From 652ca2bded3c818320d92c70d2b67f64bdbff5e5 Mon Sep 17 00:00:00 2001 From: slaren Date: Tue, 5 Mar 2024 22:27:29 +0100 Subject: [PATCH 01/32] compare-llama-bench.py : remove mul_mat_q (#5892) --- scripts/compare-llama-bench.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index 39c3e52e5..54a7771d8 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -18,7 +18,7 @@ except ImportError as e: KEY_PROPERTIES = [ "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas", "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads", - "type_k", "type_v", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen" + "type_k", "type_v", "no_kv_offload", "tensor_split", "n_prompt", "n_gen" ] # Properties that are boolean and are converted to Yes/No for the table: From 8ced9f7e3225adb8501e9821ed1bbd92e3a5c7ae Mon Sep 17 00:00:00 2001 From: Neo Zhang Jianyu Date: Wed, 6 Mar 2024 12:08:32 +0800 Subject: [PATCH 02/32] add wait() to make code stable (#5895) --- ci/run.sh | 3 ++- ggml-sycl.cpp | 59 +++++++++++++++++++++++++++++++++++++++------------ 2 files changed, 48 insertions(+), 14 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index 35eb3c7aa..51f4c74cc 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -45,7 +45,8 @@ 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" + echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:" + echo "source /opt/intel/oneapi/setvars.sh" exit 1 fi diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index 477f5cb02..ddd951dd6 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -3769,8 +3769,42 @@ void log_ggml_var_device(const char*name, float *src, size_t total_elements, boo std::ofstream logfile; logfile.open(filename); for(size_t i=0; imemcpy(local_buf, src, total_size).wait(); + } + else { + local_buf = (sycl::half *)src; + } + + std::ofstream logfile; + logfile.open(filename); + for(size_t i=0; iwait(); 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); } @@ -14159,6 +14193,7 @@ inline void ggml_sycl_op_mul_mat_sycl( dpct::get_value(&alpha, *g_sycl_handles[id]), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *g_sycl_handles[id]), dst_dd_i, ldc))); + g_sycl_handles[id]->wait(); } (void) dst; (void) src1_ddq_i; @@ -15295,8 +15330,8 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, 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; + dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float; + dpct::library_data_t cu_data_type = dpct::library_data_t::real_float; // dst strides size_t nbd2 = dst->nb[2]; @@ -15308,15 +15343,13 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, const float alpha_f32 = 1.0f; const float beta_f32 = 0.0f; - const void * alpha = &alpha_f16; - const void * beta = &beta_f16; + const void * alpha = &alpha_f32; + const void * beta = &beta_f32; // TODO: Renable (dst->op_params[0] =! GGML_PREC_DEFAULT) pathway - // once oneMKL open source supports half, half, float, float: datatypes - dst_t = (char *) dst_f16.alloc(ne_dst); + // oneMKL open source supports half, half, float, float: datatypes - nbd2 /= sizeof(float) / sizeof(sycl::half); - nbd3 /= sizeof(float) / sizeof(sycl::half); + dst_t = (char *) dst_ddf; GGML_ASSERT(ne12 % ne02 == 0); GGML_ASSERT(ne13 % ne03 == 0); @@ -15356,6 +15389,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, nb11 / nb10, nb12 / nb10, beta, (char *)dst_t, cu_data_type, ne01, nb2 / nb0, ne12 * ne13, cu_compute_type))); + g_sycl_handles[g_main_device]->wait(); } else { const int ne23 = ne12*ne13; @@ -15386,7 +15420,7 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, nb02, nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1); }); - }); + }).wait(); } SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch( *g_sycl_handles[g_main_device], oneapi::mkl::transpose::trans, @@ -15397,11 +15431,10 @@ static void ggml_sycl_mul_mat_batched_sycl(const ggml_tensor *src0, dpct::library_data_t::real_half, nb11 / nb10, beta, (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23, cu_compute_type))); + g_sycl_handles[g_main_device]->wait(); } #endif - 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_dst, main_stream); } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ From 1e35d619a6fb0b9c5e3dc955345980ff056ddbaf Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Wed, 6 Mar 2024 09:12:25 +0200 Subject: [PATCH 03/32] convert : remove AWQ remnants (#5768) --- convert.py | 13 ------------- 1 file changed, 13 deletions(-) diff --git a/convert.py b/convert.py index 6e3a0319b..c15f8c47e 100755 --- a/convert.py +++ b/convert.py @@ -1377,7 +1377,6 @@ def main(args_in: list[str] | None = None) -> None: # We currently only support Q8_0 output on little endian systems. output_choices.append("q8_0") 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") @@ -1393,18 +1392,6 @@ def main(args_in: list[str] | None = None) -> None: parser.add_argument("--skip-unknown", action="store_true", help="skip unknown tensor names instead of failing") args = parser.parse_args(args_in) - if args.awq_path: - 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.") - else: - tmp_model_path.mkdir(parents=True, exist_ok=True) - print("Saving new weighted model ...") - add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path)) - print(f"Saved weighted model at {tmp_model_path}.") - args.model = tmp_model_path if args.dump_single: model_plus = lazy_load_file(args.model) From e25fb4b18fcedb9bed6be4585cf842e9a669b28b Mon Sep 17 00:00:00 2001 From: bobqianic <129547291+bobqianic@users.noreply.github.com> Date: Wed, 6 Mar 2024 07:35:07 +0000 Subject: [PATCH 04/32] ggml : use `uint8x16_t` return type for `ggml_vqtbl1q_u8` (#5894) * use uint8x16_t * Update ggml-quants.c --- ggml-quants.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml-quants.c b/ggml-quants.c index e0c125d43..9dcb76def 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -464,8 +464,8 @@ inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) { } // NOTE: not tested -inline static int8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) { - int8x16_t res; +inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) { + uint8x16_t res; res[ 0] = a[b[ 0]]; res[ 1] = a[b[ 1]]; From e04e04f8fad549bb0b3ec1c91f0413aeb08baf29 Mon Sep 17 00:00:00 2001 From: Jared Van Bortel Date: Wed, 6 Mar 2024 15:42:23 -0500 Subject: [PATCH 05/32] ggml : use SYS_get_cpu if SYS_getcpu is not defined (#5906) Fixes #5694 Fixes ggerganov/whisper.cpp#1894 --- ggml.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/ggml.c b/ggml.c index 6a10bbcb4..92b17ee6e 100644 --- a/ggml.c +++ b/ggml.c @@ -2154,7 +2154,10 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) { getcpu_ret = getcpu(¤t_cpu, &g_state.numa.current_node); #else // old glibc doesn't have a wrapper for this call. Fall back on direct syscall - getcpu_ret = syscall(SYS_getcpu,¤t_cpu,&g_state.numa.current_node); +# if !defined(SYS_getcpu) && defined(SYS_get_cpu) +# define SYS_getcpu SYS_get_cpu // some older glibc versions use this name +# endif + getcpu_ret = syscall(SYS_getcpu, ¤t_cpu, &g_state.numa.current_node); #endif if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) { From ceca1aef0738b57951cd12c603c3477e75312dec Mon Sep 17 00:00:00 2001 From: Neo Zhang Jianyu Date: Thu, 7 Mar 2024 16:34:31 +0800 Subject: [PATCH 06/32] [SYCL] fix error when set main gpu to non-zero (#5901) * fix error when set main gpu to non-zero * fix delete condition --- ggml-sycl.cpp | 178 ++++++++++++++++++++++++++++++-------------------- ggml-sycl.h | 1 + llama.cpp | 16 +++-- 3 files changed, 119 insertions(+), 76 deletions(-) diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index ddd951dd6..221d67b8d 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -3559,12 +3559,31 @@ class sycl_gpu_mgr { int work_group_size = 0; std::string gpus_list = ""; + /* + Use all GPU with same top max compute units + */ sycl_gpu_mgr() { detect_sycl_gpu_list_with_max_cu(); get_allow_gpus(); create_context_with_gpus(); } + /* + Use the assigned GPU as only one + */ + sycl_gpu_mgr(int main_gpu_id) { + sycl::device device = dpct::dev_mgr::instance().get_device(main_gpu_id); + dpct::device_info prop; + dpct::get_device_info(prop, device); + gpus.push_back(main_gpu_id); + devices.push_back(device); + work_group_size = prop.get_max_work_group_size(); + max_compute_units = prop.get_max_compute_units(); + + get_allow_gpus(); + create_context_with_gpus(); + } + void create_context_with_gpus() { sycl::context ctx = sycl::context(devices); assert(gpus.size() > 0); @@ -3580,7 +3599,7 @@ class sycl_gpu_mgr { gpus_list += std::to_string(gpus[i]); gpus_list += ","; } - if (gpus_list.length() > 2) { + if (gpus_list.length() > 1) { gpus_list.pop_back(); } } @@ -3629,8 +3648,8 @@ class sycl_gpu_mgr { if (gpus[i] == id) return i; } - assert(false); - return -1; + printf("miss to get device index by id=%d\n", id); + GGML_ASSERT(false); } int get_next_index(int id) { @@ -3639,8 +3658,7 @@ class sycl_gpu_mgr { if (gpus[i] == id) return i; } - assert(false); - return -1; + GGML_ASSERT(false); } }; @@ -3649,6 +3667,7 @@ 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_id = -1; +static bool g_ggml_backend_sycl_buffer_type_initialized = false; static std::array g_default_tensor_split = {}; @@ -13225,7 +13244,7 @@ void ggml_backend_sycl_print_sycl_devices() { } void print_gpu_device_list() { - fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n", + fprintf(stderr, "detect %d SYCL GPUs: [%s] with top Max compute units:%d\n", g_sycl_gpu_mgr->get_gpu_count(), g_sycl_gpu_mgr->gpus_list.c_str(), g_sycl_gpu_mgr->max_compute_units); @@ -13264,23 +13283,6 @@ void ggml_init_sycl() try { #else fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); #endif - 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); - ggml_backend_sycl_print_sycl_devices(); - - if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); - - g_device_count = g_sycl_gpu_mgr->get_gpu_count(); - g_work_group_size = g_sycl_gpu_mgr->work_group_size; - - print_gpu_device_list(); - - int64_t total_vram = 0; /* NOT REMOVE, keep it for next optimize for XMX. #if defined(SYCL_USE_XMX) @@ -13289,52 +13291,71 @@ void ggml_init_sycl() try { fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); #endif */ - for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { - g_device_caps[id].vmm = 0; - g_device_caps[id].device_id = -1; - g_device_caps[id].cc = 0; - g_tensor_split[id] = 0; - g_default_tensor_split[id] = 0; + + if (CHECK_TRY_ERROR(g_all_sycl_device_count = + dpct::dev_mgr::instance().device_count()) != 0) { + initialized = true; + g_sycl_loaded = false; + return; } - - for (int i = 0; i < g_device_count; ++i) { - int device_id = g_sycl_gpu_mgr->gpus[i]; - g_device_caps[i].vmm = 0; - - dpct::device_info prop; - SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(device_id)))); - - g_default_tensor_split[i] = total_vram; - total_vram += prop.get_global_mem_size(); - - g_device_caps[i].cc = - 100 * prop.get_major_version() + 10 * prop.get_minor_version(); - } - - for (int i = 0; i < g_device_count; ++i) { - g_default_tensor_split[i] /= total_vram; - } - - for (int i = 0; i < g_device_count; ++i) { - SYCL_CHECK(ggml_sycl_set_device(i)); - - // create sycl streams - for (int is = 0; is < MAX_STREAMS; ++is) { - SYCL_CHECK(CHECK_TRY_ERROR( - g_syclStreams[i][is] = - dpct::get_current_device().create_queue( - g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device()))); - } - - const dpct::queue_ptr stream = g_syclStreams[i][0]; - // create sycl handle - SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); - } - + GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES); + ggml_backend_sycl_print_sycl_devices(); + if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); + print_gpu_device_list(); initialized = true; g_sycl_loaded = true; } + + + + g_device_count = g_sycl_gpu_mgr->get_gpu_count(); + g_work_group_size = g_sycl_gpu_mgr->work_group_size; + + int64_t total_vram = 0; + + + for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { + g_device_caps[id].vmm = 0; + g_device_caps[id].device_id = -1; + g_device_caps[id].cc = 0; + g_tensor_split[id] = 0; + g_default_tensor_split[id] = 0; + } + + for (int i = 0; i < g_device_count; ++i) { + int device_id = g_sycl_gpu_mgr->gpus[i]; + g_device_caps[i].vmm = 0; + + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(device_id)))); + + g_default_tensor_split[i] = total_vram; + total_vram += prop.get_global_mem_size(); + + g_device_caps[i].cc = + 100 * prop.get_major_version() + 10 * prop.get_minor_version(); + } + + for (int i = 0; i < g_device_count; ++i) { + g_default_tensor_split[i] /= total_vram; + } + + for (int i = 0; i < g_device_count; ++i) { + SYCL_CHECK(ggml_sycl_set_device(i)); + + // create sycl streams + for (int is = 0; is < MAX_STREAMS; ++is) { + SYCL_CHECK(CHECK_TRY_ERROR( + g_syclStreams[i][is] = + dpct::get_current_device().create_queue( + g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device()))); + } + + const dpct::queue_ptr stream = g_syclStreams[i][0]; + // create sycl handle + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); + } } catch (sycl::exception const &exc) { std::cerr << exc.what() << "Exception caught at file:" << __FILE__ @@ -16732,22 +16753,24 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = { /* .is_host = */ nullptr, }; -ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) { +ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) { + if (device_index>=g_device_count or device_index<0) { + printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n", + device_index, g_device_count-1); + GGML_ASSERT(device_indexgpus[i])}, }; } - ggml_backend_sycl_buffer_type_initialized = true; + g_ggml_backend_sycl_buffer_type_initialized = true; } - - return &ggml_backend_sycl_buffer_types[device]; + return &ggml_backend_sycl_buffer_types[device_index]; } // sycl split buffer type @@ -17496,6 +17519,17 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) { return g_sycl_gpu_mgr->get_index(device_id); } +GGML_API GGML_CALL void ggml_backend_sycl_set_single_device(int main_gpu_id) { + GGML_ASSERT(main_gpu_idbackends.push_back(backend); } else { // LLAMA_SPLIT_LAYER requires a backend for each GPU - int id_list[GGML_SYCL_MAX_DEVICES]; - ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); + for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { - int device_id = id_list[i]; ggml_backend_t backend = ggml_backend_sycl_init(i); if (backend == nullptr) { - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, device_id, i); + int id_list[GGML_SYCL_MAX_DEVICES]; + ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, id_list[i], i); llama_free(ctx); return nullptr; } From 2002bc96bf2cbf5ab981a17d7e994d817c9801f5 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 7 Mar 2024 11:41:53 +0200 Subject: [PATCH 07/32] server : refactor (#5882) * server : refactoring (wip) * server : remove llava/clip objects from build * server : fix empty prompt handling + all slots idle logic * server : normalize id vars * server : code style * server : simplify model chat template validation * server : code style * server : minor * llama : llama_chat_apply_template support null buf * server : do not process embedding requests when disabled * server : reorganize structs and enums + naming fixes * server : merge oai.hpp in utils.hpp * server : refactor system prompt update at start * server : disable cached prompts with self-extend * server : do not process more than n_batch tokens per iter * server: tests: embeddings use a real embeddings model (#5908) * server, tests : bump batch to fit 1 embedding prompt * server: tests: embeddings fix build type Debug is randomly failing (#5911) * server: tests: embeddings, use different KV Cache size * server: tests: embeddings, fixed prompt do not exceed n_batch, increase embedding timeout, reduce number of concurrent embeddings * server: tests: embeddings, no need to wait for server idle as it can timout * server: refactor: clean up http code (#5912) * server : avoid n_available var ggml-ci * server: refactor: better http codes * server : simplify json parsing + add comment about t_last * server : rename server structs * server : allow to override FQDN in tests ggml-ci * server : add comments --------- Co-authored-by: Pierrick Hymbert --- .github/workflows/server.yml | 3 +- Makefile | 5 +- examples/server-embd.py | 2 +- examples/server/CMakeLists.txt | 4 +- examples/server/README.md | 2 +- examples/server/oai.hpp | 225 - examples/server/server.cpp | 3895 ++++++++--------- .../server/tests/features/embeddings.feature | 94 + .../server/tests/features/parallel.feature | 46 - examples/server/tests/features/server.feature | 28 - examples/server/tests/features/steps/steps.py | 86 +- examples/server/tests/requirements.txt | 1 + examples/server/utils.hpp | 703 ++- llama.cpp | 6 +- 14 files changed, 2327 insertions(+), 2773 deletions(-) delete mode 100644 examples/server/oai.hpp create mode 100644 examples/server/tests/features/embeddings.feature diff --git a/.github/workflows/server.yml b/.github/workflows/server.yml index 04e3fc0c1..f9aeefaa8 100644 --- a/.github/workflows/server.yml +++ b/.github/workflows/server.yml @@ -58,7 +58,8 @@ jobs: cmake \ python3-pip \ wget \ - psmisc + psmisc \ + language-pack-en - name: Build id: cmake_build diff --git a/Makefile b/Makefile index 4f26c0463..efce10bb8 100644 --- a/Makefile +++ b/Makefile @@ -724,10 +724,9 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) -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 examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +server: examples/server/server.cpp 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 common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) - $(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual - $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2) + $(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2) gguf: examples/gguf/gguf.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) diff --git a/examples/server-embd.py b/examples/server-embd.py index c5c4ea87b..118e04271 100644 --- a/examples/server-embd.py +++ b/examples/server-embd.py @@ -13,7 +13,7 @@ async def main(): model_url = "http://127.0.0.1:6900" responses: list[requests.Response] = await asyncio.gather(*[requests_post_async( url= f"{model_url}/embedding", - json= {"content": str(i)*1024} + json= {"content": str(0)*1024} ) for i in range(n)]) for response in responses: diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index cc13b2d63..c21eba634 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -1,12 +1,12 @@ set(TARGET server) option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON) include_directories(${CMAKE_CURRENT_SOURCE_DIR}) -add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h) +add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h) install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ ) -target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) if (WIN32) TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32) endif() diff --git a/examples/server/README.md b/examples/server/README.md index 21da7a0a0..591f748f8 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -436,7 +436,7 @@ Notice that each `probs` is an array of length `n_probs`. "next_token": { "has_next_token": true, "n_remain": -1, - "num_tokens_predicted": 0, + "n_decoded": 0, "stopped_eos": false, "stopped_limit": false, "stopped_word": false, diff --git a/examples/server/oai.hpp b/examples/server/oai.hpp deleted file mode 100644 index ff4ad6994..000000000 --- a/examples/server/oai.hpp +++ /dev/null @@ -1,225 +0,0 @@ -#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 struct llama_model * model, - const json &body, /* openai api json semantics */ - const std::string &chat_template) -{ - 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_chat(model, chat_template, body["messages"]); - 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}); -} - -inline static json format_embeddings_response_oaicompat(const json &request, const json &embeddings) -{ - json res = - json{ - {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, - {"object", "list"}, - {"usage", - json{{"prompt_tokens", 0}, - {"total_tokens", 0}}}, - {"data", embeddings} - }; - return res; -} - diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 8fe5e0b19..3bdbde954 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,13 +1,8 @@ +#include "utils.hpp" + #include "common.h" #include "llama.h" #include "grammar-parser.h" -#include "utils.hpp" -#include "oai.hpp" - -#include "../llava/clip.h" -#include "../llava/llava.h" - -#include "stb_image.h" #ifndef NDEBUG // crash the server in debug mode, otherwise send an http 500 error @@ -24,46 +19,76 @@ #include "completion.js.hpp" #include "json-schema-to-grammar.mjs.hpp" -#include -#include +#include #include #include -#include +#include +#include +#include +#include #include using json = nlohmann::json; -struct server_params { - std::string hostname = "127.0.0.1"; - std::vector api_keys; - std::string public_path = "examples/server/public"; - std::string chat_template = ""; - int32_t port = 8080; - int32_t read_timeout = 600; - int32_t write_timeout = 600; - bool slots_endpoint = true; - bool metrics_endpoint = false; - int n_threads_http = -1; -}; - bool server_verbose = false; bool server_log_json = true; enum stop_type { - STOP_FULL, - STOP_PARTIAL, + STOP_TYPE_FULL, + STOP_TYPE_PARTIAL, }; -// TODO: can become bool if we can't find use of more states enum slot_state { - IDLE, - PROCESSING, + SLOT_STATE_IDLE, + SLOT_STATE_PROCESSING, }; enum slot_command { - NONE, - LOAD_PROMPT, - RELEASE, + SLOT_COMMAND_NONE, + SLOT_COMMAND_LOAD_PROMPT, + SLOT_COMMAND_RELEASE, +}; + +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 server_task_type { + SERVER_TASK_TYPE_COMPLETION, + SERVER_TASK_TYPE_CANCEL, + SERVER_TASK_TYPE_NEXT_RESPONSE, + SERVER_TASK_TYPE_METRICS +}; + +struct server_task { + int id = -1; // to be filled by server_queue + int id_multi = -1; + int id_target = -1; + + server_task_type type; + json data; + + bool infill = false; + bool embedding = false; +}; + +struct server_task_result { + int id = -1; + int id_multi = -1; + + json data; + + bool stop; + bool error; +}; + +struct server_task_multi { + int id = -1; + + std::set subtasks_remaining; + std::vector results; }; struct slot_params { @@ -80,26 +105,32 @@ struct slot_params { json input_suffix; }; -struct slot_image { - int32_t id; +struct server_params { + int32_t port = 8080; + int32_t read_timeout = 600; + int32_t write_timeout = 600; + int32_t n_threads_http = -1; - bool request_encode_image = false; - float * image_embedding = nullptr; - int32_t image_tokens = 0; + std::string hostname = "127.0.0.1"; + std::string public_path = "examples/server/public"; + std::string chat_template = ""; + std::string system_prompt = ""; - clip_image_u8 * img_data; + std::vector api_keys; - std::string prefix_prompt; // before of this image + bool slots_endpoint = true; + bool metrics_endpoint = false; }; struct server_slot { int id; - int task_id = -1; + int id_task = -1; + int id_multi = -1; struct slot_params params; - slot_state state = IDLE; - slot_command command = NONE; + slot_state state = SLOT_STATE_IDLE; + slot_command command = SLOT_COMMAND_NONE; // used to determine the slot that has been used the longest int64_t t_last_used = -1; @@ -116,27 +147,31 @@ struct server_slot { int32_t n_prompt_tokens_processed = 0; json prompt; + + // when a task is submitted, we first tokenize the prompt and store it here + std::vector prompt_tokens; + std::string generated_text; - llama_token sampled; std::vector cache_tokens; std::vector generated_token_probs; - bool infill = false; - bool embedding = false; + bool infill = false; + bool embedding = false; bool has_next_token = true; - bool truncated = false; - bool stopped_eos = false; - bool stopped_word = false; - bool stopped_limit = false; + bool truncated = false; + bool stopped_eos = false; + bool stopped_word = false; + bool stopped_limit = false; bool oaicompat = false; - std::string oaicompat_model; + std::string oaicompat_model; std::string stopping_word; // sampling + llama_token sampled; struct llama_sampling_params sparams; - llama_sampling_context *ctx_sampling = nullptr; + llama_sampling_context * ctx_sampling = nullptr; int32_t ga_i = 0; // group-attention state int32_t ga_n = 1; // group-attention factor @@ -144,48 +179,32 @@ struct server_slot { int32_t n_past_se = 0; // self-extend - // multimodal - std::vector images; - // stats size_t n_sent_text = 0; // number of sent text character size_t n_sent_token_probs = 0; int64_t t_start_process_prompt; - int64_t t_start_genereration; + int64_t t_start_generation; double t_prompt_processing; // ms double t_token_generation; // ms - // multitasks - int multitask_id = -1; - void reset() { - n_prompt_tokens = 0; - generated_text = ""; - truncated = false; - stopped_eos = false; - stopped_word = false; - stopped_limit = false; - stopping_word = ""; - n_past = 0; - n_sent_text = 0; - n_sent_token_probs = 0; - infill = false; - ga_i = 0; - n_past_se = 0; + n_prompt_tokens = 0; + generated_text = ""; + truncated = false; + stopped_eos = false; + stopped_word = false; + stopped_limit = false; + stopping_word = ""; + n_past = 0; + n_sent_text = 0; + n_sent_token_probs = 0; + infill = false; + ga_i = 0; + n_past_se = 0; generated_token_probs.clear(); - - for (slot_image & img : images) { - free(img.image_embedding); - if (img.img_data) { - clip_image_u8_free(img.img_data); - } - img.prefix_prompt = ""; - } - - images.clear(); } bool has_budget(gpt_params &global_params) { @@ -205,32 +224,29 @@ struct server_slot { } bool available() const { - return state == IDLE && command == NONE; + return state == SLOT_STATE_IDLE && command == SLOT_COMMAND_NONE; } bool is_processing() const { - return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING; + return (state == SLOT_STATE_IDLE && command == SLOT_COMMAND_LOAD_PROMPT) || state == SLOT_STATE_PROCESSING; } - void add_token_string(const completion_token_output &token) { - if (command == RELEASE) { + void add_token_string(const completion_token_output & token) { + if (command == SLOT_COMMAND_RELEASE) { return; } - cache_tokens.push_back(token.tok); generated_token_probs.push_back(token); } void release() { - if (state == PROCESSING) - { - t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3; - command = RELEASE; + if (state == SLOT_STATE_PROCESSING) { + t_token_generation = (ggml_time_us() - t_start_generation) / 1e3; + command = SLOT_COMMAND_RELEASE; } } - json get_formated_timings() { - return json - { + json get_formated_timings() const { + return json { {"prompt_n", n_prompt_tokens_processed}, {"prompt_ms", t_prompt_processing}, {"prompt_per_token_ms", t_prompt_processing / n_prompt_tokens_processed}, @@ -243,16 +259,47 @@ struct server_slot { }; } + size_t find_stopping_strings(const std::string & text, const size_t last_token_size, const stop_type type) { + size_t stop_pos = std::string::npos; + + for (const std::string & word : params.antiprompt) { + size_t pos; + + if (type == STOP_TYPE_FULL) { + const size_t tmp = word.size() + last_token_size; + const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0; + + pos = text.find(word, from_pos); + } else { + pos = find_partial_stop_string(word, text); + } + + if (pos != std::string::npos && (stop_pos == std::string::npos || pos < stop_pos)) { + if (type == STOP_TYPE_FULL) { + stopped_word = true; + stopping_word = word; + has_next_token = false; + } + stop_pos = pos; + } + } + + return stop_pos; + } + void print_timings() const { - char buffer[512]; + char buffer[512]; + double t_token = t_prompt_processing / n_prompt_tokens_processed; double n_tokens_second = 1e3 / t_prompt_processing * n_prompt_tokens_processed; - sprintf(buffer, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)", + + snprintf(buffer, 512, "prompt eval time = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)", t_prompt_processing, n_prompt_tokens_processed, t_token, n_tokens_second); + LOG_INFO(buffer, { - {"slot_id", id}, - {"task_id", task_id}, + {"id_slot", id}, + {"id_task", id_task}, {"t_prompt_processing", t_prompt_processing}, {"n_prompt_tokens_processed", n_prompt_tokens_processed}, {"t_token", t_token}, @@ -261,22 +308,25 @@ struct server_slot { t_token = t_token_generation / n_decoded; n_tokens_second = 1e3 / t_token_generation * n_decoded; - sprintf(buffer, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)", + + snprintf(buffer, 512, "generation eval time = %10.2f ms / %5d runs (%8.2f ms per token, %8.2f tokens per second)", t_token_generation, n_decoded, t_token, n_tokens_second); + LOG_INFO(buffer, { - {"slot_id", id}, - {"task_id", task_id}, + {"id_slot", id}, + {"id_task", id_task}, {"t_token_generation", t_token_generation}, {"n_decoded", n_decoded}, {"t_token", t_token}, {"n_tokens_second", n_tokens_second}, }); - sprintf(buffer, " total time = %10.2f ms", t_prompt_processing + t_token_generation); + snprintf(buffer, 512, " total time = %10.2f ms", t_prompt_processing + t_token_generation); + LOG_INFO(buffer, { - {"slot_id", id}, - {"task_id", task_id}, + {"id_slot", id}, + {"id_task", id_task}, {"t_prompt_processing", t_prompt_processing}, {"t_token_generation", t_token_generation}, {"t_total", t_prompt_processing + t_token_generation}, @@ -291,9 +341,8 @@ struct server_metrics { uint64_t n_prompt_tokens_processed = 0; uint64_t t_prompt_processing = 0; - uint64_t n_tokens_predicted = 0; - uint64_t t_tokens_generation = 0; - + uint64_t n_tokens_predicted = 0; + uint64_t t_tokens_generation = 0; void on_prompt_eval(const server_slot &slot) { n_prompt_tokens_processed_total += slot.n_prompt_tokens_processed; @@ -315,23 +364,261 @@ struct server_metrics { } }; -struct llama_server_context -{ - llama_model *model = nullptr; - llama_context *ctx = nullptr; +struct server_queue { + int id = 0; + bool running; - clip_ctx *clp_ctx = nullptr; + // queues + std::vector queue_tasks; + std::vector queue_tasks_deferred; + + std::vector queue_multitasks; + + std::mutex mutex_tasks; + std::condition_variable condition_tasks; + + // callback functions + std::function callback_new_task; + std::function callback_finish_multitask; + std::function callback_run_slots; + + // Add a new task to the end of the queue + int post(server_task task) { + std::unique_lock lock(mutex_tasks); + if (task.id == -1) { + task.id = id++; + LOG_VERBOSE("new task id", {{"new_id", task.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(server_task 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); + int new_id = id++; + LOG_VERBOSE("new task id", {{"new_id", new_id}}); + return new_id; + } + + // Register function to process a new task + void on_new_task(std::function callback) { + callback_new_task = std::move(callback); + } + + // Register function to process a multitask when it is finished + void on_finish_multitask(std::function callback) { + callback_finish_multitask = std::move(callback); + } + + // Register the function to be called when all slots data is ready to be processed + void on_run_slots(std::function callback) { + callback_run_slots = std::move(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(); + } + + // end the start_loop routine + void terminate() { + std::unique_lock lock(mutex_tasks); + running = false; + condition_tasks.notify_all(); + } + + /** + * Main loop consists of these steps: + * - Wait until a new task arrives + * - Process the task (i.e. maybe copy data into slot) + * - Check if multitask is finished + * - Run all slots + */ + void start_loop() { + running = true; + + while (true) { + LOG_VERBOSE("new task may arrive", {}); + + while (true) { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + lock.unlock(); + break; + } + server_task task = queue_tasks.front(); + queue_tasks.erase(queue_tasks.begin()); + lock.unlock(); + LOG_VERBOSE("callback_new_task", {{"id_task", task.id}}); + callback_new_task(task); + } + + LOG_VERBOSE("update_multitasks", {}); + + // check if we have any finished 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 + server_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 processed, slots data is now ready + LOG_VERBOSE("callback_run_slots", {}); + + callback_run_slots(); + + LOG_VERBOSE("wait for new task", {}); + { + std::unique_lock lock(mutex_tasks); + if (queue_tasks.empty()) { + if (!running) { + LOG_VERBOSE("ending start_loop", {}); + return; + } + condition_tasks.wait(lock, [&]{ + return (!queue_tasks.empty() || !running); + }); + } + } + } + } + + // + // functions to manage multitasks + // + + // add a multitask by specifying the id of all subtask (subtask is a server_task) + void add_multitask(int id_multi, std::vector & sub_ids) { + std::lock_guard lock(mutex_tasks); + server_task_multi multi; + multi.id = id_multi; + 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 id_multi, int id_sub, server_task_result & result) { + std::lock_guard lock(mutex_tasks); + for (auto & multitask : queue_multitasks) { + if (multitask.id == id_multi) { + multitask.subtasks_remaining.erase(id_sub); + multitask.results.push_back(result); + } + } + } +}; + +struct 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; + + // add the id_task to the list of tasks waiting for response + void add_waiting_task_id(int id_task) { + LOG_VERBOSE("waiting for task id", {{"id_task", id_task}}); + + std::unique_lock lock(mutex_results); + waiting_task_ids.insert(id_task); + } + + // when the request is finished, we can remove task associated with it + void remove_waiting_task_id(int id_task) { + LOG_VERBOSE("remove waiting for task id", {{"id_task", id_task}}); + + std::unique_lock lock(mutex_results); + waiting_task_ids.erase(id_task); + } + + // This function blocks the thread until there is a response for this id_task + server_task_result recv(int id_task) { + 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++) { + if (queue_results[i].id == id_task) { + assert(queue_results[i].id_multi == -1); + server_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 = std::move(callback); + } + + // Send a new result to a waiting id_task + void send(server_task_result result) { + LOG_VERBOSE("send new result", {{"id_task", result.id}}); + + std::unique_lock lock(mutex_results); + for (const auto & id_task : waiting_task_ids) { + // LOG_TEE("waiting task id %i \n", id_task); + // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result + if (result.id_multi == id_task) { + LOG_VERBOSE("callback_update_multitask", {{"id_task", id_task}}); + callback_update_multitask(id_task, result.id, result); + continue; + } + + if (result.id == id_task) { + LOG_VERBOSE("queue_results.push_back", {{"id_task", id_task}}); + queue_results.push_back(result); + condition_results.notify_all(); + return; + } + } + } +}; + +struct server_context { + llama_model * model = nullptr; + llama_context * ctx = nullptr; gpt_params params; llama_batch batch; - bool multimodal = false; - bool clean_kv_cache = true; - bool all_slots_are_idle = false; - bool add_bos_token = true; + bool clean_kv_cache = true; + bool add_bos_token = true; - int32_t n_ctx; // total context for all clients / slots + int32_t n_ctx; // total context for all clients / slots // system prompt bool system_need_update = false; @@ -346,60 +633,32 @@ struct llama_server_context std::vector slots; json default_generation_settings_for_props; - llama_server_queue queue_tasks; - llama_server_response queue_results; + server_queue queue_tasks; + server_response queue_results; server_metrics metrics; - ~llama_server_context() - { - if (ctx) - { + ~server_context() { + if (ctx) { llama_free(ctx); ctx = nullptr; } - if (model) - { + + if (model) { llama_free_model(model); model = nullptr; } } - bool load_model(const gpt_params ¶ms_) - { + bool load_model(const gpt_params & params_) { params = params_; - if (!params.mmproj.empty()) { - multimodal = true; - LOG_INFO("Multi Modal Mode Enabled", {}); - clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1); - if(clp_ctx == nullptr) { - LOG_ERROR("unable to load clip model", {{"model", params.mmproj}}); - return false; - } - - if (params.n_ctx < 2048) { // request larger context for the image embedding - params.n_ctx = 2048; - } - } std::tie(model, ctx) = llama_init_from_gpt_params(params); - if (model == nullptr) - { + if (model == nullptr) { LOG_ERROR("unable to load model", {{"model", params.model}}); return false; } - if (multimodal) { - const int n_embd_clip = clip_n_mmproj_embd(clp_ctx); - const int n_embd_llm = llama_n_embd(model); - if (n_embd_clip != n_embd_llm) { - LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm); - llama_free(ctx); - llama_free_model(model); - return false; - } - } - n_ctx = llama_n_ctx(ctx); add_bos_token = llama_should_add_bos_token(model); @@ -407,25 +666,19 @@ struct llama_server_context return true; } - void validate_model_chat_template(server_params & sparams) { + bool validate_model_chat_template() const { llama_chat_message chat[] = {{"user", "test"}}; - std::vector buf(1); - int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size()); - if (res < 0) { - LOG_ERROR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); - sparams.chat_template = "chatml"; - } + + const int res = llama_chat_apply_template(model, nullptr, chat, 1, true, nullptr, 0); + + return res > 0; } void initialize() { - // create slots - all_slots_are_idle = true; - const int32_t n_ctx_slot = n_ctx / params.n_parallel; LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}}); - for (int i = 0; i < params.n_parallel; i++) - { + for (int i = 0; i < params.n_parallel; i++) { server_slot slot; slot.id = i; @@ -433,7 +686,7 @@ struct llama_server_context slot.n_predict = params.n_predict; LOG_INFO("new slot", { - {"slot_id", slot.id}, + {"id_slot", slot.id}, {"n_ctx_slot", slot.n_ctx} }); @@ -447,9 +700,9 @@ struct llama_server_context //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT LOG_INFO("slot self-extend", { - {"slot_id", slot.id}, - {"ga_n", ga_n}, - {"ga_w", ga_w} + {"id_slot", slot.id}, + {"ga_n", ga_n}, + {"ga_w", ga_w} }); } @@ -468,8 +721,7 @@ struct llama_server_context batch = llama_batch_init(n_ctx, 0, params.n_parallel); } - std::vector tokenize(const json & json_prompt, bool add_bos) const - { + std::vector tokenize(const json & json_prompt, bool add_bos) const { // TODO: currently, we tokenize using special tokens by default // this is not always correct (see https://github.com/ggerganov/llama.cpp/pull/4160#issuecomment-1824826216) // but it's better compared to completely ignoring ChatML and other chat templates @@ -479,38 +731,30 @@ struct llama_server_context // or the first element of the json_prompt array is a string. std::vector prompt_tokens; - if (json_prompt.is_array()) - { + if (json_prompt.is_array()) { bool first = true; - for (const auto& p : json_prompt) - { - if (p.is_string()) - { + for (const auto & p : json_prompt) { + if (p.is_string()) { auto s = p.template get(); + std::vector p; - if (first) - { + if (first) { p = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL); first = false; - } - else - { + } else { p = ::llama_tokenize(ctx, s, false, TMP_FORCE_SPECIAL); } + prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end()); - } - else - { - if (first) - { + } else { + if (first) { first = false; } + prompt_tokens.push_back(p.template get()); } } - } - else - { + } else { auto s = json_prompt.template get(); prompt_tokens = ::llama_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL); } @@ -518,19 +762,18 @@ struct llama_server_context return prompt_tokens; } - server_slot* get_slot(int id) { + server_slot * get_slot(int id) { int64_t t_last = ggml_time_us(); - server_slot *last_used = nullptr; - for (server_slot & slot : slots) - { - if (slot.id == id && slot.available()) - { + server_slot * last_used = nullptr; + + for (server_slot & slot : slots) { + if (slot.id == id && slot.available()) { return &slot; } - if (slot.available() && slot.t_last_used < t_last) - { + // among all available slots, find the one that has been least recently used + if (slot.available() && slot.t_last_used < t_last) { last_used = &slot; t_last = slot.t_last_used; } @@ -539,295 +782,204 @@ struct llama_server_context return last_used; } - bool launch_slot_with_data(server_slot* &slot, json data) { + bool launch_slot_with_data(server_slot & slot, json data) const { slot_params default_params; llama_sampling_params default_sparams; if (data.count("__oaicompat") != 0) { - slot->oaicompat = true; - slot->oaicompat_model = json_value(data, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); + slot.oaicompat = true; + slot.oaicompat_model = json_value(data, "model", std::string(DEFAULT_OAICOMPAT_MODEL)); } else { - slot->oaicompat = false; - slot->oaicompat_model = ""; + slot.oaicompat = false; + slot.oaicompat_model = ""; } - slot->params.stream = json_value(data, "stream", false); - slot->params.cache_prompt = json_value(data, "cache_prompt", false); - slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict); - slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k); - slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p); - slot->sparams.min_p = json_value(data, "min_p", default_sparams.min_p); - slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); - slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); - slot->sparams.temp = json_value(data, "temperature", default_sparams.temp); - slot->sparams.dynatemp_range = json_value(data, "dynatemp_range", default_sparams.dynatemp_range); - slot->sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent); - slot->sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n); - slot->sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat); - slot->sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq); - slot->sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present); - slot->sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat); - slot->sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau); - slot->sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta); - slot->sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); - slot->params.n_keep = json_value(data, "n_keep", slot->params.n_keep); - slot->params.seed = json_value(data, "seed", default_params.seed); - slot->sparams.grammar = json_value(data, "grammar", default_sparams.grammar); - slot->sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); - slot->sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep); + slot.params.stream = json_value(data, "stream", false); + slot.params.cache_prompt = json_value(data, "cache_prompt", false); + slot.params.n_predict = json_value(data, "n_predict", default_params.n_predict); + slot.sparams.top_k = json_value(data, "top_k", default_sparams.top_k); + slot.sparams.top_p = json_value(data, "top_p", default_sparams.top_p); + slot.sparams.min_p = json_value(data, "min_p", default_sparams.min_p); + slot.sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); + slot.sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); + slot.sparams.temp = json_value(data, "temperature", default_sparams.temp); + slot.sparams.dynatemp_range = json_value(data, "dynatemp_range", default_sparams.dynatemp_range); + slot.sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent); + slot.sparams.penalty_last_n = json_value(data, "repeat_last_n", default_sparams.penalty_last_n); + slot.sparams.penalty_repeat = json_value(data, "repeat_penalty", default_sparams.penalty_repeat); + slot.sparams.penalty_freq = json_value(data, "frequency_penalty", default_sparams.penalty_freq); + slot.sparams.penalty_present = json_value(data, "presence_penalty", default_sparams.penalty_present); + slot.sparams.mirostat = json_value(data, "mirostat", default_sparams.mirostat); + slot.sparams.mirostat_tau = json_value(data, "mirostat_tau", default_sparams.mirostat_tau); + slot.sparams.mirostat_eta = json_value(data, "mirostat_eta", default_sparams.mirostat_eta); + slot.sparams.penalize_nl = json_value(data, "penalize_nl", default_sparams.penalize_nl); + slot.params.n_keep = json_value(data, "n_keep", slot.params.n_keep); + slot.params.seed = json_value(data, "seed", default_params.seed); + slot.sparams.grammar = json_value(data, "grammar", default_sparams.grammar); + slot.sparams.n_probs = json_value(data, "n_probs", default_sparams.n_probs); + slot.sparams.min_keep = json_value(data, "min_keep", default_sparams.min_keep); - if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) { + if (slot.params.cache_prompt && slot.ga_n != 1) { + LOG_WARNING("cache_prompt is not supported with group-attention", {}); + slot.params.cache_prompt = false; + } + + if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) { // Might be better to reject the request with a 400 ? LOG_WARNING("Max tokens to predict exceeds server configuration", { - {"params.n_predict", slot->params.n_predict}, - {"slot.n_predict", slot->n_predict}, + {"params.n_predict", slot.params.n_predict}, + {"slot.n_predict", slot.n_predict}, }); - slot->params.n_predict = slot->n_predict; + slot.params.n_predict = slot.n_predict; } // infill - if (data.count("input_prefix") != 0) - { - slot->params.input_prefix = data["input_prefix"]; - } - else - { - slot->params.input_prefix = ""; - } + slot.params.input_prefix = json_value(data, "input_prefix", default_params.input_prefix); + slot.params.input_suffix = json_value(data, "input_suffix", default_params.input_suffix); + slot.prompt = json_value(data, "prompt", std::string("")); - if (data.count("input_suffix") != 0) + // penalize user-provided tokens { - slot->params.input_suffix = data["input_suffix"]; - } - else - { - slot->params.input_suffix = ""; - } + slot.sparams.penalty_prompt_tokens.clear(); + slot.sparams.use_penalty_prompt_tokens = false; - if (data.count("prompt") != 0) - { - slot->prompt = data["prompt"]; - } - else - { - slot->prompt = ""; - } + const auto & penalty_prompt = data.find("penalty_prompt"); - slot->sparams.penalty_prompt_tokens.clear(); - slot->sparams.use_penalty_prompt_tokens = false; - const auto &penalty_prompt = data.find("penalty_prompt"); - if (penalty_prompt != data.end()) - { - if (penalty_prompt->is_string()) - { - const auto penalty_prompt_string = penalty_prompt->get(); - auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false); - slot->sparams.penalty_prompt_tokens.swap(penalty_tokens); - if (slot->params.n_predict > 0) - { - slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict); - } - slot->sparams.use_penalty_prompt_tokens = true; - } - else if (penalty_prompt->is_array()) - { - const auto n_tokens = penalty_prompt->size(); - slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict)); - const int n_vocab = llama_n_vocab(model); - for (const auto &penalty_token : *penalty_prompt) - { - if (penalty_token.is_number_integer()) - { - const auto tok = penalty_token.get(); - if (tok >= 0 && tok < n_vocab) - { - slot->sparams.penalty_prompt_tokens.push_back(tok); - } + if (penalty_prompt != data.end()) { + if (penalty_prompt->is_string()) { + const auto penalty_prompt_string = penalty_prompt->get(); + slot.sparams.penalty_prompt_tokens = llama_tokenize(model, penalty_prompt_string, false); + + if (slot.params.n_predict > 0) { + slot.sparams.penalty_prompt_tokens.reserve(slot.sparams.penalty_prompt_tokens.size() + slot.params.n_predict); } - } - slot->sparams.use_penalty_prompt_tokens = true; - } - } + slot.sparams.use_penalty_prompt_tokens = true; - slot->sparams.logit_bias.clear(); - - if (json_value(data, "ignore_eos", false)) - { - slot->sparams.logit_bias[llama_token_eos(model)] = -INFINITY; - } - - const auto &logit_bias = data.find("logit_bias"); - if (logit_bias != data.end() && logit_bias->is_array()) - { - const int n_vocab = llama_n_vocab(model); - for (const auto &el : *logit_bias) - { - if (el.is_array() && el.size() == 2) - { - float bias; - if (el[1].is_number()) - { - bias = el[1].get(); - } - else if (el[1].is_boolean() && !el[1].get()) - { - bias = -INFINITY; - } - else - { - continue; - } - - if (el[0].is_number_integer()) - { - llama_token tok = el[0].get(); - if (tok >= 0 && tok < n_vocab) - { - slot->sparams.logit_bias[tok] = bias; - } - } - else if (el[0].is_string()) - { - auto toks = llama_tokenize(model, el[0].get(), false); - for (auto tok : toks) - { - slot->sparams.logit_bias[tok] = bias; - } - } - } - } - } - - slot->params.antiprompt.clear(); - - const auto &stop = data.find("stop"); - if (stop != data.end() && stop->is_array()) - { - for (const auto &word : *stop) - { - if (!word.empty()) - { - slot->params.antiprompt.push_back(word); - } - } - } - - const auto &samplers_sequence = data.find("samplers"); - if (samplers_sequence != data.end() && samplers_sequence->is_array()) - { - std::vector sampler_names; - for (const auto &sampler_name : *samplers_sequence) - { - if (sampler_name.is_string()) - { - sampler_names.emplace_back(sampler_name); - } - } - slot->sparams.samplers_sequence = sampler_types_from_names(sampler_names, false); - } - else - { - slot->sparams.samplers_sequence = default_sparams.samplers_sequence; - } - - if (multimodal) - { - const auto &images_data = data.find("image_data"); - if (images_data != data.end() && images_data->is_array()) - { - for (const auto &img : *images_data) - { - const std::vector image_buffer = base64_decode(img["data"].get()); - - slot_image img_sl; - img_sl.id = img.count("id") != 0 ? img["id"].get() : slot->images.size(); - img_sl.img_data = clip_image_u8_init(); - if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data)) - { - LOG_ERROR("failed to load image", { - {"slot_id", slot->id}, - {"img_sl_id", img_sl.id} - }); - return false; - } - LOG_VERBOSE("image loaded", { - {"slot_id", slot->id}, - {"img_sl_id", img_sl.id} + LOG_VERBOSE("penalty_prompt_tokens", { + {"id_slot", slot.id}, + {"tokens", slot.sparams.penalty_prompt_tokens}, }); - img_sl.request_encode_image = true; - slot->images.push_back(img_sl); } - // process prompt - // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]} - if (slot->images.size() > 0 && !slot->prompt.is_array()) - { - std::string prompt = slot->prompt.get(); - size_t pos = 0, begin_prefix = 0; - std::string pattern = "[img-"; - while ((pos = prompt.find(pattern, pos)) != std::string::npos) { - size_t end_prefix = pos; - pos += pattern.length(); - size_t end_pos = prompt.find(']', pos); - if (end_pos != std::string::npos) - { - std::string image_id = prompt.substr(pos, end_pos - pos); - try - { - int img_id = std::stoi(image_id); - bool found = false; - for (slot_image &img : slot->images) - { - if (img.id == img_id) { - found = true; - img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix); - begin_prefix = end_pos + 1; - break; - } - } - if (!found) { - LOG_TEE("ERROR: Image with id: %i, not found.\n", img_id); - slot->images.clear(); - return false; - } - } catch (const std::invalid_argument& e) { - LOG_TEE("Invalid image number id in prompt\n"); - slot->images.clear(); - return false; + else if (penalty_prompt->is_array()) { + const auto n_tokens = penalty_prompt->size(); + slot.sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot.params.n_predict)); + + const int n_vocab = llama_n_vocab(model); + for (const auto & penalty_token : *penalty_prompt) { + if (penalty_token.is_number_integer()) { + const auto tok = penalty_token.get(); + if (tok >= 0 && tok < n_vocab) { + slot.sparams.penalty_prompt_tokens.push_back(tok); } } } - slot->prompt = ""; - slot->params.input_suffix = prompt.substr(begin_prefix); - slot->params.cache_prompt = false; // multimodal doesn't support cache prompt + slot.sparams.use_penalty_prompt_tokens = true; + + LOG_VERBOSE("penalty_prompt_tokens", { + {"id_slot", slot.id}, + {"tokens", slot.sparams.penalty_prompt_tokens}, + }); } } } - if (slot->ctx_sampling != nullptr) { - llama_sampling_free(slot->ctx_sampling); - } - slot->ctx_sampling = llama_sampling_init(slot->sparams); - llama_set_rng_seed(ctx, slot->params.seed); - slot->command = LOAD_PROMPT; + slot.sparams.logit_bias.clear(); - all_slots_are_idle = false; + if (json_value(data, "ignore_eos", false)) { + slot.sparams.logit_bias[llama_token_eos(model)] = -INFINITY; + } + + const auto & logit_bias = data.find("logit_bias"); + if (logit_bias != data.end() && logit_bias->is_array()) { + const int n_vocab = llama_n_vocab(model); + for (const auto & el : *logit_bias) { + if (el.is_array() && el.size() == 2) { + float bias; + if (el[1].is_number()) { + bias = el[1].get(); + } else if (el[1].is_boolean() && !el[1].get()) { + bias = -INFINITY; + } else { + continue; + } + + if (el[0].is_number_integer()) { + llama_token tok = el[0].get(); + if (tok >= 0 && tok < n_vocab) { + slot.sparams.logit_bias[tok] = bias; + } + } else if (el[0].is_string()) { + auto toks = llama_tokenize(model, el[0].get(), false); + for (auto tok : toks) { + slot.sparams.logit_bias[tok] = bias; + } + } + } + } + } + } + + { + slot.params.antiprompt.clear(); + + const auto & stop = data.find("stop"); + if (stop != data.end() && stop->is_array()) { + for (const auto & word : *stop) { + if (!word.empty()) { + slot.params.antiprompt.push_back(word); + } + } + } + } + + { + const auto & samplers_sequence = data.find("samplers"); + if (samplers_sequence != data.end() && samplers_sequence->is_array()) { + std::vector sampler_names; + for (const auto & sampler_name : *samplers_sequence) { + if (sampler_name.is_string()) { + sampler_names.emplace_back(sampler_name); + } + } + slot.sparams.samplers_sequence = sampler_types_from_names(sampler_names, false); + } else { + slot.sparams.samplers_sequence = default_sparams.samplers_sequence; + } + } + + { + if (slot.ctx_sampling != nullptr) { + llama_sampling_free(slot.ctx_sampling); + } + slot.ctx_sampling = llama_sampling_init(slot.sparams); + llama_set_rng_seed(ctx, slot.params.seed); + } + + slot.command = SLOT_COMMAND_LOAD_PROMPT; + slot.prompt_tokens.clear(); LOG_INFO("slot is processing task", { - {"slot_id", slot->id}, - {"task_id", slot->task_id}, + {"id_slot", slot.id}, + {"id_task", slot.id_task}, }); return true; } void kv_cache_clear() { + LOG_VERBOSE("clearing KV cache", {}); + // clear the entire KV cache llama_kv_cache_clear(ctx); clean_kv_cache = false; } void system_prompt_update() { + LOG_VERBOSE("system prompt update", { + {"system_prompt", system_prompt}, + }); + kv_cache_clear(); system_tokens.clear(); @@ -836,13 +988,11 @@ struct llama_server_context llama_batch_clear(batch); - for (int i = 0; i < (int)system_tokens.size(); ++i) - { + for (int i = 0; i < (int)system_tokens.size(); ++i) { llama_batch_add(batch, system_tokens[i], i, { 0 }, false); } - for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch) - { + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch) { const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i)); llama_batch batch_view = { n_tokens, @@ -854,78 +1004,42 @@ struct llama_server_context batch.logits + i, 0, 0, 0, // unused }; - if (llama_decode(ctx, batch_view) != 0) - { + + if (llama_decode(ctx, batch_view) != 0) { LOG_TEE("%s: llama_decode() failed\n", __func__); return; } } // assign the system KV cache to all parallel sequences - for (int32_t i = 1; i < params.n_parallel; ++i) - { + for (int32_t i = 1; i < params.n_parallel; ++i) { llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size()); } } - LOG_TEE("system prompt updated\n"); system_need_update = false; } - void system_prompt_notify() { + void system_prompt_set(const json & sys_props) { + system_prompt = sys_props.value("prompt", ""); + name_user = sys_props.value("anti_prompt", ""); + name_assistant = sys_props.value("assistant_name", ""); + + LOG_VERBOSE("system prompt process", { + {"system_prompt", system_prompt}, + {"name_user", name_user}, + {"name_assistant", name_assistant}, + }); + // release all slots - for (server_slot &slot : slots) - { + for (server_slot & slot : slots) { slot.release(); } system_need_update = true; } - void system_prompt_process(const json &sys_props) { - system_prompt = sys_props.value("prompt", ""); - name_user = sys_props.value("anti_prompt", ""); - name_assistant = sys_props.value("assistant_name", ""); - - - system_prompt_notify(); - } - - static size_t find_stopping_strings(const std::string &text, const size_t last_token_size, - const stop_type type, server_slot &slot) - { - size_t stop_pos = std::string::npos; - - for (const std::string &word : slot.params.antiprompt) - { - size_t pos; - if (type == STOP_FULL) - { - const size_t tmp = word.size() + last_token_size; - const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0; - pos = text.find(word, from_pos); - } - else - { - pos = find_partial_stop_string(word, text); - } - if (pos != std::string::npos && - (stop_pos == std::string::npos || pos < stop_pos)) - { - if (type == STOP_FULL) - { - slot.stopped_word = true; - slot.stopping_word = word; - slot.has_next_token = false; - } - stop_pos = pos; - } - } - - return stop_pos; - } - - bool process_token(completion_token_output &result, server_slot &slot) { + bool process_token(completion_token_output & result, server_slot & slot) { // remember which tokens were sampled - used for repetition penalties during sampling const std::string token_str = llama_token_to_piece(ctx, result.tok); slot.sampled = result.tok; @@ -934,34 +1048,26 @@ struct llama_server_context slot.generated_text += token_str; slot.has_next_token = true; - if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1) - { + if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1) { // we can change penalty_prompt_tokens because it is always created from scratch each request slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok); } // check if there is incomplete UTF-8 character at the end bool incomplete = false; - for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i) - { + for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i) { unsigned char c = slot.generated_text[slot.generated_text.size() - i]; - if ((c & 0xC0) == 0x80) - { + if ((c & 0xC0) == 0x80) { // continuation byte: 10xxxxxx continue; } - if ((c & 0xE0) == 0xC0) - { + if ((c & 0xE0) == 0xC0) { // 2-byte character: 110xxxxx ... incomplete = i < 2; - } - else if ((c & 0xF0) == 0xE0) - { + } else if ((c & 0xF0) == 0xE0) { // 3-byte character: 1110xxxx ... incomplete = i < 3; - } - else if ((c & 0xF8) == 0xF0) - { + } else if ((c & 0xF8) == 0xF0) { // 4-byte character: 11110xxx ... incomplete = i < 4; } @@ -969,206 +1075,181 @@ struct llama_server_context break; } - if (!incomplete) - { + if (!incomplete) { size_t pos = std::min(slot.n_sent_text, slot.generated_text.size()); + const std::string str_test = slot.generated_text.substr(pos); bool is_stop_full = false; - size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot); - if (stop_pos != std::string::npos) - { + + size_t stop_pos = slot.find_stopping_strings(str_test, token_str.size(), STOP_TYPE_FULL); + if (stop_pos != std::string::npos) { is_stop_full = true; slot.generated_text.erase( slot.generated_text.begin() + pos + stop_pos, slot.generated_text.end()); pos = std::min(slot.n_sent_text, slot.generated_text.size()); - } - else - { + } else { is_stop_full = false; - stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot); + stop_pos = slot.find_stopping_strings(str_test, token_str.size(), STOP_TYPE_PARTIAL); } // check if there is any token to predict - if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0)) - { + if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0)) { // no send the stop word in the response result.text_to_send = slot.generated_text.substr(pos, std::string::npos); slot.n_sent_text += result.text_to_send.size(); // add the token to slot queue and cache } + slot.add_token_string(result); - if (slot.params.stream) - { + if (slot.params.stream) { send_partial_response(slot, result); } } - if (incomplete) - { + if (incomplete) { slot.has_next_token = true; } // check the limits - if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params)) - { - slot.stopped_limit = true; + if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params)) { + slot.stopped_limit = true; slot.has_next_token = false; + + LOG_VERBOSE("stopped by limit", { + {"id_slot", slot.id}, + {"n_decoded", slot.n_decoded}, + {"n_predict", slot.params.n_predict}, + }); } - if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model)) - { - slot.stopped_eos = true; + if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model)) { + slot.stopped_eos = true; slot.has_next_token = false; + LOG_VERBOSE("eos token found", {}); } LOG_VERBOSE("next token", { - {"token", result.tok}, - {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, - {"has_next_token", slot.has_next_token}, - {"n_remain", slot.n_remaining}, - {"num_tokens_predicted", slot.n_decoded}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - }); + {"token", result.tok}, + {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, + {"has_next_token", slot.has_next_token}, + {"n_remain", slot.n_remaining}, + {"n_decoded", slot.n_decoded}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, + }); return slot.has_next_token; // continue } - bool process_images(server_slot &slot) const - { - for (slot_image &img : slot.images) - { - if (!img.request_encode_image) - { - continue; - } - - if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) { - LOG_TEE("Error processing the given image"); - return false; - } - - - img.request_encode_image = false; - } - - return slot.images.size() > 0; - } - - void send_error(task_server& task, const std::string &error) - { - LOG_TEE("task %i - error: %s\n", task.id, error.c_str()); - 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.send(res); - } - - json get_formated_generation(server_slot &slot) - { + json get_formated_generation(const server_slot & slot) const { const auto eos_bias = slot.sparams.logit_bias.find(llama_token_eos(model)); - const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && - eos_bias->second < 0.0f && std::isinf(eos_bias->second); + const bool ignore_eos = eos_bias != slot.sparams.logit_bias.end() && eos_bias->second < 0.0f && std::isinf(eos_bias->second); + std::vector samplers_sequence; - for (const auto &sampler_type : slot.sparams.samplers_sequence) - { + samplers_sequence.reserve(slot.sparams.samplers_sequence.size()); + for (const auto & sampler_type : slot.sparams.samplers_sequence) { samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type)); } return json { - {"n_ctx", slot.n_ctx}, - {"n_predict", slot.n_predict}, - {"model", params.model_alias}, - {"seed", slot.params.seed}, - {"temperature", slot.sparams.temp}, - {"dynatemp_range", slot.sparams.dynatemp_range}, - {"dynatemp_exponent", slot.sparams.dynatemp_exponent}, - {"top_k", slot.sparams.top_k}, - {"top_p", slot.sparams.top_p}, - {"min_p", slot.sparams.min_p}, - {"tfs_z", slot.sparams.tfs_z}, - {"typical_p", slot.sparams.typical_p}, - {"repeat_last_n", slot.sparams.penalty_last_n}, - {"repeat_penalty", slot.sparams.penalty_repeat}, - {"presence_penalty", slot.sparams.penalty_present}, - {"frequency_penalty", slot.sparams.penalty_freq}, - {"penalty_prompt_tokens", slot.sparams.penalty_prompt_tokens}, + {"n_ctx", slot.n_ctx}, + {"n_predict", slot.n_predict}, + {"model", params.model_alias}, + {"seed", slot.params.seed}, + {"temperature", slot.sparams.temp}, + {"dynatemp_range", slot.sparams.dynatemp_range}, + {"dynatemp_exponent", slot.sparams.dynatemp_exponent}, + {"top_k", slot.sparams.top_k}, + {"top_p", slot.sparams.top_p}, + {"min_p", slot.sparams.min_p}, + {"tfs_z", slot.sparams.tfs_z}, + {"typical_p", slot.sparams.typical_p}, + {"repeat_last_n", slot.sparams.penalty_last_n}, + {"repeat_penalty", slot.sparams.penalty_repeat}, + {"presence_penalty", slot.sparams.penalty_present}, + {"frequency_penalty", slot.sparams.penalty_freq}, + {"penalty_prompt_tokens", slot.sparams.penalty_prompt_tokens}, {"use_penalty_prompt_tokens", slot.sparams.use_penalty_prompt_tokens}, - {"mirostat", slot.sparams.mirostat}, - {"mirostat_tau", slot.sparams.mirostat_tau}, - {"mirostat_eta", slot.sparams.mirostat_eta}, - {"penalize_nl", slot.sparams.penalize_nl}, - {"stop", slot.params.antiprompt}, - {"n_predict", slot.params.n_predict}, - {"n_keep", params.n_keep}, - {"ignore_eos", ignore_eos}, - {"stream", slot.params.stream}, - {"logit_bias", slot.sparams.logit_bias}, - {"n_probs", slot.sparams.n_probs}, - {"min_keep", slot.sparams.min_keep}, - {"grammar", slot.sparams.grammar}, - {"samplers", samplers_sequence} + {"mirostat", slot.sparams.mirostat}, + {"mirostat_tau", slot.sparams.mirostat_tau}, + {"mirostat_eta", slot.sparams.mirostat_eta}, + {"penalize_nl", slot.sparams.penalize_nl}, + {"stop", slot.params.antiprompt}, + {"n_predict", slot.params.n_predict}, + {"n_keep", params.n_keep}, + {"ignore_eos", ignore_eos}, + {"stream", slot.params.stream}, + {"logit_bias", slot.sparams.logit_bias}, + {"n_probs", slot.sparams.n_probs}, + {"min_keep", slot.sparams.min_keep}, + {"grammar", slot.sparams.grammar}, + {"samplers", samplers_sequence} }; } - void send_partial_response(server_slot &slot, completion_token_output tkn) - { - task_result res; - res.id = slot.task_id; - res.multitask_id = slot.multitask_id; - res.error = false; - res.stop = false; + void send_error(const server_task & task, const std::string & error) { + LOG_TEE("task %i - error: %s\n", task.id, error.c_str()); - res.result_json = json - { + server_task_result res; + res.id = task.id; + res.id_multi = task.id_multi; + res.stop = false; + res.error = true; + res.data = { { "content", error } }; + + queue_results.send(res); + } + + void send_partial_response(server_slot & slot, completion_token_output tkn) { + server_task_result res; + res.id = slot.id_task; + res.id_multi = slot.id_multi; + res.error = false; + res.stop = false; + res.data = json { {"content", tkn.text_to_send}, {"stop", false}, - {"slot_id", slot.id}, - {"multimodal", multimodal} + {"id_slot", slot.id}, + {"multimodal", false} }; - if (slot.sparams.n_probs > 0) - { - std::vector probs_output = {}; + if (slot.sparams.n_probs > 0) { const std::vector to_send_toks = llama_tokenize(ctx, tkn.text_to_send, false); - size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size()); - size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size()); - if (probs_pos < probs_stop_pos) - { - probs_output = std::vector(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos); + const size_t probs_pos = std::min(slot.n_sent_token_probs, slot.generated_token_probs.size()); + const size_t probs_stop_pos = std::min(slot.n_sent_token_probs + to_send_toks.size(), slot.generated_token_probs.size()); + + std::vector probs_output; + if (probs_pos < probs_stop_pos) { + probs_output = std::vector( + slot.generated_token_probs.begin() + probs_pos, + slot.generated_token_probs.begin() + probs_stop_pos); } slot.n_sent_token_probs = probs_stop_pos; - res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); + + res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs_output); } - if (slot.oaicompat) - { - res.result_json["oaicompat_token_ctr"] = slot.n_decoded; - res.result_json["model"] = slot.oaicompat_model; + if (slot.oaicompat) { + res.data["oaicompat_token_ctr"] = slot.n_decoded; + res.data["model"] = slot.oaicompat_model; } queue_results.send(res); } - void send_final_response(server_slot &slot) - { - task_result res; - res.id = slot.task_id; - res.multitask_id = slot.multitask_id; - res.error = false; - res.stop = true; - - res.result_json = json - { + void send_final_response(const server_slot & slot) { + server_task_result res; + res.id = slot.id_task; + res.id_multi = slot.id_multi; + res.error = false; + res.stop = true; + res.data = json { {"content", !slot.params.stream ? slot.generated_text : ""}, - {"slot_id", slot.id}, + {"id_slot", slot.id}, {"stop", true}, {"model", params.model_alias}, {"tokens_predicted", slot.n_decoded}, @@ -1184,96 +1265,87 @@ struct llama_server_context {"timings", slot.get_formated_timings()} }; - if (slot.sparams.n_probs > 0) - { - std::vector probs = {}; - if (!slot.params.stream && slot.stopped_word) - { + if (slot.sparams.n_probs > 0) { + std::vector probs; + if (!slot.params.stream && slot.stopped_word) { const std::vector stop_word_toks = llama_tokenize(ctx, slot.stopping_word, false); - probs = std::vector(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size()); - } - else - { + probs = std::vector( - slot.generated_token_probs.begin(), - slot.generated_token_probs.end()); + slot.generated_token_probs.begin(), + slot.generated_token_probs.end() - stop_word_toks.size()); + } else { + probs = std::vector( + slot.generated_token_probs.begin(), + slot.generated_token_probs.end()); } - res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs); + + res.data["completion_probabilities"] = probs_vector_to_json(ctx, probs); } - if (slot.oaicompat) - { - res.result_json["oaicompat_token_ctr"] = slot.n_decoded; - res.result_json["model"] = slot.oaicompat_model; + if (slot.oaicompat) { + res.data["oaicompat_token_ctr"] = slot.n_decoded; + res.data["model"] = slot.oaicompat_model; } queue_results.send(res); } - void send_embedding(server_slot & slot, const llama_batch & batch) - { - task_result res; - res.id = slot.task_id; - res.multitask_id = slot.multitask_id; - res.error = false; - res.stop = true; + void send_embedding(const server_slot & slot, const llama_batch & batch) { + server_task_result res; + res.id = slot.id_task; + res.id_multi = slot.id_multi; + res.error = false; + res.stop = true; const int n_embd = llama_n_embd(model); - if (!params.embedding) - { - LOG_WARNING("embedding disabled", {{"params.embedding", params.embedding}}); - res.result_json = json - { - {"embedding", std::vector(n_embd, 0.0f)}, + for (int i = 0; i < batch.n_tokens; ++i) { + if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) { + continue; + } + + const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]); + if (embd == NULL) { + embd = llama_get_embeddings_ith(ctx, i); + } + + if (embd == NULL) { + LOG_ERROR("failed to get embeddings", { + {"token", batch.token [i]}, + {"seq_id", batch.seq_id[i][0]} + }); + + res.data = json { + {"embedding", std::vector(n_embd, 0.0f)}, + }; + + continue; + } + + res.data = json { + {"embedding", std::vector(embd, embd + n_embd)}, }; } - else - { - for (int i = 0; i < batch.n_tokens; ++i) { - if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) { - continue; - } - const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]); - if (embd == NULL) { - embd = llama_get_embeddings_ith(ctx, i); - if (embd == NULL) { - LOG_ERROR("failed to get embeddings for token", {{"token", batch.token[i]}, {"seq_id", batch.seq_id[i][0]}}); - res.result_json = json - { - {"embedding", std::vector(n_embd, 0.0f)}, - }; - continue; - } - } - - res.result_json = json - { - {"embedding", std::vector(embd, embd + n_embd)}, - }; - } - } queue_results.send(res); } - void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id) - { - task_server task; - task.id = task_id; - task.target_id = 0; - task.data = std::move(data); - task.infill_mode = infill; - task.embedding_mode = embedding; - task.type = TASK_TYPE_COMPLETION; - task.multitask_id = multitask_id; + void request_completion(int id_task, int id_multi, json data, bool infill, bool embedding) { + server_task task; + task.id = id_task; + task.id_multi = id_multi; + task.id_target = 0; + task.data = std::move(data); + task.infill = infill; + task.embedding = embedding; + task.type = SERVER_TASK_TYPE_COMPLETION; // when a completion task's prompt array is not a singleton, we split it into multiple requests // otherwise, it's a single-prompt task, we actually queue it // if there's numbers in the prompt array it will be treated as an array of tokens if (task.data.count("prompt") != 0 && task.data.at("prompt").size() > 1) { bool numbers = false; - for (const auto& e : task.data.at("prompt")) { + for (const auto & e : task.data.at("prompt")) { if (e.is_number()) { numbers = true; break; @@ -1288,106 +1360,23 @@ struct llama_server_context if (numbers) { queue_tasks.post(task); } else { - split_multiprompt_task(task_id, task); + split_multiprompt_task(id_task, task); } } else { - // an empty prompt can make slot become buggy - if (task.data.contains("prompt") && task.data["prompt"].is_string() && task.data["prompt"].get().empty()) { - task.data["prompt"] = " "; // add a space so that we have one token - } queue_tasks.post(task); } } - // for multiple images processing - bool ingest_images(server_slot &slot, int n_batch) - { - int image_idx = 0; + void request_cancel(int id_task) { + server_task task; + task.type = SERVER_TASK_TYPE_CANCEL; + task.id_target = id_task; - while (image_idx < (int) slot.images.size()) - { - slot_image &img = slot.images[image_idx]; - - // process prefix prompt - 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 - }; - if (llama_decode(ctx, batch_view)) - { - LOG_TEE("%s : failed to eval\n", __func__); - return false; - } - } - - // process image with llm - for (int i = 0; i < img.image_tokens; i += n_batch) - { - int n_eval = img.image_tokens - i; - if (n_eval > n_batch) - { - n_eval = n_batch; - } - - const int n_embd = llama_n_embd(model); - llama_batch batch_img = { - n_eval, - nullptr, - (img.image_embedding + i * n_embd), - nullptr, - nullptr, - nullptr, - nullptr, - slot.n_past, - 1, 0 - }; - if (llama_decode(ctx, batch_img)) - { - LOG_TEE("%s : failed to eval image\n", __func__); - return false; - } - slot.n_past += n_eval; - } - image_idx++; - - llama_batch_clear(batch); - - // append prefix of next image - const auto json_prompt = (image_idx >= (int) slot.images.size()) ? - slot.params.input_suffix : // no more images, then process suffix prompt - (json)(slot.images[image_idx].prefix_prompt); - - std::vector append_tokens = tokenize(json_prompt, false); // has next image - for (int i = 0; i < (int) append_tokens.size(); ++i) - { - llama_batch_add(batch, append_tokens[i], system_tokens.size() + slot.n_past, { slot.id }, true); - slot.n_past += 1; - } - } - - return true; - } - - void request_cancel(int task_id) - { - task_server task; - task.type = TASK_TYPE_CANCEL; - task.target_id = task_id; queue_tasks.post(task); } - void split_multiprompt_task(int multitask_id, task_server& multiprompt_task) - { - int prompt_count = multiprompt_task.data.at("prompt").size(); + void split_multiprompt_task(int id_multi, const server_task & multiprompt_task) { + const int prompt_count = multiprompt_task.data.at("prompt").size(); if (prompt_count <= 1) { send_error(multiprompt_task, "error while handling multiple prompts"); return; @@ -1395,127 +1384,121 @@ struct llama_server_context // generate all the ID for subtask std::vector subtask_ids(prompt_count); - for (int i = 0; i < prompt_count; i++) - { + 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); + queue_tasks.add_multitask(id_multi, subtask_ids); // add subtasks - for (int i = 0; i < prompt_count; i++) - { + 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.) - request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id); + request_completion(subtask_ids[i], id_multi, subtask_data, multiprompt_task.infill, multiprompt_task.embedding); } } - void process_single_task(task_server& task) - { - switch (task.type) - { - case TASK_TYPE_COMPLETION: { - server_slot *slot = get_slot(json_value(task.data, "slot_id", -1)); - if (slot == nullptr) + void process_single_task(const server_task & task) { + switch (task.type) { + case SERVER_TASK_TYPE_COMPLETION: { - // if no slot is available, we defer this task for processing later - LOG_VERBOSE("no slot is available", {{"task_id", task.id}}); - queue_tasks.defer(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"); + server_slot * slot = get_slot(json_value(task.data, "id_slot", -1)); + if (slot == nullptr) { + // if no slot is available, we defer this task for processing later + LOG_VERBOSE("no slot is available", {{"id_task", task.id}}); + queue_tasks.defer(task); break; } - system_prompt_process(task.data["system_prompt"]); - // reset cache_tokens for all slots - for (server_slot &slot : slots) - { - slot.cache_tokens.clear(); - slot.n_past = 0; - slot.n_past_se = 0; + if (task.data.contains("system_prompt")) { + system_prompt_set(task.data["system_prompt"]); + + for (server_slot & slot : slots) { + slot.n_past = 0; + slot.n_past_se = 0; + } } - } - slot->reset(); + slot->reset(); - slot->infill = task.infill_mode; - slot->embedding = task.embedding_mode; - slot->task_id = task.id; - slot->multitask_id = task.multitask_id; + slot->id_task = task.id; + slot->id_multi = task.id_multi; + slot->infill = task.infill; + slot->embedding = task.embedding; - 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(); + if (!launch_slot_with_data(*slot, task.data)) { + // send error result + send_error(task, "internal_error"); break; } - } - } break; - case TASK_TYPE_NEXT_RESPONSE: { - // do nothing - } break; - case TASK_TYPE_METRICS: { - json slots_data = json::array(); - int n_idle_slots = 0; - int n_processing_slots = 0; - - for (server_slot &slot: slots) { - json slot_data = get_formated_generation(slot); - slot_data["id"] = slot.id; - slot_data["task_id"] = slot.task_id; - slot_data["state"] = slot.state; - slot_data["prompt"] = slot.prompt; - slot_data["next_token"] = { - {"has_next_token", slot.has_next_token}, - {"n_remain", slot.n_remaining}, - {"num_tokens_predicted", slot.n_decoded}, - {"stopped_eos", slot.stopped_eos}, - {"stopped_word", slot.stopped_word}, - {"stopped_limit", slot.stopped_limit}, - {"stopping_word", slot.stopping_word}, - }; - if (slot_data["state"] == IDLE) { - n_idle_slots++; - } else { - n_processing_slots++; + } break; + case SERVER_TASK_TYPE_CANCEL: + { + // release slot linked with the task id + for (auto & slot : slots) { + if (slot.id_task == task.id_target) { + slot.release(); + break; + } } - slots_data.push_back(slot_data); - } - LOG_INFO("slot data", { - {"task_id", task.id}, - {"n_idle_slots", n_idle_slots}, - {"n_processing_slots", n_processing_slots} - }); - LOG_VERBOSE("slot data", { - {"task_id", task.id}, - {"n_idle_slots", n_idle_slots}, - {"n_processing_slots", n_processing_slots}, - {"slots", slots_data} - }); - task_result res; - res.id = task.id; - res.multitask_id = task.multitask_id; - res.stop = true; - res.error = false; - res.result_json = { + } break; + case SERVER_TASK_TYPE_NEXT_RESPONSE: + { + // do nothing + } break; + case SERVER_TASK_TYPE_METRICS: + { + json slots_data = json::array(); + + int n_idle_slots = 0; + int n_processing_slots = 0; + + for (server_slot & slot : slots) { + json slot_data = get_formated_generation(slot); + slot_data["id"] = slot.id; + slot_data["id_task"] = slot.id_task; + slot_data["state"] = slot.state; + slot_data["prompt"] = slot.prompt; + slot_data["next_token"] = { + {"has_next_token", slot.has_next_token}, + {"n_remain", slot.n_remaining}, + {"n_decoded", slot.n_decoded}, + {"stopped_eos", slot.stopped_eos}, + {"stopped_word", slot.stopped_word}, + {"stopped_limit", slot.stopped_limit}, + {"stopping_word", slot.stopping_word}, + }; + + if (slot_data["state"] == SLOT_STATE_IDLE) { + n_idle_slots++; + } else { + n_processing_slots++; + } + + slots_data.push_back(slot_data); + } + LOG_INFO("slot data", { + {"id_task", task.id}, + {"n_idle_slots", n_idle_slots}, + {"n_processing_slots", n_processing_slots} + }); + + LOG_VERBOSE("slot data", { + {"id_task", task.id}, + {"n_idle_slots", n_idle_slots}, + {"n_processing_slots", n_processing_slots}, + {"slots", slots_data} + }); + + server_task_result res; + res.id = task.id; + res.id_multi = task.id_multi; + res.stop = true; + res.error = false; + res.data = { { "idle", n_idle_slots }, { "processing", n_processing_slots }, { "deferred", queue_tasks.queue_tasks_deferred.size() }, @@ -1532,71 +1515,104 @@ struct llama_server_context { "kv_cache_used_cells", llama_get_kv_cache_used_cells(ctx)}, { "slots", slots_data }, - }; - metrics.reset_bucket(); - queue_results.send(res); - } break; + }; + + metrics.reset_bucket(); + queue_results.send(res); + } break; } } - void on_finish_multitask(task_multi& multitask) - { + void on_finish_multitask(const server_task_multi & multitask) { // all subtasks done == multitask is done - task_result result; - result.id = multitask.id; - result.stop = true; + server_task_result result; + result.id = multitask.id; + result.stop = true; result.error = false; // collect json results into one json result std::vector result_jsons; - for (auto& subres : multitask.results) - { - result_jsons.push_back(subres.result_json); + for (const auto & subres : multitask.results) { + result_jsons.push_back(subres.data); result.error = result.error && subres.error; } - result.result_json = json{ { "results", result_jsons } }; + result.data = json { + { "results", result_jsons } + }; + queue_results.send(result); } bool update_slots() { - if (system_need_update) - { - LOG_INFO("updating system prompt", {}); + if (system_need_update) { system_prompt_update(); } - llama_batch_clear(batch); + // release slots + for (auto & slot : slots) { + if (slot.command == SLOT_COMMAND_RELEASE) { + slot.state = SLOT_STATE_IDLE; + slot.command = SLOT_COMMAND_NONE; + slot.t_last_used = ggml_time_us(); - if (all_slots_are_idle) - { - if (system_prompt.empty() && clean_kv_cache) - { - LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {}); - kv_cache_clear(); + LOG_INFO("slot released", { + {"id_slot", slot.id}, + {"id_task", slot.id_task}, + {"n_ctx", n_ctx}, + {"n_past", slot.n_past}, + {"n_system_tokens", system_tokens.size()}, + {"n_cache_tokens", slot.cache_tokens.size()}, + {"truncated", slot.truncated} + }); + + queue_tasks.notify_slot_changed(); } - return true; } - LOG_VERBOSE("posting NEXT_RESPONSE", {}); - task_server task; - task.type = TASK_TYPE_NEXT_RESPONSE; - task.target_id = -1; - queue_tasks.post(task); - - for (server_slot &slot : slots) + // check if all slots are idle { - if (slot.ga_n == 1) - { - if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx) - { + bool all_idle = true; + + for (auto & slot : slots) { + if (slot.state != SLOT_STATE_IDLE || slot.command != SLOT_COMMAND_NONE) { + all_idle = false; + break; + } + } + + if (all_idle) { + LOG_INFO("all slots are idle", {}); + if (system_prompt.empty() && clean_kv_cache) { + kv_cache_clear(); + } + + return true; + } + } + + { + LOG_VERBOSE("posting NEXT_RESPONSE", {}); + + server_task task; + task.type = SERVER_TASK_TYPE_NEXT_RESPONSE; + task.id_target = -1; + + queue_tasks.post(task); + } + + // apply context-shift if needed + // TODO: simplify and improve + for (server_slot & slot : slots) { + if (slot.ga_n == 1) { + if (slot.is_processing() && (int) system_tokens.size() + slot.n_past >= slot.n_ctx - 1) { // Shift context const int n_keep = slot.params.n_keep + add_bos_token; const int n_left = (int) system_tokens.size() + slot.n_past - n_keep; const int n_discard = n_left / 2; LOG_INFO("slot context shift", { - {"slot_id", slot.id}, - {"task_id", slot.task_id}, + {"id_slot", slot.id}, + {"id_task", slot.id_task}, {"n_keep", n_keep}, {"n_left", n_left}, {"n_discard", n_discard}, @@ -1605,15 +1621,17 @@ struct llama_server_context {"n_system_tokens", system_tokens.size()}, {"n_cache_tokens", slot.cache_tokens.size()} }); + llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard); - for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) - { - slot.cache_tokens[i - n_discard] = slot.cache_tokens[i]; - } + if (slot.params.cache_prompt) { + for (size_t i = n_keep + 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.cache_tokens.resize(slot.cache_tokens.size() - n_discard); + } slot.n_past -= n_discard; @@ -1622,33 +1640,12 @@ struct llama_server_context } } - // decode any currently ongoing sequences - LOG_VERBOSE("decoding ongoing sequences", {}); - for (auto & slot : slots) - { - // release the slot - if (slot.command == RELEASE) - { - slot.state = IDLE; - slot.command = NONE; - slot.t_last_used = ggml_time_us(); + // start populating the batch for this iteration + llama_batch_clear(batch); - LOG_INFO("slot released", { - {"slot_id", slot.id}, - {"task_id", slot.task_id}, - {"n_ctx", n_ctx}, - {"n_past", slot.n_past}, - {"n_system_tokens", system_tokens.size()}, - {"n_cache_tokens", slot.cache_tokens.size()}, - {"truncated", slot.truncated} - }); - queue_tasks.notify_slot_changed(); - - continue; - } - - if (slot.state == IDLE) - { + // frist, add sampled tokens from any ongoing sequences + for (auto & slot : slots) { + if (slot.state == SLOT_STATE_IDLE) { continue; } @@ -1659,193 +1656,184 @@ struct llama_server_context // TODO: we always have to take into account the "system_tokens" // this is not great and needs to be improved somehow llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true); + slot.n_past += 1; + + if (slot.params.cache_prompt) { + slot.cache_tokens.push_back(slot.sampled); + } + + LOG_VERBOSE("slot decode token", { + {"id_slot", slot.id}, + {"id_task", slot.id_task}, + {"n_ctx", n_ctx}, + {"n_past", slot.n_past}, + {"n_system_tokens", system_tokens.size()}, + {"n_cache_tokens", slot.cache_tokens.size()}, + {"truncated", slot.truncated} + }); } // process in chunks of params.n_batch int32_t n_batch = params.n_batch; - // assign workload to the slots - if (params.cont_batching || batch.n_tokens == 0) - { - for (auto & slot : slots) - { - const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get().empty()) || !slot.images.empty(); + // next, batch any pending prompts without exceeding n_batch + if (params.cont_batching || batch.n_tokens == 0) { + for (auto & slot : slots) { + const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get().empty()); // empty prompt passed -> release the slot and send empty response // note: infill mode allows empty prompt - if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill) - { + if (slot.state == SLOT_STATE_IDLE && slot.command == SLOT_COMMAND_LOAD_PROMPT && !has_prompt && !slot.infill) { + slot.state = SLOT_STATE_PROCESSING; + slot.command = SLOT_COMMAND_NONE; slot.release(); slot.print_timings(); send_final_response(slot); continue; } - // need process the prompt - if (slot.state == IDLE && slot.command == LOAD_PROMPT) - { - slot.state = PROCESSING; - slot.command = NONE; - std::vector prompt_tokens; - slot.t_start_process_prompt = ggml_time_us(); - slot.t_start_genereration = 0; + // this slot still has a prompt to be processed + if (slot.state == SLOT_STATE_IDLE && slot.command == SLOT_COMMAND_LOAD_PROMPT) { + auto & prompt_tokens = slot.prompt_tokens; - if (slot.infill) - { - bool suff_rm_leading_spc = true; - if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) - { - params.input_suffix.erase(0, 1); - suff_rm_leading_spc = false; - } - auto prefix_tokens = tokenize(slot.params.input_prefix, false); - auto suffix_tokens = tokenize(slot.params.input_suffix, false); - - const int space_token = 29871; // TODO: this should not be hardcoded - if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) { - suffix_tokens.erase(suffix_tokens.begin()); - } - - prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model)); - prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS - prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model)); - prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); - prefix_tokens.push_back(llama_token_middle(model)); - prompt_tokens = prefix_tokens; - } - else - { - prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt - } - - slot.n_prompt_tokens = prompt_tokens.size(); - - if (slot.params.n_keep < 0) - { - slot.params.n_keep = slot.n_prompt_tokens; - } - slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); - - // if input prompt is too big, truncate it, if group attention self-extend is disabled - if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx) - { - const int n_left = slot.n_ctx - slot.params.n_keep; - const int n_block_size = n_left / 2; - const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; - - std::vector new_tokens( - prompt_tokens.begin(), - prompt_tokens.begin() + slot.params.n_keep); - new_tokens.insert( - new_tokens.end(), - prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, - prompt_tokens.end()); - - LOG_VERBOSE("input truncated", { - {"n_ctx", slot.n_ctx}, - {"n_keep", slot.params.n_keep}, - {"n_left", n_left}, - {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, + // we haven't tokenized the prompt yet - do it now: + if (prompt_tokens.empty()) { + LOG_VERBOSE("tokenizing prompt", { + {"id_slot", slot.id}, + {"id_task", slot.id_task} }); - slot.truncated = true; - prompt_tokens = new_tokens; - slot.n_prompt_tokens = prompt_tokens.size(); - GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx); - } + slot.t_start_process_prompt = ggml_time_us(); + slot.t_start_generation = 0; - if (!slot.params.cache_prompt) - { - llama_sampling_reset(slot.ctx_sampling); - - slot.n_past = 0; - slot.n_past_se = 0; - slot.ga_i = 0; - slot.n_prompt_tokens_processed = slot.n_prompt_tokens; - } - else - { - // push the prompt into the sampling context (do not apply grammar) - for (auto &token : prompt_tokens) - { - llama_sampling_accept(slot.ctx_sampling, ctx, token, false); - } - - slot.n_past = common_part(slot.cache_tokens, prompt_tokens); - - // the last token of the cache is not in the KV cache until the next call to llama_decode - // (it was sampled, pushed into the "cache_tokens", but not yet put in the context) - if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size()) - { - slot.n_past -= 1; - } - - slot.n_prompt_tokens_processed = slot.n_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++; + if (slot.infill) { + bool suff_rm_leading_spc = true; + if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) { + params.input_suffix.erase(0, 1); + suff_rm_leading_spc = false; } - slot.n_past_se = slot_npast; - slot.ga_i = ga_i; + + auto prefix_tokens = tokenize(slot.params.input_prefix, false); + auto suffix_tokens = tokenize(slot.params.input_suffix, false); + + const int space_token = 29871; // TODO: this should not be hardcoded + if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) { + suffix_tokens.erase(suffix_tokens.begin()); + } + + prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model)); + prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS + prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model)); + prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end()); + prefix_tokens.push_back(llama_token_middle(model)); + prompt_tokens = prefix_tokens; + } else { + prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt } - LOG_INFO("slot progression", { - { "slot_id", slot.id }, - { "task_id", slot.task_id }, - { "n_past", slot.n_past }, - { "n_past_se", slot.n_past_se }, - { "ga_i", slot.ga_i }, - { "n_prompt_tokens_processed", slot.n_prompt_tokens_processed } - }); + slot.n_past = 0; + slot.n_prompt_tokens = prompt_tokens.size(); + + if (slot.embedding) { + // this prompt is too large to process - discard it + if (slot.n_prompt_tokens > n_batch) { + slot.state = SLOT_STATE_PROCESSING; + slot.command = SLOT_COMMAND_NONE; + slot.release(); + slot.print_timings(); + send_final_response(slot); + continue; + } + } else { + if (slot.params.n_keep < 0) { + slot.params.n_keep = slot.n_prompt_tokens; + } + slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); + + // if input prompt is too big, truncate it (if group attention self-extend is disabled) + if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx) { + const int n_left = slot.n_ctx - slot.params.n_keep; + + const int n_block_size = n_left / 2; + const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + + std::vector new_tokens( + prompt_tokens.begin(), + prompt_tokens.begin() + slot.params.n_keep); + + new_tokens.insert( + new_tokens.end(), + prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, + prompt_tokens.end()); + + prompt_tokens = std::move(new_tokens); + + slot.truncated = true; + slot.n_prompt_tokens = prompt_tokens.size(); + + LOG_VERBOSE("input truncated", { + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"prompt_tokens", tokens_to_str(ctx, prompt_tokens.cbegin(), prompt_tokens.cend())}, + }); + + GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx); + } + + llama_sampling_reset(slot.ctx_sampling); + + if (!slot.params.cache_prompt) { + slot.n_past_se = 0; + slot.ga_i = 0; + } else { + GGML_ASSERT(slot.ga_n == 1); + + // reuse any previously computed tokens that are common with the new prompt + slot.n_past = common_part(slot.cache_tokens, prompt_tokens); + + // remove the non-common part from the cache + slot.cache_tokens.resize(slot.n_past); + + // push the prompt into the sampling context (do not apply grammar) + for (int i = 0; i < slot.n_past; ++i) { + llama_sampling_accept(slot.ctx_sampling, ctx, slot.cache_tokens[i], false); + } + } + } + + if (slot.n_past == slot.n_prompt_tokens && slot.n_past > 0) { + // we have to evaluate at least 1 token to generate logits. + LOG_INFO("we have to evaluate at least 1 token to generate logits", { + { "id_slot", slot.id }, + { "id_task", slot.id_task } + }); + + slot.n_past--; + if (slot.ga_i > 0) { + slot.n_past_se--; + } + } + + slot.n_prompt_tokens_processed = 0; } - slot.cache_tokens = prompt_tokens; - - if (slot.n_past == slot.n_prompt_tokens && slot.n_past > 0) - { - // we have to evaluate at least 1 token to generate logits. - LOG_INFO("we have to evaluate at least 1 token to generate logits", { - { "slot_id", slot.id }, - { "task_id", slot.task_id } - }); - slot.n_past--; - if (slot.ga_i > 0) - { - slot.n_past_se--; + if (slot.embedding) { + // cannot fit the prompt in the current batch - will try next iter + if (batch.n_tokens + slot.n_prompt_tokens > n_batch) { + continue; } } - int p0 = (int) system_tokens.size() + slot.n_past; - LOG_INFO("kv cache rm [p0, end)", { - { "slot_id", slot.id }, - { "task_id", slot.task_id }, - { "p0", p0 } - }); + const int p0 = (int) system_tokens.size() + slot.n_past; llama_kv_cache_seq_rm(ctx, slot.id, p0, -1); - LOG_VERBOSE("prompt ingested", { - {"n_past", slot.n_past}, - {"cached", tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)}, - {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())}, - }); - - const bool has_images = process_images(slot); - - // process the prefix of first image - std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens; + LOG_INFO("kv cache rm [p0, end)", { + { "id_slot", slot.id }, + { "id_task", slot.id_task }, + { "p0", p0 } + }); int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past; @@ -1853,61 +1841,82 @@ struct llama_server_context 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) - { - if (slot.ga_n != 1) - { + // add prompt tokens for processing in the current batch + // TODO: the self-extend stuff here is a mess - simplify and/or abstract it somehow + for (; slot.n_past < slot.n_prompt_tokens && batch.n_tokens < n_batch; ++slot.n_past) { + 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); + + llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id }, false); + + if (slot.params.cache_prompt) { + slot.cache_tokens.push_back(prompt_tokens[slot.n_past]); + } + + slot.n_prompt_tokens_processed++; slot_npast++; } - if (has_images && !ingest_images(slot, n_batch)) - { - LOG_ERROR("failed processing images", { - {"slot_id", slot.id}, - {"task_id", slot.task_id}, - }); - // FIXME @phymbert: to be properly tested - // early returning without changing the slot state will block the slot for ever - // no one at the moment is checking the return value - return false; - } + LOG_VERBOSE("prompt processing progress", { + {"id_slot", slot.id}, + {"n_past", slot.n_past}, + {"n_ctx", n_ctx}, + {"n_tokens", batch.n_tokens}, + {"progress", (float) slot.n_prompt_tokens_processed / slot.n_prompt_tokens}, + }); - // extract the logits only for the last token - if (batch.n_tokens > 0) - { + // entire prompt has been processed - start decoding new tokens + if (slot.n_past == slot.n_prompt_tokens) { + slot.state = SLOT_STATE_PROCESSING; + slot.command = SLOT_COMMAND_NONE; + + GGML_ASSERT(batch.n_tokens > 0); + + // extract the logits only for the last token batch.logits[batch.n_tokens - 1] = true; - } - slot.n_decoded = 0; - slot.i_batch = batch.n_tokens - 1; + slot.n_decoded = 0; + slot.i_batch = batch.n_tokens - 1; + + LOG_VERBOSE("prompt done", { + {"id_slot", slot.id}, + {"n_past", slot.n_past}, + {"n_ctx", n_ctx}, + {"n_tokens", batch.n_tokens}, + }); + } + } + + if (batch.n_tokens >= n_batch) { + break; } } } - if (batch.n_tokens == 0) - { - all_slots_are_idle = true; + if (batch.n_tokens == 0) { + LOG_VERBOSE("no tokens to decode", {}); + return true; } - for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) - { + LOG_VERBOSE("decoding batch", { + {"n_tokens", batch.n_tokens}, + }); + + // process the created batch of tokens + for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) { const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i); - for (auto & slot : slots) - { - if (slot.ga_n != 1) - { + 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) - { + // TODO: simplify and/or abstract this + 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; @@ -1918,8 +1927,8 @@ struct llama_server_context 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_add(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_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd); + 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_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd); slot.n_past_se -= bd; @@ -1927,12 +1936,12 @@ struct llama_server_context 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 = - { + llama_batch batch_view = { n_tokens, batch.token + i, nullptr, @@ -1945,10 +1954,8 @@ struct llama_server_context const int ret = llama_decode(ctx, batch_view); - if (ret != 0) - { - if (n_batch == 1 || ret < 0) - { + if (ret != 0) { + if (n_batch == 1 || ret < 0) { // if you get here, it means the KV cache is full - try increasing it via the context size LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret); return false; @@ -1959,19 +1966,17 @@ struct llama_server_context // retry with half the batch size to try to find a free slot in the KV cache n_batch /= 2; i -= n_batch; + continue; } - for (auto & slot : slots) - { - if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) - { + for (auto & slot : slots) { + if (slot.state != SLOT_STATE_PROCESSING || slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens)) { continue; } // prompt evaluated for embedding - if (slot.embedding) - { + if (slot.embedding) { send_embedding(slot, batch_view); slot.release(); slot.i_batch = -1; @@ -1984,10 +1989,9 @@ struct llama_server_context llama_sampling_accept(slot.ctx_sampling, ctx, id, true); slot.n_decoded += 1; - if (slot.n_decoded == 1) - { - slot.t_start_genereration = ggml_time_us(); - slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3; + if (slot.n_decoded == 1) { + slot.t_start_generation = ggml_time_us(); + slot.t_prompt_processing = (slot.t_start_generation - slot.t_start_process_prompt) / 1e3; metrics.on_prompt_eval(slot); } @@ -1995,19 +1999,19 @@ struct llama_server_context result.tok = id; const int32_t n_probs = slot.sparams.n_probs; - if (slot.sparams.temp <= 0 && n_probs > 0) - { + if (slot.sparams.temp <= 0 && n_probs > 0) { // for llama_sample_token_greedy we need to sort candidates llama_sample_softmax(ctx, &cur_p); } - for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i) - { - result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p}); + for (size_t i = 0; i < std::min(cur_p.size, (size_t) n_probs); ++i) { + result.probs.push_back({ + cur_p.data[i].id, + cur_p.data[i].p + }); } - if (!process_token(result, slot)) - { + if (!process_token(result, slot)) { slot.release(); slot.print_timings(); send_final_response(slot); @@ -2019,24 +2023,23 @@ struct llama_server_context } LOG_VERBOSE("slots updated", {}); + return true; } - json model_meta() { - return json{ - {"vocab_type", llama_vocab_type(model)}, - {"n_vocab", llama_n_vocab(model)}, - {"n_ctx_train", llama_n_ctx_train(model)}, - {"n_embd", llama_n_embd(model)}, - {"n_params", llama_model_n_params(model)}, - {"size", llama_model_size(model)}, + json model_meta() const { + return json { + {"vocab_type", llama_vocab_type (model)}, + {"n_vocab", llama_n_vocab (model)}, + {"n_ctx_train", llama_n_ctx_train (model)}, + {"n_embd", llama_n_embd (model)}, + {"n_params", llama_model_n_params(model)}, + {"size", llama_model_size (model)}, }; } }; -static void server_print_usage(const char *argv0, const gpt_params ¶ms, - const server_params &sparams) -{ +static void server_print_usage(const char * argv0, const gpt_params & params, const server_params & sparams) { printf("usage: %s [options]\n", argv0); printf("\n"); printf("options:\n"); @@ -2054,17 +2057,14 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n"); printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow); printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast); - printf(" --pooling {none,mean,cls}\n"); - printf(" pooling type for embeddings, use model default if unspecified\n"); + printf(" --pooling {none,mean,cls} pooling type for embeddings, use model default if unspecified\n"); printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n"); printf(" not recommended: doubles context memory required and no measurable increase in quality\n"); - if (llama_supports_mlock()) - { + if (llama_supports_mlock()) { printf(" --mlock force system to keep model in RAM rather than swapping or compressing\n"); } - if (llama_supports_mmap()) - { + if (llama_supports_mmap()) { printf(" --no-mmap do not memory-map model (slower load but may reduce pageouts if not using mlock)\n"); } printf(" --numa TYPE attempt optimizations that help on some NUMA systems\n"); @@ -2096,7 +2096,7 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" --api-key API_KEY optional api key to enhance server security. If set, requests must include this key for access.\n"); printf(" --api-key-file FNAME path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n"); printf(" -to N, --timeout N server read/write timeout in seconds (default: %d)\n", sparams.read_timeout); - printf(" --embedding enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); + printf(" --embeddings enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n"); printf(" -spf FNAME, --system-prompt-file FNAME\n"); @@ -2105,7 +2105,6 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf(" KV cache data type for K (default: f16)\n"); printf(" -ctv TYPE, --cache-type-v TYPE\n"); printf(" KV cache data type for V (default: f16)\n"); - printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA.\n"); printf(" --log-format log output format: json or text (default: json)\n"); printf(" --log-disable disables logging to a file.\n"); printf(" --slots-endpoint-disable disables slots monitoring endpoint.\n"); @@ -2123,57 +2122,41 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf("\n"); } -static void server_params_parse(int argc, char **argv, server_params &sparams, - gpt_params ¶ms, llama_server_context& llama) -{ - gpt_params default_params; +static void server_params_parse(int argc, char ** argv, server_params & sparams, gpt_params & params) { + gpt_params default_params; server_params default_sparams; + std::string arg; bool invalid_param = false; - for (int i = 1; i < argc; i++) - { + for (int i = 1; i < argc; i++) { arg = argv[i]; - if (arg == "--port") - { - if (++i >= argc) - { + if (arg == "--port") { + if (++i >= argc) { invalid_param = true; break; } sparams.port = std::stoi(argv[i]); - } - else if (arg == "--host") - { - if (++i >= argc) - { + } else if (arg == "--host") { + if (++i >= argc) { invalid_param = true; break; } sparams.hostname = argv[i]; - } - else if (arg == "--path") - { - if (++i >= argc) - { + } else if (arg == "--path") { + if (++i >= argc) { invalid_param = true; break; } sparams.public_path = argv[i]; - } - else if (arg == "--api-key") - { - if (++i >= argc) - { + } else if (arg == "--api-key") { + if (++i >= argc) { invalid_param = true; break; } sparams.api_keys.emplace_back(argv[i]); - } - else if (arg == "--api-key-file") - { - if (++i >= argc) - { + } else if (arg == "--api-key-file") { + if (++i >= argc) { invalid_param = true; break; } @@ -2190,53 +2173,36 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } } key_file.close(); - } - else if (arg == "--timeout" || arg == "-to") - { - if (++i >= argc) - { + } else if (arg == "--timeout" || arg == "-to") { + if (++i >= argc) { invalid_param = true; break; } sparams.read_timeout = std::stoi(argv[i]); sparams.write_timeout = std::stoi(argv[i]); - } - else if (arg == "-m" || arg == "--model") - { - if (++i >= argc) - { + } else if (arg == "-m" || arg == "--model") { + if (++i >= argc) { invalid_param = true; break; } params.model = argv[i]; - } - else if (arg == "-a" || arg == "--alias") - { - if (++i >= argc) - { + } else if (arg == "-a" || arg == "--alias") { + if (++i >= argc) { invalid_param = true; break; } params.model_alias = argv[i]; - } - else if (arg == "-h" || arg == "--help") - { + } else if (arg == "-h" || arg == "--help") { server_print_usage(argv[0], default_params, default_sparams); exit(0); - } - else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") - { - if (++i >= argc) - { + } else if (arg == "-c" || arg == "--ctx-size" || arg == "--ctx_size") { + if (++i >= argc) { invalid_param = true; break; } params.n_ctx = std::stoi(argv[i]); - } - else if (arg == "--rope-scaling") - { - if (++i >= argc) - { + } else if (arg == "--rope-scaling") { + if (++i >= argc) { invalid_param = true; break; } @@ -2245,59 +2211,44 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; } else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; } else { invalid_param = true; break; } - } - else if (arg == "--rope-freq-base") - { - if (++i >= argc) - { + } else if (arg == "--rope-freq-base") { + if (++i >= argc) { invalid_param = true; break; } params.rope_freq_base = std::stof(argv[i]); - } - else if (arg == "--rope-freq-scale") - { - if (++i >= argc) - { + } else if (arg == "--rope-freq-scale") { + if (++i >= argc) { invalid_param = true; break; } params.rope_freq_scale = std::stof(argv[i]); - } - else if (arg == "--yarn-ext-factor") - { + } else if (arg == "--yarn-ext-factor") { if (++i >= argc) { invalid_param = true; break; } params.yarn_ext_factor = std::stof(argv[i]); } - else if (arg == "--yarn-attn-factor") - { + else if (arg == "--yarn-attn-factor") { if (++i >= argc) { invalid_param = true; break; } params.yarn_attn_factor = std::stof(argv[i]); - } - else if (arg == "--yarn-beta-fast") - { + } else if (arg == "--yarn-beta-fast") { if (++i >= argc) { invalid_param = true; break; } params.yarn_beta_fast = std::stof(argv[i]); - } - else if (arg == "--yarn-beta-slow") - { + } else if (arg == "--yarn-beta-slow") { if (++i >= argc) { invalid_param = true; break; } params.yarn_beta_slow = std::stof(argv[i]); - } - else if (arg == "--pooling") - { + } else if (arg == "--pooling") { if (++i >= argc) { invalid_param = true; break; @@ -2307,108 +2258,79 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; } else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; } else { invalid_param = true; break; } - } - else if (arg == "--threads" || arg == "-t") - { + } else if (arg == "--threads" || arg == "-t") { if (++i >= argc) { invalid_param = true; break; } params.n_threads = std::stoi(argv[i]); - } - else if (arg == "--grp-attn-n" || arg == "-gan") - { + } 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) - { + } 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) - { + } else if (arg == "--threads-batch" || arg == "-tb") { + if (++i >= argc) { invalid_param = true; break; } params.n_threads_batch = std::stoi(argv[i]); - } - else if (arg == "--threads-http") - { - if (++i >= argc) - { + } else if (arg == "--threads-http") { + if (++i >= argc) { invalid_param = true; break; } sparams.n_threads_http = std::stoi(argv[i]); - } - else if (arg == "-b" || arg == "--batch-size") - { - if (++i >= argc) - { + } else if (arg == "-b" || arg == "--batch-size") { + if (++i >= argc) { invalid_param = true; break; } params.n_batch = std::stoi(argv[i]); - } - else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") - { - if (++i >= argc) - { + } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { invalid_param = true; break; } if (llama_supports_gpu_offload()) { params.n_gpu_layers = std::stoi(argv[i]); } else { - LOG_WARNING("Not compiled with GPU offload support, --n-gpu-layers option will be ignored. " - "See main README.md for information on enabling GPU BLAS support", - {{"n_gpu_layers", params.n_gpu_layers}}); + LOG_WARNING( + "Not compiled with GPU offload support, --n-gpu-layers option will be ignored. " + "See main README.md for information on enabling GPU BLAS support", + {{"n_gpu_layers", params.n_gpu_layers}}); } - } - else if (arg == "--split-mode" || arg == "-sm") - { + } else if (arg == "--split-mode" || arg == "-sm") { if (++i >= argc) { invalid_param = true; break; } std::string arg_next = argv[i]; - if (arg_next == "none") - { + if (arg_next == "none") { params.split_mode = LLAMA_SPLIT_MODE_NONE; - } - else if (arg_next == "layer") - { + } else if (arg_next == "layer") { params.split_mode = LLAMA_SPLIT_MODE_LAYER; - } - else if (arg_next == "row") - { + } else if (arg_next == "row") { params.split_mode = LLAMA_SPLIT_MODE_ROW; - } - else { + } else { 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 - } - else if (arg == "--tensor-split" || arg == "-ts") - { - if (++i >= argc) - { + } else if (arg == "--tensor-split" || arg == "-ts") { + if (++i >= argc) { invalid_param = true; break; } @@ -2421,25 +2343,18 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, std::vector split_arg{it, {}}; GGML_ASSERT(split_arg.size() <= llama_max_devices()); - for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device) - { - if (i_device < split_arg.size()) - { + for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device) { + if (i_device < split_arg.size()) { params.tensor_split[i_device] = std::stof(split_arg[i_device]); - } - else - { + } else { params.tensor_split[i_device] = 0.0f; } } #else LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n", {}); #endif // GGML_USE_CUBLAS - } - else if (arg == "--main-gpu" || arg == "-mg") - { - if (++i >= argc) - { + } else if (arg == "--main-gpu" || arg == "-mg") { + if (++i >= argc) { invalid_param = true; break; } @@ -2448,98 +2363,70 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, #else LOG_WARNING("llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.", {}); #endif - } - else if (arg == "--lora") - { - if (++i >= argc) - { + } else if (arg == "--lora") { + if (++i >= argc) { invalid_param = true; break; } params.lora_adapter.emplace_back(argv[i], 1.0f); params.use_mmap = false; - } - else if (arg == "--lora-scaled") - { - if (++i >= argc) - { + } else if (arg == "--lora-scaled") { + if (++i >= argc) { invalid_param = true; break; } const char * lora_adapter = argv[i]; - if (++i >= argc) - { + if (++i >= argc) { invalid_param = true; break; } params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i])); params.use_mmap = false; - } - else if (arg == "--lora-base") - { - if (++i >= argc) - { + } else if (arg == "--lora-base") { + if (++i >= argc) { invalid_param = true; break; } params.lora_base = argv[i]; - } - else if (arg == "-v" || arg == "--verbose") - { + } else if (arg == "-v" || arg == "--verbose") { #if SERVER_VERBOSE != 1 LOG_WARNING("server.cpp is not built with verbose logging.", {}); #else server_verbose = true; #endif - } - else if (arg == "--mlock") - { + } else if (arg == "--mlock") { params.use_mlock = true; - } - else if (arg == "--no-mmap") - { + } else if (arg == "--no-mmap") { params.use_mmap = false; - } - else if (arg == "--numa") { + } else if (arg == "--numa") { if (++i >= argc) { invalid_param = true; break; } else { std::string value(argv[i]); /**/ if (value == "distribute" || value == "" ) { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; } - else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } - else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } + else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; } + else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; } else { invalid_param = true; break; } } - } - else if (arg == "--embedding") - { + } else if (arg == "--embedding" || arg == "--embeddings") { params.embedding = true; - } - else if (arg == "-cb" || arg == "--cont-batching") - { + } else if (arg == "-cb" || arg == "--cont-batching") { params.cont_batching = true; - } - else if (arg == "-np" || arg == "--parallel") - { - if (++i >= argc) - { + } else if (arg == "-np" || arg == "--parallel") { + if (++i >= argc) { invalid_param = true; break; } params.n_parallel = std::stoi(argv[i]); - } else if (arg == "-n" || arg == "--n-predict") - { - if (++i >= argc) - { + } else if (arg == "-n" || arg == "--n-predict") { + if (++i >= argc) { invalid_param = true; break; } params.n_predict = std::stoi(argv[i]); - } else if (arg == "-spf" || arg == "--system-prompt-file") - { - if (++i >= argc) - { + } else if (arg == "-spf" || arg == "--system-prompt-file") { + if (++i >= argc) { invalid_param = true; break; } @@ -2549,67 +2436,39 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, invalid_param = true; break; } - std::string systm_content; + std::string system_prompt; std::copy( std::istreambuf_iterator(file), std::istreambuf_iterator(), - std::back_inserter(systm_content) + std::back_inserter(system_prompt) ); - llama.system_prompt_process(json::parse(systm_content)); - } - else if (arg == "-ctk" || arg == "--cache-type-k") { + sparams.system_prompt = system_prompt; + } else if (arg == "-ctk" || arg == "--cache-type-k") { params.cache_type_k = argv[++i]; - } - else if (arg == "-ctv" || arg == "--cache-type-v") { + } else if (arg == "-ctv" || arg == "--cache-type-v") { params.cache_type_v = argv[++i]; - } - else if(arg == "--mmproj") - { - if (++i >= argc) - { + } else if (arg == "--log-format") { + if (++i >= argc) { invalid_param = true; break; } - params.mmproj = argv[i]; - } - else if (arg == "--log-format") - { - if (++i >= argc) - { - invalid_param = true; - break; - } - if (std::strcmp(argv[i], "json") == 0) - { + if (std::strcmp(argv[i], "json") == 0) { server_log_json = true; - } - else if (std::strcmp(argv[i], "text") == 0) - { + } else if (std::strcmp(argv[i], "text") == 0) { server_log_json = false; - } - else - { + } else { invalid_param = true; break; } - } - else if (arg == "--log-disable") - { + } else if (arg == "--log-disable") { log_set_target(stdout); LOG_INFO("logging to file is disabled.", {}); - } - else if (arg == "--slots-endpoint-disable") - { + } else if (arg == "--slots-endpoint-disable") { sparams.slots_endpoint = false; - } - else if (arg == "--metrics") - { + } else if (arg == "--metrics") { sparams.metrics_endpoint = true; - } - else if (arg == "--chat-template") - { - if (++i >= argc) - { + } else if (arg == "--chat-template") { + if (++i >= argc) { invalid_param = true; break; } @@ -2620,9 +2479,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, break; } sparams.chat_template = argv[i]; - } - else if (arg == "--override-kv") - { + } else if (arg == "--override-kv") { if (++i >= argc) { invalid_param = true; break; @@ -2633,6 +2490,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, invalid_param = true; break; } + struct llama_model_kv_override kvo; std::strncpy(kvo.key, argv[i], sep - argv[i]); kvo.key[sep - argv[i]] = 0; @@ -2663,67 +2521,28 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, break; } params.kv_overrides.push_back(kvo); - } - else - { + } else { fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); server_print_usage(argv[0], default_params, default_sparams); exit(1); } } + if (!params.kv_overrides.empty()) { params.kv_overrides.emplace_back(); params.kv_overrides.back().key[0] = 0; } - if (invalid_param) - { + if (invalid_param) { fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); server_print_usage(argv[0], default_params, default_sparams); exit(1); } } -/* llama.cpp completion api semantics */ -static json format_partial_response( - llama_server_context &llama, server_slot *slot, const std::string &content, const std::vector &probs -) { - json res = json - { - {"content", content }, - {"stop", false}, - {"slot_id", slot->id }, - {"multimodal", llama.multimodal } - }; - - if (slot->sparams.n_probs > 0) - { - res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs); - } - - return res; -} - -static json format_tokenizer_response(const std::vector &tokens) -{ - return json { - {"tokens", tokens} - }; -} - -static json format_detokenized_response(std::string content) -{ - return json { - {"content", content} - }; -} - - -static void log_server_request(const httplib::Request &req, const httplib::Response &res) -{ +static void log_server_request(const httplib::Request & req, const httplib::Response & res) { // skip GH copilot requests when using default port - if (req.path == "/v1/health" || req.path == "/v1/completions") - { + if (req.path == "/v1/health" || req.path == "/v1/completions") { return; } @@ -2742,24 +2561,9 @@ static void log_server_request(const httplib::Request &req, const httplib::Respo }); } -static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama, server_slot *slot) -{ - auto & gtps = slot->generated_token_probs; - auto translator = token_translator{llama.ctx}; - auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); }; - const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen); - if (slot->generated_text.capacity() < slot->generated_text.size() + len) - { - slot->generated_text.reserve(slot->generated_text.size() + len); - } - for (const completion_token_output & cto : gtps) - { - slot->generated_text += translator(cto); - } -} - std::function shutdown_handler; std::atomic_flag is_terminating = ATOMIC_FLAG_INIT; + inline void signal_handler(int signal) { if (is_terminating.test_and_set()) { // in case it hangs, we can force terminate the server by hitting Ctrl+C twice @@ -2767,40 +2571,45 @@ inline void signal_handler(int signal) { fprintf(stderr, "Received second interrupt, terminating immediately.\n"); exit(1); } + shutdown_handler(signal); } -int main(int argc, char **argv) -{ +int main(int argc, char ** argv) { #if SERVER_VERBOSE != 1 log_disable(); #endif // own arguments required by this example - gpt_params params; + gpt_params params; server_params sparams; // struct that contains llama context and inference - llama_server_context llama; + server_context ctx_server; - server_params_parse(argc, argv, sparams, params, llama); + server_params_parse(argc, argv, sparams, params); - if (params.model_alias == "unknown") - { + if (!sparams.system_prompt.empty()) { + ctx_server.system_prompt_set(json::parse(sparams.system_prompt)); + } + + if (params.model_alias == "unknown") { params.model_alias = params.model; } llama_backend_init(); llama_numa_init(params.numa); - LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER}, - {"commit", LLAMA_COMMIT}}); + LOG_INFO("build info", { + {"build", LLAMA_BUILD_NUMBER}, + {"commit", LLAMA_COMMIT} + }); LOG_INFO("system info", { - {"n_threads", params.n_threads}, - {"n_threads_batch", params.n_threads_batch}, - {"total_threads", std::thread::hardware_concurrency()}, - {"system_info", llama_print_system_info()}, - }); + {"n_threads", params.n_threads}, + {"n_threads_batch", params.n_threads_batch}, + {"total_threads", std::thread::hardware_concurrency()}, + {"system_info", llama_print_system_info()}, + }); httplib::Server svr; @@ -2809,152 +2618,163 @@ int main(int argc, char **argv) svr.set_default_headers({{"Server", "llama.cpp"}}); // CORS preflight - svr.Options(R"(.*)", [](const httplib::Request &req, httplib::Response &res) { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + svr.Options(R"(.*)", [](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); res.set_header("Access-Control-Allow-Credentials", "true"); - res.set_header("Access-Control-Allow-Methods", "POST"); - res.set_header("Access-Control-Allow-Headers", "*"); + res.set_header("Access-Control-Allow-Methods", "POST"); + res.set_header("Access-Control-Allow-Headers", "*"); }); - svr.Get("/health", [&](const httplib::Request& req, httplib::Response& res) { + svr.Get("/health", [&](const httplib::Request & req, httplib::Response & res) { server_state current_state = state.load(); - switch(current_state) { - case SERVER_STATE_READY: { - // request slots data using task queue - task_server task; - task.id = llama.queue_tasks.get_new_id(); - task.type = TASK_TYPE_METRICS; - task.target_id = -1; + switch (current_state) { + case SERVER_STATE_READY: + { + // request slots data using task queue + server_task task; + task.id = ctx_server.queue_tasks.get_new_id(); + task.type = SERVER_TASK_TYPE_METRICS; + task.id_target = -1; - llama.queue_results.add_waiting_task_id(task.id); - llama.queue_tasks.post(task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - // get the result - task_result result = llama.queue_results.recv(task.id); - llama.queue_results.remove_waiting_task_id(task.id); + // get the result + server_task_result result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - int n_idle_slots = result.result_json["idle"]; - int n_processing_slots = result.result_json["processing"]; + const int n_idle_slots = result.data["idle"]; + const int n_processing_slots = result.data["processing"]; - json health = { + json health = { {"status", "ok"}, {"slots_idle", n_idle_slots}, - {"slots_processing", n_processing_slots}}; - res.status = 200; // HTTP OK - if (sparams.slots_endpoint && req.has_param("include_slots")) { - health["slots"] = result.result_json["slots"]; - } + {"slots_processing", n_processing_slots} + }; - if (n_idle_slots == 0) { - health["status"] = "no slot available"; - if (req.has_param("fail_on_no_slot")) { - res.status = 503; // HTTP Service Unavailable + res.status = 200; // HTTP OK + if (sparams.slots_endpoint && req.has_param("include_slots")) { + health["slots"] = result.data["slots"]; } + + if (n_idle_slots == 0) { + health["status"] = "no slot available"; + if (req.has_param("fail_on_no_slot")) { + res.status = 503; // HTTP Service Unavailable + } + } + + res.set_content(health.dump(), "application/json"); + break; } - res.set_content(health.dump(), "application/json"); - break; - } case SERVER_STATE_LOADING_MODEL: - res.set_content(R"({"status": "loading model"})", "application/json"); - res.status = 503; // HTTP Service Unavailable - break; + { + res.set_content(R"({"status": "loading model"})", "application/json"); + res.status = 503; // HTTP Service Unavailable + } break; case SERVER_STATE_ERROR: - res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json"); - res.status = 500; // HTTP Internal Server Error - break; + { + res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json"); + res.status = 500; // HTTP Internal Server Error + } break; } }); if (sparams.slots_endpoint) { - svr.Get("/slots", [&](const httplib::Request&, httplib::Response& res) { + svr.Get("/slots", [&](const httplib::Request &, httplib::Response & res) { // request slots data using task queue - task_server task; - task.id = llama.queue_tasks.get_new_id(); - task.type = TASK_TYPE_METRICS; - task.target_id = -1; + server_task task; + task.id = ctx_server.queue_tasks.get_new_id(); + task.id_multi = -1; + task.id_target = -1; + task.type = SERVER_TASK_TYPE_METRICS; - llama.queue_results.add_waiting_task_id(task.id); - llama.queue_tasks.post(task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); // get the result - task_result result = llama.queue_results.recv(task.id); - llama.queue_results.remove_waiting_task_id(task.id); + server_task_result result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - res.set_content(result.result_json["slots"].dump(), "application/json"); + res.set_content(result.data["slots"].dump(), "application/json"); res.status = 200; // HTTP OK }); } if (sparams.metrics_endpoint) { - svr.Get("/metrics", [&](const httplib::Request&, httplib::Response& res) { + svr.Get("/metrics", [&](const httplib::Request &, httplib::Response & res) { // request slots data using task queue - task_server task; - task.id = llama.queue_tasks.get_new_id(); - task.type = TASK_TYPE_METRICS; - task.target_id = -1; + server_task task; + task.id = ctx_server.queue_tasks.get_new_id(); + task.id_multi = -1; + task.id_target = -1; + task.type = SERVER_TASK_TYPE_METRICS; - llama.queue_results.add_waiting_task_id(task.id); - llama.queue_tasks.post(task); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); // get the result - task_result result = llama.queue_results.recv(task.id); - llama.queue_results.remove_waiting_task_id(task.id); + server_task_result result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - json data = result.result_json; + json data = result.data; - uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"]; - uint64_t t_prompt_processing = data["t_prompt_processing"]; + const uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"]; + const uint64_t t_prompt_processing = data["t_prompt_processing"]; - uint64_t n_tokens_predicted = data["n_tokens_predicted"]; - uint64_t t_tokens_generation = data["t_tokens_generation"]; + const uint64_t n_tokens_predicted = data["n_tokens_predicted"]; + const uint64_t t_tokens_generation = data["t_tokens_generation"]; - int32_t kv_cache_used_cells = data["kv_cache_used_cells"]; + const int32_t kv_cache_used_cells = data["kv_cache_used_cells"]; // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names json all_metrics_def = json { - {"counter", {{ - {"name", "prompt_tokens_total"}, - {"help", "Number of prompt tokens processed."}, - {"value", data["n_prompt_tokens_processed_total"]} - }, { - {"name", "tokens_predicted_total"}, - {"help", "Number of generation tokens processed."}, - {"value", data["n_tokens_predicted_total"]} - }}}, - {"gauge", {{ - {"name", "prompt_tokens_seconds"}, - {"help", "Average prompt throughput in tokens/s."}, - {"value", n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0} - },{ - {"name", "predicted_tokens_seconds"}, - {"help", "Average generation throughput in tokens/s."}, - {"value", n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0} - },{ - {"name", "kv_cache_usage_ratio"}, - {"help", "KV-cache usage. 1 means 100 percent usage."}, - {"value", 1. * kv_cache_used_cells / params.n_ctx} - },{ - {"name", "kv_cache_tokens"}, - {"help", "KV-cache tokens."}, - {"value", data["kv_cache_tokens_count"]} - },{ - {"name", "requests_processing"}, - {"help", "Number of request processing."}, - {"value", data["processing"]} - },{ - {"name", "requests_deferred"}, - {"help", "Number of request deferred."}, - {"value", data["deferred"]} - }}} + {"counter", {{ + {"name", "prompt_tokens_total"}, + {"help", "Number of prompt tokens processed."}, + {"value", data["n_prompt_tokens_processed_total"]} + }, { + {"name", "tokens_predicted_total"}, + {"help", "Number of generation tokens processed."}, + {"value", data["n_tokens_predicted_total"]} + }}}, + {"gauge", {{ + {"name", "prompt_tokens_seconds"}, + {"help", "Average prompt throughput in tokens/s."}, + {"value", n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0} + },{ + {"name", "predicted_tokens_seconds"}, + {"help", "Average generation throughput in tokens/s."}, + {"value", n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0} + },{ + {"name", "kv_cache_usage_ratio"}, + {"help", "KV-cache usage. 1 means 100 percent usage."}, + {"value", 1. * kv_cache_used_cells / params.n_ctx} + },{ + {"name", "kv_cache_tokens"}, + {"help", "KV-cache tokens."}, + {"value", data["kv_cache_tokens_count"]} + },{ + {"name", "requests_processing"}, + {"help", "Number of request processing."}, + {"value", data["processing"]} + },{ + {"name", "requests_deferred"}, + {"help", "Number of request deferred."}, + {"value", data["deferred"]} + }}} }; std::stringstream prometheus; - for (const auto& el : all_metrics_def.items()) { - const auto& type = el.key(); - const auto& metrics_def = el.value(); - for (const auto& metric_def : metrics_def) { - std::string name = metric_def["name"]; - std::string help = metric_def["help"]; + + for (const auto & el : all_metrics_def.items()) { + const auto & type = el.key(); + const auto & metrics_def = el.value(); + + for (const auto & metric_def : metrics_def) { + const std::string name = metric_def["name"]; + const std::string help = metric_def["help"]; + auto value = json_value(metric_def, "value", 0); prometheus << "# HELP llamacpp:" << name << " " << help << "\n" << "# TYPE llamacpp:" << name << " " << type << "\n" @@ -2969,49 +2789,39 @@ int main(int argc, char **argv) svr.set_logger(log_server_request); - svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep) - { - const char fmt[] = "500 Internal Server Error\n%s"; - char buf[BUFSIZ]; - try - { - std::rethrow_exception(std::move(ep)); - } - catch (std::exception &e) - { - snprintf(buf, sizeof(buf), fmt, e.what()); - } - catch (...) - { - snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); - } - res.set_content(buf, "text/plain; charset=utf-8"); - res.status = 500; - }); + svr.set_exception_handler([](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { + const char fmt[] = "500 Internal Server Error\n%s"; - svr.set_error_handler([](const httplib::Request &, httplib::Response &res) - { - if (res.status == 401) - { - res.set_content("Unauthorized", "text/plain; charset=utf-8"); - } - if (res.status == 400) - { - res.set_content("Invalid request", "text/plain; charset=utf-8"); - } - else if (res.status == 404) - { - res.set_content("File Not Found", "text/plain; charset=utf-8"); - res.status = 404; - } - }); + char buf[BUFSIZ]; + try { + std::rethrow_exception(std::move(ep)); + } catch (std::exception &e) { + snprintf(buf, sizeof(buf), fmt, e.what()); + } catch (...) { + snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); + } + + res.set_content(buf, "text/plain; charset=utf-8"); + res.status = 500; + }); + + svr.set_error_handler([](const httplib::Request &, httplib::Response & res) { + if (res.status == 401) { + res.set_content("Unauthorized", "text/plain; charset=utf-8"); + } + if (res.status == 400) { + res.set_content("Invalid request", "text/plain; charset=utf-8"); + } + if (res.status == 404) { + res.set_content("File Not Found", "text/plain; charset=utf-8"); + } + }); // set timeouts and change hostname and port svr.set_read_timeout (sparams.read_timeout); svr.set_write_timeout(sparams.write_timeout); - if (!svr.bind_to_port(sparams.hostname, sparams.port)) - { + if (!svr.bind_to_port(sparams.hostname, sparams.port)) { fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); return 1; } @@ -3020,8 +2830,9 @@ int main(int argc, char **argv) svr.set_base_dir(sparams.public_path); std::unordered_map log_data; + log_data["hostname"] = sparams.hostname; - log_data["port"] = std::to_string(sparams.port); + log_data["port"] = std::to_string(sparams.port); if (sparams.api_keys.size() == 1) { log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4); @@ -3030,20 +2841,23 @@ int main(int argc, char **argv) } // load the model - if (!llama.load_model(params)) - { + if (!ctx_server.load_model(params)) { state.store(SERVER_STATE_ERROR); return 1; } else { - llama.initialize(); + ctx_server.initialize(); state.store(SERVER_STATE_READY); - LOG_INFO("model loaded", {}); } - const auto model_meta = llama.model_meta(); + + LOG_INFO("model loaded", {}); + + const auto model_meta = ctx_server.model_meta(); if (sparams.chat_template.empty()) { // custom chat template is not supplied - // check if the template comes with the model is supported by us - llama.validate_model_chat_template(sparams); + if (!ctx_server.validate_model_chat_template()) { + LOG_ERROR("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); + sparams.chat_template = "chatml"; + } } // Middleware for API key validation @@ -3055,6 +2869,7 @@ int main(int argc, char **argv) // Check for API key in the header auto auth_header = req.get_header_value("Authorization"); + std::string prefix = "Bearer "; if (auth_header.substr(0, prefix.size()) == prefix) { std::string received_api_key = auth_header.substr(prefix.size()); @@ -3073,179 +2888,173 @@ int main(int argc, char **argv) }; // this is only called if no index.html is found in the public --path - svr.Get("/", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html; charset=utf-8"); - return false; - }); + svr.Get("/", [](const httplib::Request &, httplib::Response & res) { + res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html; charset=utf-8"); + return false; + }); // this is only called if no index.js is found in the public --path - svr.Get("/index.js", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript; charset=utf-8"); - return false; - }); + svr.Get("/index.js", [](const httplib::Request &, httplib::Response & res) { + res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript; charset=utf-8"); + return false; + }); // this is only called if no index.html is found in the public --path - svr.Get("/completion.js", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript; charset=utf-8"); - return false; - }); + svr.Get("/completion.js", [](const httplib::Request &, httplib::Response & res) { + res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript; charset=utf-8"); + return false; + }); // this is only called if no index.html is found in the public --path - svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response &res) - { - res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8"); - return false; - }); + svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response & res) { + res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8"); + return false; + }); - svr.Get("/props", [&llama](const httplib::Request & req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - json data = { - { "user_name", llama.name_user.c_str() }, - { "assistant_name", llama.name_assistant.c_str() }, - { "default_generation_settings", llama.default_generation_settings_for_props }, - { "total_slots", llama.params.n_parallel } - }; - res.set_content(data.dump(), "application/json; charset=utf-8"); - }); + svr.Get("/props", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + json data = { + { "user_name", ctx_server.name_user.c_str() }, + { "assistant_name", ctx_server.name_assistant.c_str() }, + { "default_generation_settings", ctx_server.default_generation_settings_for_props }, + { "total_slots", ctx_server.params.n_parallel } + }; - svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!validate_api_key(req, res)) { - return; - } - json data = json::parse(req.body); - 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.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"); - } - else - { - res.status = 404; - res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - } - 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.queue_results.recv(task_id); - if (!result.error) { - const std::string str = - "data: " + - result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + - "\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; - } - if (result.stop) { - break; - } - } else { - const std::string str = - "error: " + - result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + - "\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; - } - } + res.set_content(data.dump(), "application/json; charset=utf-8"); + }); - llama.queue_results.remove_waiting_task_id(task_id); - sink.done(); - return true; - }; - - auto on_complete = [task_id, &llama] (bool) - { - // 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); - } - }); - - svr.Get("/v1/models", [¶ms, &model_meta](const httplib::Request& req, httplib::Response& res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - std::time_t t = std::time(0); - - json models = { - {"object", "list"}, - {"data", { - { - {"id", params.model_alias}, - {"object", "model"}, - {"created", t}, - {"owned_by", "llamacpp"}, - {"meta", model_meta} - }, - }} - }; - - res.set_content(models.dump(), "application/json; charset=utf-8"); - }); - - const auto chat_completions = [&llama, &validate_api_key, &sparams](const httplib::Request &req, httplib::Response &res) - { + svr.Post("/completion", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; } - json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template); - 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); + json data = json::parse(req.body); + + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, data, false, false); if (!json_value(data, "stream", false)) { - std::string completion_text; - task_result result = llama.queue_results.recv(task_id); - + server_task_result result = ctx_server.queue_results.recv(id_task); if (!result.error && result.stop) { - json oaicompat_result = format_final_response_oaicompat(data, result); - - res.set_content(oaicompat_result.dump(-1, ' ', false, - json::error_handler_t::replace), - "application/json; charset=utf-8"); + res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); } else { res.status = 500; - res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); + res.set_content(result.data["content"], "text/plain; charset=utf-8"); } - 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.queue_results.recv(task_id); - if (!llama_result.error) { - std::vector result_array = format_partial_response_oaicompat( llama_result); - for (auto it = result_array.begin(); it != result_array.end(); ++it) - { + ctx_server.queue_results.remove_waiting_task_id(id_task); + } else { + const auto chunked_content_provider = [id_task, &ctx_server](size_t, httplib::DataSink & sink) { + while (true) { + server_task_result result = ctx_server.queue_results.recv(id_task); + if (!result.error) { + const std::string str = + "data: " + + result.data.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + + if (!sink.write(str.c_str(), str.size())) { + ctx_server.queue_results.remove_waiting_task_id(id_task); + return false; + } + + if (result.stop) { + break; + } + } else { + const std::string str = + "error: " + + result.data.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + + if (!sink.write(str.c_str(), str.size())) { + ctx_server.queue_results.remove_waiting_task_id(id_task); + return false; + } + + break; + } + } + + ctx_server.queue_results.remove_waiting_task_id(id_task); + sink.done(); + + return true; + }; + + auto on_complete = [id_task, &ctx_server] (bool) { + // cancel + ctx_server.request_cancel(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); + }; + + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); + } + }); + + svr.Get("/v1/models", [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + + json models = { + {"object", "list"}, + {"data", { + { + {"id", params.model_alias}, + {"object", "model"}, + {"created", std::time(0)}, + {"owned_by", "llamacpp"}, + {"meta", model_meta} + }, + }} + }; + + res.set_content(models.dump(), "application/json; charset=utf-8"); + }); + + const auto chat_completions = [&ctx_server, &validate_api_key, &sparams](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + if (!validate_api_key(req, res)) { + return; + } + + json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), sparams.chat_template); + + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, data, false, false); + + if (!json_value(data, "stream", false)) { + server_task_result result = ctx_server.queue_results.recv(id_task); + + if (!result.error && result.stop) { + json result_oai = format_final_response_oaicompat(data, result.data); + + res.set_content(result_oai.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); + } else { + res.status = 500; + res.set_content(result.data["content"], "text/plain; charset=utf-8"); + } + ctx_server.queue_results.remove_waiting_task_id(id_task); + } else { + const auto chunked_content_provider = [id_task, &ctx_server](size_t, httplib::DataSink & sink) { + while (true) { + server_task_result result = ctx_server.queue_results.recv(id_task); + if (!result.error) { + std::vector result_array = format_partial_response_oaicompat(result.data); + + for (auto it = result_array.begin(); it != result_array.end(); ++it) { if (!it->empty()) { const std::string str = "data: " + @@ -3253,251 +3062,235 @@ 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); + ctx_server.queue_results.remove_waiting_task_id(id_task); return false; } } } - if (llama_result.stop) { + if (result.stop) { break; } } else { const std::string str = "error: " + - llama_result.result_json.dump(-1, ' ', false, - json::error_handler_t::replace) + + result.data.dump(-1, ' ', false, json::error_handler_t::replace) + "\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); + ctx_server.queue_results.remove_waiting_task_id(id_task); return false; } break; } } sink.done(); - llama.queue_results.remove_waiting_task_id(task_id); + ctx_server.queue_results.remove_waiting_task_id(id_task); return true; }; - auto on_complete = [task_id, &llama](bool) { + auto on_complete = [id_task, &ctx_server](bool) { // cancel request - llama.request_cancel(task_id); - llama.queue_results.remove_waiting_task_id(task_id); + ctx_server.request_cancel(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); }; res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); } }; - svr.Post("/chat/completions", chat_completions); + svr.Post("/chat/completions", chat_completions); svr.Post("/v1/chat/completions", chat_completions); - svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!validate_api_key(req, res)) { - return; - } - json data = json::parse(req.body); - 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.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"); - } - else - { - res.status = 404; - res.set_content(result.result_json["content"], "text/plain; charset=utf-8"); - } - 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.queue_results.recv(task_id); - if (!result.error) { - const std::string str = - "data: " + - result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + - "\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; - } - if (result.stop) - { - break; - } - } - else - { - break; - } - } - - llama.queue_results.remove_waiting_task_id(task_id); - sink.done(); - return true; - }; - - auto on_complete = [task_id, &llama] (bool) - { - // cancel - llama.request_cancel(task_id); - }; - - res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); - } - }); - - svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response &res) - { return res.set_content("", "application/json; charset=utf-8"); }); - - svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - const json body = json::parse(req.body); - std::vector tokens; - if (body.count("content") != 0) - { - tokens = llama.tokenize(body["content"], false); - } - const json data = format_tokenizer_response(tokens); - return res.set_content(data.dump(), "application/json; charset=utf-8"); - }); - - svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - const json body = json::parse(req.body); - std::string content; - if (body.count("tokens") != 0) - { - const std::vector tokens = body["tokens"]; - content = tokens_to_str(llama.ctx, tokens.cbegin(), tokens.cend()); - } - - const json data = format_detokenized_response(content); - return res.set_content(data.dump(), "application/json; charset=utf-8"); - }); - - svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - const json body = json::parse(req.body); - json prompt; - if (body.count("content") != 0) - { - prompt = body["content"]; - } - else - { - prompt = ""; - } - - json image_data; - if (body.count("image_data") != 0) { - image_data = body["image_data"]; - } - else - { - image_data = ""; - } - - // 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"); - }); - - svr.Post("/v1/embeddings", [&llama](const httplib::Request &req, httplib::Response &res) - { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - const json body = json::parse(req.body); - - json prompt; - if (body.count("input") != 0) - { - prompt = body["input"]; - // batch - if(prompt.is_array()) { - json data = json::array(); - int i = 0; - for (const json &elem : prompt) { - 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", elem}, { "n_predict", 0} }, false, true, -1); - - // get the result - task_result result = llama.queue_results.recv(task_id); - llama.queue_results.remove_waiting_task_id(task_id); - - json embedding = json{ - {"embedding", json_value(result.result_json, "embedding", json::array())}, - {"index", i++}, - {"object", "embedding"} - }; - data.push_back(embedding); - } - json result = format_embeddings_response_oaicompat(body, data); - return res.set_content(result.dump(), "application/json; charset=utf-8"); - } - } - else - { - prompt = ""; - } - - // 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}}, false, true, -1); - - // get the result - task_result result = llama.queue_results.recv(task_id); - llama.queue_results.remove_waiting_task_id(task_id); - - json data = json::array({json{ - {"embedding", json_value(result.result_json, "embedding", json::array())}, - {"index", 0}, - {"object", "embedding"} - }} - ); - - json root = format_embeddings_response_oaicompat(body, data); - - // send the result - return res.set_content(root.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(); + svr.Post("/infill", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + if (!validate_api_key(req, res)) { + return; } - }*/ - //); + + json data = json::parse(req.body); + + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, data, true, false); + + if (!json_value(data, "stream", false)) { + server_task_result result = ctx_server.queue_results.recv(id_task); + if (!result.error && result.stop) { + res.set_content(result.data.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8"); + } else { + res.status = 404; + res.set_content(result.data["content"], "text/plain; charset=utf-8"); + } + + ctx_server.queue_results.remove_waiting_task_id(id_task); + } else { + const auto chunked_content_provider = [id_task, &ctx_server](size_t, httplib::DataSink & sink) { + while (true) { + server_task_result result = ctx_server.queue_results.recv(id_task); + if (!result.error) { + const std::string str = + "data: " + + result.data.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + + if (!sink.write(str.c_str(), str.size())) { + ctx_server.queue_results.remove_waiting_task_id(id_task); + return false; + } + + if (result.stop) { + break; + } + } else { + break; + } + } + + ctx_server.queue_results.remove_waiting_task_id(id_task); + sink.done(); + + return true; + }; + + auto on_complete = [id_task, &ctx_server] (bool) { + ctx_server.request_cancel(id_task); + }; + + res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); + } + }); + + svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) { + return res.set_content("", "application/json; charset=utf-8"); + }); + + svr.Post("/tokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + const json body = json::parse(req.body); + + std::vector tokens; + if (body.count("content") != 0) { + tokens = ctx_server.tokenize(body["content"], false); + } + const json data = format_tokenizer_response(tokens); + return res.set_content(data.dump(), "application/json; charset=utf-8"); + }); + + svr.Post("/detokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + const json body = json::parse(req.body); + + std::string content; + if (body.count("tokens") != 0) { + const std::vector tokens = body["tokens"]; + content = tokens_to_str(ctx_server.ctx, tokens.cbegin(), tokens.cend()); + } + + const json data = format_detokenized_response(content); + return res.set_content(data.dump(), "application/json; charset=utf-8"); + }); + + svr.Post("/embedding", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + if (!params.embedding) { + res.status = 501; + res.set_content("This server does not support embeddings. Start it with `--embeddings`", "text/plain; charset=utf-8"); + return; + } + + const json body = json::parse(req.body); + + json prompt; + if (body.count("content") != 0) { + prompt = body["content"]; + } else { + prompt = ""; + } + + // create and queue the task + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0} }, false, true); + + // get the result + server_task_result result = ctx_server.queue_results.recv(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); + + // send the result + return res.set_content(result.data.dump(), "application/json; charset=utf-8"); + }); + + svr.Post("/v1/embeddings", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + if (!params.embedding) { + res.status = 501; + res.set_content("This server does not support embeddings. Start it with `--embeddings`", "text/plain; charset=utf-8"); + return; + } + + const json body = json::parse(req.body); + + json prompt; + if (body.count("input") != 0) { + prompt = body["input"]; + if (prompt.is_array()) { + json data = json::array(); + + int i = 0; + for (const json & elem : prompt) { + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, { {"prompt", elem}, { "n_predict", 0} }, false, true); + + // get the result + server_task_result result = ctx_server.queue_results.recv(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); + + json embedding = json{ + {"embedding", json_value(result.data, "embedding", json::array())}, + {"index", i++}, + {"object", "embedding"} + }; + + data.push_back(embedding); + } + + json result = format_embeddings_response_oaicompat(body, data); + + return res.set_content(result.dump(), "application/json; charset=utf-8"); + } + } else { + prompt = ""; + } + + // create and queue the task + const int id_task = ctx_server.queue_tasks.get_new_id(); + + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0}}, false, true); + + // get the result + server_task_result result = ctx_server.queue_results.recv(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); + + json data = json::array({json{ + {"embedding", json_value(result.data, "embedding", json::array())}, + {"index", 0}, + {"object", "embedding"} + }} + ); + + json root = format_embeddings_response_oaicompat(body, data); + + return res.set_content(root.dump(), "application/json; charset=utf-8"); + }); if (sparams.n_threads_http < 1) { // +2 threads for monitoring endpoints @@ -3507,34 +3300,33 @@ int main(int argc, char **argv) svr.new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); }; LOG_INFO("HTTP server listening", log_data); + // run the HTTP server in a thread - see comment below - std::thread t([&]() - { - if (!svr.listen_after_bind()) - { - state.store(SERVER_STATE_ERROR); - return 1; - } + std::thread t([&]() { + if (!svr.listen_after_bind()) { + state.store(SERVER_STATE_ERROR); + return 1; + } - return 0; - }); + return 0; + }); - 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_run_slots(std::bind( - &llama_server_context::update_slots, &llama)); - llama.queue_results.on_multitask_update(std::bind( - &llama_server_queue::update_multitask, - &llama.queue_tasks, + ctx_server.queue_tasks.on_new_task(std::bind( + &server_context::process_single_task, &ctx_server, std::placeholders::_1)); + ctx_server.queue_tasks.on_finish_multitask(std::bind( + &server_context::on_finish_multitask, &ctx_server, std::placeholders::_1)); + ctx_server.queue_tasks.on_run_slots(std::bind( + &server_context::update_slots, &ctx_server)); + ctx_server.queue_results.on_multitask_update(std::bind( + &server_queue::update_multitask, + &ctx_server.queue_tasks, std::placeholders::_1, std::placeholders::_2, std::placeholders::_3 )); shutdown_handler = [&](int) { - llama.queue_tasks.terminate(); + ctx_server.queue_tasks.terminate(); }; #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) @@ -3549,10 +3341,13 @@ int main(int argc, char **argv) }; SetConsoleCtrlHandler(reinterpret_cast(console_ctrl_handler), true); #endif - llama.queue_tasks.start_loop(); + + ctx_server.queue_tasks.start_loop(); + svr.stop(); t.join(); llama_backend_free(); + return 0; } diff --git a/examples/server/tests/features/embeddings.feature b/examples/server/tests/features/embeddings.feature new file mode 100644 index 000000000..b47661e94 --- /dev/null +++ b/examples/server/tests/features/embeddings.feature @@ -0,0 +1,94 @@ +@llama.cpp +@embeddings +Feature: llama.cpp server + + Background: Server startup + Given a server listening on localhost:8080 + And a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models + And a model alias bert-bge-small + And 42 as server seed + And 2 slots + And 1024 as batch size + And 2048 KV cache size + And embeddings extraction + Then the server is starting + Then the server is healthy + + Scenario: Embedding + When embeddings are computed for: + """ + What is the capital of Bulgaria ? + """ + Then embeddings are generated + + Scenario: OAI Embeddings compatibility + Given a model bert-bge-small + When an OAI compatible embeddings computation request for: + """ + What is the capital of Spain ? + """ + Then embeddings are generated + + Scenario: OAI Embeddings compatibility with multiple inputs + Given a model bert-bge-small + Given a prompt: + """ + In which country Paris is located ? + """ + And a prompt: + """ + Is Madrid the capital of Spain ? + """ + When an OAI compatible embeddings computation request for multiple inputs + Then embeddings are generated + + Scenario: Multi users embeddings + Given a prompt: + """ + Write a very long story about AI. + """ + And a prompt: + """ + Write another very long music lyrics. + """ + And a prompt: + """ + Write a very long poem. + """ + And a prompt: + """ + Write a very long joke. + """ + Given concurrent embedding requests + Then the server is busy + Then the server is idle + Then all embeddings are generated + + Scenario: Multi users OAI compatibility embeddings + Given a prompt: + """ + In which country Paris is located ? + """ + And a prompt: + """ + Is Madrid the capital of Spain ? + """ + And a prompt: + """ + What is the biggest US city ? + """ + And a prompt: + """ + What is the capital of Bulgaria ? + """ + And a model bert-bge-small + Given concurrent OAI embedding requests + Then the server is busy + Then the server is idle + Then all embeddings are generated + + Scenario: All embeddings should be the same + Given 10 fixed prompts + And a model bert-bge-small + Given concurrent OAI embedding requests + Then all embeddings are the same diff --git a/examples/server/tests/features/parallel.feature b/examples/server/tests/features/parallel.feature index 86cdf7282..066698c8e 100644 --- a/examples/server/tests/features/parallel.feature +++ b/examples/server/tests/features/parallel.feature @@ -9,7 +9,6 @@ Feature: Parallel And 512 as batch size And 64 KV cache size And 2 slots - And embeddings extraction And continuous batching Then the server is starting Then the server is healthy @@ -99,48 +98,3 @@ Feature: Parallel Then the server is busy Then the server is idle Then all prompts are predicted - - Scenario: Multi users embeddings - Given a prompt: - """ - Write a very long story about AI. - """ - And a prompt: - """ - Write another very long music lyrics. - """ - And a prompt: - """ - Write a very long poem. - """ - And a prompt: - """ - Write a very long joke. - """ - Given concurrent embedding requests - Then the server is busy - Then the server is idle - Then all embeddings are generated - - Scenario: Multi users OAI compatibility embeddings - Given a prompt: - """ - In which country Paris is located ? - """ - And a prompt: - """ - Is Madrid the capital of Spain ? - """ - And a prompt: - """ - What is the biggest US city ? - """ - And a prompt: - """ - What is the capital of Bulgaria ? - """ - And a model tinyllama-2 - Given concurrent OAI embedding requests - Then the server is busy - Then the server is idle - Then all embeddings are generated diff --git a/examples/server/tests/features/server.feature b/examples/server/tests/features/server.feature index 7c977bcce..f3b758c79 100644 --- a/examples/server/tests/features/server.feature +++ b/examples/server/tests/features/server.feature @@ -49,34 +49,6 @@ Feature: llama.cpp server | llama-2 | Book | What is the best book | 8 | (Mom\|what)+ | 8 | disabled | | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64 | (thanks\|happy\|bird)+ | 32 | enabled | - Scenario: Embedding - When embeddings are computed for: - """ - What is the capital of Bulgaria ? - """ - Then embeddings are generated - - Scenario: OAI Embeddings compatibility - Given a model tinyllama-2 - When an OAI compatible embeddings computation request for: - """ - What is the capital of Spain ? - """ - Then embeddings are generated - - Scenario: OAI Embeddings compatibility with multiple inputs - Given a model tinyllama-2 - Given a prompt: - """ - In which country Paris is located ? - """ - And a prompt: - """ - Is Madrid the capital of Spain ? - """ - When an OAI compatible embeddings computation request for multiple inputs - Then embeddings are generated - Scenario: Tokenize / Detokenize When tokenizing: """ diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index 319527802..a0b2ffdfe 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -10,6 +10,7 @@ from contextlib import closing from re import RegexFlag import aiohttp +import numpy as np import openai from behave import step from behave.api.async_step import async_run_until_complete @@ -24,6 +25,9 @@ def step_server_config(context, server_fqdn, server_port): if 'PORT' in os.environ: context.server_port = int(os.environ['PORT']) print(f"$PORT set, overriding server port with to {context.server_port}") + if 'FQDN' in os.environ: + context.server_fqdn = os.environ['FQDN'] + print(f"$FQDN set, overriding server fqdn with to {context.server_fqdn}") context.base_url = f'http://{context.server_fqdn}:{context.server_port}' @@ -34,6 +38,7 @@ def step_server_config(context, server_fqdn, server_port): context.n_ga_w = None context.n_gpu_layer = None context.n_predict = None + context.n_prompts = 0 context.n_server_predict = None context.n_slots = None context.prompt_prefix = None @@ -202,6 +207,7 @@ def step_n_tokens_predicted(context, predicted_n): @step(u'a user prompt {user_prompt}') def step_user_prompt(context, user_prompt): context.prompts.append(user_prompt) + context.n_prompts = len(context.prompts) @step(u'a system prompt {system_prompt}') @@ -290,6 +296,12 @@ def step_prompt_passkey(context): context.prompt_passkey = context.text +@step(u'{n_prompts:d} fixed prompts') +def step_fixed_prompts(context, n_prompts): + context.prompts.extend([str(0)*(context.n_batch if context.n_batch is not None else 512) for i in range(n_prompts)]) + context.n_prompts = n_prompts + + @step(u'a "{passkey}" passkey challenge prompt with the passkey inserted every {i_pos:d} junk') def step_prompt_passkey(context, passkey, i_pos): prompt = "" @@ -301,6 +313,7 @@ def step_prompt_passkey(context, passkey, i_pos): passkey_highlight = "\x1b[33m" + passkey + "\x1b[0m" print(f"Passkey challenge:\n```{prompt.replace(passkey, passkey_highlight)}```\n") context.prompts.append(context.prompt_prefix + prompt + context.prompt_suffix) + context.n_prompts = len(context.prompts) @step(u'an OAI compatible chat completions request with {api_error} api error') @@ -341,11 +354,13 @@ async def step_oai_chat_completions(context, api_error): @step(u'a prompt') def step_a_prompt(context): context.prompts.append(context.text) + context.n_prompts = len(context.prompts) @step(u'a prompt {prompt}') def step_a_prompt_prompt(context, prompt): context.prompts.append(prompt) + context.n_prompts = len(context.prompts) @step(u'concurrent completion requests') @@ -430,25 +445,47 @@ async def all_prompts_are_predicted(context, expected_predicted_n=None): @step(u'embeddings are computed for') @async_run_until_complete async def step_compute_embedding(context): + context.n_prompts = 1 context.embeddings = await request_embedding(context.text, base_url=context.base_url) +@step(u'all embeddings are the same') +@async_run_until_complete +async def step_all_embeddings_are_the_same(context): + n_embedding_requests = await gather_tasks_results(context) + assert n_embedding_requests > 0 + embeddings = [] + for i in range(n_embedding_requests): + embedding = context.tasks_result.pop().pop() + embeddings.append(embedding) + assert_embeddings(embedding) + n = len(embeddings) + for i in range(n-1): + for j in range(i+1, n): + embedding1 = np.array(embeddings[i]) + embedding2 = np.array(embeddings[j]) + if context.debug: + print(f"embedding1: {embedding1[-8:]}\n") + print(f"embedding2: {embedding2[-8:]}\n") + similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2)) + msg = f"Similarity between {i} and {j}: {similarity:.10f}" + if context.debug: + print(f"{msg}\n") + assert np.isclose(similarity, 1.0, rtol=1e-05, atol=1e-08, equal_nan=False), msg + @step(u'embeddings are generated') def step_assert_embeddings(context): - if len(context.prompts) == 0: - assert_embeddings(context.embeddings) - else: - assert len(context.embeddings) == len(context.prompts), (f"unexpected response:\n" - f"context.prompts={context.prompts}\n" - f"context.embeddings={context.embeddings}") - for embedding in context.embeddings: - context.prompts.pop() - assert_embeddings(embedding) + assert context.n_prompts == len(context.embeddings), (f"unexpected response:\n" + f"context.n_prompts={context.n_prompts}\n" + f"context.embeddings={context.embeddings}") + for embedding in context.embeddings: + assert_embeddings(embedding) @step(u'an OAI compatible embeddings computation request for') @async_run_until_complete async def step_oai_compute_embeddings(context): + context.n_prompts = 1 context.embeddings = await request_oai_embeddings(context.text, base_url=context.base_url, user_api_key=context.user_api_key, @@ -462,6 +499,7 @@ async def step_oai_compute_embeddings_multiple_inputs(context): base_url=context.base_url, user_api_key=context.user_api_key, model=context.model) + context.prompts.clear() @step(u'concurrent embedding requests') @@ -488,9 +526,9 @@ async def step_concurrent_oai_embedding_requests(context): @async_run_until_complete() async def all_embeddings_are_generated(context): n_embedding_requests = await gather_tasks_results(context) - assert n_embedding_requests > 0 + assert n_embedding_requests == context.n_prompts for i in range(n_embedding_requests): - assert_embeddings(context.tasks_result.pop()) + assert_embeddings(context.tasks_result.pop().pop()) @step(u'tokenizing') @@ -588,11 +626,11 @@ def step_supported_models(context, i_model, param, preposition, param_value): async def concurrent_requests(context, f_completion, *args, **kwargs): - n_prompts = len(context.prompts) + context.n_prompts = len(context.prompts) if context.debug: - print(f"starting {n_prompts} concurrent completion requests...") - assert n_prompts > 0 - for prompt_no in range(n_prompts): + print(f"starting {context.n_prompts} concurrent completion requests...") + assert context.n_prompts > 0 + for prompt_no in range(context.n_prompts): shifted_args = [context.prompts.pop(), *args] context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs))) await asyncio.sleep(0.1) @@ -765,7 +803,7 @@ async def request_embedding(content, base_url=None): }) as response: assert response.status == 200 response_json = await response.json() - return response_json['embedding'] + return [response_json['embedding']] async def request_oai_embeddings(input, @@ -775,6 +813,7 @@ async def request_oai_embeddings(input, user_api_key = user_api_key if user_api_key is not None else 'nope' if async_client: origin = 'llama.cpp' + headers=[] if user_api_key is not None: headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin} async with aiohttp.ClientSession() as session: @@ -783,14 +822,21 @@ async def request_oai_embeddings(input, "input": input, "model": model, }, - headers=headers) as response: + headers=headers, + timeout=3600) as response: assert response.status == 200, f"received status code not expected: {response.status}" assert response.headers['Access-Control-Allow-Origin'] == origin assert response.headers['Content-Type'] == "application/json; charset=utf-8" response_json = await response.json() assert response_json['model'] == model, f"invalid model received: {response_json['model']}" assert response_json['object'] == 'list' - return response_json['data'] + if isinstance(input, collections.abc.Sequence): + embeddings = [] + for an_oai_embeddings in response_json['data']: + embeddings.append(an_oai_embeddings['embedding']) + else: + embeddings = [response_json['data']['embedding']] + return embeddings else: openai.api_key = user_api_key openai.api_base = f'{base_url}/v1' @@ -804,7 +850,7 @@ async def request_oai_embeddings(input, for an_oai_embeddings in oai_embeddings.data: embeddings.append(an_oai_embeddings.embedding) else: - embeddings = oai_embeddings.data.embedding + embeddings = [oai_embeddings.data.embedding] return embeddings @@ -899,6 +945,8 @@ def assert_embeddings(embeddings): assert len(embeddings) > 0 embeddings_computed = False for emb in embeddings: + if not isinstance(emb, float): + assert False, f"Bad embeddings: {embeddings}" if emb != 0: embeddings_computed = True assert embeddings_computed, f"Embeddings: {embeddings}" diff --git a/examples/server/tests/requirements.txt b/examples/server/tests/requirements.txt index 5d4210164..2e4f42ad2 100644 --- a/examples/server/tests/requirements.txt +++ b/examples/server/tests/requirements.txt @@ -1,5 +1,6 @@ aiohttp~=3.9.3 behave~=1.2.6 huggingface_hub~=0.20.3 +numpy~=1.24.4 openai~=0.25.0 prometheus-client~=0.20.0 diff --git a/examples/server/utils.hpp b/examples/server/utils.hpp index b6e49d8b9..df0a27782 100644 --- a/examples/server/utils.hpp +++ b/examples/server/utils.hpp @@ -1,15 +1,16 @@ #pragma once -#include -#include -#include -#include -#include -#include +#include "llama.h" +#include "common.h" #include "json.hpp" -#include "../llava/clip.h" +#include +#include +#include +#include + +#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613" using json = nlohmann::json; @@ -37,83 +38,35 @@ extern bool server_log_json; #define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__) #define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__) -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, - TASK_TYPE_METRICS -}; - -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{}; -}; - -// 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; -}; - -struct token_translator { - llama_context * ctx; - std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); } - std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); } -}; +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; +} static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) { std::stringstream ss_tid; ss_tid << std::this_thread::get_id(); json log = nlohmann::ordered_json{ - {"tid", ss_tid.str()}, + {"tid", ss_tid.str()}, {"timestamp", time(nullptr)}, }; if (server_log_json) { - log.merge_patch( - { - {"level", level}, - {"function", function}, - {"line", line}, - {"msg", message}, - }); + log.merge_patch( { + {"level", level}, + {"function", function}, + {"line", line}, + {"msg", message}, + }); + if (!extra.empty()) { log.merge_patch(extra); } - std::cout << log.dump(-1, ' ', false, json::error_handler_t::replace) << "\n" << std::flush; + printf("%s\n", log.dump(-1, ' ', false, json::error_handler_t::replace).c_str()); } else { char buf[1024]; snprintf(buf, 1024, "%4s [%24s] %s", level, function, message); @@ -136,22 +89,13 @@ static inline void server_log(const char *level, const char *function, int line, } // -// server utils +// chat template 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; -} - // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid inline bool verify_custom_template(const std::string & tmpl) { llama_chat_message chat[] = {{"user", "test"}}; - std::vector buf(1); - int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size()); + int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0); return res >= 0; } @@ -163,7 +107,7 @@ inline std::string format_chat(const struct llama_model * model, const std::stri std::vector chat(messages.size()); for (size_t i = 0; i < messages.size(); ++i) { - auto &curr_msg = messages[i]; + const auto & curr_msg = messages[i]; str[i*2 + 0] = json_value(curr_msg, "role", std::string("")); str[i*2 + 1] = json_value(curr_msg, "content", std::string("")); alloc_size += str[i*2 + 1].length(); @@ -183,261 +127,13 @@ inline std::string format_chat(const struct llama_model * model, const std::stri res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size()); } - std::string formatted_chat(buf.data(), res); + const std::string formatted_chat(buf.data(), res); + LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}}); return formatted_chat; } -// -// work queue utils -// - -struct llama_server_queue { - int id = 0; - std::mutex mutex_tasks; - bool running; - // 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_run_slots; - - // 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++; - LOG_VERBOSE("new task id", {{"new_id", task.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); - int new_id = id++; - LOG_VERBOSE("new task id", {{"new_id", new_id}}); - return new_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 when it is finished - void on_finish_multitask(std::function callback) { - callback_finish_multitask = callback; - } - - // Register the function to be called when all slots data is ready to be processed - void on_run_slots(std::function callback) { - callback_run_slots = 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(); - } - - // end the start_loop routine - void terminate() { - { - std::unique_lock lock(mutex_tasks); - running = false; - } - condition_tasks.notify_all(); - } - - /** - * Main loop consists of these steps: - * - Wait until a new task arrives - * - Process the task (i.e. maybe copy data into slot) - * - Check if multitask is finished - * - Run all slots - */ - void start_loop() { - running = true; - while (true) { - LOG_VERBOSE("new task may arrive", {}); - { - 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", {{"task_id", task.id}}); - callback_new_task(task); - } - LOG_VERBOSE("update_multitasks", {}); - // check if we have any finished 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 processed, slots data is now ready - LOG_VERBOSE("callback_run_slots", {}); - callback_run_slots(); - } - LOG_VERBOSE("wait for new task", {}); - // wait for new task - { - std::unique_lock lock(mutex_tasks); - if (queue_tasks.empty()) { - if (!running) { - LOG_VERBOSE("ending start_loop", {}); - return; - } - condition_tasks.wait(lock, [&]{ - return (!queue_tasks.empty() || !running); - }); - } - } - } - } - - // - // 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; - - // add the task_id to the list of tasks waiting for response - void add_waiting_task_id(int task_id) { - LOG_VERBOSE("waiting for task id", {{"task_id", task_id}}); - std::unique_lock lock(mutex_results); - waiting_task_ids.insert(task_id); - } - - // when the request is finished, we can remove task associated with it - void remove_waiting_task_id(int task_id) { - LOG_VERBOSE("remove waiting for task id", {{"task_id", 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(); - }); - - 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", {{"task_id", result.id}}); - 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", {{"task_id", task_id}}); - callback_update_multitask(task_id, result.id, result); - continue; - } - - if (result.id == task_id) - { - LOG_VERBOSE("queue_results.push_back", {{"task_id", task_id}}); - queue_results.push_back(result); - condition_results.notify_all(); - return; - } - } - } -}; - // // base64 utils (TODO: move to common in the future) // @@ -447,13 +143,11 @@ static const std::string base64_chars = "abcdefghijklmnopqrstuvwxyz" "0123456789+/"; -static inline bool is_base64(uint8_t c) -{ +static inline bool is_base64(uint8_t c) { return (isalnum(c) || (c == '+') || (c == '/')); } -static inline std::vector base64_decode(const std::string & encoded_string) -{ +static inline std::vector base64_decode(const std::string & encoded_string) { int i = 0; int j = 0; int in_ = 0; @@ -465,13 +159,10 @@ static inline std::vector base64_decode(const std::string & encoded_str std::vector ret; - while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) - { + 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++) - { + if (i == 4) { + for (i = 0; i < 4; i++) { char_array_4[i] = base64_chars.find(char_array_4[i]); } @@ -479,23 +170,20 @@ static inline std::vector base64_decode(const std::string & encoded_str 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++) - { + for (i = 0; (i < 3); i++) { ret.push_back(char_array_3[i]); } + i = 0; } } - if (i) - { - for (j = i; j <4; j++) - { + if (i) { + for (j = i; j < 4; j++) { char_array_4[j] = 0; } - for (j = 0; j <4; j++) - { + for (j = 0; j < 4; j++) { char_array_4[j] = base64_chars.find(char_array_4[j]); } @@ -503,8 +191,7 @@ static inline std::vector base64_decode(const std::string & encoded_str 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++) - { + for (j = 0; j < i - 1; j++) { ret.push_back(char_array_3[j]); } } @@ -516,8 +203,7 @@ static inline std::vector base64_decode(const std::string & encoded_str // random string / id // -static std::string random_string() -{ +static std::string random_string() { static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"); std::random_device rd; @@ -532,10 +218,10 @@ static std::string random_string() return result; } -static std::string gen_chatcmplid() -{ +static std::string gen_chatcmplid() { std::stringstream chatcmplid; chatcmplid << "chatcmpl-" << random_string(); + return chatcmplid.str(); } @@ -543,91 +229,316 @@ static std::string gen_chatcmplid() // other common utils // -static size_t common_part(const std::vector &a, const std::vector &b) -{ +static size_t common_part(const std::vector & a, const std::vector & b) { size_t i; - for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) - { - } + for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {} + return i; } -static bool ends_with(const std::string &str, const std::string &suffix) -{ - return str.size() >= suffix.size() && - 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix); +static bool ends_with(const std::string & str, const std::string & suffix) { + return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix); } -static size_t find_partial_stop_string(const std::string &stop, - const std::string &text) -{ - if (!text.empty() && !stop.empty()) - { +static size_t find_partial_stop_string(const std::string &stop, const std::string &text) { + if (!text.empty() && !stop.empty()) { const char text_last_char = text.back(); - for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) - { - if (stop[char_index] == text_last_char) - { + for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) { + if (stop[char_index] == text_last_char) { const std::string current_partial = stop.substr(0, char_index + 1); - if (ends_with(text, current_partial)) - { + if (ends_with(text, current_partial)) { return text.size() - char_index - 1; } } } } + return std::string::npos; } // TODO: reuse llama_detokenize template -static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end) -{ +static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) { std::string ret; - for (; begin != end; ++begin) - { + for (; begin != end; ++begin) { ret += llama_token_to_piece(ctx, *begin); } + return ret; } // format incomplete utf-8 multibyte character for output -static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token) -{ +static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) { std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token); + // if the size is 1 and first bit is 1, meaning it's a partial character // (size > 1 meaning it's already a known token) - if (out.size() == 1 && (out[0] & 0x80) == 0x80) - { + if (out.size() == 1 && (out[0] & 0x80) == 0x80) { std::stringstream ss; ss << std::hex << (out[0] & 0xff); std::string res(ss.str()); out = "byte: \\x" + res; } + return out; } +struct completion_token_output { + llama_token tok; + std::string text_to_send; + + struct token_prob { + llama_token tok; + float prob; + }; + + std::vector probs; +}; + // convert a vector of completion_token_output to json -static json probs_vector_to_json(const llama_context *ctx, const std::vector &probs) -{ +static json probs_vector_to_json(const llama_context * ctx, const std::vector & probs) { json out = json::array(); - for (const auto &prob : probs) - { + + for (const auto & prob : probs) { json probs_for_token = json::array(); - for (const auto &p : prob.probs) - { - std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok); - probs_for_token.push_back(json - { + + for (const auto & p : prob.probs) { + const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok); + probs_for_token.push_back(json { {"tok_str", tok_str}, {"prob", p.prob}, }); } - std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok); - out.push_back(json{ + + const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok); + out.push_back(json { {"content", tok_str}, {"probs", probs_for_token}, }); } + return out; } + +// +// OAI utils +// + +static json oaicompat_completion_params_parse( + const struct llama_model * model, + const json & body, /* openai api json semantics */ + const std::string & chat_template) { + 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_chat(model, chat_template, body["messages"]); + 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, json result, bool streaming = false) { + 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(json result) { + if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) { + return std::vector({result}); + } + + 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_embeddings_response_oaicompat(const json & request, const json & embeddings) { + json res = json { + {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))}, + {"object", "list"}, + {"usage", json { + {"prompt_tokens", 0}, + {"total_tokens", 0} + }}, + {"data", embeddings} + }; + + return res; +} + +static json format_tokenizer_response(const std::vector & tokens) { + return json { + {"tokens", tokens} + }; +} + +static json format_detokenized_response(const std::string & content) { + return json { + {"content", content} + }; +} diff --git a/llama.cpp b/llama.cpp index b27aa2728..478099648 100644 --- a/llama.cpp +++ b/llama.cpp @@ -13541,18 +13541,22 @@ LLAMA_API int32_t llama_chat_apply_template( curr_tmpl = std::string(model_template.data(), model_template.size()); } } + // format the chat to string std::vector chat_vec; chat_vec.resize(n_msg); for (size_t i = 0; i < n_msg; i++) { chat_vec[i] = &chat[i]; } + std::string formatted_chat; int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass); if (res < 0) { return res; } - strncpy(buf, formatted_chat.c_str(), length); + if (buf && length > 0) { + strncpy(buf, formatted_chat.c_str(), length); + } return res; } From 55a2a900ff4a02fc33708ac7858d595d289a3f2a Mon Sep 17 00:00:00 2001 From: Minsoo Cheong <54794500+mscheong01@users.noreply.github.com> Date: Thu, 7 Mar 2024 19:42:39 +0900 Subject: [PATCH 08/32] server : add `/v1/completions` endpoint (#5914) * add-`/v1/completions`-endpoint * add legacy comment to `/completion` endpoint --- examples/server/server.cpp | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 3bdbde954..f255ad764 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2923,7 +2923,7 @@ int main(int argc, char ** argv) { res.set_content(data.dump(), "application/json; charset=utf-8"); }); - svr.Post("/completion", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { + const auto completions = [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; @@ -3001,7 +3001,11 @@ int main(int argc, char ** argv) { res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); } - }); + }; + + svr.Post("/completion", completions); // legacy + svr.Post("/completions", completions); + svr.Post("/v1/completions", completions); svr.Get("/v1/models", [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); From 89fb735fcfd21781a8194b211cf32824beb3f71f Mon Sep 17 00:00:00 2001 From: Neo Zhang Jianyu Date: Thu, 7 Mar 2024 19:14:49 +0800 Subject: [PATCH 09/32] Revert "[SYCL] fix error when set main gpu to non-zero (#5901)" (#5918) This reverts commit ceca1aef0738b57951cd12c603c3477e75312dec. --- ggml-sycl.cpp | 172 ++++++++++++++++++++------------------------------ ggml-sycl.h | 1 - llama.cpp | 16 ++--- 3 files changed, 73 insertions(+), 116 deletions(-) diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index 221d67b8d..ddd951dd6 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -3559,31 +3559,12 @@ class sycl_gpu_mgr { int work_group_size = 0; std::string gpus_list = ""; - /* - Use all GPU with same top max compute units - */ sycl_gpu_mgr() { detect_sycl_gpu_list_with_max_cu(); get_allow_gpus(); create_context_with_gpus(); } - /* - Use the assigned GPU as only one - */ - sycl_gpu_mgr(int main_gpu_id) { - sycl::device device = dpct::dev_mgr::instance().get_device(main_gpu_id); - dpct::device_info prop; - dpct::get_device_info(prop, device); - gpus.push_back(main_gpu_id); - devices.push_back(device); - work_group_size = prop.get_max_work_group_size(); - max_compute_units = prop.get_max_compute_units(); - - get_allow_gpus(); - create_context_with_gpus(); - } - void create_context_with_gpus() { sycl::context ctx = sycl::context(devices); assert(gpus.size() > 0); @@ -3599,7 +3580,7 @@ class sycl_gpu_mgr { gpus_list += std::to_string(gpus[i]); gpus_list += ","; } - if (gpus_list.length() > 1) { + if (gpus_list.length() > 2) { gpus_list.pop_back(); } } @@ -3648,8 +3629,8 @@ class sycl_gpu_mgr { if (gpus[i] == id) return i; } - printf("miss to get device index by id=%d\n", id); - GGML_ASSERT(false); + assert(false); + return -1; } int get_next_index(int id) { @@ -3658,7 +3639,8 @@ class sycl_gpu_mgr { if (gpus[i] == id) return i; } - GGML_ASSERT(false); + assert(false); + return -1; } }; @@ -3667,7 +3649,6 @@ 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_id = -1; -static bool g_ggml_backend_sycl_buffer_type_initialized = false; static std::array g_default_tensor_split = {}; @@ -13244,7 +13225,7 @@ void ggml_backend_sycl_print_sycl_devices() { } void print_gpu_device_list() { - fprintf(stderr, "detect %d SYCL GPUs: [%s] with top Max compute units:%d\n", + fprintf(stderr, "detect %d SYCL GPUs: [%s] with Max compute units:%d\n", g_sycl_gpu_mgr->get_gpu_count(), g_sycl_gpu_mgr->gpus_list.c_str(), g_sycl_gpu_mgr->max_compute_units); @@ -13283,15 +13264,6 @@ void ggml_init_sycl() try { #else fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__); #endif - -/* NOT REMOVE, keep it for next optimize for XMX. -#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 -*/ - if (CHECK_TRY_ERROR(g_all_sycl_device_count = dpct::dev_mgr::instance().device_count()) != 0) { initialized = true; @@ -13300,61 +13272,68 @@ void ggml_init_sycl() try { } GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES); ggml_backend_sycl_print_sycl_devices(); + if (!g_sycl_gpu_mgr) g_sycl_gpu_mgr = new sycl_gpu_mgr(); + + g_device_count = g_sycl_gpu_mgr->get_gpu_count(); + g_work_group_size = g_sycl_gpu_mgr->work_group_size; + print_gpu_device_list(); - initialized = true; - g_sycl_loaded = true; - } + int64_t total_vram = 0; - - g_device_count = g_sycl_gpu_mgr->get_gpu_count(); - g_work_group_size = g_sycl_gpu_mgr->work_group_size; - - int64_t total_vram = 0; - - - for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { - g_device_caps[id].vmm = 0; - g_device_caps[id].device_id = -1; - g_device_caps[id].cc = 0; - g_tensor_split[id] = 0; - g_default_tensor_split[id] = 0; - } - - for (int i = 0; i < g_device_count; ++i) { - int device_id = g_sycl_gpu_mgr->gpus[i]; - g_device_caps[i].vmm = 0; - - dpct::device_info prop; - SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( - prop, dpct::dev_mgr::instance().get_device(device_id)))); - - g_default_tensor_split[i] = total_vram; - total_vram += prop.get_global_mem_size(); - - g_device_caps[i].cc = - 100 * prop.get_major_version() + 10 * prop.get_minor_version(); - } - - for (int i = 0; i < g_device_count; ++i) { - g_default_tensor_split[i] /= total_vram; - } - - for (int i = 0; i < g_device_count; ++i) { - SYCL_CHECK(ggml_sycl_set_device(i)); - - // create sycl streams - for (int is = 0; is < MAX_STREAMS; ++is) { - SYCL_CHECK(CHECK_TRY_ERROR( - g_syclStreams[i][is] = - dpct::get_current_device().create_queue( - g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device()))); +/* NOT REMOVE, keep it for next optimize for XMX. +#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 +*/ + for (int id = 0; id < GGML_SYCL_MAX_DEVICES; ++id) { + g_device_caps[id].vmm = 0; + g_device_caps[id].device_id = -1; + g_device_caps[id].cc = 0; + g_tensor_split[id] = 0; + g_default_tensor_split[id] = 0; } - const dpct::queue_ptr stream = g_syclStreams[i][0]; - // create sycl handle - SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); + for (int i = 0; i < g_device_count; ++i) { + int device_id = g_sycl_gpu_mgr->gpus[i]; + g_device_caps[i].vmm = 0; + + dpct::device_info prop; + SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info( + prop, dpct::dev_mgr::instance().get_device(device_id)))); + + g_default_tensor_split[i] = total_vram; + total_vram += prop.get_global_mem_size(); + + g_device_caps[i].cc = + 100 * prop.get_major_version() + 10 * prop.get_minor_version(); + } + + for (int i = 0; i < g_device_count; ++i) { + g_default_tensor_split[i] /= total_vram; + } + + for (int i = 0; i < g_device_count; ++i) { + SYCL_CHECK(ggml_sycl_set_device(i)); + + // create sycl streams + for (int is = 0; is < MAX_STREAMS; ++is) { + SYCL_CHECK(CHECK_TRY_ERROR( + g_syclStreams[i][is] = + dpct::get_current_device().create_queue( + g_sycl_gpu_mgr->get_co_ctx(), dpct::get_current_device()))); + } + + const dpct::queue_ptr stream = g_syclStreams[i][0]; + // create sycl handle + SYCL_CHECK(CHECK_TRY_ERROR(g_sycl_handles[i] = stream)); + } + + initialized = true; + g_sycl_loaded = true; } } catch (sycl::exception const &exc) { @@ -16753,24 +16732,22 @@ static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = { /* .is_host = */ nullptr, }; -ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device_index) { - if (device_index>=g_device_count or device_index<0) { - printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n", - device_index, g_device_count-1); - GGML_ASSERT(device_indexgpus[i])}, }; } - g_ggml_backend_sycl_buffer_type_initialized = true; + ggml_backend_sycl_buffer_type_initialized = true; } - return &ggml_backend_sycl_buffer_types[device_index]; + + return &ggml_backend_sycl_buffer_types[device]; } // sycl split buffer type @@ -17519,17 +17496,6 @@ GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id) { return g_sycl_gpu_mgr->get_index(device_id); } -GGML_API GGML_CALL void ggml_backend_sycl_set_single_device(int main_gpu_id) { - GGML_ASSERT(main_gpu_idbackends.push_back(backend); } else { // LLAMA_SPLIT_LAYER requires a backend for each GPU - + int id_list[GGML_SYCL_MAX_DEVICES]; + ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); for (int i = 0; i < ggml_backend_sycl_get_device_count(); ++i) { + int device_id = id_list[i]; ggml_backend_t backend = ggml_backend_sycl_init(i); if (backend == nullptr) { - int id_list[GGML_SYCL_MAX_DEVICES]; - ggml_sycl_get_gpu_list(id_list, GGML_SYCL_MAX_DEVICES); - LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, id_list[i], i); + LLAMA_LOG_ERROR("%s: failed to initialize SYCL%d (index %d)backend\n", __func__, device_id, i); llama_free(ctx); return nullptr; } From 6cdabe652695167263c8b447520987b11856f7ca Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Thu, 7 Mar 2024 16:32:38 +0200 Subject: [PATCH 10/32] llama-bench : add embeddings option (#5924) * llama-bench : add embeddings option * llama-bench : do not hard code embd default value --------- Co-authored-by: slaren --- examples/llama-bench/llama-bench.cpp | 30 +++++++++++++++++++++++++--- 1 file changed, 27 insertions(+), 3 deletions(-) diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index aa79d002a..2ff86ef6f 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -173,6 +173,7 @@ struct cmd_params { std::vector no_kv_offload; std::vector> tensor_split; std::vector use_mmap; + std::vector embeddings; int reps; bool verbose; output_formats output_format; @@ -192,6 +193,7 @@ static const cmd_params cmd_params_defaults = { /* no_kv_offload */ {false}, /* tensor_split */ {std::vector(llama_max_devices(), 0.0f)}, /* use_mmap */ {true}, + /* embeddings */ {false}, /* reps */ 5, /* verbose */ false, /* output_format */ MARKDOWN @@ -214,6 +216,7 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -mg, --main-gpu (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str()); printf(" -nkvo, --no-kv-offload <0|1> (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str()); printf(" -mmp, --mmap <0|1> (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str()); + printf(" -embd, --embeddings <0|1> (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str()); printf(" -ts, --tensor_split (default: 0)\n"); printf(" -r, --repetitions (default: %d)\n", cmd_params_defaults.reps); printf(" -o, --output (default: %s)\n", output_format_str(cmd_params_defaults.output_format)); @@ -382,6 +385,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { } auto p = split(argv[i], split_delim); params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end()); + } else if (arg == "-embd" || arg == "--embeddings") { + if (++i >= argc) { + invalid_param = true; + break; + } + auto p = split(argv[i], split_delim); + params.embeddings.insert(params.embeddings.end(), p.begin(), p.end()); } else if (arg == "-ts" || arg == "--tensor-split") { if (++i >= argc) { invalid_param = true; @@ -453,6 +463,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) { if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; } if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; } if (params.use_mmap.empty()) { params.use_mmap = cmd_params_defaults.use_mmap; } + if (params.embeddings.empty()) { params.embeddings = cmd_params_defaults.embeddings; } if (params.n_threads.empty()) { params.n_threads = cmd_params_defaults.n_threads; } return params; @@ -472,6 +483,7 @@ struct cmd_params_instance { bool no_kv_offload; std::vector tensor_split; bool use_mmap; + bool embeddings; llama_model_params to_llama_mparams() const { llama_model_params mparams = llama_model_default_params(); @@ -502,6 +514,7 @@ struct cmd_params_instance { cparams.type_k = type_k; cparams.type_v = type_v; cparams.offload_kqv = !no_kv_offload; + cparams.embeddings = embeddings; return cparams; } @@ -517,6 +530,7 @@ static std::vector get_cmd_params_instances(const cmd_param for (const auto & mg : params.main_gpu) for (const auto & ts : params.tensor_split) for (const auto & mmp : params.use_mmap) + for (const auto & embd : params.embeddings) for (const auto & nb : params.n_batch) for (const auto & tk : params.type_k) for (const auto & tv : params.type_v) @@ -540,6 +554,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .no_kv_offload= */ nkvo, /* .tensor_split = */ ts, /* .use_mmap = */ mmp, + /* .embeddings = */ embd, }; instances.push_back(instance); } @@ -562,6 +577,7 @@ static std::vector get_cmd_params_instances(const cmd_param /* .no_kv_offload= */ nkvo, /* .tensor_split = */ ts, /* .use_mmap = */ mmp, + /* .embeddings = */ embd, }; instances.push_back(instance); } @@ -597,6 +613,7 @@ struct test { bool no_kv_offload; std::vector tensor_split; bool use_mmap; + bool embeddings; int n_prompt; int n_gen; std::string test_time; @@ -619,6 +636,7 @@ struct test { no_kv_offload = inst.no_kv_offload; tensor_split = inst.tensor_split; use_mmap = inst.use_mmap; + embeddings = inst.embeddings; n_prompt = inst.n_prompt; n_gen = inst.n_gen; // RFC 3339 date-time format @@ -690,7 +708,7 @@ struct test { "n_batch", "n_threads", "type_k", "type_v", "n_gpu_layers", "split_mode", "main_gpu", "no_kv_offload", - "tensor_split", "use_mmap", + "tensor_split", "use_mmap", "embeddings", "n_prompt", "n_gen", "test_time", "avg_ns", "stddev_ns", "avg_ts", "stddev_ts" @@ -710,7 +728,7 @@ struct test { } if (field == "cuda" || field == "opencl" || field == "vulkan" || field == "kompute" || field == "metal" || field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" || - field == "use_mmap") { + field == "use_mmap" || field == "embeddings") { return BOOL; } if (field == "avg_ts" || field == "stddev_ts") { @@ -744,7 +762,7 @@ struct test { std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v), std::to_string(n_gpu_layers), split_mode_str(split_mode), std::to_string(main_gpu), std::to_string(no_kv_offload), - tensor_split_str, std::to_string(use_mmap), + tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings), std::to_string(n_prompt), std::to_string(n_gen), test_time, std::to_string(avg_ns()), std::to_string(stdev_ns()), std::to_string(avg_ts()), std::to_string(stdev_ts()) @@ -914,6 +932,9 @@ struct markdown_printer : public printer { if (field == "use_mmap") { return "mmap"; } + if (field == "embeddings") { + return "embd"; + } if (field == "tensor_split") { return "ts"; } @@ -957,6 +978,9 @@ struct markdown_printer : public printer { if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) { fields.emplace_back("use_mmap"); } + if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) { + fields.emplace_back("embeddings"); + } fields.emplace_back("test"); fields.emplace_back("t/s"); From 581ed5c4fe3a8909aaa8313633ac443f471ba755 Mon Sep 17 00:00:00 2001 From: "UEXTM.com" <84163508+uextm@users.noreply.github.com> Date: Fri, 8 Mar 2024 04:35:04 -0500 Subject: [PATCH 11/32] log : fix MSVC compile errors (#5643) MSVC gives the following error with the existing macros: `Error C2059 : syntax error: ','` This patch adds `##` as a prefix to `__VA_ARGS__` to address this error. --- common/log.h | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/common/log.h b/common/log.h index e4e1b9f4f..eb111e784 100644 --- a/common/log.h +++ b/common/log.h @@ -297,7 +297,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: #ifndef _MSC_VER #define LOG(...) LOG_IMPL(__VA_ARGS__, "") #else - #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "") + #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "") #endif // Main TEE macro. @@ -311,7 +311,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: #ifndef _MSC_VER #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "") #else - #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "") + #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "") #endif // LOG macro variants with auto endline. @@ -319,8 +319,8 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std:: #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n") #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n") #else - #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n") - #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n") + #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n") + #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n") #endif // INTERNAL, DO NOT USE From af37fd8b30e37ccbffdd82e6f48559e2fb7ce7dd Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 8 Mar 2024 12:40:02 +0200 Subject: [PATCH 12/32] server : fix EOS token detection with disabled cache (#5938) --- examples/server/server.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index f255ad764..1434095fc 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1123,7 +1123,7 @@ struct server_context { }); } - if (!slot.cache_tokens.empty() && result.tok == llama_token_eos(model)) { + if (result.tok == llama_token_eos(model)) { slot.stopped_eos = true; slot.has_next_token = false; From e457fb3540e0aaec47cfde0abf784c213f9216ee Mon Sep 17 00:00:00 2001 From: Don Mahurin Date: Fri, 8 Mar 2024 02:41:50 -0800 Subject: [PATCH 13/32] llama : assume tied weights if lm_head/output weights is missing (#5824) This is to support model configurations with "tie_word_embeddings" set to true. Co-authored-by: Don Mahurin <2797413+dmahurin@users.noreply.github.com> --- llama.cpp | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/llama.cpp b/llama.cpp index 4225f9555..458382b21 100644 --- a/llama.cpp +++ b/llama.cpp @@ -3888,7 +3888,13 @@ static bool llm_load_tensors( { model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); if (model.arch != LLM_ARCH_MINICPM){ - model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}); + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, false); + // if output is NULL, init from the input tok embed + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } } } From 76e868821a94072fbc87cb1fcca291694319eae8 Mon Sep 17 00:00:00 2001 From: Pierrick Hymbert Date: Fri, 8 Mar 2024 12:25:04 +0100 Subject: [PATCH 14/32] server: metrics: add llamacpp:prompt_seconds_total and llamacpp:tokens_predicted_seconds_total, reset bucket only on /metrics. Fix values cast to int. Add Process-Start-Time-Unix header. (#5937) Closes #5850 --- examples/server/server.cpp | 47 ++++++++++++++----- examples/server/tests/features/server.feature | 1 + examples/server/tests/features/steps/steps.py | 11 ++++- 3 files changed, 46 insertions(+), 13 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 1434095fc..109ff7175 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -335,8 +335,12 @@ struct server_slot { }; struct server_metrics { + const int64_t t_start = ggml_time_us(); + uint64_t n_prompt_tokens_processed_total = 0; + uint64_t t_prompt_processing_total = 0; uint64_t n_tokens_predicted_total = 0; + uint64_t t_tokens_generation_total = 0; uint64_t n_prompt_tokens_processed = 0; uint64_t t_prompt_processing = 0; @@ -348,12 +352,14 @@ struct server_metrics { n_prompt_tokens_processed_total += slot.n_prompt_tokens_processed; n_prompt_tokens_processed += slot.n_prompt_tokens_processed; t_prompt_processing += slot.t_prompt_processing; + t_prompt_processing_total += slot.t_prompt_processing; } void on_prediction(const server_slot &slot) { - n_tokens_predicted_total += slot.n_decoded; - n_tokens_predicted += slot.n_decoded; - t_tokens_generation += slot.t_token_generation; + n_tokens_predicted_total += slot.n_decoded; + n_tokens_predicted += slot.n_decoded; + t_tokens_generation += slot.t_token_generation; + t_tokens_generation_total += slot.t_token_generation; } void reset_bucket() { @@ -1502,9 +1508,12 @@ struct server_context { { "idle", n_idle_slots }, { "processing", n_processing_slots }, { "deferred", queue_tasks.queue_tasks_deferred.size() }, + { "t_start", metrics.t_start}, { "n_prompt_tokens_processed_total", metrics.n_prompt_tokens_processed_total}, + { "t_tokens_generation_total", metrics.t_tokens_generation_total}, { "n_tokens_predicted_total", metrics.n_tokens_predicted_total}, + { "t_prompt_processing_total", metrics.t_prompt_processing_total}, { "n_prompt_tokens_processed", metrics.n_prompt_tokens_processed}, { "t_prompt_processing", metrics.t_prompt_processing}, @@ -1517,7 +1526,9 @@ struct server_context { { "slots", slots_data }, }; - metrics.reset_bucket(); + if (json_value(task.data, "reset_bucket", false)) { + metrics.reset_bucket(); + } queue_results.send(res); } break; } @@ -2709,6 +2720,7 @@ int main(int argc, char ** argv) { task.id_multi = -1; task.id_target = -1; task.type = SERVER_TASK_TYPE_METRICS; + task.data.push_back({{"reset_bucket", true}}); ctx_server.queue_results.add_waiting_task_id(task.id); ctx_server.queue_tasks.post(task); @@ -2732,20 +2744,28 @@ int main(int argc, char ** argv) { {"counter", {{ {"name", "prompt_tokens_total"}, {"help", "Number of prompt tokens processed."}, - {"value", data["n_prompt_tokens_processed_total"]} + {"value", (uint64_t) data["n_prompt_tokens_processed_total"]} + }, { + {"name", "prompt_seconds_total"}, + {"help", "Prompt process time"}, + {"value", (uint64_t) data["t_prompt_processing_total"] / 1.e3} }, { {"name", "tokens_predicted_total"}, {"help", "Number of generation tokens processed."}, - {"value", data["n_tokens_predicted_total"]} + {"value", (uint64_t) data["n_tokens_predicted_total"]} + }, { + {"name", "tokens_predicted_seconds_total"}, + {"help", "Predict process time"}, + {"value", (uint64_t) data["t_tokens_generation_total"] / 1.e3} }}}, {"gauge", {{ {"name", "prompt_tokens_seconds"}, {"help", "Average prompt throughput in tokens/s."}, - {"value", n_prompt_tokens_processed ? 1e3 / t_prompt_processing * n_prompt_tokens_processed : 0} + {"value", n_prompt_tokens_processed ? 1.e3 / t_prompt_processing * n_prompt_tokens_processed : 0.} },{ {"name", "predicted_tokens_seconds"}, {"help", "Average generation throughput in tokens/s."}, - {"value", n_tokens_predicted ? 1e3 / t_tokens_generation * n_tokens_predicted : 0} + {"value", n_tokens_predicted ? 1.e3 / t_tokens_generation * n_tokens_predicted : 0.} },{ {"name", "kv_cache_usage_ratio"}, {"help", "KV-cache usage. 1 means 100 percent usage."}, @@ -2753,15 +2773,15 @@ int main(int argc, char ** argv) { },{ {"name", "kv_cache_tokens"}, {"help", "KV-cache tokens."}, - {"value", data["kv_cache_tokens_count"]} + {"value", (uint64_t) data["kv_cache_tokens_count"]} },{ {"name", "requests_processing"}, {"help", "Number of request processing."}, - {"value", data["processing"]} + {"value", (uint64_t) data["processing"]} },{ {"name", "requests_deferred"}, {"help", "Number of request deferred."}, - {"value", data["deferred"]} + {"value", (uint64_t) data["deferred"]} }}} }; @@ -2775,13 +2795,16 @@ int main(int argc, char ** argv) { const std::string name = metric_def["name"]; const std::string help = metric_def["help"]; - auto value = json_value(metric_def, "value", 0); + auto value = json_value(metric_def, "value", 0.); prometheus << "# HELP llamacpp:" << name << " " << help << "\n" << "# TYPE llamacpp:" << name << " " << type << "\n" << "llamacpp:" << name << " " << value << "\n"; } } + const int64_t t_start = data["t_start"]; + res.set_header("Process-Start-Time-Unix", std::to_string(t_start)); + res.set_content(prometheus.str(), "text/plain; version=0.0.4"); res.status = 200; // HTTP OK }); diff --git a/examples/server/tests/features/server.feature b/examples/server/tests/features/server.feature index f3b758c79..878ac1363 100644 --- a/examples/server/tests/features/server.feature +++ b/examples/server/tests/features/server.feature @@ -29,6 +29,7 @@ Feature: llama.cpp server And a completion request with no api error Then tokens are predicted matching And prometheus metrics are exposed + And metric llamacpp:tokens_predicted is Examples: Prompts | prompt | n_predict | re_content | n_predicted | diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index a0b2ffdfe..d7f005836 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -586,14 +586,24 @@ async def step_prometheus_metrics_exported(context): metric_exported = False if context.debug: print(f"/metrics answer:\n{metrics_raw}\n") + context.metrics = {} for metric in parser.text_string_to_metric_families(metrics_raw): match metric.name: case "llamacpp:kv_cache_usage_ratio": assert len(metric.samples) > 0 metric_exported = True + context.metrics[metric.name] = metric + assert int(metrics_response.headers["Process-Start-Time-Unix"]) > 0, "no header process start time" assert metric_exported, "No metrics exported" +@step(u'metric {metric_name} is {metric_value:d}') +def step_assert_metric_value(context, metric_name, metric_value): + if metric_name not in context.metrics: + assert False, f"no metric {metric_name} in {context.metrics.keys()}" + assert context.metrics[metric_name].samples[0].value == metric_value, f"metric: {context.metrics[metric_name]}" + + @step(u'available models') def step_available_models(context): # openai client always expects an api_key @@ -879,7 +889,6 @@ def assert_n_tokens_predicted(completion_response, expected_predicted_n=None, re f' {n_predicted} <> {expected_predicted_n}') - async def gather_tasks_results(context): n_tasks = len(context.concurrent_tasks) if context.debug: From 515f7d0d4fce41c752fc253acf30707c3be2531e Mon Sep 17 00:00:00 2001 From: compilade <113953597+compilade@users.noreply.github.com> Date: Fri, 8 Mar 2024 10:53:37 -0500 Subject: [PATCH 15/32] llama : fix quantization of shared token_embd (#5944) --- llama.cpp | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) diff --git a/llama.cpp b/llama.cpp index 458382b21..4a20b7928 100644 --- a/llama.cpp +++ b/llama.cpp @@ -10973,6 +10973,9 @@ struct quantize_state_internal { bool has_imatrix = false; + // used to figure out if a model shares tok_embd with the output weight + bool has_output = false; + quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params) : model(model) , params(params) @@ -11070,8 +11073,7 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty // for arches that share the same tensor between the token embeddings and the output, we quantize the token embeddings // with the quantization of the output tensor - if (name == tn(LLM_TENSOR_OUTPUT, "weight") || - (LLM_TENSOR_NAMES.at(arch).find(LLM_TENSOR_OUTPUT) == LLM_TENSOR_NAMES.at(arch).end() && name == "token_embd.weight")) { + if (name == tn(LLM_TENSOR_OUTPUT, "weight") || (!qs.has_output && name == tn(LLM_TENSOR_TOKEN_EMBD, "weight"))) { int nx = tensor->ne[0]; if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) { new_type = GGML_TYPE_Q8_0; @@ -11460,6 +11462,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s else if (name.find("ffn_up") != std::string::npos) { ++qs.n_ffn_up; } + else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) { + qs.has_output = true; + } } 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", From c2101a2e909ac7c08976d414e64e96c90ee5fa9e Mon Sep 17 00:00:00 2001 From: compilade <113953597+compilade@users.noreply.github.com> Date: Fri, 8 Mar 2024 17:31:00 -0500 Subject: [PATCH 16/32] llama : support Mamba Selective State Space Models (#5328) * mamba : begin working on support for Mamba SSM * mamba : begin figuring out how to (ab)use the kv cache for Mamba * mamba : recurrent inference almost works, but incoherent * mamba : recurrent inference WORKS!!! * convert : optionally use d_conv and d_state from config.json for Mamba * mamba : refactor recurrent conv, resulting in 20% perf increase It's still slower than I'd like, but I did not really optimize `ggml_exp` yet. I also refactored `ggml_exp` to work with tensors with more than 2 dimensions. * ggml : parallelize ggml_exp This results in 8% faster token generation for Mamba-130M. * mamba : simplify the conv step with a self-overlapping view Turns out the conv_state can be made smaller by one column. Note that this breaks existing GGUFs of Mamba, because the key_value_length field is tied to the conv_state size. Convolution with a self-overlapping view is cool! And it's much simpler than what I initially thought would be necessary to make the convolution step work with more than 1 token at a time. Next step is to make the SSM step work on batches of tokens too, and thus I need to figure out a way to make a parallel selective scan which will keep the ssm_state small and won't make it bigger by a factor of (n_layer * batch_size). * llama : fix Mamba KV self size wrongly displaying as f16 instead of f32 Relatedly, I also tried to see if other types than f32 worked for the states, but they don't, because of the operators used. It's probably better anyway to keep lots of precision there, since the states are small anyway. * mamba : fix self-overlapping view depth stride * mamba : handle batches of more than 1 token This means running Mamba no longer crashes when using the default settings! And probably also slightly faster prompt processing. Both batched and non-batched processing yield the same output. Previously, the state was not cleared when starting a sequence. Next step is to make the KV cache API work as expected for Mamba models. * ggml: add ggml_ssm_scan to help with parallel selective scan If the selective scan was implemented without a custom operator, there would be waaay too many nodes in the graph. For example, for Mamba-130M, with a batch size of 512 (the default), a naive selective scan could add at least 24*512=12288 nodes, which is more than LLAMA_MAX_NODES (8192), and that's only for the smallest Mamba model. So it's much cleaner with a custom operator. Not sure about the name, though. * ggml : in ggml_ssm_scan, merge multiple rows in the same vec operation This will help with performance on CPU if ggml_vec_mul_f32 and ggml_vec_add_f32 are ever optimized with SIMD. * mamba : very basic quantization support Mostly works, but there is currently no difference between the variants of a k-quant (e.g. Q4_K_S and Q4_K_M are the same). Most of the SSM-specific weights can be kept in f32 without affecting the size that much, since they are relatively small. (the linear projection weights are responsible for most of Mamba's size) Too much quantization seems to make the state degrade quite fast, and the model begins to output gibberish. It seems to affect bigger models to a lesser extent than small models, but I'm not sure by how much. Experimentation will be needed to figure out which weights are more important for the _M (and _L?) variants of k-quants for Mamba. * convert : fix wrong name for layer norm weight of offical Mamba models I was using Q-bert/Mamba-* models before, which have a slighlty different naming scheme for the weights. (they start with "model.layers" instead of "backbone.layers") * mamba : fuse more steps of the SSM scan in the ggml_ssm_scan operator This increases performance on CPU by around 30% for prompt processing, and by around 20% for text generation. However, it also makes the ggml_exp and ggml_soft_plus operators unused. Whether or not they should be kept will be decided later. * convert : for Mamba, also consider the "MambaLMHeadModel" arch name It's the name of the class of the official implementation, though they don't use it (yet) in the "architectures" field of config.json * mamba : fix vocab size problems with official models The perplexity was waaaay to high for models with a non-round vocab size. Not sure why, but it needed to be fixed in the metadata. Note that this breaks existing GGUF-converted Mamba models, but **only if** the vocab size was not already rounded. * ggml : remove ggml_exp and ggml_soft_plus They did not exist anyway outside of this branch, and since ggml_ssm_scan fused operations together, they are unused. It's always possible to bring them back if needed. * mamba : remove some useless comments No code change. * convert : fix flake8 linter errors * mamba : apply suggestions from code review * mamba : remove unecessary branch for row-wise ssm_state and C multiplication It was previously done to avoid permuting when only one token is processed at a time (like when generating text), but permuting is cheap, and dynamically changing the compute graph is not future-proof. * ggml : in ggml_ssm_scan, use more appropriate asserts * ggml : rename the destination pointer in ggml_compute_forward_ssm_scan_f32 * mamba : multiple sequences, but one at a time This is a step towards making this Mamba implementation usable with the server example (the way the system prompt is kept when clearing the client slots will need to be changed before this can work, though). The KV cache size for this kind of model is tied to the maximum number of sequences kept at any single time. For now, this number is obtained from n_parallel (plus one, to have an extra sequence to dedicate to the system prompt), but there might be a better way to do this which won't also make the main example use 2 cells even if only 1 is really used. (for this specific case, --parallel 0 helps) Simultaneous sequence processing will probably require changes to ggml_ssm_scan, and possibly a new operator for the conv step. * mamba : support llama_kv_cache_seq_cp This (mis)uses the logic around K shifts, because tokens in a state can't be shifted anyway, and because inp_K_shift has the right shape and type. Using ggml_get_rows is a nice way to do copies, but copy chains can't work. Fortunately, copy chains don't really seem to be used in the examples. Each KV cell is dedicated to the sequence ID corresponding to its own index. * mamba : use a state mask It's cleaner than the previous heuristic of checking for the pos of the first token in the batch. inp_KQ_mask could not be re-used for this, because it has the wrong shape and because it seems more suited to the next step of simultaneous sequence processing (helping with the problem of remembering which token belongs to which sequence(s)/state(s)). * llama : replace the usage of n_ctx with kv_self.size in many places * mamba : use n_tokens directly instead of n_tok * mamba : in comments, properly refer to KV cells instead of slots * mamba : reduce memory usage of ggml_ssm_scan From 290.37 MiB to 140.68 MiB of CPU compute buffer size with Mamba 3B with a batch size of 512. The result tensor of ggml_ssm_scan was previously a big part of the CPU compute buffer size. To make it smaller, it does not contain the intermediate ssm states anymore. Both y and the last ssm state are combined in the result tensor, because it seems only a single tensor can be returned by an operator with the way the graph is built. * mamba : simultaneous sequence processing A batch can now contain tokens from multiple sequences. This is necessary for at least the parallel example, the server example, and the HellaSwag test in the perplexity example. However, for this to be useful, uses of llama_kv_cache_seq_rm/cp will need to be changed to work on whole sequences. * ggml : add ggml_ssm_conv as a new operator for the conv step of Mamba This operator makes it possible to use and update the correct states for each token of the batch in the same way as ggml_ssm_scan. Other solutions which use existing operators would need loops which would add too many nodes to the graph (at least the ones I thought of). Using this operator further reduces the size of the CPU compute buffer from 140.68 MiB to 103.20 MiB with Mamba 3B with a batch size of 512. And (at least on CPU), it's a bit faster than before. Note that "ggml_ssm_conv" is probably not the most appropriate name, and it could be changed if a better one is found. * llama : add inp_s_seq as a new input tensor The most convenient implementation to select the correct state (for Mamba) for each token is to directly get the correct index from a tensor. This is why inp_s_seq is storing int32_t and not floats. The other, less convenient way to select the correct state would be to have inp_KQ_mask contain 1.0f for each state used by a token and 0.0f otherwise. This complicates quickly fetching the first used state of a token, and is also less efficient because a whole row of the mask would always need to be read for each token. Using indexes makes it easy to stop searching when there are no more sequences for a token, and the first sequence assigned is always very quickly available (it's the first element of each row). * mamba : support llama_kv_cache_seq_cp copy chains * mamba : support shifting and dividing the kv cache pos * mamba : make the server and parallel examples work with whole sequences A seq_id is dedicated to the system prompt in both cases. * llama : make llama_kv_cache_seq_rm return whether it succeeded or not * mamba : dedicate an input tensor for state copy indices This is cleaner and makes it easier to adapt when/if token positions (and by extension, inp_K_shift) are no longer integers. * mamba : adapt perplexity, batched, and batched-bench examples * perplexity : limit the max number of sequences This adapts to what the loaded model can provide. * llama : add llama_n_max_seq to get the upper limit for seq_ids Used by the perplexity example. * batched : pass n_parallel to the model's context params This should have been there already, but it wasn't. * batched-bench : reserve sequences to support Mamba * batched-bench : fix tokens being put in wrong sequences Generation quality isn't what's measured in there anyway, but at least using the correct sequences avoids using non-consecutive token positions. * mamba : stop abusing attention metadata This breaks existing converted-to-GGUF Mamba models, but will allow supporting mixed architectures like MambaFormer without needing to break Mamba models. This will also allow changing the size of Mamba's states without having to reconvert models in the future. (e.g. using something else than d_conv - 1 columns for the conv_states will not require breaking existing converted Mamba models again) * gguf-py : add new KV metadata key-value pairs for Mamba * llama : add new metadata key-value pairs for Mamba * llama : guard against divisions by zero when n_head is 0 * mamba : rename "unlimited" KV cache property to "recurrent" * mamba : more correctly update the "used" field of the KV cache * ggml : in ggml_ssm_scan, use a threshold for soft_plus This is how the official Mamba implementation does it, and it's also what torch.nn.Softplus does. * convert : for Mamba, fallback to internal NeoX tokenizer The resulting models are exactly the same as if the tokenizer.json and tokenizer_config.json of GPT-NeoX were there. * mamba : support state saving and restoring * ggml : implicitly pass src tensors through dst for Mamba-related ops * mamba : clarify some comments * server : fix cache_tokens not getting correctly resized Otherwise, when the "we have to evaluate at least 1 token" special case was triggered, an extra token was kept in cache_tokens even if it was removed from the KV cache. For Mamba, this caused useless prompt reprocessing when the previous request triggered the above case. * convert-hf : support new metadata keys for Mamba For the models available at https://huggingface.co/collections/state-spaces/transformers-compatible-mamba-65e7b40ab87e5297e45ae406 * mamba : rename metadata to be more similar to transformers library This breaks existing converted-to-GGUF models, but the metadata names are more "standard". * mamba : support mamba-*-hf models These models share their token_embd.weight with their output.weight * mamba : add missing spaces This is purely a formatting change. * convert-hf : omit output.weight when identical with token_embd.weight Only for Mamba for now, but it might be relevant for other models eventually. Most Mamba models actually share these two tensors, albeit implicitly. * readme : add Mamba to supported models, and add recent API changes * mamba : move state_seq and state_mask views outside layer loop A few tensors were also missing `struct` in front of `ggml_tensor`. --- README.md | 2 + common/common.cpp | 1 + convert-hf-to-gguf.py | 118 ++++ examples/batched-bench/batched-bench.cpp | 13 +- examples/batched/batched.cpp | 3 +- examples/parallel/parallel.cpp | 20 +- examples/perplexity/perplexity.cpp | 9 +- examples/server/server.cpp | 51 +- ggml.c | 379 +++++++++++- ggml.h | 19 + gguf-py/gguf/constants.py | 41 ++ gguf-py/gguf/gguf_writer.py | 12 + gguf-py/gguf/tensor_mapping.py | 46 +- llama.cpp | 698 +++++++++++++++++++++-- llama.h | 4 +- 15 files changed, 1342 insertions(+), 74 deletions(-) diff --git a/README.md b/README.md index f754022de..d7dba73e6 100644 --- a/README.md +++ b/README.md @@ -10,6 +10,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Recent API changes +- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796 - [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849 @@ -110,6 +111,7 @@ Typically finetunes of the base models below are supported as well. - [x] [InternLM2](https://huggingface.co/models?search=internlm2) - [x] [CodeShell](https://github.com/WisdomShell/codeshell) - [x] [Gemma](https://ai.google.dev/gemma) +- [x] [Mamba](https://github.com/state-spaces/mamba) **Multimodal models:** diff --git a/common/common.cpp b/common/common.cpp index c244db644..d7f650ef4 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1288,6 +1288,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param cparams.n_ctx = params.n_ctx; cparams.n_batch = params.n_batch; + cparams.n_parallel = params.n_parallel; cparams.n_threads = params.n_threads; cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; cparams.seed = params.seed; diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index f6369af38..5eee32016 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -1847,6 +1847,124 @@ class StarCoder2Model(Model): model_arch = gguf.MODEL_ARCH.STARCODER2 +@Model.register("MambaForCausalLM", "MambaLMHeadModel") +class MambaModel(Model): + model_arch = gguf.MODEL_ARCH.MAMBA + + def set_vocab(self): + vocab_size = self.hparams["vocab_size"] + # Round vocab size to next multiple of 8 + pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8) + # pad using ceiling division + # ref: https://stackoverflow.com/a/17511341/22827863 + vocab_size = -(vocab_size // -pad_vocab) * pad_vocab + self.hparams["vocab_size"] = vocab_size + + if (self.dir_model / "tokenizer.json").is_file(): + self._set_vocab_gpt2() + else: + # Use the GPT-NeoX tokenizer when no tokenizer files are present + tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf" + print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'") + neox_reader = gguf.GGUFReader(tokenizer_path, "r") + + field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL) + self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1])) + field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST) + self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size]) + field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE) + self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size]) + field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES) + self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data]) + field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID) + self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0]) + field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID) + self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0]) + field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID) + self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0]) + + def set_gguf_parameters(self): + d_model = self.find_hparam(["hidden_size", "d_model"]) + d_conv = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4 + d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model + d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16 + # ceiling division + # ref: https://stackoverflow.com/a/17511341/22827863 + # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58 + dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16) + rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5 + + # Fail early for models which don't have a block expansion factor of 2 + assert d_inner == 2 * d_model + + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default + self.gguf_writer.add_embedding_length(d_model) + self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading + self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading + self.gguf_writer.add_block_count(self.hparams["n_layer"]) + self.gguf_writer.add_ssm_conv_kernel(d_conv) + self.gguf_writer.add_ssm_inner_size(d_inner) + self.gguf_writer.add_ssm_state_size(d_state) + self.gguf_writer.add_ssm_time_step_rank(dt_rank) + self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps) + self.gguf_writer.add_file_type(self.ftype) + + def write_tensors(self): + block_count = self.hparams["n_layer"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + tok_embd = None + tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight" + output_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT] + ".weight" + + for name, data_torch in self.get_tensors(): + 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) + + # 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() + + if name.endswith(".A_log"): + print("A_log --> A ==> " + new_name) + data_torch = -torch.exp(data_torch) + + # assuming token_embd.weight is seen before output.weight + if tok_embd is not None and new_name == output_name: + if torch.equal(tok_embd, data_torch): + print(f"{output_name} is equivalent to {tok_embd_name}, omitting") + continue + if new_name == tok_embd_name: + tok_embd = data_torch + + data = data_torch.squeeze().numpy() + + 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 big float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) 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) + + ###### CONVERSION LOGIC ###### diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 19aff18ae..dff6c68ec 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -105,6 +105,9 @@ int main(int argc, char ** argv) { ctx_params.n_threads = params.n_threads; ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; + // ensure enough sequences are available + ctx_params.n_parallel = *std::max_element(n_pl.begin(), n_pl.end()); + llama_context * ctx = llama_new_context_with_model(model, ctx_params); if (ctx == NULL) { @@ -174,10 +177,10 @@ int main(int argc, char ** argv) { llama_batch_clear(batch); - const int n_tokens = is_pp_shared ? pp : pl*pp; - - for (int i = 0; i < n_tokens; ++i) { - llama_batch_add(batch, 0, i, { 0 }, false); + for (int i = 0; i < pp; ++i) { + for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) { + llama_batch_add(batch, 0, i, { j }, false); + } } batch.logits[batch.n_tokens - 1] = true; @@ -192,7 +195,7 @@ int main(int argc, char ** argv) { if (is_pp_shared) { for (int32_t i = 1; i < pl; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, 0, pp); + llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); } } diff --git a/examples/batched/batched.cpp b/examples/batched/batched.cpp index 9be7eb56b..dde4d5a06 100644 --- a/examples/batched/batched.cpp +++ b/examples/batched/batched.cpp @@ -80,6 +80,7 @@ int main(int argc, char ** argv) { ctx_params.seed = 1234; ctx_params.n_ctx = n_kv_req; ctx_params.n_batch = std::max(n_len, n_parallel); + ctx_params.n_parallel = n_parallel; ctx_params.n_threads = params.n_threads; ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; @@ -132,7 +133,7 @@ int main(int argc, char ** argv) { // assign the system KV cache to all parallel sequences // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them for (int32_t i = 1; i < n_parallel; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens); + llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); } if (n_parallel > 1) { diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index 7d11fcd59..a2ef0fb03 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -107,6 +107,9 @@ int main(int argc, char ** argv) { // number of simultaneous "clients" to simulate const int32_t n_clients = params.n_parallel; + // dedicate one sequence to the system prompt + params.n_parallel += 1; + // requests to simulate const int32_t n_seq = params.n_sequences; @@ -196,8 +199,8 @@ int main(int argc, char ** argv) { } // assign the system KV cache to all parallel sequences - for (int32_t i = 1; i < n_clients; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system); + for (int32_t i = 1; i <= n_clients; ++i) { + llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); } LOG_TEE("\n"); @@ -221,15 +224,17 @@ int main(int argc, char ** argv) { client.i_batch = batch.n_tokens; - llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true); + llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true); client.n_decoded += 1; } if (batch.n_tokens == 0) { // all sequences have ended - clear the entire KV cache - for (int i = 0; i < n_clients; ++i) { - llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1); + for (int i = 1; i <= n_clients; ++i) { + llama_kv_cache_seq_rm(ctx, i, -1, -1); + // but keep the system prompt + llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); } LOG_TEE("%s: clearing the KV cache\n", __func__); @@ -255,7 +260,7 @@ int main(int argc, char ** argv) { tokens_prompt = ::llama_tokenize(ctx, client.prompt, false); for (size_t i = 0; i < tokens_prompt.size(); ++i) { - llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false); + llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false); } // extract the logits only for the last token @@ -366,7 +371,8 @@ int main(int argc, char ** argv) { } // delete only the generated part of the sequence, i.e. keep the system prompt in the cache - llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1); + llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1); + llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1); const auto t_main_end = ggml_time_us(); diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 9ec989389..52789ee63 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -809,7 +809,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { const int n_batch = params.n_batch; const int max_tasks_per_batch = 32; - const int max_seq = 4*max_tasks_per_batch; + const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx)); llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); @@ -1086,7 +1086,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) { const int n_batch = params.n_batch; const int max_tasks_per_batch = 128; - const int max_seq = 2*max_tasks_per_batch; + const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_max_seq(ctx)); llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); @@ -1438,7 +1438,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params const int n_batch = params.n_batch; const int max_tasks_per_batch = 32; - const int max_seq = 4*max_tasks_per_batch; + const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx)); llama_batch batch = llama_batch_init(n_ctx, 0, max_seq); @@ -1815,6 +1815,9 @@ int main(int argc, char ** argv) { llama_model * model; llama_context * ctx; + // ensure there's at least enough seq_ids for HellaSwag + params.n_parallel = std::max(4, params.n_parallel); + // load the model and apply lora adapter, if any std::tie(model, ctx) = llama_init_from_gpt_params(params); if (model == NULL) { diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 109ff7175..59a59d56b 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -659,7 +659,11 @@ struct server_context { bool load_model(const gpt_params & params_) { params = params_; + // dedicate one sequence to the system prompt + params.n_parallel += 1; + std::tie(model, ctx) = llama_init_from_gpt_params(params); + params.n_parallel -= 1; // but be sneaky about it if (model == nullptr) { LOG_ERROR("unable to load model", {{"model", params.model}}); return false; @@ -1018,8 +1022,8 @@ struct server_context { } // assign the system KV cache to all parallel sequences - for (int32_t i = 1; i < params.n_parallel; ++i) { - llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size()); + for (int32_t i = 1; i <= params.n_parallel; ++i) { + llama_kv_cache_seq_cp(ctx, 0, i, -1, -1); } } @@ -1306,7 +1310,7 @@ struct server_context { const int n_embd = llama_n_embd(model); for (int i = 0; i < batch.n_tokens; ++i) { - if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) { + if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) { continue; } @@ -1633,8 +1637,8 @@ struct server_context { {"n_cache_tokens", slot.cache_tokens.size()} }); - llama_kv_cache_seq_rm (ctx, slot.id, n_keep , n_keep + n_discard); - llama_kv_cache_seq_add(ctx, slot.id, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard); + llama_kv_cache_seq_rm (ctx, slot.id + 1, n_keep , n_keep + n_discard); + llama_kv_cache_seq_add(ctx, slot.id + 1, n_keep + n_discard, system_tokens.size() + slot.n_past, -n_discard); if (slot.params.cache_prompt) { for (size_t i = n_keep + n_discard; i < slot.cache_tokens.size(); i++) { @@ -1666,7 +1670,7 @@ struct server_context { // TODO: we always have to take into account the "system_tokens" // this is not great and needs to be improved somehow - llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true); + llama_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id + 1 }, true); slot.n_past += 1; @@ -1804,9 +1808,6 @@ struct server_context { // reuse any previously computed tokens that are common with the new prompt slot.n_past = common_part(slot.cache_tokens, prompt_tokens); - // remove the non-common part from the cache - slot.cache_tokens.resize(slot.n_past); - // push the prompt into the sampling context (do not apply grammar) for (int i = 0; i < slot.n_past; ++i) { llama_sampling_accept(slot.ctx_sampling, ctx, slot.cache_tokens[i], false); @@ -1837,8 +1838,28 @@ struct server_context { } } - const int p0 = (int) system_tokens.size() + slot.n_past; - llama_kv_cache_seq_rm(ctx, slot.id, p0, -1); + // keep only the common part + int p0 = (int) system_tokens.size() + slot.n_past; + if (!llama_kv_cache_seq_rm(ctx, slot.id + 1, p0, -1)) { + // could not partially delete (likely using a non-Transformer model) + llama_kv_cache_seq_rm(ctx, slot.id + 1, -1, -1); + + p0 = (int) system_tokens.size(); + if (p0 != 0) { + // copy over the system prompt when there is one + llama_kv_cache_seq_cp(ctx, 0, slot.id + 1, -1, -1); + } + + // there is no common part left (except for the system prompt) + slot.n_past = 0; + slot.n_past_se = 0; + slot.ga_i = 0; + // TODO: is the system prompt ever in the sampling context? + llama_sampling_reset(slot.ctx_sampling); + } + + // remove the non-common part from the cache + slot.cache_tokens.resize(slot.n_past); LOG_INFO("kv cache rm [p0, end)", { { "id_slot", slot.id }, @@ -1863,7 +1884,7 @@ struct server_context { } } - llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id }, false); + llama_batch_add(batch, prompt_tokens[slot.n_past], system_tokens.size() + slot_npast, { slot.id + 1 }, false); if (slot.params.cache_prompt) { slot.cache_tokens.push_back(prompt_tokens[slot.n_past]); @@ -1937,9 +1958,9 @@ struct server_context { 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_add(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_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd); + llama_kv_cache_seq_add(ctx, slot.id + 1, slot.ga_i, slot.n_past_se, ib * bd); + llama_kv_cache_seq_div(ctx, slot.id + 1, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n); + llama_kv_cache_seq_add(ctx, slot.id + 1, slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd); slot.n_past_se -= bd; diff --git a/ggml.c b/ggml.c index 92b17ee6e..6eff98ab6 100644 --- a/ggml.c +++ b/ggml.c @@ -1841,6 +1841,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "FLASH_ATTN", "FLASH_FF", "FLASH_ATTN_BACK", + "SSM_CONV", + "SSM_SCAN", "WIN_PART", "WIN_UNPART", "GET_REL_POS", @@ -1863,7 +1865,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); +static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1929,6 +1931,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "flash_attn(x)", "flash_ff(x)", "flash_attn_back(x)", + "ssm_conv(x)", + "ssm_scan(x)", "win_part(x)", "win_unpart(x)", "get_rel_pos(x)", @@ -1951,7 +1955,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 74, "GGML_OP_COUNT != 74"); +static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -6154,6 +6158,108 @@ struct ggml_tensor * ggml_flash_attn_back( return result; } +// ggml_ssm_conv + +struct ggml_tensor * ggml_ssm_conv( + struct ggml_context * ctx, + struct ggml_tensor * s, + struct ggml_tensor * x, + struct ggml_tensor * c, + struct ggml_tensor * sq) { + GGML_ASSERT(ggml_is_3d(s)); + GGML_ASSERT(ggml_is_matrix(x)); + GGML_ASSERT(ggml_is_matrix(c)); + GGML_ASSERT(ggml_is_matrix(sq)); + GGML_ASSERT(sq->type == GGML_TYPE_I32); + + const int64_t d_conv = c->ne[0]; + const int64_t d_inner = c->ne[1]; + const int64_t n_tokens = x->ne[1]; + const int64_t n_kv = s->ne[2]; + + GGML_ASSERT( s->ne[0] == d_conv - 1); + GGML_ASSERT( s->ne[1] == d_inner); + GGML_ASSERT( x->ne[0] == d_inner); + GGML_ASSERT(sq->ne[0] == n_kv); + GGML_ASSERT(sq->ne[1] == n_tokens); + + bool is_node = false; + + if (s->grad || x->grad || c->grad || sq->grad) { + GGML_ASSERT(false); // TODO: implement + is_node = true; + } + + // 2-in-1 concatenated x and conv_states, {d_inner, n_tokens} with {d_conv, d_inner, n_kv} + struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, (d_inner*n_tokens) + (d_conv*d_inner*n_kv)); + + result->op = GGML_OP_SSM_CONV; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = s; + result->src[1] = x; + result->src[2] = c; + result->src[3] = sq; + + return result; +} + +// ggml_ssm_scan + +struct ggml_tensor * ggml_ssm_scan( + struct ggml_context * ctx, + struct ggml_tensor * s, + struct ggml_tensor * x, + struct ggml_tensor * dt, + struct ggml_tensor * A, + struct ggml_tensor * B, + struct ggml_tensor * C, + struct ggml_tensor * sq) { + GGML_ASSERT(ggml_is_contiguous(s)); + GGML_ASSERT(ggml_is_contiguous(x)); + GGML_ASSERT(ggml_is_contiguous(dt)); + GGML_ASSERT(ggml_is_contiguous(A)); + GGML_ASSERT(sq->type == GGML_TYPE_I32); + GGML_ASSERT(B->nb[0] == ggml_type_size(B->type)); + GGML_ASSERT(C->nb[0] == ggml_type_size(C->type)); + GGML_ASSERT(ggml_are_same_shape(x, dt)); + + { + const int64_t d_state = s->ne[0]; + const int64_t d_inner = s->ne[1]; + const int64_t n_tokens = x->ne[1]; + + GGML_ASSERT(x->ne[0] == d_inner); + GGML_ASSERT(A->ne[0] == d_state); + GGML_ASSERT(A->ne[1] == d_inner); + GGML_ASSERT(B->ne[0] == d_state); + GGML_ASSERT(B->ne[1] == n_tokens); + GGML_ASSERT(C->ne[0] == d_state); + GGML_ASSERT(C->ne[1] == n_tokens); + } + + bool is_node = false; + + if (s->grad || x->grad || dt->grad || A->grad || B->grad || C->grad || sq->grad) { + GGML_ASSERT(false); // TODO: implement + is_node = true; + } + + // 2-in-1 concatenated y and ssm_states, {d_inner, n_tokens} with {d_state, d_inner, n_kv} + struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s)); + + result->op = GGML_OP_SSM_SCAN; + result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; + result->src[0] = s; + result->src[1] = x; + result->src[2] = dt; + result->src[3] = A; + result->src[4] = B; + result->src[5] = C; + result->src[6] = sq; + + return result; +} + // ggml_win_part struct ggml_tensor * ggml_win_part( @@ -14771,6 +14877,257 @@ static void ggml_compute_forward_flash_attn_back( } } +// ggml_compute_forward_ssm_conv + +static void ggml_compute_forward_ssm_conv_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { + return; + } + + const struct ggml_tensor * src0 = dst->src[0]; // conv_state + const struct ggml_tensor * src1 = dst->src[1]; // x + const struct ggml_tensor * src2 = dst->src[2]; // conv1d.weight + const struct ggml_tensor * src3 = dst->src[3]; // state_seq + + const int ith = params->ith; + const int nth = params->nth; + + const int nc = src2->ne[0]; // d_conv + const int nr = src0->ne[1]; // d_inner + const int n_t = src1->ne[1]; // n_tokens + const int n_kv = src0->ne[2]; // max number of sequences in the batch + + GGML_ASSERT((nr*n_t) + (nc*nr*n_kv) == ggml_nelements(dst)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src2->nb[0] == sizeof(float)); + GGML_ASSERT(src3->nb[0] == sizeof(int32_t)); + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + // for use with the destination state offset between sequences + GGML_ASSERT(src2->nb[2] == src2->ne[1]*src2->ne[0]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + if (n_kv > 1) { + // multiple sequences means it's hard to know when it's the first time a state is read, + // so copy them all over to the destination, just to be sure. + for (int i3 = 0; i3 < n_kv; ++i3) { + float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); + float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + i3*(src2->nb[2]) + nr*n_t*sizeof(float)); + // can't use memcpy because of d_conv vs d_conv - 1 + for (int i1 = 0; i1 < ir; ++i1) { + for (int i0 = 0; i0 < nc - 1; ++i0) { + // copy s0 to last (d_conv - 1) columns of s + s[1 + i0 + i1*nc] = s0[i0 + i1*(nc - 1)]; + } + } + } + } + + for (int i2 = 0; i2 < n_t; ++i2) { + int32_t * sq = (int32_t *) ((char *) src3->data + i2*(src3->nb[1])); // {n_kv, n_tokens} + float * x = (float *) ((char *) dst->data + ir0*sizeof(float) + i2*(nr*sizeof(float))); // {d_inner, n_tokens} + float * s = (float *) ((char *) dst->data + ir0*(src2->nb[1]) + sq[0]*(src2->nb[2]) + nr*n_t*sizeof(float)); // {d_conv, d_inner, n_kv} + float * s0; // {d_conv - 1, d_inner, n_kv} + float * x0 = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} + float * c = (float *) ((char *) src2->data + ir0*(src2->nb[1])); // {d_conv, d_inner} + int ne0s0; + + GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv); + + // avoid needing to copy the state for the first token + if (i2 == 0) { + s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_conv - 1, d_inner, n_kv} + ne0s0 = src0->ne[0]; + } else { + // the source is the last (d_conv - 1) columns of the destination + s0 = s + 1; + ne0s0 = nc; + } + + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // shift state left + for (int i0 = 0; i0 < nc - 1; ++i0) { + s[i0 + i1*nc] = s0[i0 + i1*ne0s0]; + } + // insert x on the last column + s[(nc - 1) + i1*nc] = x0[i1]; + } + + // handle copies when there are multiple output states + for (int i3 = 1; i3 < n_kv; ++i3) { + int32_t seq = sq[i3]; + if (0 <= seq && seq < n_kv) { + float * s1 = s + (seq - sq[0])*nc*nr; + memcpy(s1, s, nc*ir*sizeof(float)); + } else { + // stop at negative or too big seq_ids + break; + } + } + + // it seems a little faster when this is separate from the state shift + for (int i1 = 0; i1 < ir; ++i1) { + // rowwise dot product + float sumf = 0.0f; + for (int i0 = 0; i0 < nc; ++i0) { + int i = i0 + i1*nc; + sumf += s[i] * c[i]; + } + x[i1] = sumf; + } + } +} + +static void ggml_compute_forward_ssm_conv( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_conv_f32(params, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + +// ggml_compute_forward_ssm_scan + +static void ggml_compute_forward_ssm_scan_f32( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) { + return; + } + + const struct ggml_tensor * src0 = dst->src[0]; // s + const struct ggml_tensor * src1 = dst->src[1]; // x + const struct ggml_tensor * src2 = dst->src[2]; // dt + const struct ggml_tensor * src3 = dst->src[3]; // A + const struct ggml_tensor * src4 = dst->src[4]; // B + const struct ggml_tensor * src5 = dst->src[5]; // C + const struct ggml_tensor * src6 = dst->src[6]; // sq + + const int ith = params->ith; + const int nth = params->nth; + + const int64_t nc = src0->ne[0]; // d_state + const int64_t nr = src0->ne[1]; // d_inner + const int64_t n_t = src1->ne[1]; // number of tokens in the batch + const int64_t n_kv = src0->ne[2]; // max number of sequences in the batch + + GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst)); + GGML_ASSERT(src0->nb[0] == sizeof(float)); + GGML_ASSERT(src1->nb[0] == sizeof(float)); + GGML_ASSERT(src2->nb[0] == sizeof(float)); + GGML_ASSERT(src3->nb[0] == sizeof(float)); + GGML_ASSERT(src4->nb[0] == sizeof(float)); + GGML_ASSERT(src5->nb[0] == sizeof(float)); + // required for the dot product between s and C, and when copying the states + GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float)); + // required for per-sequence offsets for states + GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float)); + // required to get correct offset for state destination (i.e. src1->nb[2]) + GGML_ASSERT(src1->nb[2] == src1->ne[0]*src1->ne[1]*sizeof(float)); + + // rows per thread + const int dr = (nr + nth - 1)/nth; + + // row range for this thread + const int ir0 = dr*ith; + const int ir1 = MIN(ir0 + dr, nr); + const int ir = ir1 - ir0; + + if (n_kv > 1) { + // it's hard to know if the source states have already been copied + // when there are multiple, so copy them already. + for (int i3 = 0; i3 < n_kv; ++i3) { + float * s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); + float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[2]); + memcpy(s, s0, nc*ir*sizeof(float)); + } + } + + for (int i2 = 0; i2 < n_t; ++i2) { + int32_t * sq = (int32_t *) ((char *) src6->data + i2*(src6->nb[1])); // {n_kv, n_tokens} + float * y = (float *) ((char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} + float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2]) + src1->nb[2]); // {d_state, d_inner, n_kv} + float * s0; + float * x = (float *) ((char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1])); // {d_inner, n_tokens} + float * dt = (float *) ((char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1])); // {d_inner, n_tokens} + float * A = (float *) ((char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner} + float * B = (float *) ((char *) src4->data + i2*(src4->nb[1])); // {d_state, n_tokens} + float * C = (float *) ((char *) src5->data + i2*(src5->nb[1])); // {d_state, n_tokens} + + GGML_ASSERT(0 <= sq[0] && sq[0] < n_kv); + + // avoid needing to copy the state for the first token + if (i2 == 0) { + s0 = (float *) ((char *) src0->data + ir0*(src0->nb[1]) + sq[0]*(src0->nb[2])); // {d_state, d_inner, n_kv} + } else { + // otherwise the source is the same as the destination + s0 = s; + } + + // d_inner + for (int i1 = 0; i1 < ir; ++i1) { + // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78 + float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1]; + float x_dt = x[i1] * dt_soft_plus; + float sumf = 0.0f; + // d_state + for (int i0 = 0; i0 < nc; ++i0) { + int i = i0 + i1*nc; + // state = prev_state * dA + dB * x + float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt); + // y = rowwise_dotprod(state, C) + sumf += state * C[i0]; + s[i] = state; + } + y[i1] = sumf; + } + + // handle copies when there are multiple output states + for (int i3 = 1; i3 < n_kv; ++i3) { + int32_t seq = sq[i3]; + if (0 <= seq && seq < n_kv) { + float * s1 = s + (seq - sq[0])*nc*nr; + memcpy(s1, s, nc*ir*sizeof(float)); + } else { + // stop at negative or too big seq_ids + break; + } + } + } +} + +static void ggml_compute_forward_ssm_scan( + const struct ggml_compute_params * params, + struct ggml_tensor * dst) { + switch (dst->src[0]->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_ssm_scan_f32(params, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_win_part static void ggml_compute_forward_win_part_f32( @@ -15830,6 +16187,14 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm bool masked = t != 0; ggml_compute_forward_flash_attn_back(params, masked, tensor); } break; + case GGML_OP_SSM_CONV: + { + ggml_compute_forward_ssm_conv(params, tensor); + } break; + case GGML_OP_SSM_SCAN: + { + ggml_compute_forward_ssm_scan(params, tensor); + } break; case GGML_OP_WIN_PART: { ggml_compute_forward_win_part(params, tensor); @@ -16884,6 +17249,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // not supported } break; + case GGML_OP_SSM_CONV: + case GGML_OP_SSM_SCAN: + { + GGML_ASSERT(false); // TODO: not implemented + } break; case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: case GGML_OP_UNARY: @@ -17590,6 +17960,11 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { { n_tasks = n_threads; } break; + case GGML_OP_SSM_CONV: + case GGML_OP_SSM_SCAN: + { + n_tasks = n_threads; + } break; case GGML_OP_WIN_PART: case GGML_OP_WIN_UNPART: case GGML_OP_GET_REL_POS: diff --git a/ggml.h b/ggml.h index 0ea4f8847..a13b0cec4 100644 --- a/ggml.h +++ b/ggml.h @@ -472,6 +472,8 @@ extern "C" { GGML_OP_FLASH_ATTN, GGML_OP_FLASH_FF, GGML_OP_FLASH_ATTN_BACK, + GGML_OP_SSM_CONV, + GGML_OP_SSM_SCAN, GGML_OP_WIN_PART, GGML_OP_WIN_UNPART, GGML_OP_GET_REL_POS, @@ -1728,6 +1730,23 @@ extern "C" { struct ggml_tensor * c0, struct ggml_tensor * c1); + GGML_API struct ggml_tensor * ggml_ssm_conv( + struct ggml_context * ctx, + struct ggml_tensor * s, + struct ggml_tensor * x, + struct ggml_tensor * c, + struct ggml_tensor * sq); + + GGML_API struct ggml_tensor * ggml_ssm_scan( + struct ggml_context * ctx, + struct ggml_tensor * s, + struct ggml_tensor * x, + struct ggml_tensor * dt, + struct ggml_tensor * A, + struct ggml_tensor * B, + struct ggml_tensor * C, + struct ggml_tensor * sq); + // partition into non-overlapping windows with padding if needed // example: // a: 768 64 64 1 diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index a62139811..b23badb10 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -61,6 +61,12 @@ class Keys: SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" + class SSM: + CONV_KERNEL = "{arch}.ssm.conv_kernel" + INNER_SIZE = "{arch}.ssm.inner_size" + STATE_SIZE = "{arch}.ssm.state_size" + TIME_STEP_RANK = "{arch}.ssm.time_step_rank" + class Tokenizer: MODEL = "tokenizer.ggml.model" LIST = "tokenizer.ggml.tokens" @@ -113,6 +119,7 @@ class MODEL_ARCH(IntEnum): MINICPM = auto() GEMMA = auto() STARCODER2 = auto() + MAMBA = auto() class MODEL_TENSOR(IntEnum): @@ -144,6 +151,13 @@ class MODEL_TENSOR(IntEnum): ATTN_Q_NORM = auto() ATTN_K_NORM = auto() LAYER_OUT_NORM = auto() + SSM_IN = auto() + SSM_CONV1D = auto() + SSM_X = auto() + SSM_DT = auto() + SSM_A = auto() + SSM_D = auto() + SSM_OUT = auto() MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { @@ -171,6 +185,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.MINICPM: "minicpm", MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.STARCODER2: "starcoder2", + MODEL_ARCH.MAMBA: "mamba", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -202,6 +217,13 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = { MODEL_TENSOR.FFN_DOWN_EXP: "blk.{bid}.ffn_down.{xid}", MODEL_TENSOR.FFN_UP_EXP: "blk.{bid}.ffn_up.{xid}", MODEL_TENSOR.LAYER_OUT_NORM: "blk.{bid}.layer_output_norm", + MODEL_TENSOR.SSM_IN: "blk.{bid}.ssm_in", + MODEL_TENSOR.SSM_CONV1D: "blk.{bid}.ssm_conv1d", + MODEL_TENSOR.SSM_X: "blk.{bid}.ssm_x", + MODEL_TENSOR.SSM_DT: "blk.{bid}.ssm_dt", + MODEL_TENSOR.SSM_A: "blk.{bid}.ssm_a", + MODEL_TENSOR.SSM_D: "blk.{bid}.ssm_d", + MODEL_TENSOR.SSM_OUT: "blk.{bid}.ssm_out", } MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { @@ -543,6 +565,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.FFN_DOWN, MODEL_TENSOR.FFN_UP, ], + MODEL_ARCH.MAMBA: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.SSM_IN, + MODEL_TENSOR.SSM_CONV1D, + MODEL_TENSOR.SSM_X, + MODEL_TENSOR.SSM_DT, + MODEL_TENSOR.SSM_A, + MODEL_TENSOR.SSM_D, + MODEL_TENSOR.SSM_OUT, + ], # TODO } @@ -734,6 +769,12 @@ KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED +# SSM +KEY_SSM_CONV_KERNEL = Keys.SSM.CONV_KERNEL +KEY_SSM_INNER_SIZE = Keys.SSM.INNER_SIZE +KEY_SSM_STATE_SIZE = Keys.SSM.STATE_SIZE +KEY_SSM_TIME_STEP_RANK = Keys.SSM.TIME_STEP_RANK + # tokenization KEY_TOKENIZER_MODEL = Keys.Tokenizer.MODEL KEY_TOKENIZER_LIST = Keys.Tokenizer.LIST diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 801160832..e49c5db68 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -382,6 +382,18 @@ class GGUFWriter: def add_rope_scaling_finetuned(self, value: bool) -> None: self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value) + def add_ssm_conv_kernel(self, value: int) -> None: + self.add_uint32(Keys.SSM.CONV_KERNEL.format(arch=self.arch), value) + + def add_ssm_inner_size(self, value: int) -> None: + self.add_uint32(Keys.SSM.INNER_SIZE.format(arch=self.arch), value) + + def add_ssm_state_size(self, value: int) -> None: + self.add_uint32(Keys.SSM.STATE_SIZE.format(arch=self.arch), value) + + def add_ssm_time_step_rank(self, value: int) -> None: + self.add_uint32(Keys.SSM.TIME_STEP_RANK.format(arch=self.arch), value) + def add_tokenizer_model(self, model: str) -> None: self.add_string(Keys.Tokenizer.MODEL, model) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index db2ec9704..ed89955d8 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -20,6 +20,9 @@ class TensorNameMap: "wte", # gpt2 "transformer.embd.wte", # phi2 "model.tok_embeddings", # internlm2 + "model.embedding", # mamba-qbert + "backbone.embedding", # mamba + "backbone.embeddings", # mamba-hf ), # Token type embeddings @@ -44,7 +47,7 @@ class TensorNameMap: # Output MODEL_TENSOR.OUTPUT: ( "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen + "lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba "output", # llama-pth bloom internlm2 "word_embeddings_for_head", # persimmon "lm_head.linear", # phi2 @@ -61,6 +64,8 @@ class TensorNameMap: "language_model.encoder.final_layernorm", # persimmon "model.final_layernorm", # persimmon "lm_head.ln", # phi2 + "model.norm_f", # mamba-qbert + "backbone.norm_f", # mamba ), # Rope frequencies @@ -86,6 +91,8 @@ class TensorNameMap: "transformer.h.{bid}.ln", # phi2 "model.layers.layers.{bid}.norm", # plamo "model.layers.{bid}.attention_norm", # internlm2 + "model.layers.{bid}.norm", # mamba-qbert + "backbone.layers.{bid}.norm", # mamba ), # Attention norm 2 @@ -282,7 +289,42 @@ class TensorNameMap: MODEL_TENSOR.LAYER_OUT_NORM: ( "encoder.layer.{bid}.output.LayerNorm", # bert "encoder.layers.{bid}.norm2", # nomic-bert - ) + ), + + MODEL_TENSOR.SSM_IN: ( + "model.layers.{bid}.in_proj", + "backbone.layers.{bid}.mixer.in_proj", + ), + + MODEL_TENSOR.SSM_CONV1D: ( + "model.layers.{bid}.conv1d", + "backbone.layers.{bid}.mixer.conv1d", + ), + + MODEL_TENSOR.SSM_X: ( + "model.layers.{bid}.x_proj", + "backbone.layers.{bid}.mixer.x_proj", + ), + + MODEL_TENSOR.SSM_DT: ( + "model.layers.{bid}.dt_proj", + "backbone.layers.{bid}.mixer.dt_proj", + ), + + MODEL_TENSOR.SSM_A: ( + "model.layers.{bid}.A_log", + "backbone.layers.{bid}.mixer.A_log", + ), + + MODEL_TENSOR.SSM_D: ( + "model.layers.{bid}.D", + "backbone.layers.{bid}.mixer.D", + ), + + MODEL_TENSOR.SSM_OUT: ( + "model.layers.{bid}.out_proj", + "backbone.layers.{bid}.mixer.out_proj", + ), } mapping: dict[str, tuple[MODEL_TENSOR, str]] diff --git a/llama.cpp b/llama.cpp index 4a20b7928..8c147a42b 100644 --- a/llama.cpp +++ b/llama.cpp @@ -213,6 +213,7 @@ enum llm_arch { LLM_ARCH_MINICPM, LLM_ARCH_GEMMA, LLM_ARCH_STARCODER2, + LLM_ARCH_MAMBA, LLM_ARCH_UNKNOWN, }; @@ -241,6 +242,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_MINICPM, "minicpm" }, { LLM_ARCH_GEMMA, "gemma" }, { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -284,6 +286,11 @@ enum llm_kv { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, LLM_KV_ROPE_SCALING_FINETUNED, + LLM_KV_SSM_INNER_SIZE, + LLM_KV_SSM_CONV_KERNEL, + LLM_KV_SSM_STATE_SIZE, + LLM_KV_SSM_TIME_STEP_RANK, + LLM_KV_TOKENIZER_MODEL, LLM_KV_TOKENIZER_LIST, LLM_KV_TOKENIZER_TOKEN_TYPE, @@ -342,6 +349,11 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, + { LLM_KV_SSM_CONV_KERNEL, "%s.ssm.conv_kernel" }, + { LLM_KV_SSM_INNER_SIZE, "%s.ssm.inner_size" }, + { LLM_KV_SSM_STATE_SIZE, "%s.ssm.state_size" }, + { LLM_KV_SSM_TIME_STEP_RANK, "%s.ssm.time_step_rank" }, + { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, { LLM_KV_TOKENIZER_TOKEN_TYPE, "tokenizer.ggml.token_type" }, @@ -399,6 +411,13 @@ enum llm_tensor { LLM_TENSOR_ATTN_Q_NORM, LLM_TENSOR_ATTN_K_NORM, LLM_TENSOR_LAYER_OUT_NORM, + LLM_TENSOR_SSM_IN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_X, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_A, + LLM_TENSOR_SSM_D, + LLM_TENSOR_SSM_OUT, }; static const std::map> LLM_TENSOR_NAMES = { @@ -801,6 +820,22 @@ static const std::map> LLM_TENSOR_NA { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, }, }, + { + LLM_ARCH_MAMBA, + { + { 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_SSM_IN, "blk.%d.ssm_in" }, + { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, + { LLM_TENSOR_SSM_X, "blk.%d.ssm_x" }, + { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, + { LLM_TENSOR_SSM_A, "blk.%d.ssm_a" }, + { LLM_TENSOR_SSM_D, "blk.%d.ssm_d" }, + { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -1613,6 +1648,12 @@ struct llama_hparams { float rope_freq_scale_train; uint32_t n_yarn_orig_ctx; + // for State Space Models + uint32_t ssm_d_conv = 0; + uint32_t ssm_d_inner = 0; + uint32_t ssm_d_state = 0; + uint32_t ssm_dt_rank = 0; + float f_clamp_kqv = 0.0f; float f_max_alibi_bias = 0.0f; @@ -1641,6 +1682,11 @@ struct llama_hparams { if (this->rope_finetuned != other.rope_finetuned) return true; if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; + if (this->ssm_d_conv != other.ssm_d_conv) return true; + if (this->ssm_d_inner != other.ssm_d_inner) return true; + if (this->ssm_d_state != other.ssm_d_state) return true; + if (this->ssm_dt_rank != other.ssm_dt_rank) return true; + const float EPSILON = 1e-9f; if (!is_float_close(this->f_norm_eps, other.f_norm_eps, EPSILON)) return true; @@ -1652,6 +1698,9 @@ struct llama_hparams { } uint32_t n_gqa() const { + if (n_head_kv == 0) { + return 0; + } return n_head/n_head_kv; } @@ -1662,6 +1711,18 @@ struct llama_hparams { uint32_t n_embd_v_gqa() const { // dimension of value embeddings across all k-v heads return n_embd_head_v * n_head_kv; } + + uint32_t n_embd_k_s() const { // dimension of the rolling state embeddings + // corresponds to Mamba's conv_states size + // TODO: maybe support other convolution strides than 1 + // NOTE: since the first column of the conv_state is shifted out each time, it's not actually needed + return (ssm_d_conv > 0 ? ssm_d_conv - 1 : 0) * ssm_d_inner; + } + + uint32_t n_embd_v_s() const { // dimension of the recurrent state embeddings + // corresponds to Mamba's ssm_states size + return ssm_d_state * ssm_d_inner; + } }; struct llama_cparams { @@ -1739,11 +1800,27 @@ struct llama_layer { struct ggml_tensor * ffn_down_b; // b2 struct ggml_tensor * ffn_up_b; // b3 struct ggml_tensor * ffn_act; + + // mamba proj + struct ggml_tensor * ssm_in; + struct ggml_tensor * ssm_x; + struct ggml_tensor * ssm_dt; + struct ggml_tensor * ssm_out; + + // mamba + struct ggml_tensor * ssm_conv1d; + struct ggml_tensor * ssm_a; + struct ggml_tensor * ssm_d; + + // mamba bias + struct ggml_tensor * ssm_conv1d_b; + struct ggml_tensor * ssm_dt_b; }; struct llama_kv_cell { llama_pos pos = -1; llama_pos delta = 0; + int32_t src = 0; // used by recurrent state models to copy states std::set seq_id; @@ -1764,6 +1841,9 @@ struct llama_kv_cell { struct llama_kv_cache { bool has_shift = false; bool do_defrag = false; + bool do_copy = false; + // with recurrent state models, a cell can hold the state for more than one past token + bool recurrent = false; // Note: The value of head isn't only used to optimize searching // for a free KV slot. llama_decode_internal also uses it, so it @@ -2003,11 +2083,14 @@ struct llama_context { 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_KQ_pos; // F32 [n_ctx] - struct ggml_tensor * inp_K_shift; // I32 [n_ctx] + struct ggml_tensor * inp_KQ_mask; // F32 [kv_size, n_batch] + struct ggml_tensor * inp_KQ_pos; // F32 [kv_size] + struct ggml_tensor * inp_K_shift; // I32 [kv_size] struct ggml_tensor * inp_mean; // F32 [n_batch, n_batch] struct ggml_tensor * inp_cls; // I32 [n_batch] + struct ggml_tensor * inp_s_copy; // I32 [kv_size] + struct ggml_tensor * inp_s_mask; // F32 [kv_size] + struct ggml_tensor * inp_s_seq; // I32 [kv_size, n_batch] #ifdef GGML_USE_MPI ggml_mpi_context * ctx_mpi = NULL; @@ -2023,25 +2106,42 @@ static bool llama_kv_cache_init( const llama_model & model, ggml_type type_k, ggml_type type_v, - uint32_t n_ctx, + uint32_t kv_size, bool offload) { const struct llama_hparams & hparams = model.hparams; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); const int64_t n_layer = hparams.n_layer; cache.has_shift = false; + // TODO: find a nicer way to add other recurrent model architectures + cache.recurrent = model.arch == LLM_ARCH_MAMBA; + + // TODO: support mixed reccurent Transformer architectues + // NOTE: (!a || b) is a logical implication (a -> b) + GGML_ASSERT(!cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_s()); + GGML_ASSERT(!cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_s()); + GGML_ASSERT( cache.recurrent || n_embd_k_gqa == hparams.n_embd_k_gqa()); + GGML_ASSERT( cache.recurrent || n_embd_v_gqa == hparams.n_embd_v_gqa()); + cache.head = 0; - cache.size = n_ctx; + cache.size = kv_size; cache.used = 0; cache.type_k = type_k; cache.type_v = type_v; cache.cells.clear(); - cache.cells.resize(n_ctx); + cache.cells.resize(kv_size); + + if (cache.recurrent) { + // init state copy sources + for (uint32_t i = 0; i < cache.size; ++i) { + cache.cells[i].src = i; + } + } #ifdef GGML_USE_CLBLAST offload = false; @@ -2080,8 +2180,8 @@ static bool llama_kv_cache_init( for (int i = 0; i < (int) n_layer; i++) { struct ggml_context * ctx = offload ? ctx_map.at(model.buft_layer[i].buft) : cache.ctxs.front(); - ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*n_ctx); - ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*n_ctx); + ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size); + ggml_tensor * v = ggml_new_tensor_1d(ctx, type_v, n_embd_v_gqa*kv_size); ggml_format_name(k, "cache_k_l%d", i); ggml_format_name(v, "cache_v_l%d", i); cache.k_l.push_back(k); @@ -2115,6 +2215,54 @@ static bool llama_kv_cache_find_slot( const uint32_t n_ctx = cache.size; const uint32_t n_tokens = batch.n_tokens; + if (cache.recurrent) { + // For recurrent state architectures (like Mamba), + // each KV cache cell can store the state for a whole sequence. + + llama_seq_id min = cache.size - 1; + llama_seq_id max = 0; + + for (uint32_t i = 0; i < n_tokens; ++i) { + for (int32_t j = 0; j < batch.n_seq_id[i]; ++j) { + llama_seq_id seq_id = batch.seq_id[i][j]; + // make sure it's a valid seq_id + if ((uint32_t) seq_id < cache.size) { + if (seq_id > max) { + max = seq_id; + } + if (seq_id < min) { + min = seq_id; + } + // Assuming the tokens are in-order + if (batch.pos[i] != cache.cells[seq_id].pos + 1) { + // What should happen when the pos backtracks or skips a value? + // Clearing the state mid-batch would require special-casing which isn't done. + LLAMA_LOG_WARN("%s: non-consecutive token position %d after %d for sequence %d\n", + __func__, batch.pos[i], cache.cells[seq_id].pos, seq_id); + } + if (cache.cells[seq_id].pos < 0 && 0 <= batch.pos[i]) { + cache.used += 1; + } + cache.cells[seq_id].pos = batch.pos[i]; + // NOTE: seq_ids are not inserted here; they are handled when the input tensors are set + } else { + // too big seq_id + // TODO: would it be possible to resize the KV cache size instead? + LLAMA_LOG_ERROR("%s: seq_id=%d >= kv_size=%d Try using a bigger --parallel value\n", __func__, seq_id, cache.size); + return false; + } + } + } + + // allow getting the range of used cells, from head to head + n + cache.head = min; + cache.n = max - min + 1; + + // sanity check + return max >= min; + } + // otherwise, one cell per token. + if (n_tokens > n_ctx) { LLAMA_LOG_ERROR("%s: n_tokens=%d > n_ctx=%d\n", __func__, n_tokens, n_ctx); return false; @@ -2184,7 +2332,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { cache.used = 0; } -static void llama_kv_cache_seq_rm( +static bool llama_kv_cache_seq_rm( struct llama_kv_cache & cache, llama_seq_id seq_id, llama_pos p0, @@ -2194,6 +2342,25 @@ static void llama_kv_cache_seq_rm( if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); + // models like Mamba can't have a state partially erased + if (cache.recurrent) { + if (seq_id >= (int64_t) cache.size) { + // could be fatal + return false; + } + if (0 <= seq_id) { + // partial intersection is invalid + if ((0 < p0 && p0 <= cache.cells[seq_id].pos) || (0 < p1 && p1 <= cache.cells[seq_id].pos)) { + return false; + } + } else { + // seq_id is negative, then the range should include everything or nothing + if (p0 != p1 && (p0 != 0 || p1 != std::numeric_limits::max())) { + return false; + } + } + } + for (uint32_t i = 0; i < cache.size; ++i) { if (cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { if (seq_id < 0) { @@ -2215,6 +2382,8 @@ static void llama_kv_cache_seq_rm( // If we freed up a slot, set head to it so searching can start there. if (new_head != cache.size && new_head < cache.head) cache.head = new_head; + + return true; } static void llama_kv_cache_seq_cp( @@ -2226,6 +2395,29 @@ static void llama_kv_cache_seq_cp( if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); + if (cache.recurrent) { + if ((uint32_t) seq_id_dst < cache.size && (uint32_t) seq_id_src < cache.size) { + seq_id_src = cache.cells[seq_id_src].src; + GGML_ASSERT((uint32_t) seq_id_src < cache.size); + // intent to "copy from" + // supports copy chains thanks to taking the source of the source + cache.cells[seq_id_dst].src = seq_id_src; + + // preserve the "keep or clear" status of the copied sequence + if (cache.cells[seq_id_src].has_seq_id(seq_id_src)) { + cache.cells[seq_id_dst].seq_id.insert(seq_id_dst); + } else { + cache.cells[seq_id_dst].seq_id.erase(seq_id_dst); + } + + cache.do_copy = true; + + cache.cells[seq_id_dst].pos = cache.cells[seq_id_src].pos; + } + return; + } + // otherwise, this is the KV cache of a Transformer-like model + cache.head = 0; for (uint32_t i = 0; i < cache.size; ++i) { @@ -2265,6 +2457,17 @@ static void llama_kv_cache_seq_add( if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); + if (cache.recurrent) { + // for Mamba-like models, only the pos needs to be shifted + if (0 <= seq_id && seq_id < (int64_t) cache.size) { + llama_kv_cell & cell = cache.cells[seq_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos += delta; + } + } + return; + } + for (uint32_t i = 0; i < cache.size; ++i) { if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { cache.has_shift = true; @@ -2298,6 +2501,17 @@ static void llama_kv_cache_seq_div( if (p0 < 0) p0 = 0; if (p1 < 0) p1 = std::numeric_limits::max(); + if (cache.recurrent) { + // for Mamba-like models, only the pos needs to be changed + if (0 <= seq_id && seq_id < (int64_t) cache.size) { + llama_kv_cell & cell = cache.cells[seq_id]; + if (cell.has_seq_id(seq_id) && p0 <= cell.pos && cell.pos < p1) { + cell.pos /= d; + } + } + return; + } + for (uint32_t i = 0; i < cache.size; ++i) { if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) { cache.has_shift = true; @@ -3117,7 +3331,7 @@ static void llm_load_hparams( // sanity check for n_rot (optional) { - hparams.n_rot = hparams.n_embd / hparams.n_head; + hparams.n_rot = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false); @@ -3130,10 +3344,10 @@ static void llm_load_hparams( // gpt-j n_rot = rotary_dim } - hparams.n_embd_head_k = hparams.n_embd / hparams.n_head; + hparams.n_embd_head_k = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k, false); - hparams.n_embd_head_v = hparams.n_embd / hparams.n_head; + hparams.n_embd_head_v = (hparams.n_head == 0) ? 0 : hparams.n_embd / hparams.n_head; ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v, false); // arch-specific KVs @@ -3383,6 +3597,36 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_MAMBA: + { + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + switch (hparams.n_layer) { + case 24: + switch (hparams.n_embd) { + case 768: model.type = e_model::MODEL_SMALL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 48: + switch (hparams.n_embd) { + case 1024: model.type = e_model::MODEL_MEDIUM; break; + case 1536: model.type = e_model::MODEL_LARGE; break; + case 2048: model.type = e_model::MODEL_XL; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + case 64: + switch (hparams.n_embd) { + case 2560: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; default: (void)0; } @@ -3702,6 +3946,10 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); LLAMA_LOG_INFO("%s: n_yarn_orig_ctx = %u\n", __func__, hparams.n_yarn_orig_ctx); LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); + LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); + LLAMA_LOG_INFO("%s: ssm_d_inner = %u\n", __func__, hparams.ssm_d_inner); + LLAMA_LOG_INFO("%s: ssm_d_state = %u\n", __func__, hparams.ssm_d_state); + LLAMA_LOG_INFO("%s: ssm_dt_rank = %u\n", __func__, hparams.ssm_dt_rank); LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); if (ml.n_elements >= 1e12) { @@ -4609,6 +4857,57 @@ 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_MAMBA: + { + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + // only an expansion factor of 2 is supported for now + GGML_ASSERT(2 * n_embd == d_inner); + + 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}, false); + // if output is NULL, init from the input tok embed, duplicated to allow offloading + if (model.output == NULL) { + model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); + ml.n_created--; // artificial tensor + ml.size_data += ggml_nbytes(model.output); + } + } + + 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]; + + // norm + layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); + + layer.ssm_in = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}); + + layer.ssm_conv1d = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}); + layer.ssm_conv1d_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}); + + layer.ssm_x = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}); + + layer.ssm_dt = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}); + layer.ssm_dt_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}); + + // no "weight" suffix for these + layer.ssm_a = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}); + layer.ssm_d = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_SSM_D, i), {d_inner}); + + // out_proj + layer.ssm_out = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -4834,6 +5133,8 @@ static void llm_build_kv_store( const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + GGML_ASSERT(kv.size == n_ctx); + // compute the transposed [n_tokens, n_embd] V matrix struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, n_embd_v_gqa, n_tokens)); //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed @@ -5043,6 +5344,8 @@ static struct ggml_tensor * llm_build_kqv( cb(kq, "kq_soft_max_ext", il); } + GGML_ASSERT(kv.size == n_ctx); + // split cached v into n_head heads struct ggml_tensor * v = ggml_view_3d(ctx, kv.v_l[il], @@ -5190,8 +5493,8 @@ struct llm_build_context { norm_eps (hparams.f_norm_eps), norm_rms_eps (hparams.f_norm_rms_eps), n_tokens (batch.n_tokens), - n_kv (worst_case ? n_ctx : kv_self.n), - kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), + n_kv (worst_case ? kv_self.size : kv_self.n), + kv_head (worst_case ? (kv_self.recurrent ? 0 : kv_self.size - n_tokens) : kv_self.head), n_orig_ctx (cparams.n_yarn_orig_ctx), pooling_type (cparams.pooling_type), rope_type (hparams.rope_type), @@ -5220,6 +5523,8 @@ struct llm_build_context { struct ggml_cgraph * build_k_shift() { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + GGML_ASSERT(kv_self.size == n_ctx); + for (int il = 0; il < n_layer; ++il) { struct ggml_tensor * tmp = // we rotate only the first n_rot dimensions @@ -5238,6 +5543,27 @@ struct llm_build_context { return gf; } + struct ggml_cgraph * build_s_copy() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + GGML_ASSERT(kv_self.recurrent); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size); + struct ggml_tensor * ssm_states = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size); + + conv_states = ggml_get_rows(ctx0, conv_states, lctx.inp_s_copy); + ssm_states = ggml_get_rows(ctx0, ssm_states, lctx.inp_s_copy); + + // TODO: name the intermediate tensors with cb() + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, conv_states, kv_self.k_l[il])); + ggml_build_forward_expand(gf, ggml_cpy(ctx0, ssm_states, kv_self.v_l[il])); + } + + return gf; + } + struct ggml_cgraph * build_defrag(const std::vector & ids) { struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); @@ -7835,6 +8161,145 @@ struct llm_build_context { return gf; } + + struct ggml_cgraph * build_mamba() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t d_model = n_embd; + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + GGML_ASSERT(2 * d_model == d_inner); + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, lctx.inp_tokens, lctx.inp_embd, cb); + cb(inpL, "inp_embd", -1); + + struct ggml_tensor * state_mask = ggml_view_2d(ctx0, lctx.inp_s_mask, 1, n_kv, lctx.inp_s_mask->nb[0], 0); + struct ggml_tensor * state_seq = ggml_view_2d(ctx0, lctx.inp_s_seq, n_kv, n_tokens, n_kv*ggml_element_size(lctx.inp_s_seq), 0); + + for (int il = 0; il < n_layer; ++il) { + // (ab)using the KV cache to store the states + struct ggml_tensor * conv_states = ggml_reshape_2d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s(), kv_self.size); + struct ggml_tensor * ssm_states = ggml_reshape_2d(ctx0, kv_self.v_l[il], hparams.n_embd_v_s(), kv_self.size); + + // clear states of sequences which are starting at the beginning of this batch + { + conv_states = ggml_mul(ctx0, + ggml_view_2d(ctx0, conv_states, conv_states->ne[0], n_kv, conv_states->nb[1], kv_head*conv_states->nb[1]), + state_mask); + ssm_states = ggml_mul(ctx0, + ggml_view_2d(ctx0, ssm_states, ssm_states->ne[0], n_kv, ssm_states->nb[1], kv_head*ssm_states->nb[1]), + state_mask); + } + + conv_states = ggml_reshape_3d(ctx0, conv_states, d_conv - 1, d_inner, n_kv); + ssm_states = ggml_reshape_3d(ctx0, ssm_states, d_state, d_inner, n_kv); + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // {n_embd, 2*d_inner} * {n_embd, n_tokens} => {2*d_inner, n_tokens} + struct ggml_tensor * xz = ggml_mul_mat(ctx0, model.layers[il].ssm_in, cur); + // split the above in two + // => {d_inner, n_tokens} + struct ggml_tensor * x = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], 0); + struct ggml_tensor * z = ggml_view_2d(ctx0, xz, d_inner, xz->ne[1], xz->nb[1], ggml_element_size(xz)*d_inner); + + // conv + { + // Custom operator which is needed only to ease simultaneous sequence processing. + // For a single sequence, the equivalent is to concatenate the columns of conv_states and x, + // then make a self-overlapping view of that over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weigth, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // The new conv_states is the last (d_conv - 1) columns + // of the last 3rd dimensional "layer" of the self-overlapping view. + // For simultaneous sequences, it's more complicated. + struct ggml_tensor * x_conv = ggml_ssm_conv(ctx0, conv_states, x, model.layers[il].ssm_conv1d, state_seq); + + // store last (d_conv - 1) columns of the conv_state part of x_conv back into the KV cache + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_2d(ctx0, x_conv, d_conv - 1, d_inner*n_kv, d_conv*ggml_element_size(x_conv), (1+d_inner*n_tokens)*ggml_element_size(x_conv)), + ggml_view_1d(ctx0, kv_self.k_l[il], (d_conv - 1)*(d_inner)*(n_kv), kv_self.head*(d_conv - 1)*(d_inner)*ggml_element_size(x_conv)))); + + // extract x from x_conv + x = ggml_view_2d(ctx0, x_conv, d_inner, n_tokens, d_inner*ggml_element_size(x_conv), 0); + + // bias + x = ggml_add(ctx0, x, model.layers[il].ssm_conv1d_b); + + x = ggml_silu(ctx0, x); + } + + // ssm + { + // {d_inner, dt_rank + 2*d_state} * {d_inner, n_tokens} => {dt_rank + 2*d_state, n_tokens} + struct ggml_tensor * x_db = ggml_mul_mat(ctx0, model.layers[il].ssm_x, x); + // split + struct ggml_tensor * dt = ggml_view_2d(ctx0, x_db, dt_rank, n_tokens, x_db->nb[1], 0); + struct ggml_tensor * B = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*dt_rank); + struct ggml_tensor * C = ggml_view_2d(ctx0, x_db, d_state, n_tokens, x_db->nb[1], ggml_element_size(x_db)*(dt_rank+d_state)); + + // {dt_rank, d_inner} * {dt_rank, n_tokens} => {d_inner, n_tokens} + dt = ggml_mul_mat(ctx0, model.layers[il].ssm_dt, dt); + dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_tokens} and {d_state, d_inner, n_kv} combined, + // because only a single tensor can be returned. + struct ggml_tensor * y_ssm_states = ggml_ssm_scan(ctx0, ssm_states, x, dt, model.layers[il].ssm_a, B, C, state_seq); + + // store last states (the second part of y_ssm_states) + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, + ggml_view_1d(ctx0, y_ssm_states, d_state*d_inner*n_kv, d_inner*n_tokens*ggml_element_size(y_ssm_states)), + ggml_view_1d(ctx0, kv_self.v_l[il], d_state*d_inner*n_kv, kv_self.head*d_state*d_inner*ggml_element_size(ssm_states)))); + + struct ggml_tensor * y = ggml_view_2d(ctx0, y_ssm_states, d_inner, n_tokens, d_inner*ggml_element_size(y_ssm_states), 0); + + // {d_inner, n_tokens} * {d_inner} => {d_inner, n_tokens} + y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); + y = ggml_mul(ctx0, y, ggml_silu(ctx0, z)); + + // {d_inner, n_embd} * {d_inner, n_tokens} => {n_embd, n_tokens} + cur = ggml_mul_mat(ctx0, model.layers[il].ssm_out, y); + } + + // residual + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = llm_build_norm(ctx0, inpL, 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; + } }; static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector & ids) { @@ -7871,6 +8336,23 @@ static struct ggml_cgraph * llama_build_graph_k_shift(llama_context & lctx) { return result; } +static struct ggml_cgraph * llama_build_graph_s_copy(llama_context & lctx) { + llama_batch dummy; + dummy.n_tokens = 0; + + llm_build_cb cb = [&](struct ggml_tensor * , const char * , int ) { }; + + struct llm_build_context llm(lctx, dummy, cb, false); + + llm.init(); + + struct ggml_cgraph * result = llm.build_s_copy(); + + llm.free(); + + return result; +} + static struct ggml_cgraph * llama_build_graph( llama_context & lctx, const llama_batch & batch, @@ -7985,6 +8467,10 @@ static struct ggml_cgraph * llama_build_graph( { result = llm.build_starcoder2(); } break; + case LLM_ARCH_MAMBA: + { + result = llm.build_mamba(); + } break; default: GGML_ASSERT(false); } @@ -7995,19 +8481,29 @@ static struct ggml_cgraph * llama_build_graph( } static void llama_set_k_shift(llama_context & lctx) { - const auto & cparams = lctx.cparams; - - const int64_t n_ctx = cparams.n_ctx; + const int64_t kv_size = lctx.kv_self.size; 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) { + for (int i = 0; i < kv_size; ++i) { data[i] = lctx.kv_self.cells[i].delta; } } +static void llama_set_s_copy(llama_context & lctx) { + const int64_t kv_size = lctx.kv_self.size; + + assert(ggml_backend_buffer_is_host(lctx.inp_s_copy->buffer)); + + int32_t * data = (int32_t *) lctx.inp_s_copy->data; + + for (int i = 0; i < kv_size; ++i) { + data[i] = lctx.kv_self.cells[i].src; + } +} + static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { // // set input data @@ -8044,6 +8540,9 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { float * data = (float *) lctx.inp_KQ_mask->data; + // For causal attention, use only the previous KV cells + // of the correct sequence for each token of the batch. + // It's assumed that if a token in the batch has multiple sequences, they are equivalent. for (int h = 0; h < 1; ++h) { for (int j = 0; j < n_tokens; ++j) { const llama_pos pos = batch.pos[j]; @@ -8149,6 +8648,53 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) { } } } + + if (kv_self.recurrent) { + const int64_t n_kv = kv_self.n; + + { + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_mask->buffer)); + float * data = (float *) lctx.inp_s_mask->data; + + // states which are not affected by the current batch are left untouched + for (int i = 0; i < n_kv; ++i) { + llama_seq_id seq_id = i + lctx.kv_self.head; + llama_kv_cell & kv_cell = lctx.kv_self.cells[seq_id]; + bool has_self_seq = kv_cell.has_seq_id(seq_id); + + data[i] = (float) has_self_seq; + + // ensure current sequences will be kept + if (!has_self_seq && kv_cell.pos >= 0) { + kv_cell.seq_id.insert(seq_id); + } + } + } + // For Mamba (and other recurrent architectures), + // update the correct state(s)/sequence(s) for each token of the batch. + // Like with the KQ_mask, if a token in the batch has multiple sequences, + // they are assumed to be equivalent (not here, but in ggml_ssm_scan and ggml_ssm_conv). + { + const int64_t n_tokens = batch.n_tokens; + + GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_s_seq->buffer)); + int32_t * data = (int32_t *) lctx.inp_s_seq->data; + + for (int j = 0; j < n_tokens; ++j) { + const int32_t n_seq = batch.n_seq_id[j]; + GGML_ASSERT(0 < n_seq); // a token should be part of at least 1 sequence + + for (int i = 0; i < n_kv; ++i) { + if (i < n_seq) { + // for this type of model, the head is the minimum seq_id of the batch + data[j*n_kv + i] = batch.seq_id[j][i] - kv_self.head; + } else { + data[j*n_kv + i] = -1; + } + } + } + } + } } static void llama_graph_compute( @@ -8271,11 +8817,13 @@ static int llama_decode_internal( return 1; } - // a heuristic, to avoid attending the full cache if it is not yet utilized - // after enough generations, the benefit from this heuristic disappears - // if we start defragmenting the cache, the benefit from this will be more important - kv_self.n = std::min(cparams.n_ctx, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); - //kv_self.n = llama_kv_cache_cell_max(kv_self); + if (!kv_self.recurrent) { + // a heuristic, to avoid attending the full cache if it is not yet utilized + // after enough generations, the benefit from this heuristic disappears + // if we start defragmenting the cache, the benefit from this will be more important + kv_self.n = std::min(kv_self.size, std::max(32u, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32))); + //kv_self.n = llama_kv_cache_cell_max(kv_self); + } } //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); @@ -8701,6 +9249,26 @@ static void llama_kv_cache_update_internal(struct llama_context & lctx) { } } + if (lctx.kv_self.recurrent && lctx.kv_self.do_copy) { + llama_set_s_copy(lctx); + + { + ggml_cgraph * gf = llama_build_graph_s_copy(lctx); + + llama_graph_compute(lctx, gf, lctx.cparams.n_threads); + } + + { + auto & kv_self = lctx.kv_self; + + kv_self.do_copy = false; + + for (uint32_t i = 0; i < kv_self.size; ++i) { + kv_self.cells[i].src = i; + } + } + } + // defragment the KV cache if needed if (lctx.kv_self.do_defrag) { llama_kv_cache_defrag_internal(lctx); @@ -11535,6 +12103,12 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_POS_EMBD, "weight"); quantize &= name != LLM_TN(model.arch)(LLM_TENSOR_TOKEN_TYPES, "weight"); + // do not quantize Mamba's small yet 2D weights + // NOTE: can't use LLM_TN here because the layer number is not known + quantize &= name.find("ssm_conv1d.weight") == std::string::npos; + quantize &= name.find("ssm_x.weight") == std::string::npos; + quantize &= name.find("ssm_dt.weight") == std::string::npos; + enum ggml_type new_type; void * new_data; size_t new_size; @@ -11985,6 +12559,7 @@ struct llama_context_params llama_context_default_params() { /*.seed =*/ LLAMA_DEFAULT_SEED, /*.n_ctx =*/ 512, /*.n_batch =*/ 512, + /*.n_parallel =*/ 1, /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED, @@ -12146,6 +12721,7 @@ struct llama_context * llama_new_context_with_model( auto & cparams = ctx->cparams; cparams.n_batch = params.n_batch; + // TODO: maybe add n_parallel here too cparams.n_threads = params.n_threads; cparams.n_threads_batch = params.n_threads_batch; cparams.yarn_ext_factor = params.yarn_ext_factor; @@ -12203,8 +12779,18 @@ struct llama_context * llama_new_context_with_model( ctx->rng = std::mt19937(params.seed); ctx->logits_all = params.logits_all; - const ggml_type type_k = params.type_k; - const ggml_type type_v = params.type_v; + uint32_t kv_size = cparams.n_ctx; + ggml_type type_k = params.type_k; + ggml_type type_v = params.type_v; + + // Mamba only needs a constant number of KV cache cells per sequence + if (model->arch == LLM_ARCH_MAMBA) { + // Mamba needs at least as many KV cells as there are sequences kept at any time + kv_size = std::max((uint32_t) 1, params.n_parallel); + // it's probably best to keep as much precision as possible for the states + type_k = GGML_TYPE_F32; // required by ggml_ssm_conv for Mamba's conv_states + type_v = GGML_TYPE_F32; // required by ggml_ssm_scan for Mamba's ssm_states + } GGML_ASSERT(hparams.n_embd_head_k % ggml_blck_size(type_k) == 0); GGML_ASSERT(hparams.n_embd_head_v % ggml_blck_size(type_v) == 0); @@ -12304,7 +12890,7 @@ struct llama_context * llama_new_context_with_model( } ctx->backends.push_back(ctx->backend_cpu); - if (!llama_kv_cache_init(ctx->kv_self, ctx->model, type_k, type_v, cparams.n_ctx, cparams.offload_kqv)) { + if (!llama_kv_cache_init(ctx->kv_self, ctx->model, type_k, type_v, kv_size, cparams.offload_kqv)) { LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__); llama_free(ctx); return nullptr; @@ -12338,7 +12924,7 @@ struct llama_context * llama_new_context_with_model( // graph inputs { ggml_init_params init_params = { - /* .mem_size */ ggml_tensor_overhead()*8, + /* .mem_size */ ggml_tensor_overhead()*(8 + 3*(ctx->kv_self.recurrent)), /* .mem_buffer */ nullptr, /* .no_alloc */ true, }; @@ -12347,11 +12933,16 @@ struct llama_context * llama_new_context_with_model( 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_KQ_pos = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_ctx); - ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_ctx); + ctx->inp_KQ_mask = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, kv_size, cparams.n_batch); + ctx->inp_KQ_pos = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_F32, kv_size); + ctx->inp_K_shift = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, kv_size); ctx->inp_mean = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_F32, cparams.n_batch, cparams.n_batch); ctx->inp_cls = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, cparams.n_batch); + if (ctx->kv_self.recurrent) { + ctx->inp_s_copy = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_I32, kv_size); + ctx->inp_s_mask = ggml_new_tensor_1d(ctx->ctx_input, GGML_TYPE_F32, kv_size); + ctx->inp_s_seq = ggml_new_tensor_2d(ctx->ctx_input, GGML_TYPE_I32, kv_size, cparams.n_batch); + } ggml_set_name(ctx->inp_tokens, "inp_tokens"); ggml_set_name(ctx->inp_embd, "inp_embd"); @@ -12361,6 +12952,11 @@ struct llama_context * llama_new_context_with_model( ggml_set_name(ctx->inp_K_shift, "inp_K_shift"); ggml_set_name(ctx->inp_mean, "inp_mean"); ggml_set_name(ctx->inp_cls, "inp_cls"); + if (ctx->kv_self.recurrent) { + ggml_set_name(ctx->inp_s_copy, "inp_s_copy"); + ggml_set_name(ctx->inp_s_mask, "inp_s_mask"); + ggml_set_name(ctx->inp_s_seq, "inp_s_seq"); + } 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__, @@ -12447,6 +13043,10 @@ uint32_t llama_n_batch(const struct llama_context * ctx) { return ctx->cparams.n_batch; } +uint32_t llama_n_max_seq(const struct llama_context * ctx) { + return ctx->kv_self.size; +} + enum llama_vocab_type llama_vocab_type(const struct llama_model * model) { return model->vocab.type; } @@ -12460,6 +13060,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { case LLM_ARCH_MPT: case LLM_ARCH_REFACT: case LLM_ARCH_BLOOM: + case LLM_ARCH_MAMBA: return LLAMA_ROPE_TYPE_NONE; // use what we call a normal RoPE, operating on pairs of consecutive head values @@ -12713,8 +13314,8 @@ void llama_kv_cache_clear(struct llama_context * ctx) { llama_kv_cache_clear(ctx->kv_self); } -void llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) { - llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1); +bool llama_kv_cache_seq_rm(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1) { + return llama_kv_cache_seq_rm(ctx->kv_self, seq_id, p0, p1); } void llama_kv_cache_seq_cp(struct llama_context * ctx, llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) { @@ -12891,8 +13492,8 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); const size_t kv_buf_size = kv_self.total_size(); const uint32_t kv_head = llama_kv_cache_cell_max(kv_self); @@ -12913,6 +13514,17 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat ggml_backend_tensor_get(kv_self.k_l[il], tmp_buf.data(), 0, tmp_buf.size()); data_ctx->write(tmp_buf.data(), tmp_buf.size()); + if (kv_self.recurrent) { + // v is contiguous for recurrent models + // TODO: use other tensors for state models than k and v + const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); + + tmp_buf.resize(v_size); + ggml_backend_tensor_get(kv_self.v_l[il], tmp_buf.data(), 0, tmp_buf.size()); + data_ctx->write(tmp_buf.data(), tmp_buf.size()); + continue; + } + // v is not contiguous, copy row by row const size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size); @@ -13005,8 +13617,8 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { const auto & hparams = ctx->model.hparams; const uint32_t n_layer = hparams.n_layer; - const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(); - const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(); + const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa() + hparams.n_embd_k_s(); + const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa() + hparams.n_embd_v_s(); size_t kv_buf_size; uint32_t kv_head; @@ -13027,6 +13639,16 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) { ggml_backend_tensor_set(kv_self.k_l[il], inp, 0, k_size); inp += k_size; + if (kv_self.recurrent) { + // v is contiguous for recurrent models + // TODO: use other tensors for state models than k and v + const size_t v_size = ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa*kv_head); + + ggml_backend_tensor_set(kv_self.v_l[il], inp, 0, v_size); + inp += v_size; + continue; + } + // v is not contiguous, copy row by row const size_t v_row_size = ggml_row_size(kv_self.v_l[il]->type, kv_head); const size_t v_row_stride = ggml_row_size(kv_self.v_l[il]->type, kv_size); diff --git a/llama.h b/llama.h index 3dc162b07..7a107c7f3 100644 --- a/llama.h +++ b/llama.h @@ -235,6 +235,7 @@ extern "C" { uint32_t seed; // RNG seed, -1 for random uint32_t n_ctx; // text context, 0 = from model uint32_t n_batch; // prompt processing maximum batch size + uint32_t n_parallel; // number of parallel sequences (i.e. distinct states for recurrent models) uint32_t n_threads; // number of threads to use for generation uint32_t n_threads_batch; // number of threads to use for batch processing @@ -376,6 +377,7 @@ extern "C" { LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); + LLAMA_API uint32_t llama_n_max_seq (const struct llama_context * ctx); LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model); LLAMA_API enum llama_rope_type llama_rope_type (const struct llama_model * model); @@ -502,7 +504,7 @@ extern "C" { // seq_id < 0 : match any sequence // p0 < 0 : [0, p1] // p1 < 0 : [p0, inf) - LLAMA_API void llama_kv_cache_seq_rm( + LLAMA_API bool llama_kv_cache_seq_rm( struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, From fd72d2d2a5e79d61ccef6af3d15f16e5e5cbc352 Mon Sep 17 00:00:00 2001 From: Pierrick Hymbert Date: Sat, 9 Mar 2024 10:30:04 +0100 Subject: [PATCH 17/32] server: tests: add truncated prompt tests, better kv cache size (#5933) * server: tests: add truncated prompt tests, better size * server, tests : update regex --------- Co-authored-by: Georgi Gerganov --- examples/server/server.cpp | 23 +++++++++-- .../server/tests/features/parallel.feature | 5 ++- examples/server/tests/features/server.feature | 41 ++++++++++++++----- examples/server/tests/features/steps/steps.py | 35 +++++++++++++--- 4 files changed, 81 insertions(+), 23 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 59a59d56b..6f4449984 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1128,6 +1128,7 @@ struct server_context { LOG_VERBOSE("stopped by limit", { {"id_slot", slot.id}, + {"id_task", slot.id_task}, {"n_decoded", slot.n_decoded}, {"n_predict", slot.params.n_predict}, }); @@ -1141,6 +1142,8 @@ struct server_context { } LOG_VERBOSE("next token", { + {"id_slot", slot.id}, + {"id_task", slot.id_task}, {"token", result.tok}, {"token_text", tokens_to_output_formatted_string(ctx, result.tok)}, {"has_next_token", slot.has_next_token}, @@ -1750,6 +1753,15 @@ struct server_context { slot.n_past = 0; slot.n_prompt_tokens = prompt_tokens.size(); + LOG_VERBOSE("prompt tokenized", { + {"id_slot", slot.id}, + {"id_task", slot.id_task}, + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_prompt_tokens", slot.n_prompt_tokens}, + {"prompt_tokens", tokens_to_str(ctx, prompt_tokens.cbegin(), prompt_tokens.cend())}, + }); + if (slot.embedding) { // this prompt is too large to process - discard it if (slot.n_prompt_tokens > n_batch) { @@ -1788,10 +1800,13 @@ struct server_context { slot.n_prompt_tokens = prompt_tokens.size(); LOG_VERBOSE("input truncated", { - {"n_ctx", slot.n_ctx}, - {"n_keep", slot.params.n_keep}, - {"n_left", n_left}, - {"prompt_tokens", tokens_to_str(ctx, prompt_tokens.cbegin(), prompt_tokens.cend())}, + {"id_slot", slot.id}, + {"id_task", slot.id_task}, + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"n_prompt_tokens", slot.n_prompt_tokens}, + {"prompt_tokens", tokens_to_str(ctx, prompt_tokens.cbegin(), prompt_tokens.cend())}, }); GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx); diff --git a/examples/server/tests/features/parallel.feature b/examples/server/tests/features/parallel.feature index 066698c8e..a66fed626 100644 --- a/examples/server/tests/features/parallel.feature +++ b/examples/server/tests/features/parallel.feature @@ -6,8 +6,8 @@ Feature: Parallel Given a server listening on localhost:8080 And a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models And 42 as server seed - And 512 as batch size - And 64 KV cache size + And 128 as batch size + And 256 KV cache size And 2 slots And continuous batching Then the server is starting @@ -76,6 +76,7 @@ Feature: Parallel | disabled | 128 | | enabled | 64 | + Scenario: Multi users with total number of tokens to predict exceeds the KV Cache size #3969 Given a prompt: """ diff --git a/examples/server/tests/features/server.feature b/examples/server/tests/features/server.feature index 878ac1363..aa132fa34 100644 --- a/examples/server/tests/features/server.feature +++ b/examples/server/tests/features/server.feature @@ -10,11 +10,10 @@ Feature: llama.cpp server # KV Cache corresponds to the total amount of tokens # that can be stored across all independent sequences: #4130 # see --ctx-size and #5568 - And 32 KV cache size - And 512 as batch size - And 1 slots - And embeddings extraction - And 32 server max tokens to predict + And 256 KV cache size + And 32 as batch size + And 2 slots + And 64 server max tokens to predict And prometheus compatible metrics exposed Then the server is starting Then the server is healthy @@ -23,18 +22,35 @@ Feature: llama.cpp server Then the server is ready And all slots are idle + Scenario Outline: Completion Given a prompt And max tokens to predict And a completion request with no api error Then tokens are predicted matching + And the completion is truncated + And prompt tokens are processed And prometheus metrics are exposed And metric llamacpp:tokens_predicted is Examples: Prompts - | prompt | n_predict | re_content | n_predicted | - | I believe the meaning of life is | 8 | (read\|going)+ | 8 | - | Write a joke about AI | 64 | (park\|friends\|scared\|always)+ | 32 | + | prompt | n_predict | re_content | n_prompt | n_predicted | truncated | + | I believe the meaning of life is | 8 | (read\|going)+ | 18 | 8 | not | + | Write a joke about AI from a very long prompt which will not be truncated | 256 | (princesses\|everyone\|kids)+ | 46 | 64 | not | + + Scenario: Completion prompt truncated + Given a prompt: + """ + Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. + Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. + Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. + Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum. + """ + And a completion request with no api error + Then 64 tokens are predicted matching fun|Annaks|popcorns + And the completion is truncated + And 109 prompt tokens are processed + Scenario Outline: OAI Compatibility Given a model @@ -44,11 +60,14 @@ Feature: llama.cpp server And streaming is Given an OAI compatible chat completions request with no api error Then tokens are predicted matching + And prompt tokens are processed + And the completion is truncated Examples: Prompts - | model | system_prompt | user_prompt | max_tokens | re_content | n_predicted | enable_streaming | - | llama-2 | Book | What is the best book | 8 | (Mom\|what)+ | 8 | disabled | - | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 64 | (thanks\|happy\|bird)+ | 32 | enabled | + | model | system_prompt | user_prompt | max_tokens | re_content | n_prompt | n_predicted | enable_streaming | truncated | + | llama-2 | Book | What is the best book | 8 | (Here\|what)+ | 77 | 8 | disabled | not | + | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128 | (thanks\|happy\|bird)+ | -1 | 64 | enabled | | + Scenario: Tokenize / Detokenize When tokenizing: diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index d7f005836..0076f805b 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -196,12 +196,30 @@ async def step_request_completion(context, api_error): @step(u'{predicted_n:d} tokens are predicted matching {re_content}') def step_n_tokens_predicted_with_content(context, predicted_n, re_content): - assert_n_tokens_predicted(context.tasks_result.pop(), predicted_n, re_content) + context.completion = context.tasks_result.pop() + assert_n_tokens_predicted(context.completion, predicted_n, re_content) @step(u'{predicted_n:d} tokens are predicted') def step_n_tokens_predicted(context, predicted_n): - assert_n_tokens_predicted(context.tasks_result.pop(), predicted_n) + context.completion = context.tasks_result.pop() + assert_n_tokens_predicted(context.completion, predicted_n) + + +@step(u'the completion is truncated') +def step_assert_completion_truncated(context): + step_assert_completion_truncated(context, '') + + +@step(u'the completion is {truncated} truncated') +def step_assert_completion_truncated(context, truncated): + truncated = truncated != "not" + assert context.completion['truncated'] == truncated, f'{context.completion}' + + +@step(u'{n_prompt:d} prompt tokens are processed') +def step_impl(context, n_prompt): + assert n_prompt < 0 or n_prompt == context.completion['timings']['prompt_n'], f"n_prompt={context.completion['timings']['prompt_n']}" @step(u'a user prompt {user_prompt}') @@ -722,7 +740,8 @@ async def oai_chat_completions(user_prompt, completion_response = { 'content': '', 'timings': { - 'predicted_n': 0 + 'predicted_n': 0, + 'prompt_n': 0 } } if async_client: @@ -763,7 +782,8 @@ async def oai_chat_completions(user_prompt, completion_response = { 'content': chat_completion_raw['choices'][0]['message'], 'timings': { - 'predicted_n': chat_completion_raw['usage']['completion_tokens'] + 'predicted_n': chat_completion_raw['usage']['completion_tokens'], + 'prompt_n': chat_completion_raw['usage']['prompt_tokens'] } } else: @@ -792,13 +812,16 @@ async def oai_chat_completions(user_prompt, if 'content' in delta: completion_response['content'] += delta['content'] completion_response['timings']['predicted_n'] += 1 + completion_response['truncated'] = chunk.choices[0].finish_reason != 'stop' else: assert len(chat_completion.choices) == 1 completion_response = { 'content': chat_completion.choices[0].message.content, 'timings': { - 'predicted_n': chat_completion.usage.completion_tokens - } + 'predicted_n': chat_completion.usage.completion_tokens, + 'prompt_n': chat_completion.usage.prompt_tokens + }, + 'truncated': chat_completion.choices[0].finish_reason != 'stop' } if debug: print("OAI response formatted to llama.cpp:", completion_response) From e1fa9569ba8ce276bc7801a3cebdcf8b1aa116ea Mon Sep 17 00:00:00 2001 From: Gabe Goodhart Date: Sat, 9 Mar 2024 02:57:09 -0700 Subject: [PATCH 18/32] server : add SSL support (#5926) * add cmake build toggle to enable ssl support in server Signed-off-by: Gabe Goodhart * add flags for ssl key/cert files and use SSLServer if set All SSL setup is hidden behind CPPHTTPLIB_OPENSSL_SUPPORT in the same way that the base httlib hides the SSL support Signed-off-by: Gabe Goodhart * Update readme for SSL support in server Signed-off-by: Gabe Goodhart * Add LLAMA_SERVER_SSL variable setup to top-level Makefile Signed-off-by: Gabe Goodhart --------- Signed-off-by: Gabe Goodhart --- Makefile | 4 ++ examples/server/CMakeLists.txt | 6 ++ examples/server/README.md | 26 ++++++++ examples/server/server.cpp | 108 ++++++++++++++++++++++----------- 4 files changed, 110 insertions(+), 34 deletions(-) diff --git a/Makefile b/Makefile index efce10bb8..aea969222 100644 --- a/Makefile +++ b/Makefile @@ -201,6 +201,10 @@ ifdef LLAMA_SERVER_VERBOSE MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE) endif +ifdef LLAMA_SERVER_SSL + MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT + MK_LDFLAGS += -lssl -lcrypto +endif ifdef LLAMA_CODE_COVERAGE MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase '' diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index c21eba634..f94de1e99 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -1,5 +1,6 @@ set(TARGET server) option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON) +option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF) include_directories(${CMAKE_CURRENT_SOURCE_DIR}) add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h) install(TARGETS ${TARGET} RUNTIME) @@ -7,6 +8,11 @@ target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ ) target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT}) +if (LLAMA_SERVER_SSL) + find_package(OpenSSL REQUIRED) + target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto) + target_compile_definitions(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT) +endif() if (WIN32) TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32) endif() diff --git a/examples/server/README.md b/examples/server/README.md index 591f748f8..bf8c450b6 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -59,6 +59,10 @@ see https://github.com/ggerganov/llama.cpp/issues/1437 - `--log-disable`: Output logs to stdout only, default: enabled. - `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json) +**If compiled with `LLAMA_SERVER_SSL=ON`** +- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key +- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate + ## Build server is build alongside everything else from the root of the project @@ -75,6 +79,28 @@ server is build alongside everything else from the root of the project cmake --build . --config Release ``` +## Build with SSL + +server can also be built with SSL support using OpenSSL 3 + +- Using `make`: + + ```bash + # NOTE: For non-system openssl, use the following: + # CXXFLAGS="-I /path/to/openssl/include" + # LDFLAGS="-L /path/to/openssl/lib" + make LLAMA_SERVER_SSL=true server + ``` + +- Using `CMake`: + + ```bash + mkdir build + cd build + cmake .. -DLLAMA_SERVER_SSL=ON + make server + ``` + ## Quick Start To get started right away, run the following command, making sure to use the correct path for the model you have: diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 6f4449984..c3b87c846 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -27,6 +27,7 @@ #include #include #include +#include using json = nlohmann::json; @@ -118,6 +119,11 @@ struct server_params { std::vector api_keys; +#ifdef CPPHTTPLIB_OPENSSL_SUPPORT + std::string ssl_key_file = ""; + std::string ssl_cert_file = ""; +#endif + bool slots_endpoint = true; bool metrics_endpoint = false; }; @@ -2142,6 +2148,10 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" --path PUBLIC_PATH path from which to serve static files (default %s)\n", sparams.public_path.c_str()); printf(" --api-key API_KEY optional api key to enhance server security. If set, requests must include this key for access.\n"); printf(" --api-key-file FNAME path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n"); +#ifdef CPPHTTPLIB_OPENSSL_SUPPORT + printf(" --ssl-key-file FNAME path to file a PEM-encoded SSL private key\n"); + printf(" --ssl-cert-file FNAME path to file a PEM-encoded SSL certificate\n"); +#endif printf(" -to N, --timeout N server read/write timeout in seconds (default: %d)\n", sparams.read_timeout); printf(" --embeddings enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled"); printf(" -np N, --parallel N number of slots for process requests (default: %d)\n", params.n_parallel); @@ -2220,7 +2230,24 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, } } key_file.close(); - } else if (arg == "--timeout" || arg == "-to") { + + } +#ifdef CPPHTTPLIB_OPENSSL_SUPPORT + else if (arg == "--ssl-key-file") { + if (++i >= argc) { + invalid_param = true; + break; + } + sparams.ssl_key_file = argv[i]; + } else if (arg == "--ssl-cert-file") { + if (++i >= argc) { + invalid_param = true; + break; + } + sparams.ssl_cert_file = argv[i]; + } +#endif + else if (arg == "--timeout" || arg == "-to") { if (++i >= argc) { invalid_param = true; break; @@ -2658,21 +2685,34 @@ int main(int argc, char ** argv) { {"system_info", llama_print_system_info()}, }); - httplib::Server svr; + std::unique_ptr svr; +#ifdef CPPHTTPLIB_OPENSSL_SUPPORT + if (sparams.ssl_key_file != "" && sparams.ssl_cert_file != "") { + LOG_INFO("Running with SSL", {{"key", sparams.ssl_key_file}, {"cert", sparams.ssl_cert_file}}); + svr.reset( + new httplib::SSLServer(sparams.ssl_cert_file.c_str(), sparams.ssl_key_file.c_str()) + ); + } else { + LOG_INFO("Running without SSL", {}); + svr.reset(new httplib::Server()); + } +#else + svr.reset(new httplib::Server()); +#endif std::atomic state{SERVER_STATE_LOADING_MODEL}; - svr.set_default_headers({{"Server", "llama.cpp"}}); + svr->set_default_headers({{"Server", "llama.cpp"}}); // CORS preflight - svr.Options(R"(.*)", [](const httplib::Request & req, httplib::Response & res) { + svr->Options(R"(.*)", [](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); res.set_header("Access-Control-Allow-Credentials", "true"); res.set_header("Access-Control-Allow-Methods", "POST"); res.set_header("Access-Control-Allow-Headers", "*"); }); - svr.Get("/health", [&](const httplib::Request & req, httplib::Response & res) { + svr->Get("/health", [&](const httplib::Request & req, httplib::Response & res) { server_state current_state = state.load(); switch (current_state) { case SERVER_STATE_READY: @@ -2728,7 +2768,7 @@ int main(int argc, char ** argv) { }); if (sparams.slots_endpoint) { - svr.Get("/slots", [&](const httplib::Request &, httplib::Response & res) { + svr->Get("/slots", [&](const httplib::Request &, httplib::Response & res) { // request slots data using task queue server_task task; task.id = ctx_server.queue_tasks.get_new_id(); @@ -2749,7 +2789,7 @@ int main(int argc, char ** argv) { } if (sparams.metrics_endpoint) { - svr.Get("/metrics", [&](const httplib::Request &, httplib::Response & res) { + svr->Get("/metrics", [&](const httplib::Request &, httplib::Response & res) { // request slots data using task queue server_task task; task.id = ctx_server.queue_tasks.get_new_id(); @@ -2846,9 +2886,9 @@ int main(int argc, char ** argv) { }); } - svr.set_logger(log_server_request); + svr->set_logger(log_server_request); - svr.set_exception_handler([](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { + svr->set_exception_handler([](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { const char fmt[] = "500 Internal Server Error\n%s"; char buf[BUFSIZ]; @@ -2864,7 +2904,7 @@ int main(int argc, char ** argv) { res.status = 500; }); - svr.set_error_handler([](const httplib::Request &, httplib::Response & res) { + svr->set_error_handler([](const httplib::Request &, httplib::Response & res) { if (res.status == 401) { res.set_content("Unauthorized", "text/plain; charset=utf-8"); } @@ -2877,16 +2917,16 @@ int main(int argc, char ** argv) { }); // set timeouts and change hostname and port - svr.set_read_timeout (sparams.read_timeout); - svr.set_write_timeout(sparams.write_timeout); + svr->set_read_timeout (sparams.read_timeout); + svr->set_write_timeout(sparams.write_timeout); - if (!svr.bind_to_port(sparams.hostname, sparams.port)) { + if (!svr->bind_to_port(sparams.hostname, sparams.port)) { fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); return 1; } // Set the base directory for serving static files - svr.set_base_dir(sparams.public_path); + svr->set_base_dir(sparams.public_path); std::unordered_map log_data; @@ -2947,30 +2987,30 @@ int main(int argc, char ** argv) { }; // this is only called if no index.html is found in the public --path - svr.Get("/", [](const httplib::Request &, httplib::Response & res) { + svr->Get("/", [](const httplib::Request &, httplib::Response & res) { res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html; charset=utf-8"); return false; }); // this is only called if no index.js is found in the public --path - svr.Get("/index.js", [](const httplib::Request &, httplib::Response & res) { + svr->Get("/index.js", [](const httplib::Request &, httplib::Response & res) { res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript; charset=utf-8"); return false; }); // this is only called if no index.html is found in the public --path - svr.Get("/completion.js", [](const httplib::Request &, httplib::Response & res) { + svr->Get("/completion.js", [](const httplib::Request &, httplib::Response & res) { res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript; charset=utf-8"); return false; }); // this is only called if no index.html is found in the public --path - svr.Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response & res) { + svr->Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response & res) { res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8"); return false; }); - svr.Get("/props", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + svr->Get("/props", [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); json data = { { "user_name", ctx_server.name_user.c_str() }, @@ -3062,11 +3102,11 @@ int main(int argc, char ** argv) { } }; - svr.Post("/completion", completions); // legacy - svr.Post("/completions", completions); - svr.Post("/v1/completions", completions); + svr->Post("/completion", completions); // legacy + svr->Post("/completions", completions); + svr->Post("/v1/completions", completions); - svr.Get("/v1/models", [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { + svr->Get("/v1/models", [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); json models = { @@ -3161,10 +3201,10 @@ int main(int argc, char ** argv) { } }; - svr.Post("/chat/completions", chat_completions); - svr.Post("/v1/chat/completions", chat_completions); + svr->Post("/chat/completions", chat_completions); + svr->Post("/v1/chat/completions", chat_completions); - svr.Post("/infill", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { + svr->Post("/infill", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!validate_api_key(req, res)) { return; @@ -3228,11 +3268,11 @@ int main(int argc, char ** argv) { } }); - svr.Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) { + svr->Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) { return res.set_content("", "application/json; charset=utf-8"); }); - svr.Post("/tokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + svr->Post("/tokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); @@ -3244,7 +3284,7 @@ int main(int argc, char ** argv) { return res.set_content(data.dump(), "application/json; charset=utf-8"); }); - svr.Post("/detokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + svr->Post("/detokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); @@ -3258,7 +3298,7 @@ int main(int argc, char ** argv) { return res.set_content(data.dump(), "application/json; charset=utf-8"); }); - svr.Post("/embedding", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { + svr->Post("/embedding", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!params.embedding) { res.status = 501; @@ -3289,7 +3329,7 @@ int main(int argc, char ** argv) { return res.set_content(result.data.dump(), "application/json; charset=utf-8"); }); - svr.Post("/v1/embeddings", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { + svr->Post("/v1/embeddings", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!params.embedding) { res.status = 501; @@ -3360,13 +3400,13 @@ int main(int argc, char ** argv) { sparams.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1); } log_data["n_threads_http"] = std::to_string(sparams.n_threads_http); - svr.new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); }; + svr->new_task_queue = [&sparams] { return new httplib::ThreadPool(sparams.n_threads_http); }; LOG_INFO("HTTP server listening", log_data); // run the HTTP server in a thread - see comment below std::thread t([&]() { - if (!svr.listen_after_bind()) { + if (!svr->listen_after_bind()) { state.store(SERVER_STATE_ERROR); return 1; } @@ -3407,7 +3447,7 @@ int main(int argc, char ** argv) { ctx_server.queue_tasks.start_loop(); - svr.stop(); + svr->stop(); t.join(); llama_backend_free(); From 950ba1ab84db199f0bbdecdb2bb911f35261b321 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Sat, 9 Mar 2024 11:27:53 +0100 Subject: [PATCH 19/32] Server: reorganize some http logic (#5939) * refactor static file handler * use set_pre_routing_handler for validate_api_key * merge embedding handlers * correct http verb for endpoints * fix embedding response * fix test case CORS Options * fix code style --- examples/server/README.md | 4 +- examples/server/server.cpp | 725 +++++++++--------- .../server/tests/features/security.feature | 5 +- examples/server/tests/features/steps/steps.py | 3 +- 4 files changed, 379 insertions(+), 358 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index bf8c450b6..3abb1abe3 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -42,7 +42,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437 - `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`. - `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`. - `--port`: Set the port to listen. Default: `8080`. -- `--path`: path from which to serve static files (default examples/server/public) +- `--path`: path from which to serve static files (default: disabled) - `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys. - `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s. - `--embedding`: Enable embedding extraction, Default: disabled. @@ -558,7 +558,7 @@ The HTTP server supports OAI-like API ### Extending or building alternative Web Front End -The default location for the static files is `examples/server/public`. You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method. +You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method. Read the documentation in `/completion.js` to see convenient ways to access llama. diff --git a/examples/server/server.cpp b/examples/server/server.cpp index c3b87c846..6e0f8328c 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -113,7 +113,7 @@ struct server_params { int32_t n_threads_http = -1; std::string hostname = "127.0.0.1"; - std::string public_path = "examples/server/public"; + std::string public_path = ""; std::string chat_template = ""; std::string system_prompt = ""; @@ -2145,7 +2145,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n"); printf(" --host ip address to listen (default (default: %s)\n", sparams.hostname.c_str()); printf(" --port PORT port to listen (default (default: %d)\n", sparams.port); - printf(" --path PUBLIC_PATH path from which to serve static files (default %s)\n", sparams.public_path.c_str()); + printf(" --path PUBLIC_PATH path from which to serve static files (default: disabled)\n"); printf(" --api-key API_KEY optional api key to enhance server security. If set, requests must include this key for access.\n"); printf(" --api-key-file FNAME path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n"); #ifdef CPPHTTPLIB_OPENSSL_SUPPORT @@ -2211,7 +2211,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, invalid_param = true; break; } - sparams.api_keys.emplace_back(argv[i]); + sparams.api_keys.push_back(argv[i]); } else if (arg == "--api-key-file") { if (++i >= argc) { invalid_param = true; @@ -2712,7 +2712,143 @@ int main(int argc, char ** argv) { res.set_header("Access-Control-Allow-Headers", "*"); }); - svr->Get("/health", [&](const httplib::Request & req, httplib::Response & res) { + svr->set_logger(log_server_request); + + svr->set_exception_handler([](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { + const char fmt[] = "500 Internal Server Error\n%s"; + + char buf[BUFSIZ]; + try { + std::rethrow_exception(std::move(ep)); + } catch (std::exception &e) { + snprintf(buf, sizeof(buf), fmt, e.what()); + } catch (...) { + snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); + } + + res.set_content(buf, "text/plain; charset=utf-8"); + res.status = 500; + }); + + svr->set_error_handler([](const httplib::Request &, httplib::Response & res) { + if (res.status == 401) { + res.set_content("Unauthorized", "text/plain; charset=utf-8"); + } + if (res.status == 400) { + res.set_content("Invalid request", "text/plain; charset=utf-8"); + } + if (res.status == 404) { + res.set_content("File Not Found", "text/plain; charset=utf-8"); + } + }); + + // set timeouts and change hostname and port + svr->set_read_timeout (sparams.read_timeout); + svr->set_write_timeout(sparams.write_timeout); + + if (!svr->bind_to_port(sparams.hostname, sparams.port)) { + fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); + return 1; + } + + std::unordered_map log_data; + + log_data["hostname"] = sparams.hostname; + log_data["port"] = std::to_string(sparams.port); + + if (sparams.api_keys.size() == 1) { + auto key = sparams.api_keys[0]; + log_data["api_key"] = "api_key: ****" + key.substr(std::max((int)(key.length() - 4), 0)); + } else if (sparams.api_keys.size() > 1) { + log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded"; + } + + // load the model + if (!ctx_server.load_model(params)) { + state.store(SERVER_STATE_ERROR); + return 1; + } else { + ctx_server.initialize(); + state.store(SERVER_STATE_READY); + } + + LOG_INFO("model loaded", {}); + + const auto model_meta = ctx_server.model_meta(); + + if (sparams.chat_template.empty()) { // custom chat template is not supplied + if (!ctx_server.validate_model_chat_template()) { + LOG_ERROR("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); + sparams.chat_template = "chatml"; + } + } + + // + // Middlewares + // + + auto middleware_validate_api_key = [&sparams](const httplib::Request & req, httplib::Response & res) { + // TODO: should we apply API key to all endpoints, including "/health" and "/models"? + static const std::set protected_endpoints = { + "/props", + "/completion", + "/completions", + "/v1/completions", + "/chat/completions", + "/v1/chat/completions", + "/infill", + "/tokenize", + "/detokenize", + "/embedding", + "/embeddings", + "/v1/embeddings", + }; + + // If API key is not set, skip validation + if (sparams.api_keys.empty()) { + return true; + } + + // If path is not in protected_endpoints list, skip validation + if (protected_endpoints.find(req.path) == protected_endpoints.end()) { + return true; + } + + // Check for API key in the header + auto auth_header = req.get_header_value("Authorization"); + + std::string prefix = "Bearer "; + if (auth_header.substr(0, prefix.size()) == prefix) { + std::string received_api_key = auth_header.substr(prefix.size()); + if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) { + return true; // API key is valid + } + } + + // API key is invalid or not provided + // TODO: make another middleware for CORS related logic + res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); + res.set_content("Unauthorized: Invalid API Key", "text/plain; charset=utf-8"); + res.status = 401; // Unauthorized + + LOG_WARNING("Unauthorized: Invalid API Key", {}); + + return false; + }; + + // register server middlewares + svr->set_pre_routing_handler([&middleware_validate_api_key](const httplib::Request & req, httplib::Response & res) { + if (!middleware_validate_api_key(req, res)) { + return httplib::Server::HandlerResponse::Handled; + } + return httplib::Server::HandlerResponse::Unhandled; + }); + + // + // Route handlers (or controllers) + // + + const auto handle_health = [&](const httplib::Request & req, httplib::Response & res) { server_state current_state = state.load(); switch (current_state) { case SERVER_STATE_READY: @@ -2765,252 +2901,136 @@ int main(int argc, char ** argv) { res.status = 500; // HTTP Internal Server Error } break; } - }); - - if (sparams.slots_endpoint) { - svr->Get("/slots", [&](const httplib::Request &, httplib::Response & res) { - // request slots data using task queue - server_task task; - task.id = ctx_server.queue_tasks.get_new_id(); - task.id_multi = -1; - task.id_target = -1; - task.type = SERVER_TASK_TYPE_METRICS; - - ctx_server.queue_results.add_waiting_task_id(task.id); - ctx_server.queue_tasks.post(task); - - // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); - ctx_server.queue_results.remove_waiting_task_id(task.id); - - res.set_content(result.data["slots"].dump(), "application/json"); - res.status = 200; // HTTP OK - }); - } - - if (sparams.metrics_endpoint) { - svr->Get("/metrics", [&](const httplib::Request &, httplib::Response & res) { - // request slots data using task queue - server_task task; - task.id = ctx_server.queue_tasks.get_new_id(); - task.id_multi = -1; - task.id_target = -1; - task.type = SERVER_TASK_TYPE_METRICS; - task.data.push_back({{"reset_bucket", true}}); - - ctx_server.queue_results.add_waiting_task_id(task.id); - ctx_server.queue_tasks.post(task); - - // get the result - server_task_result result = ctx_server.queue_results.recv(task.id); - ctx_server.queue_results.remove_waiting_task_id(task.id); - - json data = result.data; - - const uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"]; - const uint64_t t_prompt_processing = data["t_prompt_processing"]; - - const uint64_t n_tokens_predicted = data["n_tokens_predicted"]; - const uint64_t t_tokens_generation = data["t_tokens_generation"]; - - const int32_t kv_cache_used_cells = data["kv_cache_used_cells"]; - - // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names - json all_metrics_def = json { - {"counter", {{ - {"name", "prompt_tokens_total"}, - {"help", "Number of prompt tokens processed."}, - {"value", (uint64_t) data["n_prompt_tokens_processed_total"]} - }, { - {"name", "prompt_seconds_total"}, - {"help", "Prompt process time"}, - {"value", (uint64_t) data["t_prompt_processing_total"] / 1.e3} - }, { - {"name", "tokens_predicted_total"}, - {"help", "Number of generation tokens processed."}, - {"value", (uint64_t) data["n_tokens_predicted_total"]} - }, { - {"name", "tokens_predicted_seconds_total"}, - {"help", "Predict process time"}, - {"value", (uint64_t) data["t_tokens_generation_total"] / 1.e3} - }}}, - {"gauge", {{ - {"name", "prompt_tokens_seconds"}, - {"help", "Average prompt throughput in tokens/s."}, - {"value", n_prompt_tokens_processed ? 1.e3 / t_prompt_processing * n_prompt_tokens_processed : 0.} - },{ - {"name", "predicted_tokens_seconds"}, - {"help", "Average generation throughput in tokens/s."}, - {"value", n_tokens_predicted ? 1.e3 / t_tokens_generation * n_tokens_predicted : 0.} - },{ - {"name", "kv_cache_usage_ratio"}, - {"help", "KV-cache usage. 1 means 100 percent usage."}, - {"value", 1. * kv_cache_used_cells / params.n_ctx} - },{ - {"name", "kv_cache_tokens"}, - {"help", "KV-cache tokens."}, - {"value", (uint64_t) data["kv_cache_tokens_count"]} - },{ - {"name", "requests_processing"}, - {"help", "Number of request processing."}, - {"value", (uint64_t) data["processing"]} - },{ - {"name", "requests_deferred"}, - {"help", "Number of request deferred."}, - {"value", (uint64_t) data["deferred"]} - }}} - }; - - std::stringstream prometheus; - - for (const auto & el : all_metrics_def.items()) { - const auto & type = el.key(); - const auto & metrics_def = el.value(); - - for (const auto & metric_def : metrics_def) { - const std::string name = metric_def["name"]; - const std::string help = metric_def["help"]; - - auto value = json_value(metric_def, "value", 0.); - prometheus << "# HELP llamacpp:" << name << " " << help << "\n" - << "# TYPE llamacpp:" << name << " " << type << "\n" - << "llamacpp:" << name << " " << value << "\n"; - } - } - - const int64_t t_start = data["t_start"]; - res.set_header("Process-Start-Time-Unix", std::to_string(t_start)); - - res.set_content(prometheus.str(), "text/plain; version=0.0.4"); - res.status = 200; // HTTP OK - }); - } - - svr->set_logger(log_server_request); - - svr->set_exception_handler([](const httplib::Request &, httplib::Response & res, std::exception_ptr ep) { - const char fmt[] = "500 Internal Server Error\n%s"; - - char buf[BUFSIZ]; - try { - std::rethrow_exception(std::move(ep)); - } catch (std::exception &e) { - snprintf(buf, sizeof(buf), fmt, e.what()); - } catch (...) { - snprintf(buf, sizeof(buf), fmt, "Unknown Exception"); - } - - res.set_content(buf, "text/plain; charset=utf-8"); - res.status = 500; - }); - - svr->set_error_handler([](const httplib::Request &, httplib::Response & res) { - if (res.status == 401) { - res.set_content("Unauthorized", "text/plain; charset=utf-8"); - } - if (res.status == 400) { - res.set_content("Invalid request", "text/plain; charset=utf-8"); - } - if (res.status == 404) { - res.set_content("File Not Found", "text/plain; charset=utf-8"); - } - }); - - // set timeouts and change hostname and port - svr->set_read_timeout (sparams.read_timeout); - svr->set_write_timeout(sparams.write_timeout); - - if (!svr->bind_to_port(sparams.hostname, sparams.port)) { - fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port); - return 1; - } - - // Set the base directory for serving static files - svr->set_base_dir(sparams.public_path); - - std::unordered_map log_data; - - log_data["hostname"] = sparams.hostname; - log_data["port"] = std::to_string(sparams.port); - - if (sparams.api_keys.size() == 1) { - log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4); - } else if (sparams.api_keys.size() > 1) { - log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded"; - } - - // load the model - if (!ctx_server.load_model(params)) { - state.store(SERVER_STATE_ERROR); - return 1; - } else { - ctx_server.initialize(); - state.store(SERVER_STATE_READY); - } - - LOG_INFO("model loaded", {}); - - const auto model_meta = ctx_server.model_meta(); - - if (sparams.chat_template.empty()) { // custom chat template is not supplied - if (!ctx_server.validate_model_chat_template()) { - LOG_ERROR("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); - sparams.chat_template = "chatml"; - } - } - - // Middleware for API key validation - auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool { - // If API key is not set, skip validation - if (sparams.api_keys.empty()) { - return true; - } - - // Check for API key in the header - auto auth_header = req.get_header_value("Authorization"); - - std::string prefix = "Bearer "; - if (auth_header.substr(0, prefix.size()) == prefix) { - std::string received_api_key = auth_header.substr(prefix.size()); - if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) { - return true; // API key is valid - } - } - - // API key is invalid or not provided - res.set_content("Unauthorized: Invalid API Key", "text/plain; charset=utf-8"); - res.status = 401; // Unauthorized - - LOG_WARNING("Unauthorized: Invalid API Key", {}); - - return false; }; - // this is only called if no index.html is found in the public --path - svr->Get("/", [](const httplib::Request &, httplib::Response & res) { - res.set_content(reinterpret_cast(&index_html), index_html_len, "text/html; charset=utf-8"); - return false; - }); + const auto handle_slots = [&](const httplib::Request &, httplib::Response & res) { + if (!sparams.slots_endpoint) { + res.status = 501; + res.set_content("This server does not support slots endpoint.", "text/plain; charset=utf-8"); + return; + } - // this is only called if no index.js is found in the public --path - svr->Get("/index.js", [](const httplib::Request &, httplib::Response & res) { - res.set_content(reinterpret_cast(&index_js), index_js_len, "text/javascript; charset=utf-8"); - return false; - }); + // request slots data using task queue + server_task task; + task.id = ctx_server.queue_tasks.get_new_id(); + task.id_multi = -1; + task.id_target = -1; + task.type = SERVER_TASK_TYPE_METRICS; - // this is only called if no index.html is found in the public --path - svr->Get("/completion.js", [](const httplib::Request &, httplib::Response & res) { - res.set_content(reinterpret_cast(&completion_js), completion_js_len, "application/javascript; charset=utf-8"); - return false; - }); + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); - // this is only called if no index.html is found in the public --path - svr->Get("/json-schema-to-grammar.mjs", [](const httplib::Request &, httplib::Response & res) { - res.set_content(reinterpret_cast(&json_schema_to_grammar_mjs), json_schema_to_grammar_mjs_len, "application/javascript; charset=utf-8"); - return false; - }); + // get the result + server_task_result result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); - svr->Get("/props", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + res.set_content(result.data["slots"].dump(), "application/json"); + res.status = 200; // HTTP OK + }; + + const auto handle_metrics = [&](const httplib::Request &, httplib::Response & res) { + if (!sparams.metrics_endpoint) { + res.status = 501; + res.set_content("This server does not support metrics endpoint.", "text/plain; charset=utf-8"); + return; + } + + // request slots data using task queue + server_task task; + task.id = ctx_server.queue_tasks.get_new_id(); + task.id_multi = -1; + task.id_target = -1; + task.type = SERVER_TASK_TYPE_METRICS; + task.data.push_back({{"reset_bucket", true}}); + + ctx_server.queue_results.add_waiting_task_id(task.id); + ctx_server.queue_tasks.post(task); + + // get the result + server_task_result result = ctx_server.queue_results.recv(task.id); + ctx_server.queue_results.remove_waiting_task_id(task.id); + + json data = result.data; + + const uint64_t n_prompt_tokens_processed = data["n_prompt_tokens_processed"]; + const uint64_t t_prompt_processing = data["t_prompt_processing"]; + + const uint64_t n_tokens_predicted = data["n_tokens_predicted"]; + const uint64_t t_tokens_generation = data["t_tokens_generation"]; + + const int32_t kv_cache_used_cells = data["kv_cache_used_cells"]; + + // metrics definition: https://prometheus.io/docs/practices/naming/#metric-names + json all_metrics_def = json { + {"counter", {{ + {"name", "prompt_tokens_total"}, + {"help", "Number of prompt tokens processed."}, + {"value", (uint64_t) data["n_prompt_tokens_processed_total"]} + }, { + {"name", "prompt_seconds_total"}, + {"help", "Prompt process time"}, + {"value", (uint64_t) data["t_prompt_processing_total"] / 1.e3} + }, { + {"name", "tokens_predicted_total"}, + {"help", "Number of generation tokens processed."}, + {"value", (uint64_t) data["n_tokens_predicted_total"]} + }, { + {"name", "tokens_predicted_seconds_total"}, + {"help", "Predict process time"}, + {"value", (uint64_t) data["t_tokens_generation_total"] / 1.e3} + }}}, + {"gauge", {{ + {"name", "prompt_tokens_seconds"}, + {"help", "Average prompt throughput in tokens/s."}, + {"value", n_prompt_tokens_processed ? 1.e3 / t_prompt_processing * n_prompt_tokens_processed : 0.} + },{ + {"name", "predicted_tokens_seconds"}, + {"help", "Average generation throughput in tokens/s."}, + {"value", n_tokens_predicted ? 1.e3 / t_tokens_generation * n_tokens_predicted : 0.} + },{ + {"name", "kv_cache_usage_ratio"}, + {"help", "KV-cache usage. 1 means 100 percent usage."}, + {"value", 1. * kv_cache_used_cells / params.n_ctx} + },{ + {"name", "kv_cache_tokens"}, + {"help", "KV-cache tokens."}, + {"value", (uint64_t) data["kv_cache_tokens_count"]} + },{ + {"name", "requests_processing"}, + {"help", "Number of request processing."}, + {"value", (uint64_t) data["processing"]} + },{ + {"name", "requests_deferred"}, + {"help", "Number of request deferred."}, + {"value", (uint64_t) data["deferred"]} + }}} + }; + + std::stringstream prometheus; + + for (const auto & el : all_metrics_def.items()) { + const auto & type = el.key(); + const auto & metrics_def = el.value(); + + for (const auto & metric_def : metrics_def) { + const std::string name = metric_def["name"]; + const std::string help = metric_def["help"]; + + auto value = json_value(metric_def, "value", 0.); + prometheus << "# HELP llamacpp:" << name << " " << help << "\n" + << "# TYPE llamacpp:" << name << " " << type << "\n" + << "llamacpp:" << name << " " << value << "\n"; + } + } + + const int64_t t_start = data["t_start"]; + res.set_header("Process-Start-Time-Unix", std::to_string(t_start)); + + res.set_content(prometheus.str(), "text/plain; version=0.0.4"); + res.status = 200; // HTTP OK + }; + + const auto handle_props = [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); json data = { { "user_name", ctx_server.name_user.c_str() }, @@ -3020,13 +3040,10 @@ int main(int argc, char ** argv) { }; res.set_content(data.dump(), "application/json; charset=utf-8"); - }); + }; - const auto completions = [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { + const auto handle_completions = [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!validate_api_key(req, res)) { - return; - } json data = json::parse(req.body); @@ -3102,11 +3119,7 @@ int main(int argc, char ** argv) { } }; - svr->Post("/completion", completions); // legacy - svr->Post("/completions", completions); - svr->Post("/v1/completions", completions); - - svr->Get("/v1/models", [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { + const auto handle_models = [¶ms, &model_meta](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); json models = { @@ -3123,14 +3136,10 @@ int main(int argc, char ** argv) { }; res.set_content(models.dump(), "application/json; charset=utf-8"); - }); + }; - const auto chat_completions = [&ctx_server, &validate_api_key, &sparams](const httplib::Request & req, httplib::Response & res) { + const auto handle_chat_completions = [&ctx_server, &sparams](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!validate_api_key(req, res)) { - return; - } - json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), sparams.chat_template); const int id_task = ctx_server.queue_tasks.get_new_id(); @@ -3201,14 +3210,8 @@ int main(int argc, char ** argv) { } }; - svr->Post("/chat/completions", chat_completions); - svr->Post("/v1/chat/completions", chat_completions); - - svr->Post("/infill", [&ctx_server, &validate_api_key](const httplib::Request & req, httplib::Response & res) { + const auto handle_infill = [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!validate_api_key(req, res)) { - return; - } json data = json::parse(req.body); @@ -3266,13 +3269,9 @@ int main(int argc, char ** argv) { res.set_chunked_content_provider("text/event-stream", chunked_content_provider, on_complete); } - }); + }; - svr->Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) { - return res.set_content("", "application/json; charset=utf-8"); - }); - - svr->Post("/tokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + const auto handle_tokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); @@ -3282,9 +3281,9 @@ int main(int argc, char ** argv) { } const json data = format_tokenizer_response(tokens); return res.set_content(data.dump(), "application/json; charset=utf-8"); - }); + }; - svr->Post("/detokenize", [&ctx_server](const httplib::Request & req, httplib::Response & res) { + const auto handle_detokenize = [&ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); const json body = json::parse(req.body); @@ -3296,9 +3295,9 @@ int main(int argc, char ** argv) { const json data = format_detokenized_response(content); return res.set_content(data.dump(), "application/json; charset=utf-8"); - }); + }; - svr->Post("/embedding", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { + const auto handle_embeddings = [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); if (!params.embedding) { res.status = 501; @@ -3307,94 +3306,114 @@ int main(int argc, char ** argv) { } const json body = json::parse(req.body); + bool is_openai = false; - json prompt; - if (body.count("content") != 0) { - prompt = body["content"]; - } else { - prompt = ""; - } - - // create and queue the task - const int id_task = ctx_server.queue_tasks.get_new_id(); - - ctx_server.queue_results.add_waiting_task_id(id_task); - ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0} }, false, true); - - // get the result - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); - - // send the result - return res.set_content(result.data.dump(), "application/json; charset=utf-8"); - }); - - svr->Post("/v1/embeddings", [¶ms, &ctx_server](const httplib::Request & req, httplib::Response & res) { - res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin")); - if (!params.embedding) { - res.status = 501; - res.set_content("This server does not support embeddings. Start it with `--embeddings`", "text/plain; charset=utf-8"); - return; - } - - const json body = json::parse(req.body); - - json prompt; + // an input prompt can string or a list of tokens (integer) + std::vector prompts; if (body.count("input") != 0) { - prompt = body["input"]; - if (prompt.is_array()) { - json data = json::array(); - - int i = 0; - for (const json & elem : prompt) { - const int id_task = ctx_server.queue_tasks.get_new_id(); - - ctx_server.queue_results.add_waiting_task_id(id_task); - ctx_server.request_completion(id_task, -1, { {"prompt", elem}, { "n_predict", 0} }, false, true); - - // get the result - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); - - json embedding = json{ - {"embedding", json_value(result.data, "embedding", json::array())}, - {"index", i++}, - {"object", "embedding"} - }; - - data.push_back(embedding); + is_openai = true; + if (body["input"].is_array()) { + // support multiple prompts + for (const json & elem : body["input"]) { + prompts.push_back(elem); } - - json result = format_embeddings_response_oaicompat(body, data); - - return res.set_content(result.dump(), "application/json; charset=utf-8"); + } else { + // single input prompt + prompts.push_back(body["input"]); } + } else if (body.count("content") != 0) { + // only support single prompt here + std::string content = body["content"]; + prompts.push_back(content); } else { - prompt = ""; + // TODO @ngxson : should return an error here + prompts.push_back(""); } - // create and queue the task - const int id_task = ctx_server.queue_tasks.get_new_id(); + // process all prompts + json responses = json::array(); + for (auto & prompt : prompts) { + // TODO @ngxson : maybe support multitask for this endpoint? + // create and queue the task + const int id_task = ctx_server.queue_tasks.get_new_id(); - ctx_server.queue_results.add_waiting_task_id(id_task); - ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0}}, false, true); + ctx_server.queue_results.add_waiting_task_id(id_task); + ctx_server.request_completion(id_task, -1, { {"prompt", prompt}, { "n_predict", 0}}, false, true); - // get the result - server_task_result result = ctx_server.queue_results.recv(id_task); - ctx_server.queue_results.remove_waiting_task_id(id_task); - - json data = json::array({json{ - {"embedding", json_value(result.data, "embedding", json::array())}, - {"index", 0}, - {"object", "embedding"} - }} - ); - - json root = format_embeddings_response_oaicompat(body, data); + // get the result + server_task_result result = ctx_server.queue_results.recv(id_task); + ctx_server.queue_results.remove_waiting_task_id(id_task); + responses.push_back(result.data); + } + // write JSON response + json root; + if (is_openai) { + json res_oai = json::array(); + int i = 0; + for (auto & elem : responses) { + res_oai.push_back(json{ + {"embedding", json_value(elem, "embedding", json::array())}, + {"index", i++}, + {"object", "embedding"} + }); + } + root = format_embeddings_response_oaicompat(body, res_oai); + } else { + root = responses[0]; + } return res.set_content(root.dump(), "application/json; charset=utf-8"); - }); + }; + // + // Router + // + + // register static assets routes + if (!sparams.public_path.empty()) { + // Set the base directory for serving static files + svr->set_base_dir(sparams.public_path); + } + + // using embedded static files + auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) { + return [content, len, mime_type](const httplib::Request &, httplib::Response & res) { + res.set_content(reinterpret_cast(content), len, mime_type); + return false; + }; + }; + + svr->Options(R"(/.*)", [](const httplib::Request &, httplib::Response & res) { + // TODO @ngxson : I have no idea what it is... maybe this is redundant? + return res.set_content("", "application/json; charset=utf-8"); + }); + svr->Get("/", handle_static_file(index_html, index_html_len, "text/html; charset=utf-8")); + svr->Get("/index.js", handle_static_file(index_js, index_js_len, "text/javascript; charset=utf-8")); + svr->Get("/completion.js", handle_static_file(completion_js, completion_js_len, "text/javascript; charset=utf-8")); + svr->Get("/json-schema-to-grammar.mjs", handle_static_file( + json_schema_to_grammar_mjs, json_schema_to_grammar_mjs_len, "text/javascript; charset=utf-8")); + + // register API routes + svr->Get ("/health", handle_health); + svr->Get ("/slots", handle_slots); + svr->Get ("/metrics", handle_metrics); + svr->Get ("/props", handle_props); + svr->Get ("/v1/models", handle_models); + svr->Post("/completion", handle_completions); // legacy + svr->Post("/completions", handle_completions); + svr->Post("/v1/completions", handle_completions); + svr->Post("/chat/completions", handle_chat_completions); + svr->Post("/v1/chat/completions", handle_chat_completions); + svr->Post("/infill", handle_infill); + svr->Post("/embedding", handle_embeddings); // legacy + svr->Post("/embeddings", handle_embeddings); + svr->Post("/v1/embeddings", handle_embeddings); + svr->Post("/tokenize", handle_tokenize); + svr->Post("/detokenize", handle_detokenize); + + // + // Start the server + // if (sparams.n_threads_http < 1) { // +2 threads for monitoring endpoints sparams.n_threads_http = std::max(params.n_parallel + 2, (int32_t) std::thread::hardware_concurrency() - 1); diff --git a/examples/server/tests/features/security.feature b/examples/server/tests/features/security.feature index 42a6709a5..1d6aa40ea 100644 --- a/examples/server/tests/features/security.feature +++ b/examples/server/tests/features/security.feature @@ -39,8 +39,9 @@ Feature: Security Scenario Outline: CORS Options - When an OPTIONS request is sent from - Then CORS header is set to + Given a user api key llama.cpp + When an OPTIONS request is sent from + Then CORS header is set to Examples: Headers | origin | cors_header | cors_header_value | diff --git a/examples/server/tests/features/steps/steps.py b/examples/server/tests/features/steps/steps.py index 0076f805b..142048509 100644 --- a/examples/server/tests/features/steps/steps.py +++ b/examples/server/tests/features/steps/steps.py @@ -582,8 +582,9 @@ async def step_detokenize(context): @async_run_until_complete async def step_options_request(context, origin): async with aiohttp.ClientSession() as session: + headers = {'Authorization': f'Bearer {context.user_api_key}', 'Origin': origin} async with session.options(f'{context.base_url}/v1/chat/completions', - headers={"Origin": origin}) as response: + headers=headers) as response: assert response.status == 200 context.options_response = response From 9674aaf35cb81478eb38c3f3ebde713ec72fbb79 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 12:34:18 +0200 Subject: [PATCH 20/32] server : simplify logic for empty prompts (#5953) --- examples/server/server.cpp | 28 +++++++++++++++------------- 1 file changed, 15 insertions(+), 13 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 6e0f8328c..aedf0afc6 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1704,19 +1704,6 @@ struct server_context { // next, batch any pending prompts without exceeding n_batch if (params.cont_batching || batch.n_tokens == 0) { for (auto & slot : slots) { - const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get().empty()); - - // empty prompt passed -> release the slot and send empty response - // note: infill mode allows empty prompt - if (slot.state == SLOT_STATE_IDLE && slot.command == SLOT_COMMAND_LOAD_PROMPT && !has_prompt && !slot.infill) { - slot.state = SLOT_STATE_PROCESSING; - slot.command = SLOT_COMMAND_NONE; - slot.release(); - slot.print_timings(); - send_final_response(slot); - continue; - } - // this slot still has a prompt to be processed if (slot.state == SLOT_STATE_IDLE && slot.command == SLOT_COMMAND_LOAD_PROMPT) { auto & prompt_tokens = slot.prompt_tokens; @@ -1768,6 +1755,21 @@ struct server_context { {"prompt_tokens", tokens_to_str(ctx, prompt_tokens.cbegin(), prompt_tokens.cend())}, }); + // empty prompt passed -> release the slot and send empty response + if (prompt_tokens.empty()) { + LOG_INFO("empty prompt - releasing slot", { + {"id_slot", slot.id}, + {"id_task", slot.id_task} + }); + + slot.state = SLOT_STATE_PROCESSING; + slot.command = SLOT_COMMAND_NONE; + slot.release(); + slot.print_timings(); + send_final_response(slot); + continue; + } + if (slot.embedding) { // this prompt is too large to process - discard it if (slot.n_prompt_tokens > n_batch) { From 8a3012a4ad08112bb3dc3f1399afec4e93780c44 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 12:47:57 +0200 Subject: [PATCH 21/32] ggml : add ggml-common.h to deduplicate shared code (#5940) * ggml : add ggml-common.h to shared code ggml-ci * scripts : update sync scripts * sycl : reuse quantum tables ggml-ci * ggml : minor * ggml : minor * sycl : try to fix build --- CMakeLists.txt | 3 +- Makefile | 4 +- ggml-common.h | 779 ++++++++++++++++++++++++++++++++++++++++ ggml-cuda.cu | 743 +------------------------------------- ggml-metal.metal | 707 +----------------------------------- ggml-quants.c | 708 +----------------------------------- ggml-quants.h | 4 +- ggml-sycl.cpp | 384 +------------------- scripts/sync-ggml-am.sh | 2 + scripts/sync-ggml.sh | 1 + 10 files changed, 799 insertions(+), 2536 deletions(-) create mode 100644 ggml-common.h diff --git a/CMakeLists.txt b/CMakeLists.txt index 48880f720..9309ca6bb 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -199,7 +199,8 @@ if (LLAMA_METAL) # get full path to the file #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/") - # copy ggml-metal.metal to bin directory + # copy ggml-common.h and ggml-metal.metal to bin directory + configure_file(ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h COPYONLY) configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY) if (LLAMA_METAL_EMBED_LIBRARY) diff --git a/Makefile b/Makefile index aea969222..d809b8b3b 100644 --- a/Makefile +++ b/Makefile @@ -453,7 +453,7 @@ endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE ifdef LLAMA_CUDA_CCBIN MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN) endif -ggml-cuda.o: ggml-cuda.cu ggml-cuda.h +ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h ifdef JETSON_EOL_MODULE_DETECT $(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@ else @@ -630,7 +630,7 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h $(CC) $(CFLAGS) -c $< -o $@ -ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h +ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h $(CC) $(CFLAGS) -c $< -o $@ OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o diff --git a/ggml-common.h b/ggml-common.h new file mode 100644 index 000000000..4b6d248b6 --- /dev/null +++ b/ggml-common.h @@ -0,0 +1,779 @@ +#pragma once + +#if defined(GGML_COMMON_IMPL_C) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_METAL) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_CUDA) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static const __device__ __constant__ type name[size] = { +#define GGML_TABLE_END() }; + +#define GGML_COMMON_IMPL +#elif defined(GGML_COMMON_IMPL_SYCL) +#include + +#define GGML_TABLE_BEGIN(type, name, size) static dpct::global_memory name(sycl::range<1>(size), { +#define GGML_TABLE_END() }); + +#define GGML_COMMON_IMPL +#endif + +#if defined(GGML_COMMON_IMPL) + +GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8) + 1, 2, 4, 8, 16, 32, 64, 128 +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128) + 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, + 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, + 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, + 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, + 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, + 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, + 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, + 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, +GGML_TABLE_END() + +//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics +GGML_TABLE_BEGIN(uint64_t, ksigns64, 128) + 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff, + 0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff, + 0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff, + 0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff, + 0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff, + 0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff, + 0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff, + 0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff, + 0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff, + 0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff, + 0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff, + 0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff, + 0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff, + 0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff, + 0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff, + 0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff, + 0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff, + 0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff, + 0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff, + 0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff, + 0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff, + 0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff, + 0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff, + 0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff, + 0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff, + 0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff, + 0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff, + 0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff, + 0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff, + 0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff, + 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff, + 0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff, +GGML_TABLE_END() +//#endif + + +GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, + 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, + 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, + 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, + 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, + 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, + 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, + 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, + 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, + 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, + 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, + 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, + 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, + 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, + 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, + 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, + 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, + 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, + 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, + 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, + 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, + 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, + 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, + 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, + 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, + 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, + 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, + 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, + 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, + 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, + 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, + 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, + 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, + 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, + 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, + 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, + 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, + 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, + 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, + 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, + 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, + 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, + 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, + 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, + 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, + 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, + 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, + 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, + 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, + 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, + 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, + 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, + 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, + 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, + 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, + 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, + 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, + 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, + 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, + 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, + 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, + 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, + 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, + 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, + 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, + 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, + 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, + 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, + 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, + 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, + 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, + 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, + 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, + 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, + 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, + 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, + 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, + 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, + 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, + 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, + 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, + 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, + 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, + 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, + 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, + 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, + 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, + 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, + 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, + 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, + 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, + 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, + 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, + 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, + 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, + 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, + 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, + 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, + 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, + 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, + 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, + 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, + 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, + 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, + 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, + 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, + 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, + 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, + 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, + 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, + 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, + 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, + 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, + 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, + 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, + 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, + 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, + 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, + 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, + 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, + 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, + 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, + 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, + 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, + 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, + 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, + 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, + 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, + 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, + 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, + 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, + 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, + 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, + 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, + 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, + 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, + 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, + 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, + 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, + 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, + 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, + 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, + 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, + 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, + 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, + 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, + 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, + 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, + 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, + 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, + 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, + 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, + 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, + 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, + 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, + 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, + 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, + 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, + 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, + 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, + 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, + 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, + 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, + 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, + 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, + 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, + 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, + 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, + 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, + 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, + 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, + 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, + 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, + 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, + 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, + 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, + 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024) + 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, + 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, + 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, + 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, + 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, + 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b, + 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919, + 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808, + 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908, + 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b, + 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908, + 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08, + 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19, + 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819, + 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919, + 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b, + 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, + 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908, + 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919, + 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908, + 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, + 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919, + 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b, + 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, + 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908, + 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b, + 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b, + 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08, + 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, + 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819, + 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808, + 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908, + 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b, + 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908, + 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08, + 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819, + 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808, + 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08, + 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819, + 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b, + 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908, + 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919, + 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b, + 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919, + 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808, + 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819, + 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919, + 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919, + 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808, + 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819, + 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b, + 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908, + 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, + 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, + 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919, + 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b, + 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919, + 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b, + 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819, + 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919, + 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908, + 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b, + 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908, + 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b, + 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908, + 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08, + 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908, + 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819, + 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819, + 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808, + 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, + 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19, + 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819, + 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808, + 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819, + 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919, + 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808, + 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19, + 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08, + 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b, + 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908, + 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808, + 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819, + 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908, + 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819, + 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808, + 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808, + 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819, + 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908, + 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08, + 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819, + 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b, + 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b, + 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08, + 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19, + 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819, + 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919, + 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908, + 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808, + 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808, + 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908, + 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808, + 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08, + 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08, + 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908, + 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919, + 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808, + 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819, + 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908, + 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08, + 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819, + 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808, + 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808, + 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819, + 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808, + 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908, + 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b, + 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, + 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, + 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b, + 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808, + 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b, + 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19, + 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819, + 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08, + 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b, + 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, + 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b, + 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b, + 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919, + 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808, + 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819, + 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908, + 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08, + 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08, + 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819, + 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919, + 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908, + 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b, + 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908, + 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b, + 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908, + 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08, + 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819, + 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808, + 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819, + 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919, + 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808, + 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808, + 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08, + 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819, + 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919, + 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808, + 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819, + 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919, + 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808, + 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b, + 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908, + 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808, + 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908, + 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b, + 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908, + 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b, + 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908, + 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b, + 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908, + 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08, + 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908, + 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b, + 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908, + 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08, + 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819, + 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919, + 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808, + 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19, + 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b, + 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919, + 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808, + 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819, + 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908, + 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919, + 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808, + 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808, + 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b, + 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919, + 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808, + 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b, + 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808, + 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919, + 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b, + 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08, + 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919, + 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808, + 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b, + 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908, + 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808, + 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808, + 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808, + 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908, + 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808, + 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808, + 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b, + 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908, + 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808, + 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808, + 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, + 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919, + 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b, + 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808, + 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819, + 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b, + 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908, + 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08, + 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908, + 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919, + 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819, + 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908, + 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b, + 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808, + 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819, + 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908, + 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919, + 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808, + 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808, + 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808, + 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919, + 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908, + 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908, + 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08, + 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819, + 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b, + 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808, + 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819, + 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908, + 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819, + 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808, + 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808, + 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b, + 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908, + 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808, + 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908, + 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819, + 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819, + 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808, + 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b, + 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b, + 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819, + 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b, + 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b, + 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b, + 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819, + 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19, + 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819, + 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908, + 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808, + 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256) + 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, + 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, + 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, + 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, + 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, + 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, + 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, + 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, + 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, + 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, + 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, + 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, + 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, + 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, + 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, + 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, + 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, + 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, + 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, + 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, + 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, + 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, + 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, + 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, + 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, + 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, + 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, + 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, + 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, + 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, + 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, + 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, +GGML_TABLE_END() + +GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512) + 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, + 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, + 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, + 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, + 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, + 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, + 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, + 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, + 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, + 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, + 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, + 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, + 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, + 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, + 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, + 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, + 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, + 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, + 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, + 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, + 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, + 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, + 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, + 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, + 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, + 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, + 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, + 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, + 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, + 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, + 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, + 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, + 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, + 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, + 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, + 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, + 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, + 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, + 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, + 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, + 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, + 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, + 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, + 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, + 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, + 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, + 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, + 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, + 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, + 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, + 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, + 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, + 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, + 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, + 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, + 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, + 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, + 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, + 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, + 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, + 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, + 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, + 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, + 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, +GGML_TABLE_END() + +#define NGRID_IQ2XXS 512 +GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ2XXS) + 0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000, + 0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01, + 0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100, + 0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00, + 0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101, + 0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100, + 0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00, + 0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff, + 0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000, + 0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000, + 0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001, + 0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff, + 0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01, + 0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001, + 0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00, + 0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001, + 0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100, + 0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000, + 0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000, + 0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000, + 0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff, + 0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff, + 0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01, + 0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100, + 0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff, + 0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000, + 0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101, + 0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff, + 0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff, + 0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001, + 0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01, + 0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101, + 0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100, + 0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00, + 0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001, + 0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff, + 0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000, + 0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000, + 0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100, + 0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100, + 0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01, + 0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff, + 0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101, + 0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000, + 0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff, + 0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000, + 0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff, + 0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00, + 0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101, + 0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000, + 0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000, + 0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000, + 0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100, + 0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000, + 0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001, + 0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff, + 0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000, + 0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000, + 0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000, + 0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000, + 0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff, + 0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000, + 0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, + 0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01, + 0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100, + 0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000, + 0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00, + 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100, + 0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000, + 0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, + 0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00, + 0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff, + 0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100, + 0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff, + 0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000, + 0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff, + 0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff, + 0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00, + 0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001, + 0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001, + 0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01, + 0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000, + 0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101, + 0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00, + 0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100, + 0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101, + 0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101, + 0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000, + 0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff, + 0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff, + 0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101, + 0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, + 0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101, + 0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001, + 0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff, + 0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff, + 0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01, + 0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff, + 0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100, + 0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001, + 0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00, + 0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff, + 0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff, + 0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000, + 0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000, + 0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101, + 0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001, + 0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000, + 0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101, + 0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000, + 0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, + 0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000, + 0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100, + 0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000, + 0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000, + 0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100, + 0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff, + 0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff, + 0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00, + 0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101, + 0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000, + 0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00, + 0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000, + 0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff, + 0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101, + 0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff, + 0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00, + 0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff, +GGML_TABLE_END() + +#endif // GGML_COMMON_IMPL diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 72bcec8cd..c207ff87a 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -2,6 +2,9 @@ #include "ggml.h" #include "ggml-backend-impl.h" +#define GGML_COMMON_IMPL_CUDA +#include "ggml-common.h" + #include #include #include @@ -1569,746 +1572,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t #endif } -static const __device__ uint64_t iq2xxs_grid[256] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, - 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, - 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, - 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, - 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, - 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, - 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, - 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, - 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, - 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, - 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, - 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, - 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, - 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, - 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, - 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, - 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, - 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, - 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, - 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, - 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, - 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, - 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, - 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, - 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, - 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, - 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, - 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, - 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, - 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, - 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, - 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, - 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, - 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, - 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, - 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, - 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, - 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, - 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, - 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, - 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, - 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, - 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, - 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, - 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, - 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, - 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, - 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, - 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, - 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, - 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, - 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, - 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, - 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, - 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, - 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, - 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, - 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, - 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, -}; - -static const __device__ uint64_t iq2xs_grid[512] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, - 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, - 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, - 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, - 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, - 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, - 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, - 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, - 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, - 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, - 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, - 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, - 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, - 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, - 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, - 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, - 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, - 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, - 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, - 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, - 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, - 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, - 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, - 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, - 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, - 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, - 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, - 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, - 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, - 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, - 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, - 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, - 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, - 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, - 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, - 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, - 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, - 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, - 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, - 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, - 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, - 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, - 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, - 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, - 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, - 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, - 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, - 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, - 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, - 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, - 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, - 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, - 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, - 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, - 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, - 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, - 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, - 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, - 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, - 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, - 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, - 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, - 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, - 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, - 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, - 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, - 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, - 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, - 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, - 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, - 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, - 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, - 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, - 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, - 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, - 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, - 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, - 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, - 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, - 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, - 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, - 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, - 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, - 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, - 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, - 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, - 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, - 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, - 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, - 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, - 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, - 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, - 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, - 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, - 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, - 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, - 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, - 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, - 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, - 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, - 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, - 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, - 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, - 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, - 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, - 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, - 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, - 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, - 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, - 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, - 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, - 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, - 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, - 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, - 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, - 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, -}; - -static const __device__ uint64_t iq2s_grid[1024] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b, - 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919, - 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808, - 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908, - 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b, - 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908, - 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08, - 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19, - 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819, - 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919, - 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b, - 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, - 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908, - 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908, - 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, - 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919, - 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b, - 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, - 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908, - 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b, - 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b, - 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08, - 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, - 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819, - 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808, - 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908, - 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b, - 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908, - 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08, - 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808, - 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08, - 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819, - 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908, - 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919, - 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b, - 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919, - 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808, - 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819, - 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919, - 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919, - 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808, - 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819, - 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b, - 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908, - 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, - 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, - 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919, - 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b, - 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919, - 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b, - 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819, - 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919, - 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908, - 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b, - 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908, - 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b, - 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908, - 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08, - 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908, - 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819, - 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819, - 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808, - 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, - 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19, - 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819, - 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808, - 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819, - 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919, - 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808, - 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19, - 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08, - 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b, - 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908, - 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808, - 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819, - 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908, - 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808, - 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808, - 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819, - 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908, - 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08, - 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819, - 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b, - 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08, - 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19, - 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819, - 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919, - 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908, - 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808, - 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808, - 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908, - 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808, - 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08, - 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08, - 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908, - 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919, - 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808, - 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819, - 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908, - 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08, - 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819, - 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808, - 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808, - 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819, - 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808, - 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908, - 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b, - 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, - 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, - 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b, - 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808, - 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b, - 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19, - 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819, - 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08, - 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b, - 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, - 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b, - 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b, - 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919, - 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808, - 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819, - 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908, - 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08, - 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08, - 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819, - 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919, - 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908, - 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b, - 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908, - 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b, - 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908, - 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08, - 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819, - 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808, - 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819, - 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919, - 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808, - 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808, - 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08, - 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819, - 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919, - 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808, - 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819, - 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919, - 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808, - 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b, - 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908, - 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808, - 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908, - 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b, - 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908, - 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b, - 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908, - 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b, - 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908, - 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08, - 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908, - 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b, - 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908, - 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08, - 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819, - 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919, - 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808, - 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19, - 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b, - 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919, - 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808, - 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819, - 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908, - 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919, - 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808, - 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808, - 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b, - 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919, - 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808, - 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b, - 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808, - 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919, - 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b, - 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08, - 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919, - 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808, - 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b, - 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908, - 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808, - 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808, - 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808, - 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908, - 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808, - 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808, - 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b, - 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908, - 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808, - 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808, - 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, - 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b, - 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808, - 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819, - 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b, - 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908, - 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08, - 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908, - 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919, - 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819, - 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908, - 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808, - 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819, - 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908, - 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919, - 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808, - 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808, - 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808, - 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919, - 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908, - 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908, - 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08, - 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819, - 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b, - 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808, - 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819, - 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908, - 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819, - 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808, - 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808, - 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b, - 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908, - 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808, - 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908, - 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819, - 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819, - 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808, - 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b, - 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b, - 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819, - 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b, - 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b, - 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b, - 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819, - 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19, - 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819, - 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908, - 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808, - 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b, -}; - -static const __device__ uint32_t iq3xxs_grid[256] = { - 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, - 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, - 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, - 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, - 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, - 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, - 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, - 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, - 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, - 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, - 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, - 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, - 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, - 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, - 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, - 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, - 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, - 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, - 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, - 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, - 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, - 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, - 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, - 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, - 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, - 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, - 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, - 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, - 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, - 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, - 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, - 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, -}; - -static const __device__ uint32_t iq3s_grid[512] = { - 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, - 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, - 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, - 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, - 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, - 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, - 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, - 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, - 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, - 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, - 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, - 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, - 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, - 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, - 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, - 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, - 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, - 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, - 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, - 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, - 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, - 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, - 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, - 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, - 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, - 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, - 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, - 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, - 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, - 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, - 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, - 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, - 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, - 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, - 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, - 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, - 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, - 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, - 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, - 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, - 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, - 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, - 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, - 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, - 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, - 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, - 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, - 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, - 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, - 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, - 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, - 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, - 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, - 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, - 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, - 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, - 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, - 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, - 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, - 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, - 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, - 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, - 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, - 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, -}; - -static const __device__ uint64_t iq1s_grid[512] = { - 0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000, - 0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01, - 0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100, - 0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00, - 0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101, - 0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100, - 0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00, - 0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff, - 0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000, - 0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000, - 0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001, - 0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff, - 0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01, - 0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001, - 0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00, - 0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001, - 0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100, - 0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000, - 0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000, - 0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000, - 0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff, - 0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff, - 0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01, - 0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100, - 0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff, - 0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000, - 0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101, - 0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff, - 0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff, - 0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001, - 0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01, - 0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101, - 0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100, - 0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00, - 0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001, - 0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff, - 0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000, - 0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000, - 0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100, - 0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100, - 0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01, - 0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff, - 0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101, - 0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000, - 0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff, - 0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000, - 0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff, - 0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00, - 0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101, - 0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000, - 0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000, - 0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000, - 0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100, - 0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000, - 0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001, - 0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff, - 0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000, - 0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000, - 0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000, - 0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000, - 0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff, - 0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000, - 0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, - 0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01, - 0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100, - 0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000, - 0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00, - 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100, - 0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000, - 0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, - 0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00, - 0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff, - 0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100, - 0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff, - 0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000, - 0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff, - 0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff, - 0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00, - 0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001, - 0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001, - 0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01, - 0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000, - 0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101, - 0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00, - 0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100, - 0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101, - 0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101, - 0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000, - 0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff, - 0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff, - 0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101, - 0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, - 0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101, - 0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001, - 0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff, - 0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff, - 0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01, - 0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff, - 0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100, - 0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001, - 0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00, - 0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff, - 0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff, - 0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000, - 0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000, - 0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101, - 0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001, - 0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000, - 0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101, - 0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000, - 0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, - 0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000, - 0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100, - 0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000, - 0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000, - 0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100, - 0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff, - 0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff, - 0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00, - 0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101, - 0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000, - 0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00, - 0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000, - 0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff, - 0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101, - 0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff, - 0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00, - 0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff, -}; - -static const __device__ uint8_t ksigns_iq2xs[128] = { - 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, - 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, - 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, - 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, - 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, - 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, - 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, - 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, -}; - -//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics -static const __device__ uint64_t ksigns64[128] = { - 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff, - 0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff, - 0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff, - 0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff, - 0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff, - 0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff, - 0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff, - 0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff, - 0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff, - 0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff, - 0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff, - 0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff, - 0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff, - 0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff, - 0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff, - 0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff, - 0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff, - 0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff, - 0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff, - 0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff, - 0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff, - 0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff, - 0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff, - 0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff, - 0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff, - 0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff, - 0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff, - 0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff, - 0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff, - 0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff, - 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff, - 0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff, -}; -//#endif - -static const __device__ uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128}; - inline bool ggml_cuda_supports_mmq(enum ggml_type type) { switch (type) { case GGML_TYPE_Q4_0: diff --git a/ggml-metal.metal b/ggml-metal.metal index a65d12641..6ebbbd195 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -1,5 +1,8 @@ #include +#define GGML_COMMON_IMPL_METAL +#include "ggml-common.h" + using namespace metal; #define MAX(x, y) ((x) > (y) ? (x) : (y)) @@ -3638,710 +3641,6 @@ kernel void kernel_mul_mv_q6_K_f32( // ======================= "True" 2-bit -constexpr constant static uint64_t iq2xxs_grid[256] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, - 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, - 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, - 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, - 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, - 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, - 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, - 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, - 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, - 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, - 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, - 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, - 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, - 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, - 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, - 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, - 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, - 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, - 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, - 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, - 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, - 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, - 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, - 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, - 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, - 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, - 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, - 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, - 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, - 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, - 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, - 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, - 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, - 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, - 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, - 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, - 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, - 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, - 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, - 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, - 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, - 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, - 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, - 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, - 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, - 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, - 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, - 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, - 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, - 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, - 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, - 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, - 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, - 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, - 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, - 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, - 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, - 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, - 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, -}; - -constexpr constant static uint64_t iq2xs_grid[512] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, - 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, - 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, - 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, - 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, - 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, - 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, - 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, - 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, - 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, - 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, - 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, - 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, - 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, - 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, - 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, - 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, - 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, - 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, - 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, - 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, - 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, - 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, - 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, - 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, - 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, - 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, - 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, - 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, - 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, - 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, - 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, - 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, - 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, - 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, - 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, - 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, - 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, - 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, - 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, - 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, - 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, - 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, - 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, - 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, - 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, - 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, - 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, - 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, - 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, - 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, - 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, - 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, - 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, - 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, - 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, - 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, - 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, - 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, - 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, - 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, - 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, - 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, - 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, - 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, - 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, - 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, - 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, - 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, - 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, - 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, - 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, - 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, - 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, - 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, - 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, - 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, - 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, - 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, - 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, - 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, - 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, - 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, - 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, - 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, - 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, - 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, - 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, - 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, - 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, - 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, - 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, - 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, - 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, - 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, - 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, - 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, - 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, - 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, - 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, - 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, - 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, - 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, - 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, - 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, - 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, - 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, - 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, - 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, - 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, - 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, - 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, - 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, - 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, - 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, - 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, -}; - -constexpr constant static uint64_t iq2s_grid[1024] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b, - 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919, - 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808, - 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908, - 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b, - 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908, - 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08, - 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19, - 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819, - 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919, - 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b, - 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, - 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908, - 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908, - 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, - 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919, - 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b, - 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, - 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908, - 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b, - 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b, - 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08, - 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, - 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819, - 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808, - 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908, - 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b, - 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908, - 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08, - 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808, - 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08, - 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819, - 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908, - 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919, - 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b, - 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919, - 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808, - 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819, - 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919, - 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919, - 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808, - 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819, - 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b, - 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908, - 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, - 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, - 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919, - 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b, - 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919, - 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b, - 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819, - 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919, - 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908, - 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b, - 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908, - 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b, - 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908, - 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08, - 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908, - 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819, - 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819, - 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808, - 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, - 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19, - 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819, - 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808, - 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819, - 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919, - 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808, - 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19, - 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08, - 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b, - 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908, - 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808, - 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819, - 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908, - 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808, - 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808, - 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819, - 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908, - 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08, - 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819, - 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b, - 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08, - 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19, - 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819, - 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919, - 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908, - 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808, - 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808, - 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908, - 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808, - 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08, - 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08, - 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908, - 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919, - 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808, - 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819, - 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908, - 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08, - 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819, - 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808, - 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808, - 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819, - 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808, - 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908, - 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b, - 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, - 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, - 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b, - 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808, - 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b, - 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19, - 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819, - 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08, - 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b, - 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, - 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b, - 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b, - 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919, - 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808, - 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819, - 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908, - 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08, - 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08, - 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819, - 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919, - 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908, - 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b, - 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908, - 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b, - 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908, - 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08, - 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819, - 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808, - 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819, - 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919, - 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808, - 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808, - 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08, - 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819, - 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919, - 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808, - 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819, - 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919, - 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808, - 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b, - 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908, - 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808, - 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908, - 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b, - 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908, - 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b, - 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908, - 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b, - 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908, - 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08, - 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908, - 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b, - 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908, - 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08, - 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819, - 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919, - 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808, - 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19, - 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b, - 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919, - 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808, - 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819, - 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908, - 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919, - 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808, - 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808, - 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b, - 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919, - 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808, - 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b, - 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808, - 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919, - 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b, - 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08, - 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919, - 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808, - 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b, - 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908, - 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808, - 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808, - 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808, - 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908, - 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808, - 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808, - 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b, - 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908, - 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808, - 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808, - 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, - 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b, - 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808, - 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819, - 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b, - 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908, - 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08, - 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908, - 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919, - 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819, - 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908, - 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808, - 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819, - 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908, - 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919, - 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808, - 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808, - 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808, - 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919, - 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908, - 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908, - 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08, - 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819, - 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b, - 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808, - 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819, - 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908, - 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819, - 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808, - 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808, - 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b, - 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908, - 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808, - 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908, - 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819, - 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819, - 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808, - 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b, - 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b, - 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819, - 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b, - 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b, - 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b, - 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819, - 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19, - 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819, - 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908, - 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808, - 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b, -}; - -constexpr constant static uint32_t iq3xxs_grid[256] = { - 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, - 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, - 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, - 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, - 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, - 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, - 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, - 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, - 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, - 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, - 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, - 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, - 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, - 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, - 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, - 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, - 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, - 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, - 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, - 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, - 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, - 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, - 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, - 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, - 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, - 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, - 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, - 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, - 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, - 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, - 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, - 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, -}; - -constexpr constant static uint32_t iq3s_grid[512] = { - 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, - 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, - 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, - 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, - 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, - 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, - 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, - 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, - 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, - 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, - 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, - 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, - 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, - 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, - 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, - 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, - 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, - 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, - 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, - 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, - 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, - 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, - 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, - 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, - 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, - 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, - 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, - 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, - 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, - 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, - 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, - 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, - 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, - 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, - 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, - 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, - 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, - 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, - 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, - 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, - 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, - 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, - 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, - 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, - 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, - 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, - 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, - 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, - 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, - 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, - 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, - 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, - 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, - 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, - 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, - 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, - 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, - 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, - 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, - 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, - 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, - 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, - 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, - 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, -}; - -#define NGRID_IQ1S 512 -constexpr constant static uint64_t iq1s_grid[NGRID_IQ1S] = { - 0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000, - 0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01, - 0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100, - 0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00, - 0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101, - 0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100, - 0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00, - 0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff, - 0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000, - 0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000, - 0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001, - 0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff, - 0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01, - 0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001, - 0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00, - 0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001, - 0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100, - 0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000, - 0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000, - 0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000, - 0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff, - 0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff, - 0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01, - 0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100, - 0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff, - 0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000, - 0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101, - 0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff, - 0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff, - 0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001, - 0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01, - 0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101, - 0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100, - 0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00, - 0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001, - 0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff, - 0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000, - 0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000, - 0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100, - 0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100, - 0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01, - 0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff, - 0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101, - 0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000, - 0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff, - 0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000, - 0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff, - 0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00, - 0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101, - 0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000, - 0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000, - 0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000, - 0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100, - 0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000, - 0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001, - 0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff, - 0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000, - 0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000, - 0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000, - 0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000, - 0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff, - 0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000, - 0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, - 0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01, - 0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100, - 0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000, - 0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00, - 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100, - 0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000, - 0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, - 0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00, - 0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff, - 0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100, - 0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff, - 0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000, - 0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff, - 0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff, - 0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00, - 0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001, - 0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001, - 0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01, - 0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000, - 0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101, - 0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00, - 0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100, - 0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101, - 0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101, - 0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000, - 0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff, - 0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff, - 0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101, - 0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, - 0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101, - 0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001, - 0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff, - 0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff, - 0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01, - 0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff, - 0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100, - 0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001, - 0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00, - 0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff, - 0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff, - 0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000, - 0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000, - 0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101, - 0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001, - 0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000, - 0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101, - 0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000, - 0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, - 0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000, - 0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100, - 0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000, - 0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000, - 0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100, - 0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff, - 0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff, - 0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00, - 0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101, - 0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000, - 0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00, - 0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000, - 0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff, - 0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101, - 0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff, - 0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00, - 0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff, -}; - -constexpr constant static uint8_t ksigns_iq2xs[128] = { - 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, - 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, - 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, - 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, - 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, - 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, - 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, - 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, -}; - -constexpr constant static uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128}; - void kernel_mul_mv_iq2_xxs_f32_impl( device const void * src0, device const float * src1, diff --git a/ggml-quants.c b/ggml-quants.c index 9dcb76def..5bb46def3 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -1,6 +1,9 @@ #include "ggml-quants.h" #include "ggml-impl.h" +#define GGML_COMMON_IMPL_C +#include "ggml-common.h" + #include #include #include @@ -3327,711 +3330,6 @@ size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int // ====================== "True" 2-bit (de)-quantization -static const uint64_t iq2xxs_grid[256] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808, - 0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, - 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819, - 0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b, - 0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, - 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08, - 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b, - 0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, - 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08, - 0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, - 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, - 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808, - 0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808, - 0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, - 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08, - 0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, - 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819, - 0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808, - 0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, - 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908, - 0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808, - 0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, - 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908, - 0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19, - 0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, - 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b, - 0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808, - 0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, - 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08, - 0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08, - 0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, - 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819, - 0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808, - 0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, - 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19, - 0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819, - 0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, - 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b, - 0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08, - 0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, - 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908, - 0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b, - 0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, - 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08, - 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08, - 0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, - 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b, - 0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b, - 0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, - 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819, - 0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808, - 0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, - 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b, - 0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808, - 0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, - 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b, - 0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808, - 0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, - 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908, -}; - -static const uint64_t iq2xs_grid[512] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, - 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819, - 0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819, - 0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, - 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b, - 0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b, - 0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, - 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908, - 0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919, - 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, - 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908, - 0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, - 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, - 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08, - 0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, - 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808, - 0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819, - 0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, - 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808, - 0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819, - 0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819, - 0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, - 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908, - 0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19, - 0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, - 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b, - 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919, - 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, - 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819, - 0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819, - 0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, - 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908, - 0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808, - 0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, - 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808, - 0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, - 0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, - 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808, - 0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908, - 0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, - 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, - 0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b, - 0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, - 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, - 0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908, - 0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, - 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908, - 0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919, - 0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, - 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19, - 0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b, - 0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, - 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808, - 0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08, - 0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, - 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908, - 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b, - 0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, - 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, - 0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808, - 0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, - 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08, - 0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819, - 0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, - 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808, - 0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808, - 0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, - 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819, - 0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908, - 0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, - 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b, - 0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908, - 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, - 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908, - 0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808, - 0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, - 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819, - 0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819, - 0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, - 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b, - 0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819, - 0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, - 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08, - 0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808, - 0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, - 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919, - 0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808, - 0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, - 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b, - 0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808, - 0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, - 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b, - 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, - 0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808, - 0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819, - 0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, - 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808, - 0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, - 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19, - 0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08, - 0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, - 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08, - 0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08, - 0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, - 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908, - 0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b, - 0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, - 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808, - 0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b, - 0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, - 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808, - 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19, - 0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, - 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808, - 0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, - 0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, - 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b, - 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, -}; - -static const uint64_t iq2s_grid[1024] = { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08, - 0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b, - 0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b, - 0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919, - 0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b, - 0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919, - 0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808, - 0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908, - 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b, - 0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908, - 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08, - 0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19, - 0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819, - 0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919, - 0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b, - 0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, - 0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908, - 0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919, - 0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908, - 0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, - 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919, - 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b, - 0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919, - 0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908, - 0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b, - 0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b, - 0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08, - 0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, - 0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819, - 0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808, - 0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908, - 0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b, - 0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908, - 0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08, - 0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819, - 0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808, - 0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08, - 0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819, - 0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908, - 0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919, - 0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b, - 0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919, - 0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808, - 0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819, - 0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919, - 0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919, - 0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808, - 0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819, - 0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b, - 0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908, - 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, - 0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, - 0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919, - 0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b, - 0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919, - 0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b, - 0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819, - 0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919, - 0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908, - 0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b, - 0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908, - 0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b, - 0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908, - 0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08, - 0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908, - 0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819, - 0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819, - 0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808, - 0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, - 0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19, - 0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819, - 0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808, - 0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819, - 0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919, - 0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808, - 0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19, - 0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08, - 0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b, - 0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908, - 0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808, - 0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819, - 0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908, - 0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819, - 0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808, - 0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808, - 0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819, - 0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908, - 0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08, - 0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819, - 0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b, - 0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08, - 0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19, - 0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819, - 0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919, - 0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908, - 0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808, - 0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808, - 0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908, - 0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808, - 0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08, - 0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08, - 0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908, - 0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919, - 0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808, - 0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819, - 0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908, - 0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08, - 0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819, - 0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808, - 0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808, - 0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819, - 0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808, - 0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908, - 0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b, - 0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, - 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, - 0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b, - 0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808, - 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b, - 0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19, - 0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819, - 0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08, - 0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b, - 0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, - 0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b, - 0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b, - 0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919, - 0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808, - 0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819, - 0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908, - 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08, - 0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08, - 0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819, - 0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919, - 0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908, - 0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b, - 0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908, - 0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b, - 0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908, - 0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08, - 0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819, - 0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808, - 0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819, - 0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919, - 0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808, - 0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808, - 0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08, - 0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819, - 0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919, - 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808, - 0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819, - 0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919, - 0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808, - 0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b, - 0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908, - 0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808, - 0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908, - 0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b, - 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908, - 0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b, - 0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908, - 0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b, - 0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908, - 0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08, - 0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908, - 0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b, - 0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908, - 0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08, - 0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819, - 0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919, - 0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808, - 0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19, - 0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b, - 0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919, - 0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808, - 0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819, - 0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908, - 0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919, - 0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808, - 0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808, - 0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b, - 0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919, - 0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808, - 0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b, - 0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808, - 0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919, - 0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b, - 0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08, - 0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919, - 0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808, - 0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b, - 0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908, - 0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808, - 0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808, - 0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808, - 0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908, - 0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808, - 0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808, - 0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b, - 0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908, - 0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808, - 0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808, - 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, - 0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b, - 0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808, - 0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819, - 0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b, - 0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908, - 0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08, - 0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908, - 0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919, - 0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819, - 0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908, - 0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808, - 0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819, - 0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908, - 0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919, - 0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808, - 0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808, - 0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808, - 0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919, - 0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908, - 0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908, - 0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08, - 0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819, - 0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b, - 0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808, - 0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819, - 0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908, - 0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819, - 0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808, - 0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808, - 0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b, - 0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908, - 0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808, - 0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908, - 0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819, - 0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819, - 0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808, - 0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b, - 0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b, - 0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819, - 0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b, - 0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b, - 0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b, - 0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819, - 0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19, - 0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819, - 0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908, - 0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808, - 0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b, -}; - -static const uint32_t iq3xxs_grid[256] = { - 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414, - 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, - 0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, - 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e, - 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, - 0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, - 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34, - 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, - 0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, - 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04, - 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, - 0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, - 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434, - 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, - 0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, - 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24, - 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, - 0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, - 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c, - 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, - 0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, - 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e, - 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, - 0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, - 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c, - 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, - 0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, - 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c, - 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, - 0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, - 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c, - 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, -}; - -static const uint32_t iq3s_grid[512] = { - 0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305, - 0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905, - 0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09, - 0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b, - 0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b, - 0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d, - 0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03, - 0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505, - 0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03, - 0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901, - 0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d, - 0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303, - 0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501, - 0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105, - 0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505, - 0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101, - 0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707, - 0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b, - 0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01, - 0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f, - 0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305, - 0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103, - 0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509, - 0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503, - 0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b, - 0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f, - 0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f, - 0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f, - 0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109, - 0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f, - 0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509, - 0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501, - 0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303, - 0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f, - 0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907, - 0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703, - 0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03, - 0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01, - 0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01, - 0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903, - 0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505, - 0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b, - 0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107, - 0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509, - 0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303, - 0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103, - 0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05, - 0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b, - 0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f, - 0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701, - 0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909, - 0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305, - 0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d, - 0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b, - 0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d, - 0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307, - 0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09, - 0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309, - 0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709, - 0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f, - 0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303, - 0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503, - 0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b, - 0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101, -}; - -#define NGRID_IQ2XXS 512 -static const uint64_t iq1s_grid[NGRID_IQ2XXS] = { - 0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000, - 0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01, - 0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100, - 0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00, - 0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101, - 0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100, - 0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00, - 0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff, - 0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000, - 0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000, - 0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001, - 0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff, - 0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01, - 0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001, - 0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00, - 0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001, - 0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100, - 0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000, - 0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000, - 0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000, - 0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff, - 0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff, - 0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01, - 0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100, - 0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff, - 0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000, - 0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101, - 0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff, - 0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff, - 0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001, - 0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01, - 0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101, - 0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100, - 0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00, - 0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001, - 0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff, - 0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000, - 0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000, - 0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100, - 0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100, - 0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01, - 0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff, - 0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101, - 0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000, - 0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff, - 0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000, - 0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff, - 0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00, - 0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101, - 0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000, - 0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000, - 0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000, - 0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100, - 0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000, - 0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001, - 0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff, - 0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000, - 0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000, - 0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000, - 0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000, - 0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff, - 0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000, - 0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, - 0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01, - 0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100, - 0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000, - 0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00, - 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100, - 0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000, - 0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, - 0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00, - 0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff, - 0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100, - 0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff, - 0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000, - 0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff, - 0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff, - 0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00, - 0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001, - 0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001, - 0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01, - 0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000, - 0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101, - 0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00, - 0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100, - 0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101, - 0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101, - 0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000, - 0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff, - 0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff, - 0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101, - 0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, - 0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101, - 0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001, - 0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff, - 0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff, - 0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01, - 0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff, - 0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100, - 0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001, - 0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00, - 0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff, - 0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff, - 0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000, - 0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000, - 0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101, - 0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001, - 0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000, - 0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101, - 0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000, - 0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, - 0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000, - 0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100, - 0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000, - 0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000, - 0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100, - 0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff, - 0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff, - 0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00, - 0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101, - 0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000, - 0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00, - 0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000, - 0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff, - 0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101, - 0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff, - 0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00, - 0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff, - -}; - -static const uint8_t ksigns_iq2xs[128] = { - 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, 141, 142, 15, - 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, 154, 27, 156, 29, 30, 159, - 160, 33, 34, 163, 36, 165, 166, 39, 40, 169, 170, 43, 172, 45, 46, 175, - 48, 177, 178, 51, 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, - 192, 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, 78, 207, - 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, 219, 92, 221, 222, 95, - 96, 225, 226, 99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111, - 240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, -}; - -static const uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128}; - void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) { assert(k % QK_K == 0); const int nb = k / QK_K; diff --git a/ggml-quants.h b/ggml-quants.h index 316e35687..d2d25eb44 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -1,9 +1,9 @@ #pragma once -#include "ggml-impl.h" - // GGML internal header +#include "ggml-impl.h" + #include #include diff --git a/ggml-sycl.cpp b/ggml-sycl.cpp index ddd951dd6..6fafa5282 100644 --- a/ggml-sycl.cpp +++ b/ggml-sycl.cpp @@ -3144,6 +3144,8 @@ namespace dpct } // COPY from DPCT head files +#define GGML_COMMON_IMPL_SYCL +#include "ggml-common.h" static int g_ggml_sycl_debug=0; #define GGML_SYCL_DEBUG(...) do{if(g_ggml_sycl_debug) printf(__VA_ARGS__);}while(0) @@ -4745,388 +4747,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri #endif } -static dpct::global_memory - iq2xxs_grid(sycl::range<1>(256), - { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, - 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, - 0x0808080808191908, 0x08080808082b0808, 0x08080808082b082b, - 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819, - 0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, - 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, - 0x080808082b08082b, 0x080808082b082b2b, 0x080808082b2b082b, - 0x0808081908080819, 0x0808081908081908, 0x0808081908190808, - 0x0808081908191919, 0x0808081919080808, 0x080808192b081908, - 0x080808192b192b08, 0x0808082b08080808, 0x0808082b0808082b, - 0x0808082b082b082b, 0x0808082b2b08082b, 0x0808190808080819, - 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819, - 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, - 0x0808190819082b08, 0x08081908192b0808, 0x080819082b080819, - 0x080819082b081908, 0x080819082b190808, 0x080819082b2b1908, - 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08, - 0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, - 0x080819192b080808, 0x080819192b190819, 0x0808192b08082b19, - 0x0808192b08190808, 0x0808192b19080808, 0x0808192b2b081908, - 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919, - 0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, - 0x08082b0819080819, 0x08082b0819081908, 0x08082b0819190808, - 0x08082b081919082b, 0x08082b082b082b08, 0x08082b1908081908, - 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908, - 0x0819080808080819, 0x0819080808081908, 0x0819080808190808, - 0x08190808082b0819, 0x0819080819080808, 0x08190808192b0808, - 0x081908082b081908, 0x081908082b190808, 0x081908082b191919, - 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808, - 0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, - 0x0819082b082b1908, 0x0819082b19081919, 0x0819190808080808, - 0x0819190808082b08, 0x08191908082b0808, 0x08191908082b1919, - 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08, - 0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, - 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808, - 0x08192b0819080808, 0x08192b082b080819, 0x08192b1908080808, - 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819, - 0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, - 0x082b080819081908, 0x082b0808192b0819, 0x082b08082b080808, - 0x082b08082b08082b, 0x082b0819082b2b19, 0x082b081919082b08, - 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819, - 0x082b190808081908, 0x082b190808190808, 0x082b190819080808, - 0x082b19081919192b, 0x082b191908080808, 0x082b191919080819, - 0x082b1919192b1908, 0x082b192b2b190808, 0x082b2b0808082b08, - 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908, - 0x1908080808080819, 0x1908080808081908, 0x1908080808190808, - 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, - 0x1908080819080808, 0x1908080819082b08, 0x190808081919192b, - 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908, - 0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, - 0x19080819192b0819, 0x190808192b080808, 0x190808192b081919, - 0x1908082b08080819, 0x1908082b08190808, 0x1908082b19082b08, - 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808, - 0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, - 0x190819082b192b19, 0x190819190819082b, 0x19081919082b1908, - 0x1908192b08080808, 0x19082b0808080819, 0x19082b0808081908, - 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, - 0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, - 0x19082b192b08082b, 0x19082b2b19081919, 0x19082b2b2b190808, - 0x1919080808080808, 0x1919080808082b08, 0x1919080808190819, - 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808, - 0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, - 0x191908192b2b1908, 0x1919082b2b190819, 0x191919082b190808, - 0x191919082b19082b, 0x1919191908082b2b, 0x1919192b08080819, - 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819, - 0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, - 0x19192b2b08082b08, 0x192b080808081908, 0x192b080808190808, - 0x192b080819080808, 0x192b0808192b2b08, 0x192b081908080808, - 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808, - 0x192b190808080808, 0x192b190808081919, 0x192b191908190808, - 0x192b19190819082b, 0x192b19192b081908, 0x192b2b081908082b, - 0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808082b2b, - 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908, - 0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, - 0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, - 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, - 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908, - 0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, - 0x2b082b080808082b, 0x2b082b1908081908, 0x2b082b2b08190819, - 0x2b19080808081908, 0x2b19080808190808, 0x2b190808082b1908, - 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b, - 0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, - 0x2b191908082b082b, 0x2b19190819081908, 0x2b19191919190819, - 0x2b192b082b080819, 0x2b192b19082b0808, 0x2b2b08080808082b, - 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19, - 0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, - 0x2b2b2b1908081908, - }); - -static dpct::global_memory - iq2xs_grid(sycl::range<1>(512), - { - 0x0808080808080808, 0x080808080808082b, 0x0808080808081919, - 0x0808080808082b08, 0x0808080808082b2b, 0x0808080808190819, - 0x0808080808191908, 0x080808080819192b, 0x0808080808192b19, - 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919, - 0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, - 0x080808081908192b, 0x0808080819082b19, 0x0808080819190808, - 0x080808081919082b, 0x0808080819191919, 0x0808080819192b08, - 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808, - 0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, - 0x080808082b190819, 0x080808082b191908, 0x080808082b192b19, - 0x080808082b2b0808, 0x0808081908080819, 0x0808081908081908, - 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, - 0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, - 0x0808081908192b2b, 0x08080819082b0819, 0x08080819082b1908, - 0x0808081919080808, 0x080808191908082b, 0x0808081919081919, - 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908, - 0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, - 0x080808192b081908, 0x080808192b190808, 0x0808082b08080808, - 0x0808082b0808082b, 0x0808082b08081919, 0x0808082b08082b08, - 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808, - 0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, - 0x0808082b19191919, 0x0808082b2b080808, 0x0808082b2b082b2b, - 0x0808190808080819, 0x0808190808081908, 0x080819080808192b, - 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, - 0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, - 0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, - 0x0808190819081919, 0x0808190819082b08, 0x0808190819190819, - 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808, - 0x080819082b080819, 0x080819082b081908, 0x080819082b190808, - 0x0808191908080808, 0x080819190808082b, 0x0808191908081919, - 0x0808191908082b08, 0x0808191908190819, 0x0808191908191908, - 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908, - 0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, - 0x0808192b08080819, 0x0808192b08081908, 0x0808192b08190808, - 0x0808192b082b192b, 0x0808192b19080808, 0x0808192b1908082b, - 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b, - 0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, - 0x08082b0808190819, 0x08082b0808191908, 0x08082b08082b0808, - 0x08082b08082b1919, 0x08082b0819080819, 0x08082b0819081908, - 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808, - 0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, - 0x08082b1908081908, 0x08082b1908190808, 0x08082b1919080808, - 0x08082b192b080819, 0x08082b192b082b19, 0x08082b2b08080808, - 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b, - 0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, - 0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, - 0x081908080819082b, 0x0819080808191919, 0x0819080808192b08, - 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808, - 0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, - 0x0819080819190819, 0x0819080819191908, 0x08190808192b0808, - 0x08190808192b2b2b, 0x081908082b080819, 0x081908082b081908, - 0x081908082b190808, 0x0819081908080808, 0x081908190808082b, - 0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, - 0x0819081908191908, 0x08190819082b0808, 0x0819081919080819, - 0x0819081919081908, 0x0819081919190808, 0x081908192b080808, - 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819, - 0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, - 0x0819082b19080808, 0x0819082b192b0808, 0x0819190808080808, - 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08, - 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808, - 0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, - 0x0819190819190808, 0x08191908192b1908, 0x081919082b080808, - 0x0819191908080819, 0x0819191908081908, 0x0819191908190808, - 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908, - 0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, - 0x08192b0808190808, 0x08192b080819082b, 0x08192b0819080808, - 0x08192b0819191908, 0x08192b082b08192b, 0x08192b1908080808, - 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819, - 0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, - 0x082b080808081919, 0x082b080808082b08, 0x082b080808082b2b, - 0x082b080808190819, 0x082b080808191908, 0x082b0808082b0808, - 0x082b080819080819, 0x082b080819081908, 0x082b080819190808, - 0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, - 0x082b081908081908, 0x082b081908190808, 0x082b081919080808, - 0x082b081919082b08, 0x082b0819192b1919, 0x082b082b08080808, - 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08, - 0x082b190808080819, 0x082b190808081908, 0x082b190808190808, - 0x082b1908082b2b19, 0x082b190819080808, 0x082b191908080808, - 0x082b191919080819, 0x082b19191919082b, 0x082b19192b192b19, - 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b, - 0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, - 0x082b2b08082b0808, 0x082b2b0819191919, 0x082b2b082b082b08, - 0x082b2b082b2b082b, 0x082b2b19192b2b08, 0x082b2b192b190808, - 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b, - 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, - 0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, - 0x1908080808190808, 0x190808080819082b, 0x1908080808191919, - 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908, - 0x1908080819080808, 0x190808081908082b, 0x1908080819081919, - 0x1908080819082b08, 0x1908080819082b2b, 0x1908080819190819, - 0x1908080819191908, 0x19080808192b0808, 0x19080808192b1919, - 0x190808082b080819, 0x190808082b081908, 0x190808082b190808, - 0x1908081908080808, 0x190808190808082b, 0x1908081908081919, - 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908, - 0x19080819082b0808, 0x1908081919080819, 0x1908081919081908, - 0x1908081919190808, 0x190808192b080808, 0x190808192b081919, - 0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, - 0x1908082b08190808, 0x1908082b0819082b, 0x1908082b082b2b19, - 0x1908082b19080808, 0x1908190808080808, 0x190819080808082b, - 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819, - 0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, - 0x1908190819080819, 0x1908190819081908, 0x1908190819190808, - 0x190819082b080808, 0x190819082b191908, 0x1908191908080819, - 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908, - 0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, - 0x1908192b08082b2b, 0x1908192b19081908, 0x1908192b19190808, - 0x19082b0808080819, 0x19082b0808081908, 0x19082b0808190808, - 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908, - 0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, - 0x19082b1919081908, 0x19082b1919190808, 0x19082b19192b2b19, - 0x19082b2b08081908, 0x1919080808080808, 0x191908080808082b, - 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, - 0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, - 0x1919080819080819, 0x1919080819081908, 0x1919080819190808, - 0x191908082b080808, 0x1919081908080819, 0x1919081908081908, - 0x1919081908190808, 0x1919081908191919, 0x1919081919080808, - 0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, - 0x1919082b2b2b2b2b, 0x1919190808080819, 0x1919190808081908, - 0x1919190808190808, 0x19191908082b0819, 0x1919190819080808, - 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819, - 0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, - 0x191919192b082b08, 0x1919192b082b0819, 0x1919192b192b2b08, - 0x1919192b2b2b0819, 0x19192b0808080808, 0x19192b0808191908, - 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19, - 0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, - 0x19192b2b2b081919, 0x192b080808080819, 0x192b080808081908, - 0x192b080808190808, 0x192b080819080808, 0x192b080819191908, - 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19, - 0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, - 0x192b082b2b19082b, 0x192b190808080808, 0x192b19080819192b, - 0x192b191908190808, 0x192b191919080808, 0x192b191919081919, - 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b, - 0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, - 0x192b2b2b192b082b, 0x2b08080808080808, 0x2b0808080808082b, - 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819, - 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b, - 0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, - 0x2b0808082b080808, 0x2b0808082b08082b, 0x2b0808082b2b2b08, - 0x2b0808082b2b2b2b, 0x2b08081908080819, 0x2b08081908081908, - 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808, - 0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, - 0x2b08082b082b0808, 0x2b08082b2b080808, 0x2b08082b2b08082b, - 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08, 0x2b08190808080819, - 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b, - 0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, - 0x2b0819082b082b19, 0x2b08191908080808, 0x2b08191919081908, - 0x2b0819192b2b1919, 0x2b08192b08192b08, 0x2b08192b192b2b2b, - 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919, - 0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, - 0x2b082b082b2b2b08, 0x2b082b190808192b, 0x2b082b2b082b082b, - 0x2b082b2b2b080808, 0x2b082b2b2b082b08, 0x2b082b2b2b19192b, - 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908, - 0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, - 0x2b1908082b081908, 0x2b19081908080808, 0x2b190819082b082b, - 0x2b190819192b1908, 0x2b19082b1919192b, 0x2b19082b2b082b19, - 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908, - 0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, - 0x2b1919192b190808, 0x2b1919192b19082b, 0x2b19192b19080819, - 0x2b192b0819190819, 0x2b192b082b2b192b, 0x2b192b1919082b19, - 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808, - 0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, - 0x2b2b0808082b0808, 0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, - 0x2b2b081919190819, 0x2b2b081919192b19, 0x2b2b08192b2b192b, - 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08, - 0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, - 0x2b2b190819080808, 0x2b2b19082b191919, 0x2b2b192b192b1919, - 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b, 0x2b2b2b08082b0808, - 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808, - 0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, - 0x2b2b2b192b08192b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, - 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b, - }); - -static dpct::global_memory iq3xxs_grid( - sycl::range<1>(256), - { - 0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, - 0x04041404, 0x04041414, 0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, - 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14, 0x040c140c, 0x040c142c, - 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404, - 0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, - 0x04141c1c, 0x04141c3e, 0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, - 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c, 0x041c3e04, 0x04240c1c, - 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c, - 0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, - 0x043e0c24, 0x043e0c34, 0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, - 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c, 0x0c041c04, 0x0c041c14, - 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c, - 0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, - 0x0c14140c, 0x0c141c04, 0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, - 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c, 0x0c24042c, 0x0c242c04, - 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414, - 0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, - 0x14041414, 0x14041434, 0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, - 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c, 0x140c1c04, 0x140c341c, - 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e, - 0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, - 0x141c0c04, 0x141c0c24, 0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, - 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24, 0x143e040c, 0x143e041c, - 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c, - 0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, - 0x1c0c1404, 0x1c0c1c0c, 0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, - 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14, 0x1c1c0c0c, 0x1c1c1c1c, - 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414, - 0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, - 0x24040424, 0x24040c3e, 0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, - 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404, 0x24143404, 0x24143434, - 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c, - 0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, - 0x2c040c14, 0x2c04240c, 0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, - 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14, 0x2c1c0414, 0x2c1c2c1c, - 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c, - 0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, - 0x34043424, 0x340c140c, 0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, - 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14, 0x34341c1c, 0x343e041c, - 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14, - 0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, - 0x3e1c0404, 0x3e1c0c2c, 0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, - 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04, - }); - -static dpct::global_memory ksigns_iq2xs( - sycl::range<1>(128), - { - 0, 129, 130, 3, 132, 5, 6, 135, 136, 9, 10, 139, 12, - 141, 142, 15, 144, 17, 18, 147, 20, 149, 150, 23, 24, 153, - 154, 27, 156, 29, 30, 159, 160, 33, 34, 163, 36, 165, 166, - 39, 40, 169, 170, 43, 172, 45, 46, 175, 48, 177, 178, 51, - 180, 53, 54, 183, 184, 57, 58, 187, 60, 189, 190, 63, 192, - 65, 66, 195, 68, 197, 198, 71, 72, 201, 202, 75, 204, 77, - 78, 207, 80, 209, 210, 83, 212, 85, 86, 215, 216, 89, 90, - 219, 92, 221, 222, 95, 96, 225, 226, 99, 228, 101, 102, 231, - 232, 105, 106, 235, 108, 237, 238, 111, 240, 113, 114, 243, 116, - 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255, - }); - -static dpct::global_memory - ksigns64(sycl::range<1>(128), - { - 0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, - 0x000000000000ffff, 0xff00000000ff0000, 0x0000000000ff00ff, - 0x0000000000ffff00, 0xff00000000ffffff, 0xff000000ff000000, - 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff, - 0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, - 0x00000000ffffffff, 0xff0000ff00000000, 0x000000ff000000ff, - 0x000000ff0000ff00, 0xff0000ff0000ffff, 0x000000ff00ff0000, - 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff, - 0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, - 0x000000ffff00ffff, 0xff0000ffffff0000, 0x000000ffffff00ff, - 0x000000ffffffff00, 0xff0000ffffffffff, 0xff00ff0000000000, - 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff, - 0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, - 0x0000ff0000ffffff, 0x0000ff00ff000000, 0xff00ff00ff0000ff, - 0xff00ff00ff00ff00, 0x0000ff00ff00ffff, 0xff00ff00ffff0000, - 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff, - 0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, - 0x0000ffff0000ffff, 0xff00ffff00ff0000, 0x0000ffff00ff00ff, - 0x0000ffff00ffff00, 0xff00ffff00ffffff, 0xff00ffffff000000, - 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff, - 0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, - 0x0000ffffffffffff, 0xffff000000000000, 0x00ff0000000000ff, - 0x00ff00000000ff00, 0xffff00000000ffff, 0x00ff000000ff0000, - 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff, - 0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, - 0x00ff0000ff00ffff, 0xffff0000ffff0000, 0x00ff0000ffff00ff, - 0x00ff0000ffffff00, 0xffff0000ffffffff, 0x00ff00ff00000000, - 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff, - 0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, - 0xffff00ff00ffffff, 0xffff00ffff000000, 0x00ff00ffff0000ff, - 0x00ff00ffff00ff00, 0xffff00ffff00ffff, 0x00ff00ffffff0000, - 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff, - 0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, - 0x00ffff000000ffff, 0xffffff0000ff0000, 0x00ffff0000ff00ff, - 0x00ffff0000ffff00, 0xffffff0000ffffff, 0xffffff00ff000000, - 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff, - 0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, - 0x00ffff00ffffffff, 0xffffffff00000000, 0x00ffffff000000ff, - 0x00ffffff0000ff00, 0xffffffff0000ffff, 0x00ffffff00ff0000, - 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff, - 0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, - 0x00ffffffff00ffff, 0xffffffffffff0000, 0x00ffffffffff00ff, - 0x00ffffffffffff00, 0xffffffffffffffff, - }); -//#endif - -static dpct::global_memory - kmask_iq2xs(sycl::range<1>(8), {1, 2, 4, 8, 16, 32, 64, 128}); - template static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy, const sycl::nd_item<3> &item_ct1, diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh index 2c391e641..83e6359b1 100755 --- a/scripts/sync-ggml-am.sh +++ b/scripts/sync-ggml-am.sh @@ -94,6 +94,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then # src/ggml-alloc.c -> ggml-alloc.c # src/ggml-backend-impl.h -> ggml-backend-impl.h # src/ggml-backend.c -> ggml-backend.c + # src/ggml-common.h -> ggml-common.h # src/ggml-cuda.cu -> ggml-cuda.cu # src/ggml-cuda.h -> ggml-cuda.h # src/ggml-impl.h -> ggml-impl.h @@ -126,6 +127,7 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then -e 's/src\/ggml-alloc\.c/ggml-alloc.c/g' \ -e 's/src\/ggml-backend-impl\.h/ggml-backend-impl.h/g' \ -e 's/src\/ggml-backend\.c/ggml-backend.c/g' \ + -e 's/src\/ggml-common\.h/ggml-common.h/g' \ -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \ -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \ -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \ diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index feb34bbc8..f1e6f0e57 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -4,6 +4,7 @@ cp -rpv ../ggml/src/ggml.c ./ggml.c cp -rpv ../ggml/src/ggml-alloc.c ./ggml-alloc.c cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml-backend-impl.h cp -rpv ../ggml/src/ggml-backend.c ./ggml-backend.c +cp -rpv ../ggml/src/ggml-common.h ./ggml-common.h cp -rpv ../ggml/src/ggml-cuda.cu ./ggml-cuda.cu cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h cp -rpv ../ggml/src/ggml-impl.h ./ggml-impl.h From 0db32beaf09d90b8959d3d0cc493ed1e45685353 Mon Sep 17 00:00:00 2001 From: Alexey Parfenov Date: Sat, 9 Mar 2024 11:16:53 +0000 Subject: [PATCH 22/32] server : fix passing prompt as tokens (#5955) * server: fix passing prompt as tokens * Update examples/server/server.cpp --------- Co-authored-by: Georgi Gerganov --- examples/server/server.cpp | 11 ++++++++++- 1 file changed, 10 insertions(+), 1 deletion(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index aedf0afc6..8cff514f2 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -852,7 +852,16 @@ struct server_context { // infill slot.params.input_prefix = json_value(data, "input_prefix", default_params.input_prefix); slot.params.input_suffix = json_value(data, "input_suffix", default_params.input_suffix); - slot.prompt = json_value(data, "prompt", std::string("")); + + // get prompt + { + const auto & prompt = data.find("prompt"); + if (prompt == data.end()) { + slot.prompt = ""; + } else { + slot.prompt = *prompt; + } + } // penalize user-provided tokens { From 2c4f566c88322ebf2f9bd11b01b5ebdaa0130b89 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 14:17:11 +0200 Subject: [PATCH 23/32] tests : gitignore ggml-common.h --- tests/.gitignore | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/.gitignore b/tests/.gitignore index 9427cf13d..620a48ee4 100644 --- a/tests/.gitignore +++ b/tests/.gitignore @@ -1,3 +1,4 @@ * !*.* *.o +ggml-common.h From fb215c3832236fec7380c4fb618bd7154cb196ef Mon Sep 17 00:00:00 2001 From: SeungWon Jeong <65549245+redlion0929@users.noreply.github.com> Date: Sat, 9 Mar 2024 21:27:58 +0900 Subject: [PATCH 24/32] server : normalize embeddings (#5956) * output normalize embedding in '/v1/embeddings' * common : reuse llama_embd_normalize * common : better normalize impl --------- Co-authored-by: Georgi Gerganov --- common/common.cpp | 15 +++++++++++++++ common/common.h | 7 +++++++ examples/embedding/embedding.cpp | 14 +------------- examples/server/server.cpp | 8 +++++++- 4 files changed, 30 insertions(+), 14 deletions(-) diff --git a/common/common.cpp b/common/common.cpp index d7f650ef4..16ef4d7f7 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1852,3 +1852,18 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) { printf("\n=== Done dumping\n"); } + +void llama_embd_normalize(const float * inp, float * out, int n) { + double sum = 0.0; + for (int i = 0; i < n; i++) { + sum += inp[i] * inp[i]; + } + sum = sqrt(sum); + + const float norm = sum > 0.0 ? 1.0f / sum : 0.0f; + + for (int i = 0; i < n; i++) { + out[i] = inp[i] * norm; + } +} + diff --git a/common/common.h b/common/common.h index 977ce419f..f8d82b871 100644 --- a/common/common.h +++ b/common/common.h @@ -260,3 +260,10 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80); // Dump the KV cache view showing individual sequences in each cell (long output). void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40); + +// +// Embedding utils +// + +void llama_embd_normalize(const float * inp, float * out, int n); + diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index ff5883da6..a553ae1c3 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -23,17 +23,6 @@ static void batch_add_seq(llama_batch & batch, const std::vector & toke } } -static void normalize(const float * vec, float * out, int n) { - float norm = 0; - for (int i = 0; i < n; i++) { - norm += vec[i] * vec[i]; - } - norm = sqrt(norm); - for (int i = 0; i < n; i++) { - out[i] = vec[i] / norm; - } -} - static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) { // clear previous kv_cache values (irrelevant for embeddings) llama_kv_cache_clear(ctx); @@ -44,7 +33,6 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu fprintf(stderr, "%s : failed to decode\n", __func__); } - // normalize on copy for (int i = 0; i < batch.n_tokens; i++) { if (!batch.logits[i]) { continue; @@ -61,7 +49,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu } float * out = output + batch.seq_id[i][0] * n_embd; - normalize(embd, out, n_embd); + llama_embd_normalize(embd, out, n_embd); } } diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 8cff514f2..796f3499c 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1327,6 +1327,8 @@ struct server_context { const int n_embd = llama_n_embd(model); + std::vector embd_res(n_embd, 0.0f); + for (int i = 0; i < batch.n_tokens; ++i) { if (!batch.logits[i] || batch.seq_id[i][0] != slot.id + 1) { continue; @@ -1350,8 +1352,10 @@ struct server_context { continue; } + llama_embd_normalize(embd, embd_res.data(), n_embd); + res.data = json { - {"embedding", std::vector(embd, embd + n_embd)}, + {"embedding", embd_res}, }; } @@ -3354,6 +3358,8 @@ int main(int argc, char ** argv) { // get the result server_task_result result = ctx_server.queue_results.recv(id_task); ctx_server.queue_results.remove_waiting_task_id(id_task); + + // append to the responses responses.push_back(result.data); } From 97c09585d65a95864773b4d25d66d0f708baf38d Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 15:47:47 +0200 Subject: [PATCH 25/32] server : clarify some items in the readme (#5957) * server : clarify some items in the readme * server : fix typo --- examples/server/README.md | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-) diff --git a/examples/server/README.md b/examples/server/README.md index 3abb1abe3..23606b32a 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -195,7 +195,11 @@ node index.js *Options:* - `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does. + `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, if `cache_prompt` is `true`, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. A `BOS` token is inserted at the start, if all of the following conditions are true: + + - The prompt is a string or an array with the first element given as a string + - The model's `tokenizer.ggml.add_bos_token` metadata is `true` + - The system prompt is empty `temperature`: Adjust the randomness of the generated text (default: 0.8). @@ -308,7 +312,7 @@ Notice that each `probs` is an array of length `n_probs`. `content`: Set the text to tokenize. - Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`. + Note that a special `BOS` token is never inserted. - **POST** `/detokenize`: Convert tokens to text. From 5b09797321430f08caf0473143a962916ab2ea89 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 15:53:59 +0200 Subject: [PATCH 26/32] ggml : remove old quantization functions (#5942) * ggml : remove old quantization functions ggml-ci * ggml : simplify ggml_quantize_chunk ggml-ci * ggml : restrict correctness ggml-ci * ggml : remove hist data from the quantization API ggml-ci * tests : remove hist usage in test-backend-ops ggml-ci * vulkan : remove hist and fix typo --- examples/benchmark/benchmark-matmult.cpp | 6 +- examples/llava/clip.cpp | 55 +--- ggml-quants.c | 143 +++------- ggml-quants.h | 42 +-- ggml-vulkan.cpp | 40 +-- ggml.c | 339 +++-------------------- ggml.h | 23 +- llama.cpp | 48 +--- tests/test-backend-ops.cpp | 3 +- 9 files changed, 131 insertions(+), 568 deletions(-) diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index e89f3de2f..47cb16c69 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -189,12 +189,10 @@ int main(int argc, char ** argv) { int32_t nelements = sizex*sizey; - std::vector hist_cur(1 << 4, 0); - // Set up a the benchmark matrices // printf("Creating new tensor q11 & Running quantize\n"); struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr); + ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr); // Set up a the compute graph // printf("Creating new tensor q31\n"); @@ -207,7 +205,7 @@ int main(int argc, char ** argv) { // Set up a second graph computation to make sure we override the CPU cache lines // printf("Creating new tensor q12 & Running quantize\n"); struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey); - ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr); + ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr); // printf("Creating new tensor q32\n"); struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index ef9e4ba7a..6653b815d 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -1862,7 +1862,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i std::vector work(512); std::vector conv_buf(512); - std::vector hist_all(1 << 4, 0); size_t total_size_org = 0; size_t total_size_new = 0; @@ -1917,48 +1916,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i } new_data = work.data(); - std::vector hist_cur(1 << 4, 0); - - switch (new_type) { - case GGML_TYPE_Q4_0: { - new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q4_1: { - new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_0: { - new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_1: { - new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q8_0: { - new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q2_K: { - new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q3_K: { - new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q4_K: { - new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q5_K: { - new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - case GGML_TYPE_Q6_K: { - new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data()); - } break; - default: { - fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type); - return false; - } - } - - for (size_t j = 0; j < hist_cur.size(); ++j) { - hist_all[j] += hist_cur[j]; - } + new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr); } else { new_type = cur->type; new_data = cur->data; @@ -1993,17 +1951,6 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i { printf("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0); printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0); - - int64_t sum_all = 0; - for (size_t i = 0; i < hist_all.size(); ++i) { - sum_all += hist_all[i]; - } - - printf("%s: hist: ", __func__); - for (size_t i = 0; i < hist_all.size(); ++i) { - printf("%5.3f ", hist_all[i] / (float)sum_all); - } - printf("\n"); } return true; diff --git a/ggml-quants.c b/ggml-quants.c index 5bb46def3..4ee4e0604 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -1704,16 +1704,6 @@ void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) { quantize_row_q2_K_reference(x, vy, k); } -size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K; - quantize_row_q2_K_reference(src + j, y, k); - } - return (n/QK_K*sizeof(block_q2_K)); -} - static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights, uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux, float rmin, float rdelta, int nstep, bool use_mad) { @@ -1966,8 +1956,7 @@ 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; +size_t quantize_q2_K(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { 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); @@ -2186,16 +2175,6 @@ void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) { quantize_row_q3_K_reference(x, vy, k); } -size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K; - quantize_row_q3_K_reference(src + j, y, k); - } - return (n/QK_K*sizeof(block_q3_K)); -} - static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int n_per_row, const float * restrict quant_weights) { #if QK_K != 256 (void)quant_weights; @@ -2285,8 +2264,7 @@ static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restri #endif } -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; +size_t quantize_q3_K(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { 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); @@ -2456,17 +2434,6 @@ void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) { quantize_row_q4_K_reference(x, y, k); } -size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K; - quantize_row_q4_K_reference(src + j, y, k); - } - return (n/QK_K*sizeof(block_q4_K)); -} - static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int n_per_row, const float * quant_weights) { #if QK_K != 256 (void)quant_weights; @@ -2545,8 +2512,7 @@ static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restri #endif } -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; +size_t quantize_q4_K(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { 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); @@ -2757,17 +2723,6 @@ void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) { quantize_row_q5_K_reference(x, y, k); } -size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K; - quantize_row_q5_K_reference(src + j, y, k); - } - return (n/QK_K*sizeof(block_q5_K)); -} - static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int n_per_row, const float * quant_weights) { #if QK_K != 256 (void)quant_weights; @@ -2866,8 +2821,7 @@ static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restri #endif } -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; +size_t quantize_q5_K(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { 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); @@ -3020,17 +2974,6 @@ void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) { quantize_row_q6_K_reference(x, y, k); } -size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK_K == 0); - (void)hist; // TODO: collect histograms - - for (int j = 0; j < n; j += k) { - block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K; - quantize_row_q6_K_reference(src + j, y, k); - } - return (n/QK_K*sizeof(block_q6_K)); -} - static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int n_per_row, const float * quant_weights) { #if QK_K != 256 (void)quant_weights; @@ -3120,8 +3063,7 @@ static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restri #endif } -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; +size_t quantize_q6_K(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { 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); @@ -3165,9 +3107,10 @@ static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restri } } -size_t quantize_q4_0(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { +size_t quantize_q4_0(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { if (!quant_weights) { - return ggml_quantize_q4_0(src, dst, nrow*n_per_row, n_per_row, hist); + quantize_row_q4_0_reference(src, dst, nrow*n_per_row); + return nrow * 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; @@ -3209,9 +3152,10 @@ static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restri } } -size_t quantize_q4_1(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { +size_t quantize_q4_1(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { if (!quant_weights) { - return ggml_quantize_q4_1(src, dst, nrow*n_per_row, n_per_row, hist); + quantize_row_q4_1_reference(src, dst, nrow*n_per_row); + return nrow * 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; @@ -3262,9 +3206,10 @@ static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restri } } -size_t quantize_q5_0(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { +size_t quantize_q5_0(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { if (!quant_weights) { - return ggml_quantize_q5_0(src, dst, nrow*n_per_row, n_per_row, hist); + quantize_row_q5_0_reference(src, dst, nrow*n_per_row); + return nrow * 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; @@ -3314,9 +3259,10 @@ static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restri } } -size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { +size_t quantize_q5_1(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { if (!quant_weights) { - return ggml_quantize_q5_1(src, dst, nrow*n_per_row, n_per_row, hist); + quantize_row_q5_1_reference(src, dst, nrow*n_per_row); + return nrow * 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; @@ -3328,6 +3274,13 @@ size_t quantize_q5_1(const float * src, void * dst, int nrow, int n_per_row, int return nrow * row_size; } +size_t quantize_q8_0(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { + (void)quant_weights; // not used + const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row); + quantize_row_q8_0_reference(src, dst, nrow*n_per_row); + return nrow * row_size; +} + // ====================== "True" 2-bit (de)-quantization void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) { @@ -9373,7 +9326,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void #endif } -void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { +void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -9620,7 +9573,7 @@ static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) { } #endif -void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { +void ggml_vec_dot_iq1_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) { assert(n % QK_K == 0); assert(nrc == 1); UNUSED(nrc); @@ -10220,7 +10173,7 @@ void iq2xs_init_impl(enum ggml_type type) { int * kmap_q2xs; uint16_t * kneighbors_q2xs; - printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); + //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); uint64_t * the_grid = (uint64_t *)malloc(grid_size*sizeof(uint64_t)); for (int k = 0; k < grid_size; ++k) { int8_t * pos = (int8_t *)(the_grid + k); @@ -10275,7 +10228,7 @@ void iq2xs_init_impl(enum ggml_type type) { } num_neighbors += n; } - printf("%s: %d neighbours in total\n", __func__, num_neighbors); + //printf("%s: %d neighbours in total\n", __func__, num_neighbors); kneighbors_q2xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t)); iq2_data[gindex].neighbours = kneighbors_q2xs; int counter = 0; @@ -10698,8 +10651,7 @@ static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict v } } -size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -10711,8 +10663,7 @@ size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row, return nrow * nblock * sizeof(block_iq2_xxs); } -size_t quantize_iq2_xs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq2_xs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -10816,7 +10767,7 @@ void iq3xs_init_impl(int grid_size) { int * kmap_q3xs; uint16_t * kneighbors_q3xs; - printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); + //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size); uint32_t * the_grid = (uint32_t *)malloc(grid_size*sizeof(uint32_t)); for (int k = 0; k < grid_size; ++k) { int8_t * pos = (int8_t *)(the_grid + k); @@ -10871,7 +10822,7 @@ void iq3xs_init_impl(int grid_size) { } num_neighbors += n; } - printf("%s: %d neighbours in total\n", __func__, num_neighbors); + //printf("%s: %d neighbours in total\n", __func__, num_neighbors); kneighbors_q3xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t)); iq3_data[gindex].neighbours = kneighbors_q3xs; int counter = 0; @@ -11154,8 +11105,7 @@ static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, v } } -size_t quantize_iq3_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -11361,8 +11311,7 @@ static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, vo } #define IQ3S_BLOCK_SIZE 32 -size_t quantize_iq3_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; float scales[QK_K/IQ3S_BLOCK_SIZE]; @@ -11392,7 +11341,7 @@ void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int k) { void quantize_row_iq3_s_reference(const float * restrict x, block_iq3_s * restrict y, int k) { assert(k % QK_K == 0); - quantize_iq3_s(x, y, 1, k, NULL, NULL); + quantize_iq3_s(x, y, 1, k, NULL); } @@ -11587,8 +11536,7 @@ static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy } } -size_t quantize_iq1_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -11613,7 +11561,7 @@ static inline int best_index_int8(int n, const int8_t * val, float x) { return x - val[mu-1] < val[mu] - x ? mu-1 : mu; } -static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x, +static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * restrict x, ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l, float * scales, float * weight, uint8_t * L, const int8_t * values, @@ -11721,8 +11669,7 @@ static void quantize_row_iq4_nl_impl(const int super_block_size, const int block } } -size_t quantize_iq4_nl(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK4_NL == 0); int nblock = n_per_row/QK4_NL; char * qrow = (char *)dst; @@ -11752,14 +11699,13 @@ void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int k) { void quantize_row_iq4_nl_reference(const float * restrict x, block_iq4_nl * restrict y, int k) { assert(k % QK4_NL == 0); - quantize_iq4_nl(x, y, 1, k, NULL, NULL); + quantize_iq4_nl(x, y, 1, k, NULL); } -size_t quantize_iq4_xs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { +size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { #if QK_K == 64 - return quantize_iq4_nl(src, dst, nrow, n_per_row, hist, quant_weights); + return quantize_iq4_nl(src, dst, nrow, n_per_row, quant_weights); #else - (void)hist; GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -11788,7 +11734,7 @@ void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int k) { void quantize_row_iq4_xs_reference(const float * restrict x, block_iq4_xs * restrict y, int k) { assert(k % QK_K == 0); - quantize_iq4_xs(x, y, 1, k, NULL, NULL); + quantize_iq4_xs(x, y, 1, k, NULL); } // =============================== 2.5625 bpw @@ -11961,8 +11907,7 @@ static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy } } -size_t quantize_iq2_s(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) { - (void)hist; +size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int nrow, int n_per_row, const float * quant_weights) { GGML_ASSERT(n_per_row%QK_K == 0); int nblock = n_per_row/QK_K; char * qrow = (char *)dst; @@ -11976,7 +11921,7 @@ size_t quantize_iq2_s(const float * src, void * dst, int nrow, int n_per_row, in void quantize_row_iq2_s_reference(const float * restrict x, block_iq2_s * restrict y, int k) { assert(k % QK_K == 0); - quantize_iq2_s(x, y, 1, k, NULL, NULL); + quantize_iq2_s(x, y, 1, k, NULL); } void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int k) { diff --git a/ggml-quants.h b/ggml-quants.h index d2d25eb44..47dd52856 100644 --- a/ggml-quants.h +++ b/ggml-quants.h @@ -261,6 +261,7 @@ void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGM void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int k); void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k); void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k); + void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k); void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int k); void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int k); @@ -280,6 +281,7 @@ void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); + void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k); @@ -300,6 +302,7 @@ void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRI void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); + void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int k); @@ -321,6 +324,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); + void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); @@ -330,26 +334,26 @@ void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc); -// // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") -// -size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq2_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq1_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq4_nl (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq4_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_iq3_s (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q2_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q3_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q4_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q5_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q6_K (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -size_t quantize_q4_0 (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix); -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); +size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); + +size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); +size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix); void iq2xs_init_impl(enum ggml_type type); void iq2xs_free_impl(enum ggml_type type); diff --git a/ggml-vulkan.cpp b/ggml-vulkan.cpp index 5a1b3f477..d41aa7d22 100644 --- a/ggml-vulkan.cpp +++ b/ggml-vulkan.cpp @@ -4102,45 +4102,7 @@ static void ggml_vk_test_transfer(ggml_backend_vk_context * ctx, size_t ne, bool } static void ggml_vk_quantize_data(const float * from, void * to, size_t ne, ggml_type quant) { - std::vector hist_cur(1 << 4, 0); - - switch(quant) { - case GGML_TYPE_F32: - memcpy(to, from, sizeof(float) * ne); - break; - case GGML_TYPE_Q4_0: - ggml_quantize_q4_0(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q4_1: - ggml_quantize_q4_1(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_0: - ggml_quantize_q5_0(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_1: - ggml_quantize_q5_1(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q8_0: - ggml_quantize_q8_0(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q2_K: - ggml_quantize_q2_K(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q3_K: - ggml_quantize_q3_K(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q4_K: - ggml_quantize_q4_K(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q5_K: - ggml_quantize_q5_K(from, to, ne, ne, hist_cur.data()); - break; - case GGML_TYPE_Q6_K: - ggml_quantize_q6_K(from, to, ne, ne, hist_cur.data()); - break; - default: - GGML_ASSERT(false); - } + ggml_quantize_chunk(quant, from, to, 0, 1, ne, nullptr); } static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) { diff --git a/ggml.c b/ggml.c index 6eff98ab6..80efa6f2a 100644 --- a/ggml.c +++ b/ggml.c @@ -20159,133 +20159,6 @@ void ggml_quantize_free(void) { 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; - - for (int b = 0; b < n; b += k) { - block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0; - - quantize_row_q4_0_reference(src + b, y, k); - - for (int i = 0; i < nb; i++) { - for (int j = 0; j < QK4_0; j += 2) { - const uint8_t vi0 = y[i].qs[j/2] & 0x0F; - const uint8_t vi1 = y[i].qs[j/2] >> 4; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK4_0*sizeof(block_q4_0)); -} - -size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; - - for (int b = 0; b < n; b += k) { - block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1; - - quantize_row_q4_1_reference(src + b, y, k); - - for (int i = 0; i < nb; i++) { - for (int j = 0; j < QK4_1; j += 2) { - const uint8_t vi0 = y[i].qs[j/2] & 0x0F; - const uint8_t vi1 = y[i].qs[j/2] >> 4; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK4_1*sizeof(block_q4_1)); -} - -size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK5_0 == 0); - const int nb = k / QK5_0; - - for (int b = 0; b < n; b += k) { - block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0; - - quantize_row_q5_0_reference(src + b, y, k); - - for (int i = 0; i < nb; i++) { - uint32_t qh; - memcpy(&qh, &y[i].qh, sizeof(qh)); - - for (int j = 0; j < QK5_0; j += 2) { - const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4; - const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12)); - - // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK5_0*sizeof(block_q5_0)); -} - -size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK5_1 == 0); - const int nb = k / QK5_1; - - for (int b = 0; b < n; b += k) { - block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1; - - quantize_row_q5_1_reference(src + b, y, k); - - for (int i = 0; i < nb; i++) { - uint32_t qh; - memcpy(&qh, &y[i].qh, sizeof(qh)); - - for (int j = 0; j < QK5_1; j += 2) { - const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4; - const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12)); - - // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK5_1*sizeof(block_q5_1)); -} - -size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - for (int b = 0; b < n; b += k) { - block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0; - - quantize_row_q8_0_reference(src + b, y, k); - - for (int i = 0; i < nb; i++) { - for (int j = 0; j < QK8_0; ++j) { - const int8_t vi = y[i].qs[j]; - - hist[vi/16 + 8]++; - } - } - } - - return (n/QK8_0*sizeof(block_q8_0)); -} - bool ggml_quantize_requires_imatrix(enum ggml_type type) { return type == GGML_TYPE_IQ2_XXS || @@ -20293,177 +20166,52 @@ bool ggml_quantize_requires_imatrix(enum ggml_type type) { type == GGML_TYPE_IQ1_S; } -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) { +size_t ggml_quantize_chunk( + enum ggml_type type, + const float * src, + void * dst, + int start, + int nrows, + int n_per_row, + const float * imatrix) { + const int n = nrows * n_per_row; + + if (ggml_quantize_requires_imatrix(type)) { + GGML_ASSERT(imatrix != NULL); + } + + GGML_ASSERT(start % type_traits[type].blck_size == 0); + GGML_ASSERT(start % n_per_row == 0); + ggml_quantize_init(type); // this is noop if already initialized + + const size_t start_row = start / n_per_row; + const size_t row_size = ggml_row_size(type, n_per_row); + size_t result = 0; - int n = nrows * n_per_row; + switch (type) { - case GGML_TYPE_Q4_0: - { - GGML_ASSERT(start % QK4_0 == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q4_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q4_1: - { - GGML_ASSERT(start % QK4_1 == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q4_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q5_0: - { - GGML_ASSERT(start % QK5_0 == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q5_0(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q5_1: - { - GGML_ASSERT(start % QK5_1 == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q5_1(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q8_0: - { - GGML_ASSERT(start % QK8_0 == 0); - block_q8_0 * block = (block_q8_0*)dst + start / QK8_0; - result = ggml_quantize_q8_0(src + start, block, n, n, hist); - } break; - case GGML_TYPE_Q2_K: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q2_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q3_K: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q3_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q4_K: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q4_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q5_K: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q5_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_Q6_K: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_q6_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ2_XXS: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - GGML_ASSERT(imatrix); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq2_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ2_XS: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - GGML_ASSERT(imatrix); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ3_XXS: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq3_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ3_S: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq3_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ2_S: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq2_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ1_S: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq1_s(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; - case GGML_TYPE_IQ4_NL: + case GGML_TYPE_Q4_0: result = quantize_q4_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_1: result = quantize_q4_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q5_0: result = quantize_q5_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q5_1: result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q8_0: result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q2_K: result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q3_K: result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q4_K: result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q5_K: result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_Q6_K: result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ2_XS: result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ3_S: result = quantize_iq3_s (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ2_S: result = quantize_iq2_s (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ1_S: result = quantize_iq1_s (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; + case GGML_TYPE_IQ4_NL: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; #if QK_K == 64 - case GGML_TYPE_IQ4_XS: -#endif - { - GGML_ASSERT(start % QK4_NL == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq4_nl(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; -#if QK_K != 64 - case GGML_TYPE_IQ4_XS: - { - GGML_ASSERT(start % QK_K == 0); - GGML_ASSERT(start % n_per_row == 0); - size_t start_row = start / n_per_row; - size_t row_size = ggml_row_size(type, n_per_row); - result = quantize_iq4_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix); - GGML_ASSERT(result == row_size * nrows); - } break; + case GGML_TYPE_IQ4_XS: result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; +#else + case GGML_TYPE_IQ4_XS: result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break; #endif case GGML_TYPE_F16: { @@ -20480,6 +20228,9 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i default: assert(false); } + + GGML_ASSERT(result == nrows * row_size); + return result; } diff --git a/ggml.h b/ggml.h index a13b0cec4..1171088a9 100644 --- a/ggml.h +++ b/ggml.h @@ -2194,25 +2194,18 @@ extern "C" { 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); - GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist); - - GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist); - GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist); - 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); + GGML_API size_t ggml_quantize_chunk( + enum ggml_type type, + const float * src, + void * dst, + int start, + int nrows, + int n_per_row, + const float * imatrix); // // gguf diff --git a/llama.cpp b/llama.cpp index 8c147a42b..c58a029f7 100644 --- a/llama.cpp +++ b/llama.cpp @@ -11890,17 +11890,16 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty return new_type; } -static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, int64_t * hist_cur, const float * imatrix, std::vector & workers, const int nthread) { +static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, const float * imatrix, std::vector & workers, const int nthread) { std::mutex mutex; int counter = 0; size_t new_size = 0; if (nthread < 2) { // single-thread - return ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur, imatrix); + return ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, imatrix); } - auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size, + auto compute = [&mutex, &counter, &new_size, new_type, f32_data, new_data, chunk_size, nrows, n_per_row, imatrix]() { - std::array local_hist = {}; const int nrows_per_chunk = chunk_size / n_per_row; size_t local_size = 0; while (true) { @@ -11908,17 +11907,13 @@ static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const flo int first_row = counter; counter += nrows_per_chunk; if (first_row >= nrows) { if (local_size > 0) { - for (int j=0; j hist_all(1 << 4, 0); std::vector workers; workers.reserve(nthread); @@ -12175,7 +12169,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s work.resize(nelements * 4); // upper bound on size } new_data = work.data(); - std::array hist_cur = {}; const int n_per_row = tensor->ne[0]; const int nrows = nelements / n_per_row; @@ -12185,22 +12178,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s const int nchunk = (nelements + chunk_size - 1)/chunk_size; const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1; - new_size = llama_tensor_quantize_internal(new_type, f32_data, new_data, chunk_size, nrows, n_per_row, hist_cur.data(), imatrix, workers, nthread_use); + new_size = llama_tensor_quantize_internal(new_type, f32_data, new_data, chunk_size, nrows, n_per_row, imatrix, workers, nthread_use); - LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); - int64_t tot_count = 0; - for (size_t i = 0; i < hist_cur.size(); i++) { - hist_all[i] += hist_cur[i]; - tot_count += hist_cur[i]; - } - - if (tot_count > 0) { - LLAMA_LOG_INFO(" | hist: "); - for (size_t i = 0; i < hist_cur.size(); i++) { - LLAMA_LOG_INFO("%5.3f ", hist_cur[i] / float(nelements)); - } - } - LLAMA_LOG_INFO("\n"); + LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); } total_size_org += ggml_nbytes(tensor); total_size_new += new_size; @@ -12229,22 +12209,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s LLAMA_LOG_INFO("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0); LLAMA_LOG_INFO("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0); - // print histogram for all tensors - { - int64_t sum_all = 0; - for (size_t i = 0; i < hist_all.size(); i++) { - sum_all += hist_all[i]; - } - - if (sum_all > 0) { - LLAMA_LOG_INFO("%s: hist: ", __func__); - for (size_t i = 0; i < hist_all.size(); i++) { - LLAMA_LOG_INFO("%5.3f ", hist_all[i] / float(sum_all)); - } - LLAMA_LOG_INFO("\n"); - } - } - if (qs.n_fallback > 0) { LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n", __func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index 8a6999f21..fc5edcc4b 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -53,7 +53,6 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m } else if (ggml_is_quantized(tensor->type) || tensor->type == GGML_TYPE_F16) { GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0); std::vector dataq(ggml_row_size(tensor->type, size)); - int64_t hist[16]; std::vector imatrix(tensor->ne[0], 1.0f); // dummy importance matrix const float * im = imatrix.data(); if (!ggml_quantize_requires_imatrix(tensor->type)) { @@ -63,7 +62,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m im = nullptr; } } - ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], hist, im); + ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], 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. From 58308a0ecce7cc261b802f4803c38d420063db21 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 17:34:15 +0200 Subject: [PATCH 27/32] server : fix metrics init (#5964) --- examples/server/server.cpp | 17 ++++++++++++----- 1 file changed, 12 insertions(+), 5 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 796f3499c..2374b7e4a 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -341,7 +341,7 @@ struct server_slot { }; struct server_metrics { - const int64_t t_start = ggml_time_us(); + int64_t t_start = 0; uint64_t n_prompt_tokens_processed_total = 0; uint64_t t_prompt_processing_total = 0; @@ -354,14 +354,18 @@ struct server_metrics { uint64_t n_tokens_predicted = 0; uint64_t t_tokens_generation = 0; - void on_prompt_eval(const server_slot &slot) { + void init() { + t_start = ggml_time_us(); + } + + void on_prompt_eval(const server_slot & slot) { n_prompt_tokens_processed_total += slot.n_prompt_tokens_processed; n_prompt_tokens_processed += slot.n_prompt_tokens_processed; t_prompt_processing += slot.t_prompt_processing; t_prompt_processing_total += slot.t_prompt_processing; } - void on_prediction(const server_slot &slot) { + void on_prediction(const server_slot & slot) { n_tokens_predicted_total += slot.n_decoded; n_tokens_predicted += slot.n_decoded; t_tokens_generation += slot.t_token_generation; @@ -690,10 +694,11 @@ struct server_context { return res > 0; } - void initialize() { + void init() { const int32_t n_ctx_slot = n_ctx / params.n_parallel; LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}}); + for (int i = 0; i < params.n_parallel; i++) { server_slot slot; @@ -735,6 +740,8 @@ struct server_context { default_generation_settings_for_props["seed"] = -1; batch = llama_batch_init(n_ctx, 0, params.n_parallel); + + metrics.init(); } std::vector tokenize(const json & json_prompt, bool add_bos) const { @@ -2783,7 +2790,7 @@ int main(int argc, char ** argv) { state.store(SERVER_STATE_ERROR); return 1; } else { - ctx_server.initialize(); + ctx_server.init(); state.store(SERVER_STATE_READY); } From 8380ecfb219c2e73cc706fcc83935b7c806cc7c3 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 17:36:20 +0200 Subject: [PATCH 28/32] ggml : fix unnecessary f32 -> f16 -> f32 casts (mmla) (#5951) --- ggml-quants.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/ggml-quants.c b/ggml-quants.c index 4ee4e0604..807c5e391 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -4059,10 +4059,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16); // mmla into int32x4_t - float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d), - GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)}; + float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d, + GGML_FP16_TO_FP32(b_x0->d)*b_y1->d, + GGML_FP16_TO_FP32(b_x1->d)*b_y0->d, + GGML_FP16_TO_FP32(b_x1->d)*b_y1->d}; int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l))); From 098dbaab449f5309a54871ba7e5acef72ae696de Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 18:14:13 +0200 Subject: [PATCH 29/32] readme : update hot topics --- README.md | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index d7dba73e6..e3ec0817a 100644 --- a/README.md +++ b/README.md @@ -8,6 +8,11 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++ +> [!IMPORTANT] +> **Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962** +> +> Vote for which quantization type provides better responses, all other parameters being the same. + ### Recent API changes - [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328 @@ -16,11 +21,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) ### Hot topics -- The `api_like_OAI.py` script has been removed - use `server` instead ([#5766](https://github.com/ggerganov/llama.cpp/issues/5766#issuecomment-1969037761)) -- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) -- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631 -- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590 -- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216 +- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328 ---- From d894f352bf433157232dc8dc54eacd50014e898e Mon Sep 17 00:00:00 2001 From: slaren Date: Sat, 9 Mar 2024 19:55:54 +0100 Subject: [PATCH 30/32] perplexity : support using multiple sequences to allow larger batch sizes (#5946) * perplexity : support using multiple sequences to allow larger batch sizes ggml-ci * set cparams.n_parallel to the number of sequences * print tested n_ctx, add assert --- examples/perplexity/perplexity.cpp | 137 +++++++++++++++++++---------- llama.cpp | 22 +++-- 2 files changed, 107 insertions(+), 52 deletions(-) diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index 52789ee63..293eb52c3 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -442,7 +442,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & return {tokens, std::exp(nll / count), logit_history, prob_history}; } -static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) { +static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) { if (params.ppl_stride > 0) { return perplexity_v2(ctx, params); } @@ -453,7 +453,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // BOS tokens will be added for each chunk before eval 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()) { @@ -499,13 +498,19 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par double nll2 = 0.0; const int num_batches = (n_ctx + n_batch - 1) / n_batch; + const int n_seq = std::max(1, n_batch / n_ctx); + + GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0); + GGML_ASSERT(params.n_ctx == n_seq * n_ctx); + + llama_batch batch = llama_batch_init(std::min(n_batch, n_ctx*n_seq), 0, 1); 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); + fprintf(stderr, "%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq); std::vector workers(std::thread::hardware_concurrency() - 1); @@ -518,10 +523,26 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par log_probs.resize(n_ctx * nv); } - for (int i = 0; i < n_chunk; ++i) { + // We get the logits for all the tokens in the context window (params.n_ctx) + // from llama_eval above. Now, based on https://huggingface.co/docs/transformers/perplexity, + // calculate the perplexity over the last half of the window (so the model always has + // some context to predict the token). + // + // We rely on the fact that attention in the forward pass only looks at previous + // tokens here, so the logits returned for each token are an accurate representation + // of what the model would have predicted at that point. + // + // Example, we have a context window of 512, we will compute perplexity for each of the + // 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; + + for (int i = 0; i < n_chunk; i += n_seq) { const int start = i * n_ctx; const int end = start + n_ctx; + const int n_seq_batch = std::min(n_seq, n_chunk - i); + const auto t_start = std::chrono::high_resolution_clock::now(); // clear the KV cache @@ -531,22 +552,37 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par 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]; + batch.n_tokens = 0; + for (int seq = 0; seq < n_seq_batch; seq++) { + int seq_start = batch_start + seq*n_ctx; - // 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)); + // save original token and restore it after eval + const auto token_org = tokens[seq_start]; + + // add BOS token for the first batch of each chunk + if (add_bos && j == 0) { + tokens[seq_start] = llama_token_bos(llama_get_model(ctx)); + } + + for (int k = 0; k < batch_size; ++k) { + const int idx = seq*n_ctx + k; + batch.token[idx] = tokens[seq_start + k]; + batch.pos[idx] = j*n_batch + k; + batch.n_seq_id[idx] = 1; + batch.seq_id[idx][0] = seq; + batch.logits[idx] = batch.pos[idx] >= first ? 1 : 0; + } + batch.n_tokens += batch_size; + + // restore the original token in case it was set to BOS + tokens[seq_start] = token_org; } - if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) { + if (llama_decode(ctx, batch)) { fprintf(stderr, "%s : failed to eval\n", __func__); return {tokens, -1, logit_history, prob_history}; } - // 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); @@ -558,7 +594,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par 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); + int total_seconds = (int)(t_total*n_chunk/n_seq); if (total_seconds >= 60*60) { fprintf(stderr, "%d hours ", total_seconds / (60*60)); total_seconds = total_seconds % (60*60); @@ -566,37 +602,31 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0); } - // We get the logits for all the tokens in the context window (params.n_ctx) - // from llama_eval above. Now, based on https://huggingface.co/docs/transformers/perplexity, - // calculate the perplexity over the last half of the window (so the model always has - // some context to predict the token). - // - // We rely on the fact that attention in the forward pass only looks at previous - // tokens here, so the logits returned for each token are an accurate representation - // of what the model would have predicted at that point. - // - // Example, we have a context window of 512, we will compute perplexity for each of the - // 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; - const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx); - 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; + for (int seq = 0; seq < n_seq_batch; seq++) { + const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx); + llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first; + if (!params.logits_file.empty()) { + process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, + tokens_data, n_ctx - 1 - first, + workers, log_probs, nll, nll2); + } else { + process_logits(n_vocab, all_logits + first*n_vocab, + tokens_data, n_ctx - 1 - first, + workers, nll, nll2, + logit_history.data() + start + seq*n_ctx + first, + prob_history.data() + start + seq*n_ctx + first); + } + count += n_ctx - first - 1; - // perplexity is e^(average negative log-likelihood) - if (params.ppl_output_type == 0) { - printf("[%d]%.4lf,", i + 1, std::exp(nll / count)); - } else { - double av = nll/count; - double av2 = nll2/count - av*av; - if (av2 > 0) av2 = sqrt(av2/(count-1)); - printf("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2); + // perplexity is e^(average negative log-likelihood) + if (params.ppl_output_type == 0) { + printf("[%d]%.4lf,", i + seq + 1, std::exp(nll / count)); + } else { + double av = nll/count; + double av2 = nll2/count - av*av; + if (av2 > 0) av2 = sqrt(av2/(count-1)); + printf("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2); + } } fflush(stdout); @@ -615,6 +645,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par printf("Unexpected negative standard deviation of log(prob)\n"); } + llama_batch_free(batch); + return {tokens, ppl, logit_history, prob_history}; } @@ -1782,13 +1814,24 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) { int main(int argc, char ** argv) { gpt_params params; - params.n_batch = 512; if (!gpt_params_parse(argc, argv, params)) { return 1; } params.logits_all = true; - params.n_batch = std::min(params.n_batch, params.n_ctx); + + const int32_t n_ctx = params.n_ctx; + + const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence; + if (ppl) { + int n_seq = std::max(1, params.n_batch / n_ctx); + int32_t n_kv = n_seq * n_ctx; + params.n_parallel = n_seq; + params.n_ctx = n_kv; + params.n_batch = std::min(params.n_batch, n_kv); + } else { + params.n_batch = std::min(params.n_batch, params.n_ctx); + } if (params.ppl_stride > 0) { fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n", @@ -1847,7 +1890,7 @@ int main(int argc, char ** argv) { } else if (params.kl_divergence) { kl_divergence(ctx, params); } else { - results = perplexity(ctx, params); + results = perplexity(ctx, params, n_ctx); } llama_print_timings(ctx); diff --git a/llama.cpp b/llama.cpp index c58a029f7..b19616e8f 100644 --- a/llama.cpp +++ b/llama.cpp @@ -8925,17 +8925,29 @@ static int llama_decode_internal( if (batch.logits) { logits_out.resize(n_vocab * n_tokens); + int32_t i_first = -1; for (uint32_t i = 0; i < n_tokens; i++) { - if (batch.logits[i] == 0) { - continue; + if (batch.logits[i] && i_first == -1) { + i_first = (int32_t) i; + } + if (batch.logits[i] == 0 || i == n_tokens - 1) { + if (i_first != -1) { + int i_last = batch.logits[i] == 0 ? i : i + 1; + // extract logits for the range [i_first, i_last) + // group the requests to minimize the number of calls to the backend + ggml_backend_tensor_get_async(backend_res, res, + logits_out.data() + (n_vocab*i_first), + (n_vocab*i_first)*sizeof(float), + (i_last - i_first)*n_vocab*sizeof(float)); + i_first = -1; + } } - ggml_backend_tensor_get_async(backend_res, res, logits_out.data() + (n_vocab*i), (n_vocab*i)*sizeof(float), n_vocab*sizeof(float)); #ifndef NDEBUG - logits_valid[i] = true; + logits_valid[i] = batch.logits[i] != 0; #endif } } else if (lctx.logits_all) { - logits_out.resize(n_vocab * n_tokens); + logits_out.resize(n_vocab*n_tokens); ggml_backend_tensor_get_async(backend_res, res, logits_out.data(), 0, n_vocab*n_tokens*sizeof(float)); #ifndef NDEBUG std::fill(logits_valid.begin(), logits_valid.end(), true); From 77d1ac7e00bf049b9f2bba1b5a310a78318c49c4 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 9 Mar 2024 22:04:00 +0200 Subject: [PATCH 31/32] server : print chat template info --- examples/server/server.cpp | 22 ++++++++++++++++++++-- 1 file changed, 20 insertions(+), 2 deletions(-) diff --git a/examples/server/server.cpp b/examples/server/server.cpp index 2374b7e4a..b14cca61b 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2197,7 +2197,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co 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`\n"); printf(" --chat-template JINJA_TEMPLATE\n"); printf(" set custom jinja chat template (default: template taken from model's metadata)\n"); - printf(" Note: only commonly used templates are accepted, since we don't have jinja parser\n"); + printf(" only commonly used templates are accepted:\n"); + printf(" https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template\n"); printf("\n"); } @@ -2798,13 +2799,30 @@ int main(int argc, char ** argv) { const auto model_meta = ctx_server.model_meta(); - if (sparams.chat_template.empty()) { // custom chat template is not supplied + // if a custom chat template is not supplied, we will use the one that comes with the model (if any) + if (sparams.chat_template.empty()) { if (!ctx_server.validate_model_chat_template()) { LOG_ERROR("The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", {}); sparams.chat_template = "chatml"; } } + // print sample chat example to make it clear which template is used + { + json chat; + chat.push_back({{"role", "system"}, {"content", "You are a helpful assistant"}}); + chat.push_back({{"role", "user"}, {"content", "Hello"}}); + chat.push_back({{"role", "assistant"}, {"content", "Hi there"}}); + chat.push_back({{"role", "user"}, {"content", "How are you?"}}); + + const std::string chat_example = format_chat(ctx_server.model, sparams.chat_template, chat); + + LOG_INFO("chat template", { + {"chat_example", chat_example}, + {"built_in", sparams.chat_template.empty()}, + }); + } + // // Middlewares // From 621e86b331f8b0e71f79fd82a4ae1cd54c3e4396 Mon Sep 17 00:00:00 2001 From: Pierrick Hymbert Date: Sat, 9 Mar 2024 23:41:49 +0100 Subject: [PATCH 32/32] server: benchmark: chat/completions scenario and other llm servers comparison (#5941) * server: bench: Init a bench scenario with K6 See #5827 * server: bench: EOL EOF * server: bench: PR feedback and improved k6 script configuration * server: bench: remove llamacpp_completions_tokens_seconds as it include prompt processing time and it's misleading server: bench: add max_tokens from SERVER_BENCH_MAX_TOKENS server: bench: increase truncated rate to 80% before failing * server: bench: fix doc * server: bench: change gauge custom metrics to trend * server: bench: change gauge custom metrics to trend server: bench: add trend custom metrics for total tokens per second average * server: bench: doc add an option to debug http request * server: bench: filter dataset too short and too long sequences * server: bench: allow to filter out conversation in the dataset based on env variable * server: bench: fix assistant message sent instead of user message * server: bench: fix assistant message sent instead of user message * server : add defrag thold parameter * server: bench: select prompts based on the current iteration id not randomly to make the bench more reproducible --------- Co-authored-by: Georgi Gerganov --- examples/server/bench/README.md | 88 +++++++++++++++++++++++ examples/server/bench/script.js | 120 ++++++++++++++++++++++++++++++++ examples/server/server.cpp | 8 +++ 3 files changed, 216 insertions(+) create mode 100644 examples/server/bench/README.md create mode 100644 examples/server/bench/script.js diff --git a/examples/server/bench/README.md b/examples/server/bench/README.md new file mode 100644 index 000000000..a53ad64d7 --- /dev/null +++ b/examples/server/bench/README.md @@ -0,0 +1,88 @@ +### Server benchmark tools + +Benchmark is using [k6](https://k6.io/). + +##### Install k6 + +Follow instruction from: https://k6.io/docs/get-started/installation/ + +Example for ubuntu: +```shell +snap install k6 +``` + +#### Download a dataset + +This dataset was originally proposed in [vLLM benchmarks](https://github.com/vllm-project/vllm/blob/main/benchmarks/README.md). + +```shell +wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json +``` + +#### Download a model +Example for PHI-2 + +```shell +../../../scripts/hf.sh --repo ggml-org/models --file phi-2/ggml-model-q4_0.gguf +``` + +#### Start the server +The server must answer OAI Chat completion requests on `http://localhost:8080/v1` or according to the environment variable `SERVER_BENCH_URL`. + +Example: +```shell +server --host localhost --port 8080 \ + --model ggml-model-q4_0.gguf \ + --cont-batching \ + --metrics \ + --parallel 8 \ + --batch-size 512 \ + --ctx-size 4096 \ + --log-format text \ + -ngl 33 +``` + +#### Run the benchmark + +For 500 chat completions request with 8 concurrent users during maximum 10 minutes, run: +```shell +k6 run script.js --duration 10m --iterations 500 --vus 8 +``` + +The benchmark values can be overridden with: +- `SERVER_BENCH_URL` server url prefix for chat completions, default `http://localhost:8080/v1` +- `SERVER_BENCH_N_PROMPTS` total prompts to randomly select in the benchmark, default `480` +- `SERVER_BENCH_MODEL_ALIAS` model alias to pass in the completion request, default `my-model` +- `SERVER_BENCH_MAX_TOKENS` max tokens to predict, default: `512` +- `SERVER_BENCH_DATASET` path to the benchmark dataset file +- `SERVER_BENCH_MAX_PROMPT_TOKENS` maximum prompt tokens to filter out in the dataset: default `1024` +- `SERVER_BENCH_MAX_CONTEXT` maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens, default `2048` + +Note: the local tokenizer is just a string space split, real number of tokens will differ. + +Or with [k6 options](https://k6.io/docs/using-k6/k6-options/reference/): + +```shell +SERVER_BENCH_N_PROMPTS=500 k6 run script.js --duration 10m --iterations 500 --vus 8 +``` + +To [debug http request](https://k6.io/docs/using-k6/http-debugging/) use `--http-debug="full"`. + +#### Metrics + +Following metrics are available computed from the OAI chat completions response `usage`: +- `llamacpp_tokens_second` Trend of `usage.total_tokens / request duration` +- `llamacpp_prompt_tokens` Trend of `usage.prompt_tokens` +- `llamacpp_prompt_tokens_total_counter` Counter of `usage.prompt_tokens` +- `llamacpp_completion_tokens` Trend of `usage.completion_tokens` +- `llamacpp_completion_tokens_total_counter` Counter of `usage.completion_tokens` +- `llamacpp_completions_truncated_rate` Rate of completions truncated, i.e. if `finish_reason === 'length'` +- `llamacpp_completions_stop_rate` Rate of completions stopped by the model, i.e. if `finish_reason === 'stop'` + +The script will fail if too many completions are truncated, see `llamacpp_completions_truncated_rate`. + +K6 metrics might be compared against [server metrics](../README.md), with: + +```shell +curl http://localhost:8080/metrics +``` diff --git a/examples/server/bench/script.js b/examples/server/bench/script.js new file mode 100644 index 000000000..a4f5ac5ab --- /dev/null +++ b/examples/server/bench/script.js @@ -0,0 +1,120 @@ +import http from 'k6/http' +import {check, sleep} from 'k6' +import {SharedArray} from 'k6/data' +import {Counter, Rate, Trend} from 'k6/metrics' +import exec from 'k6/execution'; + +// Server chat completions prefix +const server_url = __ENV.SERVER_BENCH_URL ? __ENV.SERVER_BENCH_URL : 'http://localhost:8080/v1' + +// Number of total prompts in the dataset - default 10m / 10 seconds/request * number of users +const n_prompt = __ENV.SERVER_BENCH_N_PROMPTS ? parseInt(__ENV.SERVER_BENCH_N_PROMPTS) : 600 / 10 * 8 + +// Model name to request +const model = __ENV.SERVER_BENCH_MODEL_ALIAS ? __ENV.SERVER_BENCH_MODEL_ALIAS : 'my-model' + +// Dataset path +const dataset_path = __ENV.SERVER_BENCH_DATASET ? __ENV.SERVER_BENCH_DATASET : './ShareGPT_V3_unfiltered_cleaned_split.json' + +// Max tokens to predict +const max_tokens = __ENV.SERVER_BENCH_MAX_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_TOKENS) : 512 + +// Max prompt tokens +const n_prompt_tokens = __ENV.SERVER_BENCH_MAX_PROMPT_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_PROMPT_TOKENS) : 1024 + +// Max slot context +const n_ctx_slot = __ENV.SERVER_BENCH_MAX_CONTEXT ? parseInt(__ENV.SERVER_BENCH_MAX_CONTEXT) : 2048 + +export function setup() { + console.info(`Benchmark config: server_url=${server_url} n_prompt=${n_prompt} model=${model} dataset_path=${dataset_path} max_tokens=${max_tokens}`) +} + +const data = new SharedArray('conversations', function () { + const tokenizer = (message) => message.split(/[\s,'".?]/) + + return JSON.parse(open(dataset_path)) + // Filter out the conversations with less than 2 turns. + .filter(data => data["conversations"].length >= 2) + .filter(data => data["conversations"][0]["from"] === "human") + .map(data => { + return { + prompt: data["conversations"][0]["value"], + n_prompt_tokens: tokenizer(data["conversations"][0]["value"]).length, + n_completion_tokens: tokenizer(data["conversations"][1]["value"]).length, + } + }) + // Filter out too short sequences + .filter(conv => conv.n_prompt_tokens >= 4 && conv.n_completion_tokens >= 4) + // Filter out too long sequences. + .filter(conv => conv.n_prompt_tokens <= n_prompt_tokens && conv.n_prompt_tokens + conv.n_completion_tokens <= n_ctx_slot) + // Keep only first n prompts + .slice(0, n_prompt) +}) + +const llamacpp_prompt_tokens = new Trend('llamacpp_prompt_tokens') +const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens') +const llamacpp_tokens_second = new Trend('llamacpp_tokens_second') + +const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter') +const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter') + +const llamacpp_completions_truncated_rate = new Rate('llamacpp_completions_truncated_rate') +const llamacpp_completions_stop_rate = new Rate('llamacpp_completions_stop_rate') + +export const options = { + thresholds: { + llamacpp_completions_truncated_rate: [ + // more than 80% of truncated input will abort the test + {threshold: 'rate < 0.8', abortOnFail: true, delayAbortEval: '1m'}, + ], + }, + duration: '10m', + vus: 8, +} + +export default function () { + const conversation = data[exec.scenario.iterationInInstance % data.length] + const payload = { + "messages": [ + { + "role": "system", + "content": "You are ChatGPT, an AI assistant.", + }, + { + "role": "user", + "content": conversation.prompt, + } + ], + "model": model, + "stream": false, + "max_tokens": max_tokens + } + + const body = JSON.stringify(payload) + + let res = http.post(`${server_url}/chat/completions`, body, { + headers: {'Content-Type': 'application/json'}, + timeout: '300s' + }) + + check(res, {'success completion': (r) => r.status === 200}) + + if (res.status === 200) { + const completions = res.json() + + llamacpp_prompt_tokens.add(completions.usage.prompt_tokens) + llamacpp_prompt_tokens_total_counter.add(completions.usage.prompt_tokens) + + llamacpp_completion_tokens.add(completions.usage.completion_tokens) + llamacpp_completion_tokens_total_counter.add(completions.usage.completion_tokens) + + llamacpp_completions_truncated_rate.add(completions.choices[0].finish_reason === 'length') + llamacpp_completions_stop_rate.add(completions.choices[0].finish_reason === 'stop') + + llamacpp_tokens_second.add(completions.usage.total_tokens / res.timings.duration * 1.e3) + } else { + console.error(`response: ${res.body} request=${payload}`) + } + + sleep(0.3) +} diff --git a/examples/server/server.cpp b/examples/server/server.cpp index b14cca61b..c7d3ed01b 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -2133,6 +2133,8 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow); printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast); printf(" --pooling {none,mean,cls} pooling type for embeddings, use model default if unspecified\n"); + printf(" -dt N, --defrag-thold N\n"); + printf(" KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold); printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n"); printf(" not recommended: doubles context memory required and no measurable increase in quality\n"); @@ -2355,6 +2357,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams, else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; } else if (value == "cls") { params.pooling_type = LLAMA_POOLING_TYPE_CLS; } else { invalid_param = true; break; } + } else if (arg == "--defrag-thold" || arg == "-dt") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.defrag_thold = std::stof(argv[i]); } else if (arg == "--threads" || arg == "-t") { if (++i >= argc) {