avoid possible NaN values + if (size > original_size) { + SYCL_CHECK(CHECK_TRY_ERROR( + (*stream) + .memset(buf + original_size, 0, size - original_size) + .wait())); + } + + SYCL_CHECK(CHECK_TRY_ERROR((*stream) + .memcpy(buf, buf_host, original_size) + .wait())); + + extra->data_device[id] = buf; + + if (backend == GGML_BACKEND_GPU_SPLIT) { + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + SYCL_CHECK(CHECK_TRY_ERROR(extra->events[id][is] = + new sycl::event())); + } + } + } + + tensor->extra = extra; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_free_data(struct ggml_tensor *tensor) try { + if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) { + return; + } + + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + + for (int64_t id = 0; id < g_device_count; ++id) { + const dpct::queue_ptr stream = g_syclStreams[id][0]; + if (extra->data_device[id] != nullptr) { + SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(extra->data_device[id], *stream))); + } + + for (int64_t is = 0; is < MAX_STREAMS; ++is) { + if (extra->events[id][is] != nullptr) { + SYCL_CHECK(ggml_sycl_set_device(get_device_id_by_index(id))); + SYCL_CHECK(CHECK_TRY_ERROR( + dpct::destroy_event(extra->events[id][is]))); + } + } + } + + delete extra; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static ggml_tensor_extra_gpu * g_temp_tensor_extras = nullptr; +static size_t g_temp_tensor_extra_index = 0; + +static ggml_tensor_extra_gpu * ggml_sycl_alloc_temp_tensor_extra() { + if (g_temp_tensor_extras == nullptr) { + g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_SYCL_MAX_NODES]; + } + + size_t alloc_index = g_temp_tensor_extra_index; + g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_SYCL_MAX_NODES; + ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index]; + memset(extra, 0, sizeof(*extra)); + + return extra; +} + +static void ggml_sycl_assign_buffers_impl(struct ggml_tensor *tensor, + bool scratch, bool force_inplace, + bool no_alloc) try { + if (scratch && g_scratch_size == 0) { + return; + } + + tensor->backend = GGML_BACKEND_GPU; + + if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) { + const ggml_op src0_op = tensor->src[0]->op; + if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) { + ggml_sycl_assign_buffers_impl(tensor->src[0], scratch, force_inplace, no_alloc); + } + } + if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) { + ggml_sycl_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc); + } + + if (scratch && no_alloc) { + return; + } + + ggml_tensor_extra_gpu * extra; + + const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) || + tensor->op == GGML_OP_VIEW || + force_inplace; + const size_t size = ggml_nbytes(tensor); + + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + + if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) { + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra; + char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; + size_t offset = 0; + if (tensor->op == GGML_OP_VIEW) { + memcpy(&offset, tensor->op_params, sizeof(size_t)); + } + extra = ggml_sycl_alloc_temp_tensor_extra(); + extra->data_device[g_main_device_index] = src0_ddc + offset; + } else if (tensor->op == GGML_OP_CPY) { + ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra; + void * src1_ddv = src1_extra->data_device[g_main_device_index]; + extra = ggml_sycl_alloc_temp_tensor_extra(); + extra->data_device[g_main_device_index] = src1_ddv; + } else if (scratch) { + GGML_ASSERT(size <= g_scratch_size); + if (g_scratch_offset + size > g_scratch_size) { + g_scratch_offset = 0; + } + + char * data = (char *) g_scratch_buffer; + if (data == nullptr) { + SYCL_CHECK(CHECK_TRY_ERROR( + data = (char *)sycl::malloc_device( + g_scratch_size, *stream))); + g_scratch_buffer = data; + } + extra = ggml_sycl_alloc_temp_tensor_extra(); + extra->data_device[g_main_device_index] = data + g_scratch_offset; + + g_scratch_offset += size; + + GGML_ASSERT(g_scratch_offset <= g_scratch_size); + } else { // allocate new buffers outside of scratch + void * data; + SYCL_CHECK(CHECK_TRY_ERROR(data = (void *)sycl::malloc_device( + size, *stream))); + SYCL_CHECK(CHECK_TRY_ERROR( + (*stream).memset(data, 0, size).wait())); + extra = new ggml_tensor_extra_gpu; + memset(extra, 0, sizeof(*extra)); + extra->data_device[g_main_device_index] = data; + } + + tensor->extra = extra; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_assign_scratch_offset(struct ggml_tensor *tensor, + size_t offset) try { + if (g_scratch_size == 0) { + return; + } + if (g_scratch_buffer == nullptr) { + ggml_sycl_set_device(g_main_device); + const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + SYCL_CHECK( + CHECK_TRY_ERROR(g_scratch_buffer = (void *)sycl::malloc_device( + g_scratch_size, *stream))); + } + + ggml_tensor_extra_gpu * extra = ggml_sycl_alloc_temp_tensor_extra(); + + const bool inplace = tensor->view_src != nullptr; + + if (inplace && (tensor->view_src->backend == GGML_BACKEND_GPU || tensor->view_src->backend == GGML_BACKEND_GPU_SPLIT)) { + ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra; + char * src0_ddc = (char *) src0_extra->data_device[g_main_device_index]; + size_t view_offset = 0; + if (tensor->op == GGML_OP_VIEW) { + memcpy(&view_offset, tensor->op_params, sizeof(size_t)); + } + extra->data_device[g_main_device_index] = src0_ddc + view_offset; + } else { + extra->data_device[g_main_device_index] = (char *) g_scratch_buffer + offset; + } + + tensor->extra = extra; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_copy_to_device(struct ggml_tensor *tensor) try { + GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); + GGML_ASSERT(ggml_is_contiguous(tensor)); + + ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra; + SYCL_CHECK(ggml_sycl_set_device(g_main_device)); + const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + SYCL_CHECK(CHECK_TRY_ERROR((*stream) + .memcpy(extra->data_device[g_main_device_index], + tensor->data, ggml_nbytes(tensor)) + .wait())); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_assign_buffers(struct ggml_tensor * tensor) { + ggml_sycl_assign_buffers_impl(tensor, true, false, false); +} + +void ggml_sycl_assign_buffers_no_alloc(struct ggml_tensor * tensor) { + ggml_sycl_assign_buffers_impl(tensor, true, false, true); +} + +void ggml_sycl_assign_buffers_no_scratch(struct ggml_tensor * tensor) { + ggml_sycl_assign_buffers_impl(tensor, false, false, false); +} + +void ggml_sycl_assign_buffers_force_inplace(struct ggml_tensor * tensor) { + ggml_sycl_assign_buffers_impl(tensor, false, true, false); +} + +void ggml_sycl_set_main_device(const int main_device) try { + + if (main_device >= g_all_sycl_device_count) { + fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n", + main_device, g_all_sycl_device_count, g_main_device); + return; + } + + if (g_main_device != main_device && g_device_count >= 1) { + g_main_device = main_device; + g_main_device_index = get_device_index_by_id(g_main_device); + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(g_main_device)))); + fprintf(stderr, "Using device %d (%s) as main device\n", + g_main_device, prop.get_name()); + } +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_set_scratch_size(const size_t scratch_size) { + // this is a hack to not completely break llama.cpp when using multiple models or contexts simultaneously + // it still won't always work as expected, but it's better than nothing + if (scratch_size > g_scratch_size) { + ggml_sycl_free_scratch(); + } + g_scratch_size = std::max(g_scratch_size, scratch_size); +} + +void ggml_sycl_free_scratch() try { + if (g_scratch_buffer == nullptr) { + return; + } + ggml_sycl_set_device(g_main_device); + const dpct::queue_ptr stream = g_syclStreams[g_main_device_index][0]; + + SYCL_CHECK(CHECK_TRY_ERROR( + sycl::free(g_scratch_buffer, *stream))); + g_scratch_buffer = nullptr; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { + if (!g_sycl_loaded) return false; + + ggml_sycl_func_t func; + const bool any_on_device = tensor->backend == GGML_BACKEND_GPU + || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) + || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU); + + if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) { + return false; + } + + if (tensor->op == GGML_OP_MUL_MAT) { + if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) { +#ifndef NDEBUG + fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]); +#endif + return false; + } + } + + switch (tensor->op) { + case GGML_OP_REPEAT: + func = ggml_sycl_repeat; + break; + case GGML_OP_GET_ROWS: + func = ggml_sycl_get_rows; + break; + case GGML_OP_DUP: + func = ggml_sycl_dup; + break; + case GGML_OP_ADD: + func = ggml_sycl_add; + break; + case GGML_OP_ACC: + func = ggml_sycl_acc; + break; + case GGML_OP_MUL: + func = ggml_sycl_mul; + break; + case GGML_OP_DIV: + func = ggml_sycl_div; + break; + case GGML_OP_UNARY: + switch (ggml_get_unary_op(tensor)) { + case GGML_UNARY_OP_GELU: + func = ggml_sycl_gelu; + break; + case GGML_UNARY_OP_SILU: + func = ggml_sycl_silu; + break; + case GGML_UNARY_OP_GELU_QUICK: + func = ggml_sycl_gelu_quick; + break; + case GGML_UNARY_OP_TANH: + func = ggml_sycl_tanh; + break; + case GGML_UNARY_OP_RELU: + func = ggml_sycl_relu; + break; + default: + return false; + } + break; + case GGML_OP_NORM: + func = ggml_sycl_norm; + break; + case GGML_OP_GROUP_NORM: + func = ggml_sycl_group_norm; + break; + case GGML_OP_CONCAT: + func = ggml_sycl_concat; + break; + case GGML_OP_UPSCALE: + func = ggml_sycl_upscale; + break; + case GGML_OP_PAD: + func = ggml_sycl_pad; + break; + case GGML_OP_LEAKY_RELU: + func = ggml_sycl_leaky_relu; + break; + case GGML_OP_RMS_NORM: + func = ggml_sycl_rms_norm; + break; + case GGML_OP_MUL_MAT: + if (!any_on_device && !ggml_sycl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) { + return false; + } + func = ggml_sycl_mul_mat; + break; + case GGML_OP_MUL_MAT_ID: + if (!any_on_device && !ggml_sycl_can_mul_mat(tensor->src[2], tensor->src[1], tensor)) { + return false; + } + func = ggml_sycl_mul_mat_id; + break; + case GGML_OP_SCALE: + func = ggml_sycl_scale; + break; + case GGML_OP_SQR: + func = ggml_sycl_sqr; + break; + case GGML_OP_CLAMP: + func = ggml_sycl_clamp; + break; + case GGML_OP_CPY: + func = ggml_sycl_cpy; + break; + case GGML_OP_CONT: + func = ggml_sycl_dup; + break; + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + func = ggml_sycl_nop; + break; + case GGML_OP_DIAG_MASK_INF: + func = ggml_sycl_diag_mask_inf; + break; + case GGML_OP_SOFT_MAX: + func = ggml_sycl_soft_max; + break; + case GGML_OP_ROPE: + func = ggml_sycl_rope; + break; + case GGML_OP_ALIBI: + func = ggml_sycl_alibi; + break; + case GGML_OP_IM2COL: + func = ggml_sycl_im2col; + break; + case GGML_OP_SUM_ROWS: + func = ggml_sycl_sum_rows; + break; + case GGML_OP_ARGSORT: + func = ggml_sycl_argsort; + break; + default: + return false; + } + + if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) { + ggml_sycl_set_peer_access(tensor->src[1]->ne[1]); + } + + if (params->ith != 0) { + return true; + } + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return true; + } + func(tensor->src[0], tensor->src[1], tensor); + return true; +} + +int ggml_sycl_get_device_count() try { + int device_count; + if (CHECK_TRY_ERROR(device_count = + dpct::dev_mgr::instance().device_count()) != 0) { + return 0; + } + return device_count; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +void ggml_sycl_get_device_description(int device, char *description, + size_t description_size) try { + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(device)))); + snprintf(description, description_size, "%s", prop.get_name()); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +//////////////////////////////////////////////////////////////////////////////// + +// backend interface + +#define UNUSED GGML_UNUSED + +struct ggml_backend_sycl_context { + int device; + std::string name; +}; + +// sycl buffer + +struct ggml_backend_sycl_buffer_context { + int device; + void * dev_ptr = nullptr; + ggml_tensor_extra_gpu * temp_tensor_extras = nullptr; + size_t temp_tensor_extra_index = 0; + std::string name; + + ggml_backend_sycl_buffer_context(int device, void * dev_ptr) : device(device), dev_ptr(dev_ptr) {} + + ~ ggml_backend_sycl_buffer_context() { + delete[] temp_tensor_extras; + } + + ggml_tensor_extra_gpu * ggml_sycl_alloc_temp_tensor_extra() { + if (temp_tensor_extras == nullptr) { + temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_SYCL_MAX_NODES]; + } + + size_t alloc_index = temp_tensor_extra_index; + temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_SYCL_MAX_NODES; + ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index]; + memset(extra, 0, sizeof(*extra)); + + return extra; + } +}; + +GGML_CALL static const char * ggml_backend_sycl_buffer_get_name(ggml_backend_buffer_t buffer) { + ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context; + return ctx->name.c_str(); +} + +GGML_CALL static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) { + return buffer->iface.get_name == ggml_backend_sycl_buffer_get_name; +} + +static void +ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try { + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + ggml_sycl_set_device(ctx->device); + int device_index = get_device_index_by_id(ctx->device); + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + + SYCL_CHECK( + CHECK_TRY_ERROR(sycl::free(ctx->dev_ptr, *stream))); + delete ctx; +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) { + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + return ctx->dev_ptr; +} + +static void ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer, + ggml_tensor *tensor) try { + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + + if (tensor->view_src != NULL && tensor->view_offs == 0) { + assert(tensor->view_src->buffer->buft == buffer->buft); + tensor->backend = tensor->view_src->backend; + tensor->extra = tensor->view_src->extra; + return; + } + + ggml_tensor_extra_gpu * extra = ctx->ggml_sycl_alloc_temp_tensor_extra(); + + extra->data_device[ctx->device] = tensor->data; + + tensor->backend = GGML_BACKEND_GPU; + tensor->extra = extra; + + if (ggml_is_quantized(tensor->type)) { + // initialize padding to 0 to avoid possible NaN values + int64_t row_low = 0; + int64_t row_high = ggml_nrows(tensor); + int64_t nrows_split = row_high - row_low; + + size_t original_size = ggml_nbytes_split(tensor, nrows_split); + size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor); + + if (padded_size > original_size && tensor->view_src == nullptr) { + SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[ctx->device][0]->memset( + (char *)tensor->data + original_size, 0, + padded_size - original_size))); + } + } + + UNUSED(buffer); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer, + ggml_tensor *tensor, + const void *data, size_t offset, + size_t size) try { + GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); + + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + + ggml_sycl_set_device(ctx->device); + int device_index = get_device_index_by_id(ctx->device); + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + SYCL_CHECK( + CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); + + SYCL_CHECK( + CHECK_TRY_ERROR((*stream) + .memcpy((char *)tensor->data + offset, data, size) + .wait())); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer, + const ggml_tensor *tensor, + void *data, size_t offset, + size_t size) try { + GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); + + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + + ggml_sycl_set_device(ctx->device); + int device_index = get_device_index_by_id(ctx->device); + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + + SYCL_CHECK( + CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); + + SYCL_CHECK(CHECK_TRY_ERROR( + (*stream) + .memcpy(data, (const char *)tensor->data + offset, size) + .wait())); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer, + uint8_t value) try { + ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; + + ggml_sycl_set_device(ctx->device); + int device_index = get_device_index_by_id(ctx->device); + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + SYCL_CHECK( + CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw())); + + SYCL_CHECK(CHECK_TRY_ERROR((*stream) + .memset(ctx->dev_ptr, value, buffer->size) + .wait())); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static struct ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = { + /* .get_name = */ ggml_backend_sycl_buffer_get_name, + /* .free_buffer = */ ggml_backend_sycl_buffer_free_buffer, + /* .get_base = */ ggml_backend_sycl_buffer_get_base, + /* .init_tensor = */ ggml_backend_sycl_buffer_init_tensor, + /* .set_tensor = */ ggml_backend_sycl_buffer_set_tensor, + /* .get_tensor = */ ggml_backend_sycl_buffer_get_tensor, + /* .cpy_tensor = */ NULL, + /* .clear = */ ggml_backend_sycl_buffer_clear, + /* .reset = */ NULL, +}; + +// sycl buffer type +struct ggml_backend_sycl_buffer_type_context { + int device; + std::string name; +}; + +GGML_CALL static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) { + ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context; + + return ctx->name.c_str(); +} + +static ggml_backend_buffer_t +ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, + size_t size) try { + int device = (int) (intptr_t) buft->context; + + ggml_sycl_set_device(device); + int device_index = get_device_index_by_id(device); + const dpct::queue_ptr stream = g_syclStreams[device_index][0]; + size = std::max(size, (size_t)1); // syclMalloc returns null for size 0 + + void * dev_ptr; + SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device( + size, *stream))); + + ggml_backend_sycl_buffer_context * ctx = new ggml_backend_sycl_buffer_context(device, dev_ptr); + + return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { + return 128; + + UNUSED(buft); +} + +static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) { + int64_t row_low = 0; + int64_t row_high = ggml_nrows(tensor); + int64_t nrows_split = row_high - row_low; + + size_t size = ggml_nbytes_split(tensor, nrows_split); + + int64_t ne0 = tensor->ne[0]; + + if (ggml_is_quantized(tensor->type)) { + if (ne0 % MATRIX_ROW_PADDING != 0) { + size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); + } + } + + return size; + + UNUSED(buft); +} + +static bool ggml_backend_sycl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) { + return ggml_backend_is_sycl(backend); + + UNUSED(buft); +} + +static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = { + /* .get_name = */ ggml_backend_sycl_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_sycl_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_sycl_buffer_type_get_alignment, + /* .get_alloc_size = */ ggml_backend_sycl_buffer_type_get_alloc_size, + /* .supports_backend = */ ggml_backend_sycl_buffer_type_supports_backend, + /* .is_host = */ nullptr, +}; + +ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { + static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES]; + + static bool ggml_backend_sycl_buffer_type_initialized = false; + + if (!ggml_backend_sycl_buffer_type_initialized) { + for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) { + ggml_backend_sycl_buffer_types[i] = { + /* .iface = */ ggml_backend_sycl_buffer_type_interface, + /* .context = */ (ggml_backend_buffer_type_context_t) (intptr_t) i, + }; + } + ggml_backend_sycl_buffer_type_initialized = true; + } + + return &ggml_backend_sycl_buffer_types[device]; +} + +// host buffer type + +GGML_CALL static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) { + return GGML_SYCL_NAME "_Host"; + + UNUSED(buft); +} + +GGML_CALL static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) { + return GGML_SYCL_NAME "_Host"; + + UNUSED(buffer); +} + +static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) { + ggml_sycl_host_free(buffer->context); +} + +static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { + void * ptr = ggml_sycl_host_malloc(size); + + if (ptr == nullptr) { + // fallback to cpu buffer + return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); + } + + // FIXME: this is a hack to avoid having to implement a new buffer type + ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size); + buffer->buft = buft; + buffer->iface.free_buffer = ggml_backend_sycl_host_buffer_free_buffer; + + return buffer; +} + +ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() { + static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_type_host = { + /* .iface = */ { + /* .get_name = */ ggml_backend_sycl_host_buffer_type_name, + /* .alloc_buffer = */ ggml_backend_sycl_host_buffer_type_alloc_buffer, + /* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment, + /* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size, + /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend, + /* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host, + }, + /* .context = */ nullptr, + }; + + return &ggml_backend_sycl_buffer_type_host; +} + +// backend + +struct ggml_backend_context_sycl { + int device; +}; + +static const char * ggml_backend_sycl_name(ggml_backend_t backend) { + return GGML_SYCL_NAME; + + UNUSED(backend); +} + +static void ggml_backend_sycl_free(ggml_backend_t backend) { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + delete sycl_ctx; + delete backend; +} + +static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + return ggml_backend_sycl_buffer_type(sycl_ctx->device); +} + +static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend, + ggml_tensor *tensor, + const void *data, size_t offset, + size_t size) try { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type"); + GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); + + SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy( + (char *)tensor->data + offset, data, size))); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend, + const ggml_tensor *tensor, + void *data, size_t offset, + size_t size) try { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + GGML_ASSERT(tensor->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type"); + GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU); + + SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->memcpy( + data, (const char *)tensor->data + offset, size))); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + SYCL_CHECK(CHECK_TRY_ERROR(g_syclStreams[sycl_ctx->device][0]->wait())); + + UNUSED(backend); +} +catch (sycl::exception const &exc) { + std::cerr << exc.what() << "Exception caught at file:" << __FILE__ + << ", line:" << __LINE__ << std::endl; + std::exit(1); +} + +static ggml_backend_graph_plan_t ggml_backend_sycl_graph_plan_create(ggml_backend_t backend, const ggml_cgraph * cgraph) { + GGML_ASSERT(!"not implemented"); + + return nullptr; + + UNUSED(backend); + UNUSED(cgraph); +} + +static void ggml_backend_sycl_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { + GGML_ASSERT(!"not implemented"); + + UNUSED(backend); + UNUSED(plan); +} + +static void ggml_backend_sycl_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { + GGML_ASSERT(!"not implemented"); + + UNUSED(backend); + UNUSED(plan); +} + +static bool ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) { + ggml_backend_context_sycl * sycl_ctx = (ggml_backend_context_sycl *)backend->context; + + ggml_sycl_set_main_device(sycl_ctx->device); + + ggml_compute_params params = {}; + params.type = GGML_TASK_COMPUTE; + params.ith = 0; + for (int i = 0; i < cgraph->n_nodes; i++) { + ggml_tensor * node = cgraph->nodes[i]; + + if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE) + continue; + + assert(node->backend == GGML_BACKEND_GPU); + assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device)); + assert(node->extra != nullptr); + + for (int j = 0; j < GGML_MAX_SRC; j++) { + if (node->src[j] != nullptr) { + assert(node->src[j]->backend == GGML_BACKEND_GPU); + assert(node->src[j]->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device)); + assert(node->src[j]->extra != nullptr); + } + } + + bool ok = ggml_sycl_compute_forward(¶ms, node); + if (!ok) { + fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op)); + } + GGML_ASSERT(ok); + +#if 0 + if (node->type == GGML_TYPE_F32) { + syclDeviceSynchronize(); + std::vector tmp(ggml_nelements(node), 0.0f); + syclMemcpy(tmp.data(), node->data, ggml_nelements(node)*sizeof(float), syclMemcpyDeviceToHost); + printf("\n%s (%s) (%s %s) (%s %s): ", node->name, ggml_op_name(node->op), + ggml_type_name(node->src[0]->type), + node->src[1] ? ggml_type_name(node->src[1]->type) : "none", + node->src[0]->name, + node->src[1] ? node->src[1]->name : "none"); + double sum = 0.0; + double sq_sum = 0.0; + for (int i = 0; i < ggml_nelements(node); i++) { + printf("%f ", tmp[i]); + sum += tmp[i]; + sq_sum += tmp[i]*tmp[i]; + } + printf("\n"); + printf("sum: %f, ", sum); + printf("sq_sum: %f\n", sq_sum); + } +#endif + } + + UNUSED(backend); + return true; +} + +static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) { + switch (op->op) { + case GGML_OP_UNARY: + switch (ggml_get_unary_op(op)) { + case GGML_UNARY_OP_GELU: + case GGML_UNARY_OP_SILU: + case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_GELU_QUICK: + case GGML_UNARY_OP_TANH: + return true; + default: + return false; + } + break; + case GGML_OP_MUL_MAT: + case GGML_OP_MUL_MAT_ID: + { + struct ggml_tensor * a; + struct ggml_tensor * b; + if (op->op == GGML_OP_MUL_MAT) { + a = op->src[0]; + b = op->src[1]; + } else { + a = op->src[2]; + b = op->src[1]; + } + if (a->ne[3] != b->ne[3]) { + return false; + } + + if (a->type == GGML_TYPE_IQ2_XXS) { + return false; + } + if (a->type == GGML_TYPE_IQ2_XS) { + return false; + } + + return true; + } break; + case GGML_OP_GET_ROWS: + { + switch (op->src[0]->type) { + case GGML_TYPE_F16: + case GGML_TYPE_F32: + case GGML_TYPE_Q4_0: + case GGML_TYPE_Q4_1: + case GGML_TYPE_Q5_0: + case GGML_TYPE_Q5_1: + case GGML_TYPE_Q8_0: + return true; + default: + return false; + } + } break; + case GGML_OP_CPY: + { + ggml_type src0_type = op->src[0]->type; + ggml_type src1_type = op->src[1]->type; + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) { + return true; + } + if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) { + return true; + } + if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) { + return true; + } + return false; + } break; + case GGML_OP_CONCAT: + { + ggml_type src0_type = op->src[0]->type; + if (src0_type == GGML_TYPE_F32) { + return true; + } else { + return false; + } + } break; + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + case GGML_OP_NORM: + case GGML_OP_REPEAT: + case GGML_OP_DUP: + case GGML_OP_ADD: + case GGML_OP_MUL: + case GGML_OP_DIV: + case GGML_OP_RMS_NORM: + case GGML_OP_SCALE: + case GGML_OP_SQR: + case GGML_OP_CLAMP: + case GGML_OP_CONT: + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_SOFT_MAX: + case GGML_OP_ROPE: + case GGML_OP_ALIBI: + case GGML_OP_IM2COL: + case GGML_OP_SUM_ROWS: + case GGML_OP_ARGSORT: + case GGML_OP_ACC: + case GGML_OP_GROUP_NORM: + case GGML_OP_UPSCALE: + case GGML_OP_PAD: + case GGML_OP_LEAKY_RELU: + return true; + default: + return false; + } + + UNUSED(backend); +} + +static ggml_backend_i ggml_backend_sycl_interface = { + /* .get_name = */ ggml_backend_sycl_name, + /* .free = */ ggml_backend_sycl_free, + /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type, + /* .set_tensor_async = */ ggml_backend_sycl_set_tensor_async, + /* .get_tensor_async = */ ggml_backend_sycl_get_tensor_async, + /* .cpy_tensor_async = */ NULL, + /* .synchronize = */ ggml_backend_sycl_synchronize, + /* .graph_plan_create = */ ggml_backend_sycl_graph_plan_create, + /* .graph_plan_free = */ ggml_backend_sycl_graph_plan_free, + /* .graph_plan_compute = */ ggml_backend_sycl_graph_plan_compute, + /* .graph_compute = */ ggml_backend_sycl_graph_compute, + /* .supports_op = */ ggml_backend_sycl_supports_op, +}; + +ggml_backend_t ggml_backend_sycl_init(int device) { + ggml_init_sycl(); // TODO: remove from ggml.c + + if (device < 0 || device >= ggml_sycl_get_device_count()) { + fprintf(stderr, "%s: error: invalid device %d\n", __func__, device); + return nullptr; + } + + // not strictly necessary, but it may reduce the overhead of the first graph_compute + ggml_sycl_set_main_device(device); + + ggml_backend_context_sycl * ctx = new ggml_backend_context_sycl { + /* .device = */ device + }; + + ggml_backend_t sycl_backend = new ggml_backend { + /* .interface = */ ggml_backend_sycl_interface, + /* .context = */ ctx + }; + + return sycl_backend; +} + +bool ggml_backend_is_sycl(ggml_backend_t backend) { + return backend->iface.get_name == ggml_backend_sycl_name; +} + +static ggml_backend_t ggml_backend_reg_sycl_init(const char * params, void * user_data) { + ggml_backend_t sycl_backend = ggml_backend_sycl_init((int) (intptr_t) user_data); + return sycl_backend; + + UNUSED(params); +} + +extern "C" int ggml_backend_sycl_reg_devices(); + +int ggml_backend_sycl_reg_devices() { + int device_count = ggml_sycl_get_device_count(); + + for (int i = 0; i < device_count; i++) { + char name[128]; + snprintf(name, sizeof(name), "%s%d", GGML_SYCL_NAME, i); + ggml_backend_register(name, ggml_backend_reg_sycl_init, ggml_backend_sycl_buffer_type(i), (void *) (intptr_t) i); + } + return device_count; +} diff --git a/ggml-sycl.h b/ggml-sycl.h new file mode 100644 index 000000000..0eabb53cc --- /dev/null +++ b/ggml-sycl.h @@ -0,0 +1,27 @@ +/*MIT license + Copyright (C) 2024 Intel Corporation + SPDX-License-Identifier: MIT +*/ + +#pragma once + +#include "ggml.h" +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + +#define GGML_SYCL_MAX_DEVICES 16 +#define GGML_SYCL_NAME "SYCL" + +GGML_API void ggml_init_sycl(void); +GGML_API bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor); +GGML_API ggml_backend_t ggml_backend_sycl_init(int device); +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device); +GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void); +GGML_API void ggml_backend_sycl_print_sycl_devices(void); + +#ifdef __cplusplus +} +#endif diff --git a/ggml.c b/ggml.c index 11a3114e7..1187f84a2 100644 --- a/ggml.c +++ b/ggml.c @@ -248,6 +248,8 @@ inline static void * ggml_aligned_malloc(size_t size) { #include "ggml-cuda.h" #elif defined(GGML_USE_CLBLAST) #include "ggml-opencl.h" +#elif defined(GGML_USE_SYCL) +#include "ggml-sycl.h" #endif // floating point type used to accumulate sums @@ -2293,6 +2295,8 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { ggml_init_cublas(); #elif defined(GGML_USE_CLBLAST) ggml_cl_init(); +#elif defined(GGML_USE_SYCL) + ggml_init_sycl(); #endif ggml_setup_op_has_task_pass(); @@ -14701,6 +14705,12 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU); #endif // GGML_USE_CUBLAS +#ifdef GGML_USE_SYCL + bool skip_cpu = ggml_sycl_compute_forward(params, tensor); + if (skip_cpu) { + return; + } +#endif // GGML_USE_SYCL switch (tensor->op) { case GGML_OP_DUP: { @@ -20280,7 +20290,7 @@ int ggml_cpu_has_wasm_simd(void) { } int ggml_cpu_has_blas(void) { -#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_SYCL) return 1; #else return 0; @@ -20303,8 +20313,16 @@ int ggml_cpu_has_clblast(void) { #endif } +int ggml_cpu_has_sycl(void) { +#if defined(GGML_USE_SYCL) + return 1; +#else + return 0; +#endif +} + int ggml_cpu_has_gpublas(void) { - return ggml_cpu_has_cublas() || ggml_cpu_has_clblast(); + return ggml_cpu_has_cublas() || ggml_cpu_has_clblast() || ggml_cpu_has_sycl(); } int ggml_cpu_has_sse3(void) { diff --git a/ggml.h b/ggml.h index 1c4976271..3d8d6f2ae 100644 --- a/ggml.h +++ b/ggml.h @@ -2266,6 +2266,7 @@ extern "C" { GGML_API int ggml_cpu_has_gpublas (void); GGML_API int ggml_cpu_has_sse3 (void); GGML_API int ggml_cpu_has_ssse3 (void); + GGML_API int ggml_cpu_has_sycl (void); GGML_API int ggml_cpu_has_vsx (void); // diff --git a/llama.cpp b/llama.cpp index 45eeadbe8..09d799f71 100644 --- a/llama.cpp +++ b/llama.cpp @@ -11,6 +11,8 @@ # include "ggml-cuda.h" #elif defined(GGML_USE_CLBLAST) # include "ggml-opencl.h" +#elif defined(GGML_USE_SYCL) +# include "ggml-sycl.h" #endif #ifdef GGML_USE_METAL @@ -1278,6 +1280,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_cpu(bool host_buffer if (host_buffer) { buft = ggml_backend_cuda_host_buffer_type(); } +#elif defined(GGML_USE_SYCL) + buft = ggml_backend_sycl_host_buffer_type(); #elif defined(GGML_USE_CPU_HBM) buft = ggml_backend_cpu_hbm_buffer_type(); #endif @@ -1297,6 +1301,8 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_offload(int gpu) { buft = ggml_backend_metal_buffer_type(); #elif defined(GGML_USE_CUBLAS) buft = ggml_backend_cuda_buffer_type(gpu); +#elif defined(GGML_USE_SYCL) + buft = ggml_backend_sycl_buffer_type(gpu); #elif defined(GGML_USE_CLBLAST) buft = ggml_backend_opencl_buffer_type(); #endif @@ -10225,6 +10231,16 @@ struct llama_context * llama_new_context_with_model( } } } +#elif defined(GGML_USE_SYCL) + if (model->n_gpu_layers > 0) { + ggml_backend_t backend = ggml_backend_sycl_init(model->main_gpu); + if (backend == nullptr) { + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d backend\n", __func__, model->main_gpu); + llama_free(ctx); + return nullptr; + } + ctx->backends.push_back(backend); + } #endif ctx->backend_cpu = ggml_backend_cpu_init(); if (ctx->backend_cpu == nullptr) { diff --git a/llama.h b/llama.h index 7b3634aa6..d11cba815 100644 --- a/llama.h +++ b/llama.h @@ -6,6 +6,9 @@ #ifdef GGML_USE_CUBLAS #include "ggml-cuda.h" #define LLAMA_MAX_DEVICES GGML_CUDA_MAX_DEVICES +#elif defined(GGML_USE_SYCL) +#include "ggml-sycl.h" +#define LLAMA_MAX_DEVICES GGML_SYCL_MAX_DEVICES #else #define LLAMA_MAX_DEVICES 1 #endif // GGML_USE_CUBLAS @@ -46,7 +49,7 @@ #define LLAMA_SESSION_MAGIC LLAMA_FILE_MAGIC_GGSN #define LLAMA_SESSION_VERSION 4 -#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) +#if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL) || defined(GGML_USE_SYCL) // Defined when llama.cpp is compiled with support for offloading model layers to GPU. #define LLAMA_SUPPORTS_GPU_OFFLOAD #endif diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index e3c656f56..015939105 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -239,10 +239,17 @@ static std::string var_to_str(ggml_type type) { #define VARS_TO_STR10(a, b, c, d, e, f, g, h, i, j) VAR_TO_STR(a) + "," + VARS_TO_STR9(b, c, d, e, f, g, h, i, j) #define VARS_TO_STR11(a, b, c, d, e, f, g, h, i, j, k) VAR_TO_STR(a) + "," + VARS_TO_STR10(b, c, d, e, f, g, h, i, j, k) +#ifdef GGML_USE_SYCL +static bool inline _isinf(float f) { + return (*(uint32_t *)&f & 0x7fffffff) == 0x7f800000; +} +#else +static bool inline _isinf(float f) { return std::isinf(f); } +#endif // accept FLT_MAX as infinity static bool isinf_or_max(float f) { - return std::isinf(f) || f == FLT_MAX || f == -FLT_MAX; + return _isinf(f) || f == FLT_MAX || f == -FLT_MAX; } static bool ggml_is_view_op(enum ggml_op op) {