Merge branch 'master' into add-stablelm-hash

.github/workflows/build.yml

@@ -898,9 +898,9 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
-
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
     steps:
       - name: Clone
         id: checkout
@@ -932,6 +932,17 @@ jobs:
         id: pack_artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts

CMakeLists.txt

@@ -296,7 +296,7 @@ if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
     endif()
-    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+    if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
         set(BLA_SIZEOF_INTEGER 8)
     endif()
 
@@ -431,7 +431,7 @@ if (LLAMA_CUDA)
 
         if (LLAMA_STATIC)
             if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
                 set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
             else ()
                 set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -1281,17 +1281,6 @@ install(
         WORLD_READ
         WORLD_EXECUTE
     DESTINATION ${CMAKE_INSTALL_BINDIR})
-install(
-    FILES convert-lora-to-ggml.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
 if (LLAMA_METAL)
     install(
         FILES ggml-metal.metal

README.md

@@ -2,7 +2,7 @@
 
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
-[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
@@ -140,6 +140,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
+- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 
 **HTTP server**
 
@@ -175,6 +176,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
+- [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)

ci/run.sh

@@ -365,47 +365,6 @@ function gg_run_open_llama_3b_v2 {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/3B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
     set +e
 }
 
@@ -416,7 +375,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -429,11 +387,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # open_llama_7b_v2
@@ -549,48 +502,6 @@ function gg_run_open_llama_7b_v2 {
 
     cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
 
-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/7B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # currently not supported by the CUDA backend
-    # q8_0
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
     set +e
 }
 
@@ -601,7 +512,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
     gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
     gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -614,11 +524,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
     gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
 # bge-small

common/common.cpp

@@ -901,6 +901,10 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
         params.interactive = true;
         return true;
     }
+    if (arg == "--interactive-specials") {
+        params.interactive_specials = true;
+        return true;
+    }
     if (arg == "--embedding") {
         params.embedding = true;
         return true;
@@ -1367,15 +1371,13 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
             std::replace(arg.begin(), arg.end(), '_', '-');
         }
-
         if (!gpt_params_find_arg(argc, argv, arg, params, i, invalid_param)) {
             throw std::invalid_argument("error: unknown argument: " + arg);
         }
-    }
-
         if (invalid_param) {
             throw std::invalid_argument("error: invalid parameter for argument: " + arg);
         }
+    }
 
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
@@ -1422,6 +1424,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -h, --help            show this help message and exit\n");
     printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
+    printf("  --interactive-specials allow special tokens in user text, in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -cnv, --conversation  run in conversation mode (does not print special tokens and suffix/prefix)\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -2652,6 +2655,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     dump_string_yaml_multiline(stream, "in_suffix", params.input_prefix.c_str());
     fprintf(stream, "instruct: %s # default: false\n", params.instruct ? "true" : "false");
     fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
+    fprintf(stream, "interactive_specials: %s # default: false\n", params.interactive_specials ? "true" : "false");
     fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
     fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
     fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());

common/common.h

@@ -140,6 +140,7 @@ struct gpt_params {
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
+    bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
     bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
     bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache

common/grammar-parser.cpp

@@ -142,6 +142,9 @@ namespace grammar_parser {
                 pos++;
                 last_sym_start = out_elements.size();
                 while (*pos != '"') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
                     auto char_pair = parse_char(pos);
                          pos       = char_pair.second;
                     out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
@@ -156,6 +159,9 @@ namespace grammar_parser {
                 }
                 last_sym_start = out_elements.size();
                 while (*pos != ']') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
                     auto char_pair = parse_char(pos);
                          pos       = char_pair.second;
                     enum llama_gretype type = last_sym_start < out_elements.size()
@@ -164,6 +170,9 @@ namespace grammar_parser {
 
                     out_elements.push_back({type, char_pair.first});
                     if (pos[0] == '-' && pos[1] != ']') {
+                        if (!pos[1]) {
+                            throw std::runtime_error("unexpected end of input");
+                        }
                         auto endchar_pair = parse_char(pos + 1);
                              pos          = endchar_pair.second;
                         out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});

common/sampling.cpp

@@ -35,7 +35,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
-    result->n_considered = 0;
+    result->n_valid = 0;
 
     llama_sampling_set_rng_seed(result, params.seed);
 
@@ -66,7 +66,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
 
     std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
     ctx->cur.clear();
-    ctx->n_considered = 0;
+    ctx->n_valid = 0;
 }
 
 void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
@@ -256,7 +256,7 @@ static llama_token llama_sampling_sample_impl(
         }
     }
 
-    ctx_sampling->n_considered = cur_p.size;
+    ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;
 
     return id;
 }

common/sampling.h

@@ -81,7 +81,7 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
-    size_t n_considered;
+    size_t n_valid; // Number of correct top tokens with correct probabilities.
 
     std::mt19937 rng;
 };

convert-hf-to-gguf-update.py

@@ -49,6 +49,10 @@ chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍
 
 if len(sys.argv) == 2:
     token = sys.argv[1]
+    if not token.startswith("hf_"):
+        logger.info("Huggingface token seems invalid")
+        logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
+        sys.exit(1)
 else:
     logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
     sys.exit(1)
@@ -77,6 +81,9 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-en",        "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-es",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-de",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
 ]
 
 # make directory "models/tokenizers" if it doesn't exist
@@ -175,7 +182,17 @@ for model in models:
     if tokt == TOKENIZER_TYPE.SPM:
         continue
 
+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
+    # create the tokenizer
+    try:
         tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+        continue  # Skip to the next model if the tokenizer can't be loaded
 
     chktok = tokenizer.encode(chktxt)
     chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -193,6 +210,8 @@ for model in models:
         logger.info("normalizer: " + json.dumps(normalizer, indent=4))
         pre_tokenizer = cfg["pre_tokenizer"]
         logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+        if "ignore_merges" in cfg["model"]:
+            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
 
     logger.info("")
 
@@ -293,6 +312,7 @@ tests = [
     "3333333",
     "33333333",
     "333333333",
+    # "Cửa Việt", # llama-bpe fails on this
     chktxt,
 ]
 
@@ -313,7 +333,17 @@ for model in models:
     name = model["name"]
     tokt = model["tokt"]
 
+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
+    # create the tokenizer
+    try:
         tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
+        continue  # Skip this model and continue with the next one in the loop
 
     with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
         for text in tests:

convert-hf-to-gguf.py

@@ -11,7 +11,6 @@ import re
 import sys
 from enum import IntEnum
 from hashlib import sha256
-from pathlib import Path
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -22,7 +21,16 @@ from typing import (
     Sequence,
     TypeVar,
     cast,
-    overload,
+)
+    TYPE_CHECKING,
+    Any,
+    Callable,
+    ContextManager,
+    Iterable,
+    Iterator,
+    Sequence,
+    TypeVar,
+    cast,
 )
 
 import numpy as np
@@ -59,7 +67,6 @@ class Model:
 
     dir_model: Path
     ftype: int
-    fname_out: Path
     is_big_endian: bool
     endianess: gguf.GGUFEndian
     use_temp_file: bool
@@ -67,20 +74,20 @@ class Model:
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
-    gguf_writer: gguf.GGUFWriter
     block_count: int
     tensor_map: gguf.TensorNameMap
     tensor_names: set[str] | None
+    fname_out: Path
+    gguf_writer: gguf.GGUFWriter
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
+    def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
-        if self.__class__ == Model:
+        if type(self) is Model:
-            raise TypeError(f"{self.__class__.__name__!r} should not be directly instantiated")
+            raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
         self.dir_model = dir_model
         self.ftype = ftype
-        self.fname_out = fname_out
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
         self.use_temp_file = use_temp_file
@@ -90,10 +97,23 @@ class Model:
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, ".bin")
         self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
+        if self.ftype == gguf.LlamaFileType.GUESSED:
+            # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
+            _, first_tensor = next(self.get_tensors())
+            if first_tensor.dtype == torch.float16:
+                logger.info(f"choosing --outtype f16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_F16
+            else:
+                logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
+                self.ftype = gguf.LlamaFileType.MOSTLY_BF16
+        ftype_up: str = self.ftype.name.partition("_")[2].upper()
+        ftype_lw: str = ftype_up.lower()
+        # allow templating the file name with the output ftype, useful with the "auto" ftype
+        self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up)
+        self.gguf_writer = gguf.GGUFWriter(self.fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
 
     @classmethod
     def __init_subclass__(cls):
@@ -153,14 +173,27 @@ class Model:
             raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
 
     def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
-        name: str = gguf.TENSOR_NAMES[key]
         if key not in gguf.MODEL_TENSORS[self.model_arch]:
             raise ValueError(f"Missing {key!r} for MODEL_TENSORS of {self.model_arch!r}")
+        name: str = gguf.TENSOR_NAMES[key]
         if "{bid}" in name:
             assert bid is not None
             name = name.format(bid=bid)
         return name + suffix
 
+    def match_model_tensor_name(self, name: str, key: gguf.MODEL_TENSOR, bid: int | None, suffix: str = ".weight") -> bool:
+        if key not in gguf.MODEL_TENSORS[self.model_arch]:
+            return False
+        key_name: str = gguf.TENSOR_NAMES[key]
+        if "{bid}" in key_name:
+            if bid is None:
+                return False
+            key_name = key_name.format(bid=bid)
+        else:
+            if bid is not None:
+                return False
+        return name == (key_name + suffix)
+
     def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
         new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes)
         if new_name is None:
@@ -226,6 +259,23 @@ class Model:
         return False
 
     def write_tensors(self):
+        # same as ggml_compute_fp32_to_bf16 in ggml-impl.h
+        def np_fp32_to_bf16(n: np.ndarray):
+            # force nan to quiet
+            n = np.where((n & 0x7fffffff) > 0x7f800000, (n & 0xffff0000) | (64 << 16), n)
+            # flush subnormals to zero
+            n = np.where((n & 0x7f800000) == 0, n & 0x80000000, n)
+            # round to nearest even
+            n = (n + (0x7fff + ((n >> 16) & 1))) >> 16
+            return n.astype(np.int16)
+
+        # Doing this row-wise is much, much faster than element-wise, hence the signature
+        v_fp32_to_bf16 = np.vectorize(np_fp32_to_bf16, otypes=[np.int16], signature="(n)->(n)")
+        if self.lazy:
+            # TODO: find a way to implicitly wrap np.vectorize functions
+            # NOTE: the type is changed to reflect otypes passed to np.vectorize above
+            v_fp32_to_bf16 = gguf.LazyNumpyTensor._wrap_fn(v_fp32_to_bf16, meta_noop=np.int16)
+
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
 
         for name, data_torch in self.get_tensors():
@@ -250,35 +300,60 @@ class Model:
                 data: np.ndarray = data  # type hint
                 n_dims = len(data.shape)
                 data_dtype = data.dtype
-
+                data_qtype: gguf.GGMLQuantizationType | None = None
-                # if f32 desired, convert any float16 to float32
-                if self.ftype == 0 and data_dtype == np.float16:
-                    data = data.astype(np.float32)
 
                 # when both are True, f32 should win
                 extra_f32 = self.extra_f32_tensors(name, new_name, bid, n_dims)
                 extra_f16 = self.extra_f16_tensors(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
-                extra_f32 = extra_f32 or n_dims == 1 or new_name.endswith("_norm.weight")
+                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
+                extra_f32 = any(cond for cond in (
+                    extra_f32,
+                    n_dims == 1,
+                    new_name.endswith("_norm.weight"),
+                ))
+
+                # Some tensor types are always in float32
+                extra_f32 = extra_f32 or any(self.match_model_tensor_name(new_name, key, bid) for key in (
+                    gguf.MODEL_TENSOR.FFN_GATE_INP,
+                    gguf.MODEL_TENSOR.POS_EMBD,
+                    gguf.MODEL_TENSOR.TOKEN_TYPES,
+                ))
 
                 # if f16 desired, convert any float32 2-dim weight tensors to float16
-                extra_f16 = extra_f16 or (name.endswith(".weight") and n_dims >= 2)
+                extra_f16 = any(cond for cond in (
+                    extra_f16,
+                    (name.endswith(".weight") and n_dims >= 2),
+                ))
 
-                # when both extra_f32 and extra_f16 are False, convert to float32 by default
+                if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
-                if self.ftype == 1 and data_dtype == np.float16 and (extra_f32 or not extra_f16):
+                    if self.ftype == gguf.LlamaFileType.MOSTLY_F16:
-                    data = data.astype(np.float32)
+                        if data_dtype != np.float16:
-
-                if self.ftype == 1 and data_dtype == np.float32 and extra_f16 and not extra_f32:
                             data = data.astype(np.float16)
+                        data_qtype = gguf.GGMLQuantizationType.F16
+
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
+                        if data_dtype != np.float32:
+                            data = data.astype(np.float32)
+                        data = v_fp32_to_bf16(data.view(np.int32))
+                        assert data.dtype == np.int16
+                        data_qtype = gguf.GGMLQuantizationType.BF16
+
+                else:  # by default, convert to float32
+                    if data_dtype != np.float32:
+                        data = data.astype(np.float32)
+                    data_qtype = gguf.GGMLQuantizationType.F32
+
+                assert data_qtype is not None
 
                 # reverse shape to make it similar to the internal ggml dimension order
                 shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
 
                 # n_dims is implicit in the shape
-                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data.dtype}, shape = {shape_str}")
+                logger.info(f"{f'%-{max_name_len}s' % f'{new_name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
 
-                self.gguf_writer.add_tensor(new_name, data)
+                self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)
 
     def write(self):
         self.write_tensors()
@@ -430,8 +505,17 @@ class Model:
             # ref: https://huggingface.co/allenai/OLMo-1.7-7B-hf
             res = "olmo"
         if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
-            # ref: https://huggingface.co/databricks/dbrx-instruct
+            # ref: https://huggingface.co/databricks/dbrx-base
             res = "dbrx"
+        if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
+            res = "jina-en"
+        if chkhsh == "171aeeedd6fb548d418a7461d053f11b6f1f1fc9b387bd66640d28a4b9f5c643":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-es
+            res = "jina-es"
+        if chkhsh == "27949a2493fc4a9f53f5b9b029c82689cfbe5d3a1929bb25e043089e28466de6":
+            # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-de
+            res = "jina-de"
 
         if res is None:
             logger.warning("\n")
@@ -1039,6 +1123,18 @@ class StarCoderModel(Model):
 class RefactModel(Model):
     model_arch = gguf.MODEL_ARCH.REFACT
 
+    def set_vocab(self):
+        super().set_vocab()
+
+        # TODO: how to determine special FIM tokens automatically?
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False,
+                                          special_token_types = ['prefix', 'suffix', 'middle', 'fsep', 'eot'])
+        special_vocab._set_special_token("prefix", 1)
+        special_vocab._set_special_token("suffix", 3)
+        special_vocab._set_special_token("middle", 2)
+        special_vocab._set_special_token("fsep",   4) # is this correct?
+        special_vocab.add_to_gguf(self.gguf_writer)
+
     def set_gguf_parameters(self):
         hidden_dim = self.hparams["n_embd"]
         inner_dim = 4 * hidden_dim
@@ -2049,12 +2145,6 @@ class BertModel(Model):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def extra_f32_tensors(self, name: str, new_name: str, bid: int | None, n_dims: int) -> bool:
-        del new_name, bid, n_dims  # unused
-
-        # not used with get_rows, must be F32
-        return name == "embeddings.token_type_embeddings.weight"
-
 
 @Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
@@ -2303,96 +2393,81 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("JinaBertModel", "JinaBertForMaskedLM")
+class JinaBertV2Model(BertModel):
+    model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.intermediate_size = self.hparams["intermediate_size"]
+
+    def get_tensors(self):
+        for name, data in super().get_tensors():
+            if 'gated_layers' in name:
+                d1 = data[:self.intermediate_size, :]
+                name1 = name.replace('gated_layers', 'gated_layers_w')
+                d2 = data[self.intermediate_size:, :]
+                name2 = name.replace('gated_layers', 'gated_layers_v')
+                yield name1, d1
+                yield name2, d2
+                continue
+
+            yield name, data
+
+    def set_vocab(self, *args, **kwargs):
+        tokenizer_class = 'BertTokenizer'
+        with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
+            tokenizer_class = json.load(f)['tokenizer_class']
+
+        if tokenizer_class == 'BertTokenizer':
+            super().set_vocab()
+        elif tokenizer_class == 'RobertaTokenizer':
+            self._set_vocab_gpt2()
+            self.gguf_writer.add_token_type_count(2)
+        else:
+            raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
+        self.gguf_writer.add_add_bos_token(True)
+        self.gguf_writer.add_add_eos_token(True)
+
+
 ###### CONVERSION LOGIC ######
 
 
 # tree of lazy tensors
-class LazyTorchTensor:
+class LazyTorchTensor(gguf.LazyBase):
-    _meta: Tensor
+    _tensor_type = torch.Tensor
-    _data: Tensor | None
+    # to keep the type-checker happy
-    _args: tuple
+    dtype: torch.dtype
-    _func: Callable[[tuple], Tensor] | None
+    shape: torch.Size
-
-    def __init__(self, *, meta: Tensor, data: Tensor | None = None, args: tuple = (), func: Callable[[tuple], Tensor] | None = None):
-        self._meta = meta
-        self._data = data
-        self._args = args
-        self._func = func
-
-    @staticmethod
-    def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
-        # TODO: dict and set
-        if isinstance(o, (list, tuple)):
-            L = []
-            for item in o:
-                L.append(LazyTorchTensor._recurse_apply(item, fn))
-            if isinstance(o, tuple):
-                L = tuple(L)
-            return L
-        elif isinstance(o, LazyTorchTensor):
-            return fn(o)
-        else:
-            return o
-
-    def _wrap_fn(self, fn: Callable, use_self: bool = False) -> Callable[[Any], LazyTorchTensor]:
-        def wrapped_fn(*args, **kwargs):
-            if kwargs is None:
-                kwargs = {}
-            args = ((self,) if use_self else ()) + args
-
-            meta_args = LazyTorchTensor._recurse_apply(args, lambda t: t._meta)
-
-            return LazyTorchTensor(meta=fn(*meta_args, **kwargs), args=args, func=lambda a: fn(*a, **kwargs))
-        return wrapped_fn
-
-    def __getattr__(self, __name: str) -> Any:
-        meta_attr = getattr(self._meta, __name)
-        if callable(meta_attr):
-            return self._wrap_fn(getattr(torch.Tensor, __name), use_self=True)
-        elif isinstance(meta_attr, torch.Tensor):
-            # for things like self.T
-            return self._wrap_fn(lambda s: getattr(s, __name))(self)
-        else:
-            return meta_attr
 
+    # only used when converting a torch.Tensor to a np.ndarray
     _dtype_map: dict[torch.dtype, type] = {
         torch.float16: np.float16,
         torch.float32: np.float32,
     }
 
-    def numpy(self) -> gguf.LazyTensor:
+    def numpy(self) -> gguf.LazyNumpyTensor:
         dtype = self._dtype_map[self.dtype]
-        return gguf.LazyTensor(lambda: LazyTorchTensor.to_eager(self).numpy(), dtype=dtype, shape=self.shape)
+        return gguf.LazyNumpyTensor(
+            meta=np.lib.stride_tricks.as_strided(np.zeros(1, dtype), self.shape, (0 for _ in self.shape)),
+            lazy=self._lazy,
+            args=(self,),
+            func=(lambda s: s[0].numpy())
+        )
 
-    @overload
+    @classmethod
-    @staticmethod
+    def eager_to_meta(cls, t: Tensor) -> Tensor:
-    def to_eager(t: Tensor | LazyTorchTensor) -> Tensor: ...
+        if t.is_meta:
-
-    @overload
-    @staticmethod
-    def to_eager(t: tuple) -> tuple: ...
-
-    @staticmethod
-    def to_eager(t: Any) -> Any:
-        def simple_to_eager(_t: LazyTorchTensor) -> Tensor:
-            # wake up the lazy tensor
-            if _t._data is None and _t._func is not None:
-                # recurse into its arguments
-                _t._args = LazyTorchTensor.to_eager(_t._args)
-                _t._data = _t._func(_t._args)
-            if _t._data is not None:
-                return _t._data
-            else:
-                raise ValueError(f"Could not compute lazy tensor {_t!r} with args {_t._args!r}")
-
-        # recurse into lists and/or tuples, keeping their structure
-        return LazyTorchTensor._recurse_apply(t, simple_to_eager)
-
-    @staticmethod
-    def from_eager(t: Tensor) -> Tensor:
-        if (t.__class__ == LazyTorchTensor):
             return t
-        return LazyTorchTensor(meta=t.detach().to("meta"), data=t)  # type: ignore
+        return t.detach().to("meta")
+
+    @classmethod
+    def meta_with_dtype(cls, m: Tensor, dtype: torch.dtype) -> Tensor:
+        m = m.detach()
+        if not m.is_meta:
+            m = m.to("meta")
+        m.dtype = dtype
+        return m
 
     @classmethod
     def __torch_function__(cls, func, types, args=(), kwargs=None):
@@ -2403,28 +2478,8 @@ class LazyTorchTensor:
 
         if func is torch.Tensor.numpy:
             return args[0].numpy()
-        if func is torch.equal:
-            eager_args = LazyTorchTensor.to_eager(args)
-            return func(*eager_args, **kwargs)
 
-        return LazyTorchTensor._wrap_fn(args[0], func)(*args, **kwargs)
+        return LazyTorchTensor._wrap_fn(func)(*args, **kwargs)
-
-    # special methods bypass __getattr__, so they need to be added manually
-    # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
-    # NOTE: LazyTorchTensor can't be a subclass of Tensor (and then be used
-    #       as self._meta is currently used), because then the following
-    #       operations would by default not be wrapped, and so not propagated
-    #       when the tensor is made eager.
-    #       It's better to get non-silent errors for not-yet-supported operators.
-    # TODO: add more when needed to avoid clutter, or find a more concise way
-    def __neg__(self, *args):  # mamba
-        return self._wrap_fn(torch.Tensor.__neg__)(self, *args)
-
-    def __add__(self, *args):  # gemma
-        return self._wrap_fn(torch.Tensor.__add__)(self, *args)
-
-    def __getitem__(self, *args):  # bloom falcon refact internlm2
-        return self._wrap_fn(torch.Tensor.__getitem__)(self, *args)
 
 
 def parse_args() -> argparse.Namespace:
@@ -2440,11 +2495,11 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--outfile", type=Path,
-        help="path to write to; default: based on input",
+        help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16"], default="f16",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -2498,16 +2553,18 @@ def main() -> None:
         logger.error(f'Error: {args.model} is not a directory')
         sys.exit(1)
 
-    ftype_map = {
+    ftype_map: dict[str, gguf.LlamaFileType] = {
-        "f32": gguf.GGMLQuantizationType.F32,
+        "f32": gguf.LlamaFileType.ALL_F32,
-        "f16": gguf.GGMLQuantizationType.F16,
+        "f16": gguf.LlamaFileType.MOSTLY_F16,
+        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
+        "auto": gguf.LlamaFileType.GUESSED,
     }
 
     if args.outfile is not None:
         fname_out = args.outfile
     else:
         # output in the same directory as the model by default
-        fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
+        fname_out = dir_model / 'ggml-model-{ftype}.gguf'
 
     logger.info(f"Loading model: {dir_model.name}")
 
@@ -2523,14 +2580,16 @@ def main() -> None:
         logger.info("Set model tokenizer")
         model_instance.set_vocab()
 
+        model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
+
         if args.vocab_only:
-            logger.info(f"Exporting model vocab to '{fname_out}'")
+            logger.info(f"Exporting model vocab to '{model_instance.fname_out}'")
             model_instance.write_vocab()
         else:
-            logger.info(f"Exporting model to '{fname_out}'")
+            logger.info(f"Exporting model to '{model_instance.fname_out}'")
             model_instance.write()
 
-        logger.info(f"Model successfully exported to '{fname_out}'")
+        logger.info(f"Model successfully exported to '{model_instance.fname_out}'")
 
 
 if __name__ == '__main__':

convert-lora-to-ggml.py

@@ -1,150 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any, BinaryIO, Sequence
-
-import numpy as np
-import torch
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-logging.basicConfig(level=logging.DEBUG)
-logger = logging.getLogger("lora-to-gguf")
-
-NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
-
-
-def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
-    fout.write(b"ggla"[::-1])  # magic (ggml lora)
-    fout.write(struct.pack("i", 1))  # file version
-    fout.write(struct.pack("i", params["r"]))
-    # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
-    # but some models ship a float value instead
-    # let's convert to int, but fail if lossless conversion is not possible
-    assert (
-        int(params["lora_alpha"]) == params["lora_alpha"]
-    ), "cannot convert float to int losslessly"
-    fout.write(struct.pack("i", int(params["lora_alpha"])))
-
-
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
-    sname = name.encode("utf-8")
-    fout.write(
-        struct.pack(
-            "iii",
-            len(shape),
-            len(sname),
-            NUMPY_TYPE_TO_FTYPE[data_type.name],
-        )
-    )
-    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-    fout.write(sname)
-    fout.seek((fout.tell() + 31) & -32)
-
-
-if __name__ == '__main__':
-    if len(sys.argv) < 2:
-        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
-        logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
-        logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-        sys.exit(1)
-
-    input_json = os.path.join(sys.argv[1], "adapter_config.json")
-    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
-
-    if os.path.exists(input_model):
-        model = torch.load(input_model, map_location="cpu")
-    else:
-        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-        model = load_file(input_model, device="cpu")
-
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
-
-    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-        logger.error(f"Error: unsupported architecture {arch_name}")
-        sys.exit(1)
-
-    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
-
-    with open(input_json, "r") as f:
-        params = json.load(f)
-
-    if params["peft_type"] != "LORA":
-        logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-        sys.exit(1)
-
-    if params["fan_in_fan_out"] is True:
-        logger.error("Error: param fan_in_fan_out is not supported")
-        sys.exit(1)
-
-    if params["bias"] is not None and params["bias"] != "none":
-        logger.error("Error: param bias is not supported")
-        sys.exit(1)
-
-    # TODO: these seem to be layers that have been trained but without lora.
-    # doesn't seem widely used but eventually should be supported
-    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-        logger.error("Error: param modules_to_save is not supported")
-        sys.exit(1)
-
-    with open(output_path, "wb") as fout:
-        fout.truncate()
-
-        write_file_header(fout, params)
-        for k, v in model.items():
-            orig_k = k
-            if k.endswith(".default.weight"):
-                k = k.replace(".default.weight", ".weight")
-            if k in ["llama_proj.weight", "llama_proj.bias"]:
-                continue
-            if k.endswith("lora_A.weight"):
-                if v.dtype != torch.float16 and v.dtype != torch.float32:
-                    v = v.float()
-                v = v.T
-            else:
-                v = v.float()
-
-            t = v.detach().numpy()
-
-            prefix = "base_model.model."
-            if k.startswith(prefix):
-                k = k[len(prefix) :]
-
-            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-            if k.endswith(lora_suffixes):
-                suffix = k[-len(lora_suffixes[0]):]
-                k = k[: -len(lora_suffixes[0])]
-            else:
-                logger.error(f"Error: unrecognized tensor name {orig_k}")
-                sys.exit(1)
-
-            tname = name_map.get_name(k)
-            if tname is None:
-                logger.error(f"Error: could not map tensor name {orig_k}")
-                logger.error(" Note: the arch parameter must be specified if the model is not llama")
-                sys.exit(1)
-
-            if suffix == ".lora_A.weight":
-                tname += ".weight.loraA"
-            elif suffix == ".lora_B.weight":
-                tname += ".weight.loraB"
-            else:
-                assert False
-
-            logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-            write_tensor_header(fout, tname, t.shape, t.dtype)
-            t.tofile(fout)
-
-    logger.info(f"Converted {input_json} and {input_model} to {output_path}")

docs/debugging-tests.md

@@ -0,0 +1,88 @@
+# Debugging Tests Tips
+
+## How to run & debug a specific test without anything else to keep the feedback loop short?
+
+There is a script called debug-test.sh in the scripts folder whose parameter takes a REGEX and an optional test number.
+
+For example, running the following command will output an interactive list from which you can select a test. It takes this form:
+
+`debug-test.sh [OPTION]... <test_regex> <test_number>`
+
+It will then build & run in the debugger for you.
+
+```bash
+./scripts/debug-test.sh test-tokenizer
+
+# Once in the debugger, i.e. at the chevrons prompt, setting a breakpoint could be as follows:
+>>> b main
+```
+
+For further reference use `debug-test.sh -h` to print help.
+
+&nbsp;
+
+### How does the script work?
+If you want to be able to use the concepts contained in the script separately, the important ones are briefly outlined below.
+
+#### Step 1: Reset and Setup folder context
+
+From base of this repository, let's create `build-ci-debug` as our build context.
+
+```bash
+rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+```
+
+#### Step 2: Setup Build Environment and Compile Test Binaries
+
+Setup and trigger a build under debug mode. You may adapt the arguments as needed, but in this case these are sane defaults.
+
+```bash
+cmake -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON ..
+make -j
+```
+
+#### Step 3.1: Identify Test Command for Debugging
+
+The output of this command will give you the command & arguments needed to run GDB.
+
+* `-R test-tokenizer` : looks for all the test files named `test-tokenizer*` (R=Regex)
+* `-N` : "show-only" disables test execution & shows test commands that you can feed to GDB.
+* `-V` : Verbose Mode
+
+```bash
+ctest -R "test-tokenizer" -V -N
+```
+
+This may return output similar to below (focusing on key lines to pay attention to):
+
+```bash
+...
+1: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+1: Working Directory: .
+Labels: main
+  Test  #1: test-tokenizer-0-llama-spm
+...
+4: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-falcon.gguf"
+4: Working Directory: .
+Labels: main
+  Test  #4: test-tokenizer-0-falcon
+...
+```
+
+So for test #1 we can tell these two pieces of relevant information:
+* Test Binary: `~/llama.cpp/build-ci-debug/bin/test-tokenizer-0`
+* Test GGUF Model: `~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf`
+
+#### Step 3.2: Run GDB on test command
+
+Based on the ctest 'test command' report above we can then run a gdb session via this command below:
+
+```bash
+gdb --args ${Test Binary} ${Test GGUF Model}
+```
+
+Example:
+
+```bash
+gdb --args ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+```

examples/convert-llama2c-to-ggml/README.md

@@ -2,7 +2,7 @@
 
 This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.
 
-To convert the model first download the models from the [llma2.c](https://github.com/karpathy/llama2.c) repository:
+To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository:
 
 `$ make -j`
 

examples/embedding/embedding.cpp

@@ -49,6 +49,12 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         }
 
         float * out = output + batch.seq_id[i][0] * n_embd;
+        //TODO: I would also add a parameter here to enable normalization or not.
+        /*fprintf(stdout, "unnormalized_embedding:");
+        for (int hh = 0; hh < n_embd; hh++) {
+            fprintf(stdout, "%9.6f ", embd[hh]);
+        }
+        fprintf(stdout, "\n");*/
         llama_embd_normalize(embd, out, n_embd);
     }
 }
@@ -123,10 +129,12 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // add SEP if not present
+    // check if the last token is SEP
+    // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
         if (inp.empty() || inp.back() != llama_token_sep(model)) {
-            inp.push_back(llama_token_sep(model));
+            fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
+            fprintf(stderr, "%s:          'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
 

examples/eval-callback/eval-callback.cpp

@@ -52,15 +52,15 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
                     float v;
                     if (type == GGML_TYPE_F16) {
-                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) data + i);
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                     } else if (type == GGML_TYPE_F32) {
-                        v = *(float *) data + i;
+                        v = *(float *) &data[i];
                     } else if (type == GGML_TYPE_I32) {
-                        v = (float) *(int32_t *) data + i;
+                        v = (float) *(int32_t *) &data[i];
                     } else if (type == GGML_TYPE_I16) {
-                        v = (float) *(int16_t *) data + i;
+                        v = (float) *(int16_t *) &data[i];
                     } else if (type == GGML_TYPE_I8) {
-                        v = (float) *(int8_t *) data + i;
+                        v = (float) *(int8_t *) &data[i];
                     } else {
                         GGML_ASSERT(false);
                     }

examples/llama-bench/README.md

@@ -26,16 +26,21 @@ options:
   -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
   -p, --n-prompt <n>                  (default: 512)
   -n, --n-gen <n>                     (default: 128)
-  -b, --batch-size <n>                (default: 512)
+  -pg <pp,tg>                         (default: 512,128)
-  -ctk <t>, --cache-type-k <t>        (default: f16)
+  -b, --batch-size <n>                (default: 2048)
-  -ctv <t>, --cache-type-v <t>        (default: f16)
+  -ub, --ubatch-size <n>              (default: 512)
-  -t, --threads <n>                   (default: 112)
+  -ctk, --cache-type-k <t>            (default: f16)
+  -ctv, --cache-type-v <t>            (default: f16)
+  -t, --threads <n>                   (default: 16)
   -ngl, --n-gpu-layers <n>            (default: 99)
   -sm, --split-mode <none|layer|row>  (default: layer)
   -mg, --main-gpu <i>                 (default: 0)
   -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -fa, --flash-attn <0|1>             (default: 0)
   -mmp, --mmap <0|1>                  (default: 1)
-  -ts, --tensor_split <ts0/ts1/..>    (default: 0)
+  --numa <distribute|isolate|numactl> (default: disabled)
+  -embd, --embeddings <0|1>           (default: 0)
+  -ts, --tensor-split <ts0/ts1/..>    (default: 0)
   -r, --repetitions <n>               (default: 5)
   -o, --output <csv|json|md|sql>      (default: md)
   -v, --verbose                       (default: 0)
@@ -43,10 +48,11 @@ options:
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```
 
-llama-bench can perform two types of tests:
+llama-bench can perform three types of tests:
 
 - Prompt processing (pp): processing a prompt in batches (`-p`)
 - Text generation (tg): generating a sequence of tokens (`-n`)
+- Prompt processing + text generation (pg): processing a prompt followed by generating a sequence of tokens (`-pg`)
 
 With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).
 

examples/llama-bench/llama-bench.cpp

@@ -161,10 +161,17 @@ static const char * split_mode_str(llama_split_mode mode) {
     }
 }
 
+static std::string pair_str(const std::pair<int, int> & p) {
+    static char buf[32];
+    snprintf(buf, sizeof(buf), "%d,%d", p.first, p.second);
+    return buf;
+}
+
 struct cmd_params {
     std::vector<std::string> model;
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
+    std::vector<std::pair<int, int>> n_pg;
     std::vector<int> n_batch;
     std::vector<int> n_ubatch;
     std::vector<ggml_type> type_k;
@@ -188,6 +195,7 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
+    /* n_pg          */ {{512, 128}},
     /* n_batch       */ {2048},
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
@@ -215,10 +223,11 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
     printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -pg <pp,tg>                         (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
     printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
+    printf("  -ub, --ubatch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
-    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctk, --cache-type-k <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -ctv, --cache-type-v <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
     printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
     printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
@@ -304,6 +313,17 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
+        } else if (arg == "-pg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], ',');
+            if (p.size() != 2) {
+                invalid_param = true;
+                break;
+            }
+            params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])});
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -493,6 +513,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.model.empty())        { params.model = cmd_params_defaults.model; }
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
+    if (params.n_pg.empty())         { params.n_pg = cmd_params_defaults.n_pg; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
     if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
     if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
@@ -632,6 +653,31 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
             };
             instances.push_back(instance);
         }
+
+        for (const auto & n_pg : params.n_pg) {
+            if (n_pg.first == 0 && n_pg.second == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_pg.first,
+                /* .n_gen        = */ n_pg.second,
+                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
+                /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
+                /* .tensor_split = */ ts,
+                /* .use_mmap     = */ mmp,
+                /* .embeddings   = */ embd,
+            };
+            instances.push_back(instance);
+        }
     }
 
     return instances;
@@ -965,6 +1011,9 @@ struct markdown_printer : public printer {
         if (field == "n_gpu_layers") {
             return 3;
         }
+        if (field == "test") {
+            return 13;
+        }
 
         int width = std::max((int)field.length(), 10);
 
@@ -1095,8 +1144,7 @@ struct markdown_printer : public printer {
                 } else if (t.n_gen > 0 && t.n_prompt == 0) {
                     snprintf(buf, sizeof(buf), "tg%d", t.n_gen);
                 } else {
-                    assert(false);
+                    snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
-                    exit(1);
                 }
                 value = buf;
             } else if (field == "t/s") {
@@ -1297,6 +1345,7 @@ int main(int argc, char ** argv) {
             llama_kv_cache_clear(ctx);
 
             uint64_t t_start = get_time_ns();
+
             if (t.n_prompt > 0) {
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
             }

examples/llava/clip.cpp

@@ -104,6 +104,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
@@ -425,6 +426,7 @@ struct clip_vision_model {
     // embeddings
     struct ggml_tensor * class_embedding;
     struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_bias;
     struct ggml_tensor * position_embeddings;
 
     struct ggml_tensor * pre_ln_w;
@@ -501,6 +503,11 @@ struct clip_ctx {
     bool use_gelu = false;
     int32_t ftype = 1;
 
+    bool has_class_embedding = true;
+    bool has_pre_norm = true;
+    bool has_post_norm = false;
+    bool has_patch_bias = false;
+
     struct gguf_context * ctx_gguf;
     struct ggml_context * ctx_data;
 
@@ -526,7 +533,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const int patch_size           = hparams.patch_size;
     const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
     const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
-    const int num_positions        = num_patches + 1;
+    const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
@@ -557,16 +564,23 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
     inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
 
+    if (ctx->has_patch_bias) {
+        // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
     // concat class_embeddings and patch_embeddings
-    struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+    struct ggml_tensor * embeddings = inp;
+    if (ctx->has_class_embedding) {
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
         ggml_set_name(embeddings, "embeddings");
         ggml_set_input(embeddings);
-
         embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-
         embeddings = ggml_acc(ctx0, embeddings, inp,
                 embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
+
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
@@ -576,7 +590,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
     // pre-layernorm
-    {
+    if (ctx->has_pre_norm) {
         embeddings = ggml_norm(ctx0, embeddings, eps);
         ggml_set_name(embeddings, "pre_ln");
 
@@ -664,6 +678,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         embeddings = cur;
     }
 
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
     // llava projector
     {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@@ -1149,11 +1171,38 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         try {
-            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
             vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
-            vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
+            new_clip->has_class_embedding = true;
+        } catch (const std::exception& e) {
+            new_clip->has_class_embedding = false;
+        }
+
+        try {
             vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
             vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+            new_clip->has_pre_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_pre_norm = false;
+        }
+
+        try {
+            vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
+            vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
+            new_clip->has_post_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_post_norm = false;
+        }
+
+        try {
+            vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
+            new_clip->has_patch_bias = true;
+        } catch (std::exception & e) {
+            new_clip->has_patch_bias = false;
+        }
+
+        try {
+            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+            vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& e) {
             LOG_TEE("%s: failed to load vision model tensors\n", __func__);
         }
@@ -1797,7 +1846,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int image_size    = hparams.image_size;
     const int patch_size    = hparams.patch_size;
     const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
-    const int num_positions = num_patches + 1;
+    const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
 
     {
         struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
@@ -1825,6 +1874,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     }
 
     {
+        if (ctx->has_class_embedding) {
             struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
 
             void* zero_mem = malloc(ggml_nbytes(embeddings));
@@ -1832,6 +1882,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
             free(zero_mem);
         }
+    }
 
     {
         struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");

examples/llava/llava-cli.cpp

@@ -189,6 +189,11 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     LOG_TEE("\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    if (!ctx_sampling) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
     std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);

examples/main/main.cpp

@@ -523,6 +523,10 @@ int main(int argc, char ** argv) {
     }
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+    if (!ctx_sampling) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
@@ -879,7 +883,7 @@ int main(int argc, char ** argv) {
                     }
 
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, params.interactive_specials);
                     const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
 
                     LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());

examples/server/server.cpp

@@ -651,9 +651,6 @@ struct server_context {
     std::string              system_prompt;
     std::vector<llama_token> system_tokens;
 
-    std::string name_user;      // this should be the antiprompt
-    std::string name_assistant;
-
     // slots / clients
     std::vector<server_slot> slots;
     json default_generation_settings_for_props;
@@ -673,6 +670,8 @@ struct server_context {
             llama_free_model(model);
             model = nullptr;
         }
+
+        llama_batch_free(batch);
     }
 
     bool load_model(const gpt_params & params_) {
@@ -1098,15 +1097,11 @@ struct server_context {
         system_need_update = false;
     }
 
-    void system_prompt_set(const json & sys_props) {
+    bool system_prompt_set(const std::string & sys_prompt) {
-        system_prompt  = sys_props.value("prompt", "");
+        system_prompt = sys_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
@@ -1115,6 +1110,7 @@ struct server_context {
         }
 
         system_need_update = true;
+        return true;
     }
 
     bool process_token(completion_token_output & result, server_slot & slot) {
@@ -1534,7 +1530,8 @@ struct server_context {
                     }
 
                     if (task.data.contains("system_prompt")) {
-                        system_prompt_set(task.data.at("system_prompt"));
+                        std::string sys_prompt = json_value(task.data, "system_prompt", std::string());
+                        system_prompt_set(sys_prompt);
 
                         for (server_slot & slot : slots) {
                             slot.n_past    = 0;
@@ -2270,10 +2267,10 @@ struct server_context {
 
                 const size_t n_probs = std::min(cur_p.size, (size_t) slot.sparams.n_probs);
                 if (n_probs > 0) {
-                    const size_t n_considered = slot.ctx_sampling->n_considered;
+                    const size_t n_valid = slot.ctx_sampling->n_valid;
 
                     // Make sure at least n_probs top tokens are at the front of the vector:
-                    if (slot.sparams.temp == 0.0f && n_probs > n_considered) {
+                    if (slot.sparams.temp == 0.0f && n_probs > n_valid) {
                         llama_sample_top_k(ctx, &cur_p, n_probs, 0);
                     }
 
@@ -2289,7 +2286,7 @@ struct server_context {
                         for (size_t i = 0; i < n_probs; ++i) {
                             result.probs.push_back({
                                 cur_p.data[i].id,
-                                i >= n_considered ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
+                                i >= n_valid ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
                             });
                         }
                     }
@@ -2918,7 +2915,7 @@ int main(int argc, char ** argv) {
     server_params_parse(argc, argv, sparams, params);
 
     if (!sparams.system_prompt.empty()) {
-        ctx_server.system_prompt_set(json::parse(sparams.system_prompt));
+        ctx_server.system_prompt_set(sparams.system_prompt);
     }
 
     if (params.model_alias == "unknown") {
@@ -3407,8 +3404,7 @@ int main(int argc, char ** argv) {
     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() },
+            { "system_prompt",               ctx_server.system_prompt.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 }
         };

ggml-backend.c

@@ -1182,9 +1182,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 static char * fmt_size(size_t size) {
     static char buffer[128];
     if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
     } else {
-        sprintf(buffer, "%zuK", size/1024);
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
     }
     return buffer;
 }

ggml-cuda.cu

@@ -4,7 +4,6 @@
 
 #include "ggml-cuda/common.cuh"
 #include "ggml-cuda/acc.cuh"
-#include "ggml-cuda/alibi.cuh"
 #include "ggml-cuda/arange.cuh"
 #include "ggml-cuda/argsort.cuh"
 #include "ggml-cuda/binbcast.cuh"
@@ -2205,6 +2204,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_UNARY_OP_RELU:
                     ggml_cuda_op_relu(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SIGMOID:
+                    ggml_cuda_op_sigmoid(ctx, dst);
+                    break;
                 case GGML_UNARY_OP_HARDSIGMOID:
                     ggml_cuda_op_hardsigmoid(ctx, dst);
                     break;
@@ -2277,9 +2279,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_ROPE:
             ggml_cuda_op_rope(ctx, dst);
             break;
-        case GGML_OP_ALIBI:
-            ggml_cuda_op_alibi(ctx, dst);
-            break;
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
@@ -2714,12 +2713,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 }
 
 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_SILU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_GELU_QUICK:
@@ -2829,7 +2830,6 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ROPE:
-        case GGML_OP_ALIBI:
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
@@ -2841,8 +2841,16 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_LEAKY_RELU:
-        case GGML_OP_FLASH_ATTN_EXT:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+#else
+            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
         default:
             return false;
     }

ggml-cuda/alibi.cu

@@ -1,63 +0,0 @@
-#include "alibi.cuh"
-
-static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
-                                 const int n_heads_log2_floor, const float m0, const float m1) {
-    const int col = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i = row*ncols + col;
-
-    const int k = row/k_rows;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = powf(m0, k + 1);
-    } else {
-        m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    dst[i] = col * m_k + x[i];
-}
-
-static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
-                           const int k_rows, const int n_heads_log2_floor, const float m0,
-                           const float m1, cudaStream_t stream) {
-    const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
-    const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
-    const dim3 block_nums(num_blocks_x, nrows, 1);
-    alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
-}
-
-void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    //GGML_ASSERT(ne01 + n_past == ne00);
-    GGML_ASSERT(n_head == ne02);
-
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    alibi_f32_cuda(src0_d, dst_d, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, stream);
-}

ggml-cuda/alibi.cuh

@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_ALIBI_BLOCK_SIZE 32
-
-void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/common.cuh

@@ -234,122 +234,6 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif //GGML_CUDA_F16
 
-[[noreturn]]
-static __device__ void no_device_code(
-    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
-           file_name, line, function_name, arch);
-    GGML_UNUSED(arch_list);
-#else
-    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
-           file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    __trap();
-
-    GGML_UNUSED(no_device_code); // suppress unused function warning
-}
-
-#ifdef __CUDA_ARCH__
-#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
-#else
-#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
-#endif // __CUDA_ARCH__
-
-static __device__ __forceinline__ float warp_reduce_sum(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
-    }
-    return x;
-}
-
-static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
-    }
-    return a;
-}
-
-static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-   }
-   return a;
-#else
-   GGML_UNUSED(a);
-   NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
-
-static __device__ __forceinline__ float warp_reduce_max(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-    }
-    return x;
-}
-
-static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-
-#if CUDART_VERSION >= CUDART_HMAX
-    return __hmax(a, b);
-#else
-    return __half2float(a) > __half2float(b) ? a : b;
-#endif // CUDART_VERSION >= CUDART_HMAX
-
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-}
-static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-
-#if CUDART_VERSION >= CUDART_HMAX
-    return __hmax2(a, b);
-#else
-    half2 ret;
-    reinterpret_cast<half&>(ret.x) =  __low2float(a) >  __low2float(b) ?  __low2half(a) :  __low2half(b);
-    reinterpret_cast<half&>(ret.y) = __high2float(a) > __high2float(b) ? __high2half(a) : __high2half(b);
-    return ret;
-#endif // CUDART_VERSION >= CUDART_HMAX
-
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-}
-
-static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-   }
-   return x;
-#else
-   GGML_UNUSED(x);
-   NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
-
-#if CUDART_VERSION < CUDART_HMASK
-static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
-    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
-    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
-    return mask_low | mask_high;
-}
-#endif // CUDART_VERSION < 12000
-
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300
 
@@ -433,11 +317,147 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 }
 #endif // defined(GGML_USE_HIPBLAS)
 
-#define FP16_AVAILABLE     defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) ? \
+#define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
-    defined(RDNA1) || defined(RDNA2) || defined(RDNA3) : __CUDA_ARCH__ >= CC_PASCAL
 
 #define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
 
+static bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
+static bool fp16_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+}
+
+[[noreturn]]
+static __device__ void no_device_code(
+    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
+           file_name, line, function_name, arch);
+    GGML_UNUSED(arch_list);
+#else
+    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
+           file_name, line, function_name, arch, arch_list);
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    __trap();
+
+    GGML_UNUSED(no_device_code); // suppress unused function warning
+}
+
+#ifdef __CUDA_ARCH__
+#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
+#else
+#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#endif // __CUDA_ARCH__
+
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if FP16_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+        reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
+        reinterpret_cast<half&>(a.y) += __high2half(a_other);
+    }
+    return a;
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ float warp_reduce_max(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+}
+
+static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
+#if FP16_AVAILABLE
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+    return __float2half(fmaxf(__half2float(a), __half2float(b)));
+#else
+    return __hmax(a, b);
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+
+#else
+   NO_DEVICE_CODE;
+   GGML_UNUSED(b);
+   return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+#if CUDART_VERSION >= CUDART_HMAX
+    return __hmax2(a, b);
+#else
+    half2 ret;
+    reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
+    reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
+    return ret;
+#endif // CUDART_VERSION >= CUDART_HMAX
+
+#else
+    GGML_UNUSED(a);
+    GGML_UNUSED(b);
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+   for (int mask = 16; mask > 0; mask >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   }
+   return x;
+#else
+   GGML_UNUSED(x);
+   NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+
+#if CUDART_VERSION < CUDART_HMASK
+static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
+    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
+    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
+    return mask_low | mask_high;
+}
+#endif // CUDART_VERSION < 12000
+
 // TODO: move to ggml-common.h
 static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
 

ggml-cuda/fattn-common.cuh

@@ -0,0 +1,47 @@
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}

ggml-cuda/fattn-vec-f16.cu

@@ -0,0 +1,430 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    half  slopeh = __float2half(1.0f);
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slopeh = __float2half(powf(base, exph));
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            half2 sum2[ncols] = {{0.0f, 0.0f}};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum2[j] = warp_reduce_sum(sum2[j]);
+                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
+            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f16(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    constexpr int cols_per_block = 1;
+    constexpr int parallel_blocks = 4;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 256:
+            launch_fattn_vec_f16<256, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f16< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f16<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f16.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn-vec-f32.cu

@@ -0,0 +1,384 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f32.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = powf(base, exph);
+    }
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    float2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE].x *= scale;
+            Q_h2[j][i0/WARP_SIZE].y *= scale;
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            float sum[ncols] = {0.0f};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum[j] +=  __low2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].x;
+                    sum[j] += __high2float(K_ik) * Q_h2[j][k_KQ_0/WARP_SIZE].y;
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum[j] = warp_reduce_sum(sum[j]);
+                sum[j] += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum[j]);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum[j];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = __half2float(V_h[(k_VKQ_0 + k)*stride_KV + tid]);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid != 0) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            dst_meta[(ic0 + j)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j], kqsum[j]);
+        }
+    }
+}
+
+template <int D, int cols_per_block, int parallel_blocks> void launch_fattn_vec_f32(
+        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
+        ggml_cuda_pool & pool, cudaStream_t main_stream
+) {
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
+    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const     dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const     int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks>
+        <<<blocks_num, block_dim, shmem, main_stream>>> (
+                (const char *) Q->data,
+                (const char *) K->data,
+                (const char *) V->data,
+                mask ? ((const char *) mask->data) : nullptr,
+                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+                scale, max_bias, m0, m1, n_head_log2,
+                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+                Q->nb[1], Q->nb[2], Q->nb[3],
+                K->nb[1], K->nb[2], K->nb[3],
+                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+                );
+    CUDA_CHECK(cudaGetLastError());
+
+    if (parallel_blocks == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->ne[0] == 64 || Q->ne[0] == 128 && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block = 1;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block = 2;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block = 4;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block = 8;
+        constexpr int parallel_blocks = 4;
+        switch (Q->ne[0]) {
+            case 64:
+                launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            case 128:
+                launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 8;
+    constexpr int parallel_blocks = 1;
+    switch (Q->ne[0]) {
+        case 64:
+            launch_fattn_vec_f32< 64, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        case 128:
+            launch_fattn_vec_f32<128, cols_per_block, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
+            break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}

ggml-cuda/fattn-vec-f32.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-cuda/fattn.cu

@@ -1,4 +1,7 @@
 #include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
 #include "fattn.cuh"
 
 #include <cstdint>
@@ -7,191 +10,11 @@
 #include <mma.h>
 #endif
 
-#define FATTN_KQ_STRIDE       256
-#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
-#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
-
-template<int D, int parallel_blocks> // D == head size
-__launch_bounds__(((D + WARP_SIZE - 1) / WARP_SIZE)*WARP_SIZE, 1)
-static __global__ void flash_attn_vec_ext_f16(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        float      * __restrict__ dst,
-        float2     * __restrict__ dst_meta,
-        const float scale,
-        const int ne00,
-        const int ne01,
-        const int ne02,
-        const int ne03,
-        const int ne10,
-        const int ne11,
-        const int ne12,
-        const int ne13,
-        const int ne31,
-        const int nb31,
-        const int nb01,
-        const int nb02,
-        const int nb03,
-        const int nb11,
-        const int nb12,
-        const int nb13,
-        const int ne0,
-        const int ne1,
-        const int ne2,
-        const int ne3) {
-#if FP16_AVAILABLE
-    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic = blockIdx.x / parallel_blocks; // Index of the Q/QKV column to work on.
-    const int ip = blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
-
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic);
-    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
-    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
-    const half   * maskh = (const half   *)  mask + ne11*ic;
-
-    const int stride_KV  = nb11 / sizeof(half);
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    constexpr int nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
-    __builtin_assume(tid < nwarps*WARP_SIZE);
-
-    __shared__ half KQ[nwarps*WARP_SIZE];
-    KQ[tid] = -INFINITY;
-    half2 * KQ2 = (half2 *) KQ;
-
-    half kqmax = -HALF_MAX_HALF;
-    half kqsum = 0.0f;
-
-    __shared__ half kqmax_shared[WARP_SIZE];
-    __shared__ half kqsum_shared[WARP_SIZE];
-    if (threadIdx.y == 0) {
-        kqmax_shared[threadIdx.x] = -HALF_MAX_HALF;
-        kqsum_shared[threadIdx.x] = 0.0f;
-    }
-    __syncthreads();
-
-    // Convert Q to half2 and store in registers:
-    half2 Q_h2[(D/2 + WARP_SIZE - 1) / WARP_SIZE];
-#pragma unroll
-    for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
-        const int i = i0 + threadIdx.x;
-        if (i0 + WARP_SIZE > D/2 && i >= D/2) {
-            break;
-        }
-
-        Q_h2[i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(Q_f2[i].x, Q_f2[i].y);
-    }
-
-    half2 VKQ = make_half2(0.0f, 0.0f); // Each thread calculates a single VKQ value.
-
-    const int k_start  = parallel_blocks == 1 ? 0 : ip*D;
-    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-        half kqmax_new = kqmax;
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
-            const int i_KQ = i_KQ_0 + threadIdx.y;
-
-            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
-                break;
-            }
-
-            half2 sum2 = make_half2(0.0f, 0.0f);
-#pragma unroll
-            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
-                const int k_KQ = k_KQ_0 + threadIdx.x;
-                if (k_KQ_0 + WARP_SIZE > D/2 && k_KQ >= D/2) {
-                    break;
-                }
-
-                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
-                sum2 += K_ik * Q_h2[k_KQ_0/WARP_SIZE];
-            }
-
-            sum2 = warp_reduce_sum(sum2);
-            half sum = __low2half(sum2) + __high2half(sum2);
-            sum += mask ? maskh[k_VKQ_0 + i_KQ] : __float2half(0.0f);
-            kqmax_new = ggml_cuda_hmax(kqmax_new, sum);
-            if (threadIdx.x == 0) {
-                KQ[i_KQ] = sum;
-            }
-        }
-
-        kqmax_new = warp_reduce_max(kqmax_new);
-        if (threadIdx.x == 0) {
-            kqmax_shared[threadIdx.y] = kqmax_new;
-        }
-        __syncthreads();
-        kqmax_new = kqmax_shared[threadIdx.x];
-        kqmax_new = warp_reduce_max(kqmax_new);
-
-        const half KQ_max_scale = hexp(kqmax - kqmax_new);
-        kqmax = kqmax_new;
-
-        const half val = hexp(KQ[tid] - kqmax);
-        kqsum = kqsum*KQ_max_scale + val;
-        KQ[tid] = val;
-
-        VKQ *= __half2half2(KQ_max_scale);
-
-        __syncthreads();
-
-        if (tid < D) {
-#pragma unroll
-            for (int k0 = 0; k0 < D; k0 += 2) {
-                if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
-                    break;
-                }
-
-                half2 V_k;
-                reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
-                reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
-                VKQ += V_k*KQ2[k0/2];
-            }
-        }
-
-        __syncthreads();
-    }
-
-    if (tid >= D) {
-        kqsum = 0.0f;
-    }
-
-    kqsum = warp_reduce_sum(kqsum);
-    if (threadIdx.x == 0) {
-        kqsum_shared[threadIdx.y] = kqsum;
-    }
-    __syncthreads();
-    kqsum = kqsum_shared[threadIdx.x];
-    kqsum = warp_reduce_sum(kqsum);
-
-    if (tid >= D) {
-        return;
-    }
-
-    half dst_val = (__low2half(VKQ) + __high2half(VKQ));
-    if (parallel_blocks == 1) {
-        dst_val /= kqsum;
-    }
-    dst[D*gridDim.y*blockIdx.x + D*blockIdx.y + tid] = dst_val;
-
-    if (parallel_blocks == 1 || tid != 0) {
-        return;
-    }
-    dst_meta[ic*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax, kqsum);
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -200,6 +23,10 @@ static __global__ void flash_attn_ext_f16(
         float      * __restrict__ dst,
         float2     * __restrict__ dst_meta,
         const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
         const int ne00,
         const int ne01,
         const int ne02,
@@ -256,6 +83,20 @@ static __global__ void flash_attn_ext_f16(
     const int stride_Q  = nb01 / sizeof(float);
     const int stride_KV = nb11 / sizeof(half);
 
+    half  slopeh = __float2half(1.0f);
+    half2 slope2 = make_half2(1.0f, 1.0f);
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int h = blockIdx.y;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slopeh = __float2half(powf(base, exph));
+        slope2 = make_half2(slopeh, slopeh);
+    }
+
     frag_b Q_b[D/16][ncols/frag_n];
 
     // A single buffer for temporarily holding tiles of KQ and VKQ parts:
@@ -372,7 +213,7 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
                     KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]);
                 }
                 KQ_max_new = warp_reduce_max(KQ_max_new);
@@ -415,7 +256,7 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ2_tmp[k0/WARP_SIZE] += mask ? mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ2_tmp[k0/WARP_SIZE] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
                     KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]);
                 }
                 KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
@@ -572,52 +413,6 @@ static __global__ void flash_attn_ext_f16(
 #endif // FP16_MMA_AVAILABLE
 }
 
-template<int D, int parallel_blocks> // D == head size
-__launch_bounds__(D, 1)
-static __global__ void flash_attn_combine_results(
-        const float  * __restrict__ VKQ_parts,
-        const float2 * __restrict__ VKQ_meta,
-        float * __restrict__ dst) {
-#if FP16_AVAILABLE
-    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
-    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
-    dst       +=                 D * gridDim.y*blockIdx.x;
-
-    const int tid = threadIdx.x;
-    __builtin_assume(tid < D);
-
-    __shared__ float2 meta[parallel_blocks];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
-    }
-
-    __syncthreads();
-
-    float kqmax = meta[0].x;
-#pragma unroll
-    for (int l = 1; l < parallel_blocks; ++l) {
-        kqmax = max(kqmax, meta[l].x);
-    }
-
-    float VKQ_numerator   = 0.0f;
-    float VKQ_denominator = 0.0f;
-#pragma unroll
-    for (int l = 0; l < parallel_blocks; ++l) {
-        const float diff = meta[l].x - kqmax;
-        const float KQ_max_scale = expf(diff);
-        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
-        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
-
-        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
-        VKQ_denominator += KQ_max_scale * meta[l].y;
-    }
-
-    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
-#else
-   NO_DEVICE_CODE;
-#endif // FP16_AVAILABLE
-}
-
 constexpr int get_max_power_of_2(int x) {
     return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
 }
@@ -642,57 +437,6 @@ static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
 static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
 
-template <int D, int parallel_blocks> void launch_fattn_vec_f16(
-        const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
-        ggml_cuda_pool & pool, cudaStream_t main_stream
-) {
-    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
-    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
-
-    if (parallel_blocks > 1) {
-        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
-        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
-    }
-
-    constexpr int  nwarps = (D + WARP_SIZE - 1) / WARP_SIZE;
-    const     dim3 block_dim(WARP_SIZE, nwarps, 1);
-    const     dim3 blocks_num(parallel_blocks*Q->ne[1], Q->ne[2], Q->ne[3]);
-    const     int  shmem = 0;
-
-    float scale;
-    memcpy(&scale, KQV->op_params, sizeof(float));
-
-    flash_attn_vec_ext_f16<D, parallel_blocks>
-        <<<blocks_num, block_dim, shmem, main_stream>>> (
-                (const char *) Q->data,
-                (const char *) K->data,
-                (const char *) V->data,
-                mask ? ((const char *) mask->data) : nullptr,
-                parallel_blocks == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale,
-                Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
-                K->ne[0], K->ne[1], K->ne[2], K->ne[3],
-                mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
-                Q->nb[1], Q->nb[2], Q->nb[3],
-                K->nb[1], K->nb[2], K->nb[3],
-                KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
-                );
-    CUDA_CHECK(cudaGetLastError());
-
-    if (parallel_blocks == 1) {
-        return;
-    }
-
-    const dim3 block_dim_combine(D, 1, 1);
-    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
-    const int  shmem_combine = 0;
-
-    flash_attn_combine_results<D, parallel_blocks>
-        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
-        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
-    CUDA_CHECK(cudaGetLastError());
-}
-
 template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename KQ_acc_t> void launch_fattn_f16_impl(
         const ggml_tensor * Q, const ggml_tensor * K, const ggml_tensor * V, ggml_tensor * KQV, const ggml_tensor * mask,
         ggml_cuda_pool & pool, cudaStream_t main_stream
@@ -710,8 +454,17 @@ template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename K
     const     dim3 blocks_num(parallel_blocks*(Q->ne[1] + cols_per_block - 1) / cols_per_block, Q->ne[2], Q->ne[3]);
     const     int  shmem = 0;
 
-    float scale;
+    float scale    = 1.0f;
-    memcpy(&scale, KQV->op_params, sizeof(float));
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     flash_attn_ext_f16<D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t>
         <<<blocks_num, block_dim, shmem, main_stream>>> (
@@ -720,7 +473,7 @@ template <int D, int cols_per_block, int nwarps, int parallel_blocks, typename K
                 (const char *) V->data,
                 mask ? ((const char *) mask->data) : nullptr,
                 (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
-                scale,
+                scale, max_bias, m0, m1, n_head_log2,
                 Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
                 K->ne[0], K->ne[1], K->ne[2], K->ne[3],
                 mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
@@ -783,11 +536,27 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
 
     ggml_cuda_set_device(ctx.device);
 
+    const int cc  = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
 
-    const int32_t precision = KQV->op_params[1];
+    const int32_t precision = KQV->op_params[2];
+
+    if (!fast_fp16_available(cc)) {
+        ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        return;
+    }
+
+    if (!fp16_mma_available(cc)) {
+        ggml_cuda_flash_attn_ext_vec_f16_no_mma(ctx, dst);
+        return;
+    }
 
     if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] == 1 && (Q->ne[0] == 64 || Q->ne[0] == 128)) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+
         if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
             constexpr int cols_per_block = 16;
             constexpr int nwarps         =  4;
@@ -845,21 +614,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     }
 
     if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
-        constexpr int parallel_blocks = 4;
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
-        switch (Q->ne[0]) {
-            case 64:
-                launch_fattn_vec_f16< 64, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 128:
-                launch_fattn_vec_f16<128, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            case 256:
-                launch_fattn_vec_f16<256, parallel_blocks>(Q, K, V, KQV, mask, ctx.pool(), ctx.stream());
-                break;
-            default:
-                GGML_ASSERT(false);
-                break;
-        }
         return;
     }
 

ggml-cuda/softmax.cu

@@ -11,7 +11,7 @@ __device__ float __forceinline__ t2f32<half>(half val) {
 }
 
 template <bool vals_smem, int ncols_template, int block_size_template, typename T>
-static __global__ void soft_max_f32(const float * x, const T * mask, const T * pos, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
+static __global__ void soft_max_f32(const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
     const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
 
     const int tid  = threadIdx.x;
@@ -23,16 +23,16 @@ static __global__ void soft_max_f32(const float * x, const T * mask, const T * p
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    float slope = 0.0f;
+    float slope = 1.0f;
 
     // ALiBi
     if (max_bias > 0.0f) {
         const int h = rowx/nrows_y; // head index
 
         const float base = h < n_head_log2 ? m0 : m1;
-        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+        const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
 
-        slope = powf(base, exp);
+        slope = powf(base, exph);
     }
 
     extern __shared__ float data_soft_max_f32[];
@@ -53,7 +53,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, const T * p
         const int64_t ix = (int64_t)rowx*ncols + col;
         const int64_t iy = (int64_t)rowy*ncols + col;
 
-        const float val = x[ix]*scale + (mask ? t2f32(mask[iy]) : 0.0f) + (pos ? slope*t2f32(pos[col]) : 0.0f);
+        const float val = x[ix]*scale + (mask ? slope*t2f32(mask[iy]) : 0.0f);
 
         vals[col] = val;
         max_val = max(max_val, val);
@@ -125,7 +125,7 @@ static __global__ void soft_max_f32(const float * x, const T * mask, const T * p
 }
 
 template<typename T>
-static void soft_max_f32_cuda(const float * x, const T * mask, const T * pos, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
     int nth = WARP_SIZE;
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
@@ -133,8 +133,8 @@ static void soft_max_f32_cuda(const float * x, const T * mask, const T * pos, fl
     const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
     static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
 
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t n_head      = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
@@ -142,43 +142,42 @@ static void soft_max_f32_cuda(const float * x, const T * mask, const T * pos, fl
     if (shmem < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
         switch (ncols_x) {
             case 32:
-                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 64:
-                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 128:
-                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 256:
-                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 512:
-                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 1024:
-                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 2048:
-                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             case 4096:
-                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
             default:
-                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
                 break;
         }
     } else {
         const size_t shmem_low = WARP_SIZE*sizeof(float);
-        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, pos, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
     }
 }
 
 void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
-    const ggml_tensor * src2 = dst->src[2];
 
     const float * src0_d = (const float *)src0->data;
     const void  * src1_d = src1 ? (const void *)src1->data : nullptr;
@@ -190,7 +189,6 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
-    GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F16 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional
 
     const int64_t ne00    = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
@@ -202,26 +200,15 @@ void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
-    // positions tensor
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
-    void * src2_d = nullptr;
-
-    const bool use_src2 = src2 != nullptr;
-
-    if (use_src2) {
-        src2_d = (void *)src2->data;
-    }
-
-    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16);
 
     if (use_f16) {
         const half * src1_dd = (const half *)src1_d;
-        const half * src2_dd = (const half *)src2_d;
 
-        soft_max_f32_cuda(src0_d, src1_dd, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
     } else {
         const float * src1_dd = (const float *)src1_d;
-        const float * src2_dd = (const float *)src2_d;
 
-        soft_max_f32_cuda(src0_d, src1_dd, src2_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
     }
 }

ggml-cuda/unary.cu

@@ -48,6 +48,15 @@ static __global__ void relu_f32(const float * x, float * dst, const int k) {
     dst[i] = fmaxf(x[i], 0);
 }
 
+static __global__ void sigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = 1.0f / (1.0f + expf(-x[i]));
+}
+
 static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
     const int i = blockDim.x*blockIdx.x + threadIdx.x;
 
@@ -108,6 +117,11 @@ static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_
     relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
 }
 
+static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
+    sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
 static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
     hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
@@ -188,6 +202,18 @@ void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
 }
 
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
 void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const float * src0_d = (const float *)src0->data;

ggml-cuda/unary.cuh

@@ -4,6 +4,7 @@
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_TANH_BLOCK_SIZE 256
 #define CUDA_RELU_BLOCK_SIZE 256
+#define CUDA_SIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSIGMOID_BLOCK_SIZE 256
 #define CUDA_HARDSWISH_BLOCK_SIZE 256
 #define CUDA_SQR_BLOCK_SIZE 256
@@ -18,6 +19,8 @@ void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml-kompute.cpp

@@ -1559,12 +1559,18 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                 case GGML_OP_SOFT_MAX:
                     {
                         float scale;
-                        memcpy(&scale, dst->op_params, sizeof(float));
+                        float max_bias;
 
-#pragma message("TODO: add ggml_vk_soft_max() F16/F32 src1 and src2 support")
+                        memcpy(&scale,    (float *)dst->op_params + 0, sizeof(float));
+                        memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
+
+#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
 #pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
                         GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
-                        GGML_ASSERT(src2 == nullptr);
+
+#pragma message("TODO: add ALiBi support")
+#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/7192")
+                        GGML_ASSERT(max_bias == 0.0f);
 
                         ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
                     } break;

ggml-metal.m

@@ -40,6 +40,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_CLAMP,
     GGML_METAL_KERNEL_TYPE_TANH,
     GGML_METAL_KERNEL_TYPE_RELU,
+    GGML_METAL_KERNEL_TYPE_SIGMOID,
     GGML_METAL_KERNEL_TYPE_GELU,
     GGML_METAL_KERNEL_TYPE_GELU_4,
     GGML_METAL_KERNEL_TYPE_GELU_QUICK,
@@ -169,7 +170,6 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
-    GGML_METAL_KERNEL_TYPE_ALIBI_F32,
     GGML_METAL_KERNEL_TYPE_IM2COL_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
@@ -494,6 +494,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                         clamp,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                          tanh,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                          relu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID,                       sigmoid,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                          gelu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                        gelu_4,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                    gelu_quick,                     true);
@@ -623,7 +624,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,          mul_mm_id_iq4_xs_f32,           ctx->support_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                      rope_f32,                       true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                      rope_f16,                       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                     alibi_f32,                      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                    im2col_f16,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
@@ -633,14 +633,14 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,          argsort_f32_i32_desc,           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,                leaky_relu_f32,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,        flash_attn_ext_f16_h64,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,        flash_attn_ext_f16_h64,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,        flash_attn_ext_f16_h80,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,        flash_attn_ext_f16_h80,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,        flash_attn_ext_f16_h96,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,        flash_attn_ext_f16_h96,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    ctx->support_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,                   cpy_f32_f16,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                   cpy_f32_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,                  cpy_f32_q8_0,                   true);
@@ -732,6 +732,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
             switch (ggml_get_unary_op(op)) {
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
@@ -759,7 +760,6 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_GROUP_NORM:
             return ctx->support_simdgroup_reduction;
         case GGML_OP_NORM:
-        case GGML_OP_ALIBI:
         case GGML_OP_ROPE:
         case GGML_OP_IM2COL:
             return true;
@@ -772,8 +772,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
         case GGML_OP_LEAKY_RELU:
-        case GGML_OP_FLASH_ATTN_EXT:
             return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+            return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             return ctx->support_simdgroup_reduction &&
@@ -1239,6 +1240,18 @@ static enum ggml_status ggml_metal_graph_compute(
 
                                 const int64_t n = ggml_nelements(dst);
 
+                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                            } break;
+                        case GGML_UNARY_OP_SIGMOID:
+                            {
+                                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
+
+                                [encoder setComputePipelineState:pipeline];
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                const int64_t n = ggml_nelements(dst);
+
                                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                             } break;
                         case GGML_UNARY_OP_GELU:
@@ -1357,13 +1370,12 @@ static enum ggml_status ggml_metal_graph_compute(
                 case GGML_OP_SOFT_MAX:
                     {
                         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
-                        GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F16 || src2->type == GGML_TYPE_F32);
 
                         int nth = 32; // SIMD width
 
                         id<MTLComputePipelineState> pipeline = nil;
 
-                        const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16);
+                        const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
                         if (ne00%4 == 0) {
                             while (nth < ne00/4 && nth < 256) {
@@ -1394,8 +1406,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         const int64_t nrows_x = ggml_nrows(src0);
                         const int64_t nrows_y = src0->ne[1];
 
-                        const uint32_t n_head_kv   = nrows_x/nrows_y;
+                        const uint32_t n_head      = nrows_x/nrows_y;
-                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
                         const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                         const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
@@ -1407,20 +1419,15 @@ static enum ggml_status ggml_metal_graph_compute(
                         } else {
                             [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
                         }
-                        if (id_src2) {
+                        [encoder setBuffer:id_dst      offset:offs_dst            atIndex:2];
-                            [encoder setBuffer:id_src2 offset:offs_src2   atIndex:2];
+                        [encoder setBytes:&ne00        length:sizeof(ne00)        atIndex:3];
-                        } else {
+                        [encoder setBytes:&ne01        length:sizeof(ne01)        atIndex:4];
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:2];
+                        [encoder setBytes:&ne02        length:sizeof(ne02)        atIndex:5];
-                        }
+                        [encoder setBytes:&scale       length:sizeof(scale)       atIndex:6];
-                        [encoder setBuffer:id_dst   offset:offs_dst          atIndex:3];
+                        [encoder setBytes:&max_bias    length:sizeof(max_bias)    atIndex:7];
-                        [encoder setBytes:&ne00     length:sizeof(ne00)      atIndex:4];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
-                        [encoder setBytes:&ne01     length:sizeof(ne01)      atIndex:5];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
-                        [encoder setBytes:&ne02     length:sizeof(ne02)      atIndex:6];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
-                        [encoder setBytes:&scale    length:sizeof(scale)     atIndex:7];
-                        [encoder setBytes:&max_bias length:sizeof(max_bias)  atIndex:8];
-                        [encoder setBytes:&m0       length:sizeof(m0)        atIndex:9];
-                        [encoder setBytes:&m1       length:sizeof(m1)        atIndex:10];
-                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:11];
                         [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -2225,49 +2232,6 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                     } break;
-                case GGML_OP_ALIBI:
-                    {
-                        GGML_ASSERT((src0t == GGML_TYPE_F32));
-
-                        const int nth = MIN(1024, ne00);
-
-                        //const int n_past = ((int32_t *) dst->op_params)[0];
-                        const int n_head = ((int32_t *) dst->op_params)[1];
-
-                        float max_bias;
-                        memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-                        const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-                        const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ALIBI_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                        [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                        [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                        [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                        [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                        [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                        [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                        [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                        [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                        [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                        [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                        [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                        [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                        [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                        [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                        [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                        [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
-                        [encoder setBytes:&m0   length:sizeof(   float) atIndex:18];
-                        [encoder setBytes:&m1   length:sizeof(   float) atIndex:19];
-                        [encoder setBytes:&n_heads_log2_floor   length:sizeof(int) atIndex:20];
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                    } break;
                 case GGML_OP_ROPE:
                     {
                         GGML_ASSERT(ne10 == ne02);
@@ -2565,7 +2529,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
 
                         const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
-                        const int64_t  ne31 = src3 ? src3->ne[1] : 0;
+                      //const int64_t  ne31 = src3 ? src3->ne[1] : 0;
                         const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
                         const int64_t  ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
 
@@ -2577,7 +2541,16 @@ static enum ggml_status ggml_metal_graph_compute(
                         const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
 
                         float scale;
-                        memcpy(&scale, dst->op_params, sizeof(float));
+                        float max_bias;
+
+                        memcpy(&scale,    ((int32_t *) dst->op_params) + 0, sizeof(scale));
+                        memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+
+                        const uint32_t n_head      = src0->ne[2];
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
                         id<MTLComputePipelineState> pipeline = nil;
 
@@ -2635,13 +2608,16 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setBytes:&nb11        length:sizeof(uint64_t)    atIndex:18];
                         [encoder setBytes:&nb12        length:sizeof(uint64_t)    atIndex:19];
                         [encoder setBytes:&nb13        length:sizeof(uint64_t)    atIndex:20];
-                        [encoder setBytes:&ne31    length:sizeof( int64_t) atIndex:21];
+                        [encoder setBytes:&nb31        length:sizeof(uint64_t)    atIndex:21];
-                        [encoder setBytes:&nb31    length:sizeof(uint64_t) atIndex:22];
+                        [encoder setBytes:&ne0         length:sizeof( int64_t)    atIndex:22];
-                        [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:23];
+                        [encoder setBytes:&ne1         length:sizeof( int64_t)    atIndex:23];
-                        [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:24];
+                        [encoder setBytes:&ne2         length:sizeof( int64_t)    atIndex:24];
-                        [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:25];
+                        [encoder setBytes:&ne3         length:sizeof( int64_t)    atIndex:25];
-                        [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:26];
+                        [encoder setBytes:&scale       length:sizeof(   float)    atIndex:26];
-                        [encoder setBytes:&scale   length:sizeof(   float) atIndex:27];
+                        [encoder setBytes:&max_bias    length:sizeof(   float)    atIndex:27];
+                        [encoder setBytes:&m0          length:sizeof(m0)          atIndex:28];
+                        [encoder setBytes:&m1          length:sizeof(m1)          atIndex:29];
+                        [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:30];
 
                         if (!use_vec_kernel) {
                             // half8x8 kernel

ggml-metal.metal

@@ -229,6 +229,13 @@ kernel void kernel_relu(
     dst[tpig] = max(0.0f, src0[tpig]);
 }
 
+kernel void kernel_sigmoid(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = 1.0f / (1.0f + exp(-src0[tpig]));
+}
+
 kernel void kernel_tanh(
         device const float * src0,
         device       float * dst,
@@ -356,7 +363,6 @@ template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,
         device const  char * src1,
-        device const  char * src2,
         device        char * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
@@ -378,10 +384,9 @@ kernel void kernel_soft_max(
 
     device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
     device const     T * pmask = src1 != src0 ? (device const    T *) src1         + i01*ne00 : nullptr;
-    device const     T * ppos  = src2 != src0 ? (device const    T *) src2                    : nullptr;
     device       float * pdst  = (device       float *) dst  + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
 
-    float slope = 0.0f;
+    float slope = 1.0f;
 
     // ALiBi
     if (max_bias > 0.0f) {
@@ -397,7 +402,7 @@ kernel void kernel_soft_max(
     float lmax = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f));
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
     }
 
     // find the max value in the block
@@ -422,7 +427,7 @@ kernel void kernel_soft_max(
     // parallel sum
     float lsum = 0.0f;
     for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)) - max_val);
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
         lsum += exp_psrc0;
         pdst[i00] = exp_psrc0;
     }
@@ -461,7 +466,6 @@ template<typename T>
 kernel void kernel_soft_max_4(
         device const  char * src0,
         device const  char * src1,
-        device const  char * src2,
         device        char * dst,
         constant   int64_t & ne00,
         constant   int64_t & ne01,
@@ -483,10 +487,9 @@ kernel void kernel_soft_max_4(
 
     device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
     device const      T * pmask = src1 != src0 ? (device const     T *) src1         + i01*ne00/4 : nullptr;
-    device const      T * ppos  = src2 != src0 ? (device const     T *) src2                      : nullptr;
     device       float4 * pdst4 = (device       float4 *) dst  + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
 
-    float slope = 0.0f;
+    float slope = 1.0f;
 
     if (max_bias > 0.0f) {
         const int64_t h = i02;
@@ -501,7 +504,7 @@ kernel void kernel_soft_max_4(
     float4 lmax4 = -INFINITY;
 
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f)));
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
     }
 
     const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
@@ -527,7 +530,7 @@ kernel void kernel_soft_max_4(
     // parallel sum
     float4 lsum4 = 0.0f;
     for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? pmask[i00] : 0.0f) + (ppos ? slope*ppos[i00] : 0.0f))) - max_val);
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
         lsum4 += exp_psrc4;
         pdst4[i00] = exp_psrc4;
     }
@@ -1595,60 +1598,6 @@ kernel void kernel_mul_mv_f16_f32_l4(
     }
 }
 
-kernel void kernel_alibi_f32(
-        device const float * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        constant     float & m0,
-        constant     float & m1,
-        constant       int & n_heads_log2_floor,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
-
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
-
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-  //const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
-
-    const int64_t k = i3*ne3 + i2;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = pow(m0, k + 1);
-    } else {
-        m_k = pow(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    device       char * dst_row = (device char *) dst + i3*nb3 + i2*nb2 + i1*nb1;
-    device const char * src_row = (device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        const  float   src_v = *(device float *)(src_row + i00*nb00);
-        device float * dst_v =  (device float *)(dst_row + i00*nb0);
-        *dst_v = i00 * m_k + src_v;
-    }
-}
-
 static float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -2116,13 +2065,16 @@ typedef void (flash_attn_ext_f16_t)(
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
-        constant   int64_t & ne31,
         constant  uint64_t & nb31,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
         constant   int64_t & ne3,
         constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
         threadgroup   half * shared,
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2154,13 +2106,16 @@ kernel void kernel_flash_attn_ext_f16(
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
-        constant   int64_t & ne31,
         constant  uint64_t & nb31,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
         constant   int64_t & ne3,
         constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
         threadgroup   half * shared [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2257,6 +2212,19 @@ kernel void kernel_flash_attn_ext_f16(
         // prepare diagonal scale matrix
         simdgroup_float8x8 mscale(scale);
 
+        // prepare diagonal slope matrix
+        simdgroup_float8x8 mslope(1.0f);
+
+        // ALiBi
+        if (max_bias > 0.0f) {
+            const uint32_t h = iq2;
+
+            const float base = h < n_head_log2 ? m0 : m1;
+            const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+            mslope = simdgroup_float8x8(pow(base, exph));
+        }
+
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
         for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
@@ -2279,9 +2247,10 @@ kernel void kernel_flash_attn_ext_f16(
                         simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
                     }
 
-                    // mqk = mqk*scale + mask
+                    // mqk = mqk*scale + mask*slope
                     simdgroup_half8x8 mm;
                     simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
+                    simdgroup_multiply(mm, mslope, mm);
                     simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
 
                     simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
@@ -2472,13 +2441,16 @@ kernel void kernel_flash_attn_ext_vec_f16(
         constant  uint64_t & nb11,
         constant  uint64_t & nb12,
         constant  uint64_t & nb13,
-        constant   int64_t & ne31,
         constant  uint64_t & nb31,
         constant   int64_t & ne0,
         constant   int64_t & ne1,
         constant   int64_t & ne2,
         constant   int64_t & ne3,
         constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
         threadgroup   half * shared [[threadgroup(0)]],
         uint3  tgpig[[threadgroup_position_in_grid]],
         uint3  tpitg[[thread_position_in_threadgroup]],
@@ -2497,6 +2469,18 @@ kernel void kernel_flash_attn_ext_vec_f16(
 
     const short T  = D + 2*nsg*SH; // shared memory size per query in (half)
 
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const uint32_t h = iq2;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
   //threadgroup half   * sq  = (threadgroup half   *) (shared +              0*D); // holds the query data
     threadgroup half4  * sq4 = (threadgroup half4  *) (shared +              0*D); // same as above but in half4
     threadgroup float  * ss  = (threadgroup float  *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
@@ -2603,10 +2587,10 @@ kernel void kernel_flash_attn_ext_vec_f16(
                     mqk += simd_shuffle_down(mqk,  2);
                     mqk += simd_shuffle_down(mqk,  1);
 
-                    // mqk = mqk*scale + mask
+                    // mqk = mqk*scale + mask*slope
                     if (tiisg == 0) {
                         float4 mm = (float4) mp4[ic/4 + cc];
-                        mqk = mqk*scale + mm;
+                        mqk = mqk*scale + mm*slope;
 
                         ss4[cc] = mqk;
                     }
@@ -2840,7 +2824,8 @@ kernel void kernel_cpy_f32_f16(
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
         device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
 
-        dst_data[i00] = src[0];
+        // TODO: is there a better way to handle -INFINITY?
+        dst_data[i00] = src[0] == -INFINITY ? -MAXHALF : src[0];
     }
 }
 

ggml-opencl.cpp

@@ -2119,6 +2119,7 @@ static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_
     if (alignment == (cl_uint)-1) {
         ggml_cl_init();
         clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &alignment, NULL);
+        alignment /= 8; // bits to bytes
     }
     return alignment;
 

ggml-quants.c

@@ -14,6 +14,12 @@
 #include <stdlib.h> // for qsort
 #include <stdio.h>  // for GGML_ASSERT
 
+#if defined(_MSC_VER)
+// disable "possible loss of data" to avoid warnings for hundreds of casts
+// we should just be careful :)
+#pragma warning(disable: 4244 4267)
+#endif
+
 #define UNUSED GGML_UNUSED
 
 // some compilers don't provide _mm256_set_m128i, e.g. gcc 7

ggml-sycl.cpp

@@ -3154,7 +3154,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
 #define SYCL_SCALE_BLOCK_SIZE 256
 #define SYCL_CLAMP_BLOCK_SIZE 256
 #define SYCL_ROPE_BLOCK_SIZE 256
-#define SYCL_ALIBI_BLOCK_SIZE 32
 #define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
 #define SYCL_QUANTIZE_BLOCK_SIZE 256
 #define SYCL_DEQUANTIZE_BLOCK_SIZE 256
@@ -8330,13 +8329,15 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
     const int blocks_per_row = ncols / qk;
     const int blocks_per_warp = vdr * WARP_SIZE / qi;
 
+    const int qi_vdr = (qi / vdr); // N_threads processing 1 qk block
+
     // partial sum for each thread
     float tmp = 0.0f;
 
     const block_q_t  * x = (const block_q_t  *) vx;
     const block_q8_1 * y = (const block_q8_1 *) vy;
 
-    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+    for (int i = item_ct1.get_local_id(2) / qi_vdr; i < blocks_per_row;
          i += blocks_per_warp) {
       const int ibx = row * blocks_per_row + i; // x block index
 
@@ -8344,8 +8345,8 @@ static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict_
 
       const int iqs =
           vdr *
-            (item_ct1.get_local_id(2) %
+          (item_ct1.get_local_id(2) -
-             (qi / vdr)); // x block quant index when casting the quants to int
+           i * qi_vdr); // x block quant index when casting the quants to int
 
       tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs);
     }
@@ -9314,32 +9315,6 @@ static void rope_glm_f32(
     dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
 }
 
-static void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
-                                 const int n_heads_log2_floor, const float m0, const float m1,
-                                 const sycl::nd_item<3> &item_ct1) {
-    const int col = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1);
-    const int i = row*ncols + col;
-
-    const int k = row/k_rows;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = dpct::pow(m0, k + 1);
-    } else {
-        m_k = dpct::pow(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    dst[i] = col * m_k + x[i];
-}
-
 static void k_sum_rows_f32(const float * x, float * dst, const int ncols,
                            const sycl::nd_item<3> &item_ct1) {
     const int row = item_ct1.get_group(1);
@@ -9441,7 +9416,7 @@ static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, con
 
 
 template <bool vals_smem, int ncols_template, int block_size_template>
-static void soft_max_f32(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
+static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par,
                          const int nrows_y, const float scale, const float max_bias, const float m0,
                          const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
     const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
@@ -9455,7 +9430,7 @@ static void soft_max_f32(const float * x, const float * mask, const float *pos,
     const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
     const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
 
-    float slope = 0.0f;
+    float slope = 1.0f;
 
     // ALiBi
     if (max_bias > 0.0f) {
@@ -9480,7 +9455,7 @@ static void soft_max_f32(const float * x, const float * mask, const float *pos,
         const int ix = rowx*ncols + col;
         const int iy = rowy*ncols + col;
 
-        const float val = x[ix]*scale + (mask ? mask[iy] : 0.0f) + (pos ? slope*pos[col] : 0.0f);
+        const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f);
 
         vals[col] = val;
         max_val = sycl::max(max_val, val);
@@ -12962,20 +12937,6 @@ static void rope_glm_f32_sycl(const float *x, float *dst, int ncols, int nrows,
                          });
 }
 
-static void alibi_f32_sycl(const float *x, float *dst, const int ncols,
-                           const int nrows, const int k_rows,
-                           const int n_heads_log2_floor, const float m0,
-                           const float m1, dpct::queue_ptr stream) {
-    const sycl::range<3> block_dims(1, 1, SYCL_ALIBI_BLOCK_SIZE);
-    const int num_blocks_x = (ncols + SYCL_ALIBI_BLOCK_SIZE - 1) / (SYCL_ALIBI_BLOCK_SIZE);
-    const sycl::range<3> block_nums(1, nrows, num_blocks_x);
-    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             alibi_f32(x, dst, ncols, k_rows,
-                                       n_heads_log2_floor, m0, m1, item_ct1);
-                         });
-}
-
 static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols,
                               const int nrows, dpct::queue_ptr stream) {
     const sycl::range<3> block_dims(1, 1, WARP_SIZE);
@@ -13056,7 +13017,7 @@ static void diag_mask_inf_f32_sycl(const float *x, float *dst,
 }
 
 template <bool vals_smem, int ncols_template, int block_size_template>
-static void soft_max_f32_submitter(const float * x, const float * mask, const float *pos, float * dst, const int ncols_par,
+static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par,
                                    const int nrows_y, const float scale, const float max_bias, const float m0,
                                    const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
                                    const size_t n_local_scratch, dpct::queue_ptr stream) {
@@ -13066,7 +13027,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, const fl
         cgh.parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(32)]] {
-                soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, pos, dst, ncols_par,
+                soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
                                                                              nrows_y, scale, max_bias, m0,
                                                                              m1, n_head_log2, item_ct1,
                                                                              local_buf_acc.get_pointer());
@@ -13074,7 +13035,7 @@ static void soft_max_f32_submitter(const float * x, const float * mask, const fl
     });
 }
 
-static void soft_max_f32_sycl(const float * x, const float * mask, const float * pos,
+static void soft_max_f32_sycl(const float * x, const float * mask,
                               float * dst, const int ncols_x, const int nrows_x,
                               const int nrows_y, const float scale, const float max_bias,
                               dpct::queue_ptr stream) {
@@ -13096,60 +13057,60 @@ static void soft_max_f32_sycl(const float * x, const float * mask, const float *
     const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
     if (n_local_scratch*sizeof(float) < local_mem_size) {
         if (ncols_x > max_block_size) {
-            soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+            soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
                                                max_bias, m0, m1, n_head_log2, block_nums,
                                                block_dims, n_local_scratch, stream);
             return;
         }
         switch (ncols_x) {
             case 32:
-                soft_max_f32_submitter<true, 32, 32>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 32, 32>(x, mask, dst, ncols_x, nrows_y, scale,
                                                      max_bias, m0, m1, n_head_log2, block_nums,
                                                      block_dims, n_local_scratch, stream);
                 break;
             case 64:
-                soft_max_f32_submitter<true, 64, 64>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 64, 64>(x, mask, dst, ncols_x, nrows_y, scale,
                                                      max_bias, m0, m1, n_head_log2, block_nums,
                                                      block_dims, n_local_scratch, stream);
                 break;
             case 128:
-                soft_max_f32_submitter<true, 128, 128>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 128, 128>(x, mask, dst, ncols_x, nrows_y, scale,
                                                        max_bias, m0, m1, n_head_log2, block_nums,
                                                        block_dims, n_local_scratch, stream);
                 break;
             case 256:
-                soft_max_f32_submitter<true, 256, 256>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 256, 256>(x, mask, dst, ncols_x, nrows_y, scale,
                                                        max_bias, m0, m1, n_head_log2, block_nums,
                                                        block_dims, n_local_scratch, stream);
                 break;
             case 512:
-                soft_max_f32_submitter<true, 512, 512>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 512, 512>(x, mask, dst, ncols_x, nrows_y, scale,
                                                        max_bias, m0, m1, n_head_log2, block_nums,
                                                        block_dims, n_local_scratch, stream);
                 break;
             case 1024:
-                soft_max_f32_submitter<true, 1024, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 1024, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
                                                          max_bias, m0, m1, n_head_log2, block_nums,
                                                          block_dims, n_local_scratch, stream);
                 break;
             case 2048:
-                soft_max_f32_submitter<true, 2048, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 2048, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
                                                          max_bias, m0, m1, n_head_log2, block_nums,
                                                          block_dims, n_local_scratch, stream);
                 break;
             case 4096:
-                soft_max_f32_submitter<true, 4096, 1024>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 4096, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
                                                          max_bias, m0, m1, n_head_log2, block_nums,
                                                          block_dims, n_local_scratch, stream);
                 break;
             default:
-                soft_max_f32_submitter<true, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+                soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
                                                    max_bias, m0, m1, n_head_log2, block_nums,
                                                    block_dims, n_local_scratch, stream);
                 break;
         }
     } else {
-        soft_max_f32_submitter<false, 0, 0>(x, mask, pos, dst, ncols_x, nrows_y, scale,
+        soft_max_f32_submitter<false, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
                                             max_bias, m0, m1, n_head_log2, block_nums,
                                             block_dims, WARP_SIZE, stream);
     }
@@ -14560,36 +14521,6 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
     (void) src1_dd;
 }
 
-inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1,
-                               ggml_tensor *dst, const float *src0_dd,
-                               const float *src1_dd, float *dst_dd,
-                               const dpct::queue_ptr &main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    GGML_TENSOR_LOCALS_3(int64_t, ne0, src0, ne);
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    //GGML_ASSERT(ne01 + n_past == ne00);
-    GGML_ASSERT(n_head == ne02);
-
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    alibi_f32_sycl(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream);
-
-    (void) src1;
-    (void) src1_dd;
-}
-
 static void ggml_sycl_op_pool2d(const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
                                 const float *src0_dd, const float *src1_dd,
@@ -14744,12 +14675,9 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const ggml_tensor * src2 = dst->src[2];
+#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support")
-
-#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 and src2 support")
 #pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
     GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
-    GGML_ASSERT(!src2 || src2->type == GGML_TYPE_F32); // src2 contains positions and it is optional
 
     const int64_t ne00 = src0->ne[0];
     const int64_t nrows_x = ggml_nrows(src0);
@@ -14761,25 +14689,7 @@ inline void ggml_sycl_op_soft_max(const ggml_tensor *src0,
     memcpy(&scale, dst->op_params + 0, sizeof(float));
     memcpy(&max_bias, dst->op_params + 1, sizeof(float));
 
-    // positions tensor
+    soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00,
-    float * src2_dd = nullptr;
-    sycl_pool_alloc<float> src2_f;
-
-    const bool use_src2 = src2 != nullptr;
-
-    if (use_src2) {
-        const bool src2_on_device = src2->backend == GGML_BACKEND_TYPE_GPU;
-
-        if (src2_on_device) {
-            ggml_tensor_extra_gpu * src2_extra = (ggml_tensor_extra_gpu *) src2->extra;
-            src2_dd = (float *) src2_extra->data_device[g_main_device];
-        } else {
-            src2_dd = src2_f.alloc(ggml_nelements(src2));
-            SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src2_dd, src2, 0, 0, 0, 1, main_stream));
-        }
-    }
-
-    soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, src2_dd, dst_dd, ne00,
                       nrows_x, nrows_y, scale, max_bias, main_stream);
 }
 
@@ -16230,10 +16140,6 @@ static void ggml_sycl_rope(const ggml_tensor * src0, const ggml_tensor * src1, g
     ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_rope);
 }
 
-static void ggml_sycl_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_alibi);
-}
-
 static void ggml_sycl_pool2d(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_sycl_op_flatten(src0, src1, dst, ggml_sycl_op_pool2d);
 }
@@ -16610,9 +16516,6 @@ bool ggml_sycl_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_ROPE:
             func = ggml_sycl_rope;
             break;
-        case GGML_OP_ALIBI:
-            func = ggml_sycl_alibi;
-            break;
         case GGML_OP_IM2COL:
             func = ggml_sycl_im2col;
             break;
@@ -17742,7 +17645,6 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_ROPE:
-        case GGML_OP_ALIBI:
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:

ggml.c

@@ -4,7 +4,6 @@
 #include "ggml-impl.h"
 #include "ggml-quants.h"
 #include "ggml.h"
-#include "sgemm.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -37,6 +36,10 @@
 #undef GGML_USE_LLAMAFILE
 #endif
 
+#ifdef GGML_USE_LLAMAFILE
+#include "sgemm.h"
+#endif
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -1949,6 +1952,7 @@ inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) {
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
 // TODO: optimize performance
 inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
@@ -2185,7 +2189,6 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "SOFT_MAX_BACK",
     "ROPE",
     "ROPE_BACK",
-    "ALIBI",
     "CLAMP",
     "CONV_TRANSPOSE_1D",
     "IM2COL",
@@ -2227,7 +2230,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 77, "GGML_OP_COUNT != 77");
+static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -2276,7 +2279,6 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "soft_max_back(x)",
     "rope(x)",
     "rope_back(x)",
-    "alibi(x)",
     "clamp(x)",
     "conv_transpose_1d(x)",
     "im2col(x)",
@@ -2318,7 +2320,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 77, "GGML_OP_COUNT != 77");
+static_assert(GGML_OP_COUNT == 76, "GGML_OP_COUNT != 76");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -2331,6 +2333,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "TANH",
     "ELU",
     "RELU",
+    "SIGMOID",
     "GELU",
     "GELU_QUICK",
     "SILU",
@@ -2338,7 +2341,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "HARDSIGMOID",
 };
 
-static_assert(GGML_UNARY_OP_COUNT == 12, "GGML_UNARY_OP_COUNT != 12");
+static_assert(GGML_UNARY_OP_COUNT == 13, "GGML_UNARY_OP_COUNT != 13");
 
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
@@ -4563,6 +4566,20 @@ struct ggml_tensor * ggml_leaky_relu(
     return result;
 }
 
+// ggml_sigmoid
+
+struct ggml_tensor * ggml_sigmoid(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
+struct ggml_tensor * ggml_sigmoid_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
 // ggml_gelu
 
 struct ggml_tensor * ggml_gelu(
@@ -5646,7 +5663,6 @@ static struct ggml_tensor * ggml_soft_max_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * mask,
-        struct ggml_tensor  * pos,
         float                 scale,
         float                 max_bias,
         bool                  inplace) {
@@ -5660,18 +5676,8 @@ static struct ggml_tensor * ggml_soft_max_impl(
         GGML_ASSERT(mask->ne[1] >= a->ne[1]);
     }
 
-    if (pos) {
-        GGML_ASSERT(ggml_is_vector(pos));
-        GGML_ASSERT(pos->type == GGML_TYPE_F16 || pos->type == GGML_TYPE_F32);
-        GGML_ASSERT(pos->ne[0] == a->ne[0]);
-    }
-
-    if (pos && mask) {
-        GGML_ASSERT(pos->type == mask->type);
-    }
-
     if (max_bias > 0.0f) {
-        GGML_ASSERT(pos);
+        GGML_ASSERT(mask);
     }
 
     bool is_node = false;
@@ -5689,7 +5695,6 @@ static struct ggml_tensor * ggml_soft_max_impl(
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = a;
     result->src[1] = mask;
-    result->src[2] = pos;
 
     return result;
 }
@@ -5697,23 +5702,22 @@ static struct ggml_tensor * ggml_soft_max_impl(
 struct ggml_tensor * ggml_soft_max(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, NULL, NULL, 1.0f, 0.0f, false);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, false);
 }
 
 struct ggml_tensor * ggml_soft_max_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, NULL, NULL, 1.0f, 0.0f, true);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, true);
 }
 
 struct ggml_tensor * ggml_soft_max_ext(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * mask,
-        struct ggml_tensor  * pos,
         float                 scale,
         float                 max_bias) {
-    return ggml_soft_max_impl(ctx, a, mask, pos, scale, max_bias, false);
+    return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
 }
 
 // ggml_soft_max_back
@@ -5928,37 +5932,6 @@ struct ggml_tensor * ggml_rope_back(
     return result;
 }
 
-// ggml_alibi
-
-struct ggml_tensor * ggml_alibi(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int                   n_past,
-        int                   n_head,
-        float                 bias_max) {
-    GGML_ASSERT(n_past >= 0);
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    // TODO: when implement backward, fix this:
-    //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
-    int32_t op_params[3] = { n_past, n_head };
-    memcpy(op_params + 2, &bias_max, sizeof(float));
-    ggml_set_op_params(result, op_params, sizeof(op_params));
-
-    result->op   = GGML_OP_ALIBI;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-
-    return result;
-}
-
 // ggml_clamp
 
 struct ggml_tensor * ggml_clamp(
@@ -6486,9 +6459,11 @@ struct ggml_tensor * ggml_flash_attn_ext(
         struct ggml_tensor  * k,
         struct ggml_tensor  * v,
         struct ggml_tensor  * mask,
-        float                 scale) {
+        float                 scale,
+        float                 max_bias) {
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
+
     if (mask) {
         GGML_ASSERT(ggml_is_contiguous(mask));
         GGML_ASSERT(mask->ne[2] == 1);
@@ -6498,6 +6473,10 @@ struct ggml_tensor * ggml_flash_attn_ext(
         //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
     }
 
+    if (max_bias > 0.0f) {
+        GGML_ASSERT(mask);
+    }
+
     bool is_node = false;
 
     if (q->grad || k->grad || v->grad) {
@@ -6508,7 +6487,7 @@ struct ggml_tensor * ggml_flash_attn_ext(
     int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    float params[] = { scale };
+    float params[] = { scale, max_bias };
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op   = GGML_OP_FLASH_ATTN_EXT;
@@ -6528,7 +6507,7 @@ void ggml_flash_attn_ext_set_prec(
 
     const int32_t prec_i32 = (int32_t) prec;
 
-    ggml_set_op_params_i32(a, 1, prec_i32); // scale is on first pos
+    ggml_set_op_params_i32(a, 2, prec_i32); // scale is on first pos, max_bias on second
 }
 
 // ggml_flash_ff
@@ -10892,6 +10871,52 @@ static void ggml_compute_forward_relu(
     }
 }
 
+// ggml_compute_forward_sigmoid
+
+static void ggml_compute_forward_sigmoid_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(params->ith == 0);
+    assert(ggml_are_same_shape(src0, dst));
+
+    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
+        return;
+    }
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    assert(dst->nb[0]  == sizeof(float));
+    assert(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sigmoid_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_sigmoid(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_sigmoid_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_gelu
 
 static void ggml_compute_forward_gelu_f32(
@@ -13333,7 +13358,6 @@ static void ggml_compute_forward_soft_max_f32(
 
     const struct ggml_tensor * src0 = dst->src[0];
     const struct ggml_tensor * src1 = dst->src[1];
-    const struct ggml_tensor * src2 = dst->src[2];
 
     assert(ggml_is_contiguous(dst));
     assert(ggml_are_same_shape(src0, dst));
@@ -13359,8 +13383,8 @@ static void ggml_compute_forward_soft_max_f32(
 
     // TODO: is this supposed to be ceil instead of floor?
     //       https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
-    const uint32_t n_head_kv   = ne02;
+    const uint32_t n_head      = ne02;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head_kv));
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
 
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
@@ -13377,13 +13401,13 @@ static void ggml_compute_forward_soft_max_f32(
 
     float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
 
-    // when max_bias <= 0.0f, src2 is not used and we default it to src0 to avoid branching
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
-    ggml_fp16_t * pos_f16 = src2 ? (ggml_fp16_t *) src2->data : src0->data;
-    float       * pos_f32 = src2 ? (float       *) src2->data : src0->data;
-
-    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16) || (src2 && src2->type == GGML_TYPE_F16);
 
     for (int i1 = ir0; i1 < ir1; i1++) {
+        // ALiBi
+        const uint32_t h = (i1/ne01)%ne02; // head
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
         float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
         float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
 
@@ -13396,27 +13420,11 @@ static void ggml_compute_forward_soft_max_f32(
         if (mp_f32) {
             if (use_f16) {
                 for (int i = 0; i < nc; ++i) {
-                    wp[i] += GGML_FP16_TO_FP32(mp_f16[i]);
+                    wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]);
                 }
             } else {
                 for (int i = 0; i < nc; ++i) {
-                    wp[i] += mp_f32[i];
+                    wp[i] += slope*mp_f32[i];
-                }
-            }
-        }
-
-        // ALiBi bias
-        if (max_bias > 0.0f) {
-            const uint32_t h  = (i1/ne01)%ne02; // head
-            const float slope = h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1);
-
-            if (use_f16) {
-                for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*GGML_FP16_TO_FP32(pos_f16[i]);
-                }
-            } else {
-                for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*pos_f32[i];
                 }
             }
         }
@@ -13578,178 +13586,6 @@ static void ggml_compute_forward_soft_max_back(
     }
 }
 
-// ggml_compute_forward_alibi
-
-static void ggml_compute_forward_alibi_f32(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * src0 = dst->src[0];
-
-    assert(params->ith == 0);
-
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
-        return;
-    }
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    const int64_t ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int64_t ne1 = src0->ne[1]; // seq_len_without_past
-    const int64_t ne2 = src0->ne[2]; // n_head -> this is k
-    //const int64_t ne3 = src0->ne[3]; // 1 -> bsz
-
-    const int64_t n  = ggml_nrows(src0);
-    const int64_t ne2_ne3 = n/ne1; // ne2*ne3
-
-    const size_t nb0 = src0->nb[0];
-    const size_t nb1 = src0->nb[1];
-    const size_t nb2 = src0->nb[2];
-    //const int nb3 = src0->nb[3];
-
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(n_head == ne2);
-
-    // add alibi to src0 (KQ_scaled)
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    for (int64_t k = 0; k < ne2_ne3; k++) {
-        // TODO: k*nb2 or k*nb3
-        float m_k;
-
-        if (k < n_heads_log2_floor) {
-            m_k = powf(m0, k + 1);
-        } else {
-            m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-        }
-
-        for (int64_t i = 0; i < ne0; i++) {
-            for (int64_t j = 0; j < ne1; j++) {
-                float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
-                float *      pdst = (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
-                pdst[0] = i * m_k + src[0];
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_alibi_f16(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * src0 = dst->src[0];
-
-    assert(params->ith == 0);
-
-    if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
-        return;
-    }
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int ne1 = src0->ne[1]; // seq_len_without_past
-    const int ne2 = src0->ne[2]; // n_head -> this is k
-    //const int ne3 = src0->ne[3]; // 1 -> bsz
-
-    const int n  = ggml_nrows(src0);
-    const int ne2_ne3 = n/ne1; // ne2*ne3
-
-    const int nb0 = src0->nb[0];
-    const int nb1 = src0->nb[1];
-    const int nb2 = src0->nb[2];
-    //const int nb3 = src0->nb[3];
-
-    GGML_ASSERT(nb0 == sizeof(ggml_fp16_t));
-    //GGML_ASSERT(ne1 + n_past == ne0); (void) n_past;
-    GGML_ASSERT(n_head == ne2);
-
-    // add alibi to src0 (KQ_scaled)
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    for (int k = 0; k < ne2_ne3; k++) {
-        // TODO: k*nb2 or k*nb3
-        float m_k;
-
-        if (k < n_heads_log2_floor) {
-            m_k = powf(m0, k + 1);
-        } else {
-            m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-        }
-
-        for (int i = 0; i < ne0; i++) {
-            for (int j = 0; j < ne1; j++) {
-                ggml_fp16_t * const src  = (ggml_fp16_t *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
-                float       *      pdst  =       (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
-
-                // we return F32
-                pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]);
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_alibi(
-        const struct ggml_compute_params * params,
-        struct ggml_tensor * dst) {
-
-    const struct ggml_tensor * src0 = dst->src[0];
-
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_alibi_f16(params, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_alibi_f32(params, dst);
-            } break;
-        case GGML_TYPE_BF16:
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-        case GGML_TYPE_Q8_1:
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-        case GGML_TYPE_IQ2_XXS:
-        case GGML_TYPE_IQ2_XS:
-        case GGML_TYPE_IQ3_XXS:
-        case GGML_TYPE_IQ1_S:
-        case GGML_TYPE_IQ1_M:
-        case GGML_TYPE_IQ4_NL:
-        case GGML_TYPE_IQ4_XS:
-        case GGML_TYPE_IQ3_S:
-        case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q8_K:
-        case GGML_TYPE_I8:
-        case GGML_TYPE_I16:
-        case GGML_TYPE_I32:
-        case GGML_TYPE_I64:
-        case GGML_TYPE_F64:
-        case GGML_TYPE_COUNT:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
 // ggml_compute_forward_clamp
 
 static void ggml_compute_forward_clamp_f32(
@@ -15764,7 +15600,16 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const int ir1 = MIN(ir0 + dr, nr);
 
     float scale    = 1.0f;
+    float max_bias = 0.0f;
+
     memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
@@ -15773,6 +15618,9 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         const int iq2 = (ir - iq3*neq2*neq1)/neq1;
         const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
 
+        const uint32_t h = iq2; // head
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
         float S = 0.0f;
         float M = -INFINITY;
 
@@ -15796,7 +15644,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
         for (int64_t ic = 0; ic < nek1; ++ic) {
-            const float mv = mp ? GGML_FP16_TO_FP32(mp[ic]) : 0.0f;
+            const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
             }
@@ -15867,7 +15715,7 @@ static void ggml_compute_forward_flash_attn_ext(
         const struct ggml_tensor * v,
         const struct ggml_tensor * mask,
         struct ggml_tensor * dst) {
-    switch (dst->op_params[1]) {
+    switch (dst->op_params[2]) {
         case GGML_PREC_DEFAULT:
         case GGML_PREC_F32:
             {
@@ -16834,6 +16682,10 @@ static void ggml_compute_forward_unary(
             {
                 ggml_compute_forward_relu(params, dst);
             } break;
+        case GGML_UNARY_OP_SIGMOID:
+            {
+                ggml_compute_forward_sigmoid(params, dst);
+            } break;
         case GGML_UNARY_OP_GELU:
             {
                 ggml_compute_forward_gelu(params, dst);
@@ -17630,10 +17482,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_rope_back(params, tensor);
             } break;
-        case GGML_OP_ALIBI:
-            {
-                ggml_compute_forward_alibi(params, tensor);
-            } break;
         case GGML_OP_CLAMP:
             {
                 ggml_compute_forward_clamp(params, tensor);
@@ -18652,10 +18500,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             zero_table);
                 }
             } break;
-        case GGML_OP_ALIBI:
-            {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
         case GGML_OP_CLAMP:
             {
                 GGML_ASSERT(false); // TODO: not implemented
@@ -18826,6 +18670,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                         zero_table);
                             }
                         } break;
+                    case GGML_UNARY_OP_SIGMOID:
+                        {
+                            GGML_ASSERT(false); // TODO: not implemented
+                        } break;
                     case GGML_UNARY_OP_GELU:
                         {
                             GGML_ASSERT(false); // TODO: not implemented
@@ -19355,6 +19203,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                 case GGML_UNARY_OP_HARDSWISH: // to opt for multiple threads
                 case GGML_UNARY_OP_HARDSIGMOID: // to opt for multiple threads
                     {
@@ -19428,10 +19277,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads, int n_cur_
             {
                 n_tasks = n_threads;
             } break;
-        case GGML_OP_ALIBI:
-            {
-                n_tasks = 1; //TODO
-            } break;
         case GGML_OP_CLAMP:
             {
                 n_tasks = 1; //TODO

ggml.h

@@ -468,7 +468,6 @@ extern "C" {
         GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
-        GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
@@ -520,6 +519,7 @@ extern "C" {
         GGML_UNARY_OP_TANH,
         GGML_UNARY_OP_ELU,
         GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_SIGMOID,
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
@@ -1074,6 +1074,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1428,15 +1436,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
+    // fused soft_max(a*scale + mask*(ALiBi slope))
     // mask is optional
-    // pos is required when max_bias > 0.0f
     // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * mask,
-            struct ggml_tensor  * pos,
             float                 scale,
             float                 max_bias);
 
@@ -1538,16 +1544,6 @@ extern "C" {
             float                 xpos_base,
             bool                  xpos_down);
 
-    // alibi position embedding
-    // in-place, returns view(a)
-    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   n_past,
-            int                   n_head,
-            float                 bias_max),
-        "use ggml_soft_max_ext instead (will be removed in Mar 2024)");
-
     // clamp
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_clamp(
@@ -1744,7 +1740,8 @@ extern "C" {
             struct ggml_tensor  * k,
             struct ggml_tensor  * v,
             struct ggml_tensor  * mask,
-            float                 scale);
+            float                 scale,
+            float                 max_bias);
 
     GGML_API void ggml_flash_attn_ext_set_prec(
             struct ggml_tensor * a,

gguf-py/gguf/__init__.py

@@ -1,4 +1,5 @@
 from .constants import *
+from .lazy import *
 from .gguf_reader import *
 from .gguf_writer import *
 from .tensor_mapping import *

gguf-py/gguf/constants.py

@@ -10,6 +10,7 @@ from typing import Any
 GGUF_MAGIC             = 0x46554747  # "GGUF"
 GGUF_VERSION           = 3
 GGUF_DEFAULT_ALIGNMENT = 32
+GGML_QUANT_VERSION     = 2  # GGML_QNT_VERSION from ggml.h
 
 #
 # metadata keys
@@ -118,6 +119,7 @@ class MODEL_ARCH(IntEnum):
     REFACT     = auto()
     BERT       = auto()
     NOMIC_BERT = auto()
+    JINA_BERT_V2 = auto()
     BLOOM      = auto()
     STABLELM   = auto()
     QWEN       = auto()
@@ -195,6 +197,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.REFACT:         "refact",
     MODEL_ARCH.BERT:           "bert",
     MODEL_ARCH.NOMIC_BERT:     "nomic-bert",
+    MODEL_ARCH.JINA_BERT_V2:   "jina-bert-v2",
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
@@ -380,6 +383,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_UP,
         MODEL_TENSOR.LAYER_OUT_NORM,
     ],
+    MODEL_ARCH.JINA_BERT_V2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -820,6 +839,49 @@ class GGMLQuantizationType(IntEnum):
     BF16    = 30
 
 
+# TODO: add GGMLFileType from ggml_ftype in ggml.h
+
+
+# from llama_ftype in llama.h
+# ALL VALUES SHOULD BE THE SAME HERE AS THEY ARE OVER THERE.
+class LlamaFileType(IntEnum):
+    ALL_F32              = 0
+    MOSTLY_F16           = 1   # except 1d tensors
+    MOSTLY_Q4_0          = 2   # except 1d tensors
+    MOSTLY_Q4_1          = 3   # except 1d tensors
+    MOSTLY_Q4_1_SOME_F16 = 4   # tok_embeddings.weight and output.weight are F16
+    # MOSTLY_Q4_2        = 5   # support has been removed
+    # MOSTLY_Q4_3        = 6   # support has been removed
+    MOSTLY_Q8_0          = 7   # except 1d tensors
+    MOSTLY_Q5_0          = 8   # except 1d tensors
+    MOSTLY_Q5_1          = 9   # except 1d tensors
+    MOSTLY_Q2_K          = 10  # except 1d tensors
+    MOSTLY_Q3_K_S        = 11  # except 1d tensors
+    MOSTLY_Q3_K_M        = 12  # except 1d tensors
+    MOSTLY_Q3_K_L        = 13  # except 1d tensors
+    MOSTLY_Q4_K_S        = 14  # except 1d tensors
+    MOSTLY_Q4_K_M        = 15  # except 1d tensors
+    MOSTLY_Q5_K_S        = 16  # except 1d tensors
+    MOSTLY_Q5_K_M        = 17  # except 1d tensors
+    MOSTLY_Q6_K          = 18  # except 1d tensors
+    MOSTLY_IQ2_XXS       = 19  # except 1d tensors
+    MOSTLY_IQ2_XS        = 20  # except 1d tensors
+    MOSTLY_Q2_K_S        = 21  # except 1d tensors
+    MOSTLY_IQ3_XS        = 22  # except 1d tensors
+    MOSTLY_IQ3_XXS       = 23  # except 1d tensors
+    MOSTLY_IQ1_S         = 24  # except 1d tensors
+    MOSTLY_IQ4_NL        = 25  # except 1d tensors
+    MOSTLY_IQ3_S         = 26  # except 1d tensors
+    MOSTLY_IQ3_M         = 27  # except 1d tensors
+    MOSTLY_IQ2_S         = 28  # except 1d tensors
+    MOSTLY_IQ2_M         = 29  # except 1d tensors
+    MOSTLY_IQ4_XS        = 30  # except 1d tensors
+    MOSTLY_IQ1_M         = 31  # except 1d tensors
+    MOSTLY_BF16          = 32  # except 1d tensors
+
+    GUESSED              = 1024  # not specified in the model file
+
+
 class GGUFEndian(IntEnum):
     LITTLE = 0
     BIG = 1

gguf-py/gguf/gguf_writer.py

@@ -7,7 +7,7 @@ import struct
 import tempfile
 from enum import Enum, auto
 from io import BufferedWriter
-from typing import IO, Any, Callable, Sequence, Mapping
+from typing import IO, Any, Sequence, Mapping
 from string import ascii_letters, digits
 
 import numpy as np
@@ -28,47 +28,6 @@ from .constants import (
 logger = logging.getLogger(__name__)
 
 
-class LazyTensor:
-    data: Callable[[], np.ndarray[Any, Any]]
-    # to avoid too deep recursion
-    functions: list[Callable[[np.ndarray[Any, Any]], np.ndarray[Any, Any]]]
-    dtype: np.dtype[Any]
-    shape: tuple[int, ...]
-
-    def __init__(self, data: Callable[[], np.ndarray[Any, Any]], *, dtype: type, shape: tuple[int, ...]):
-        self.data = data
-        self.functions = []
-        self.dtype = np.dtype(dtype)
-        self.shape = shape
-
-    def astype(self, dtype: type, **kwargs) -> LazyTensor:
-        self.functions.append(lambda n: n.astype(dtype, **kwargs))
-        self.dtype = np.dtype(dtype)
-        return self
-
-    @property
-    def nbytes(self) -> int:
-        size = 1
-        for n in self.shape:
-            size *= n
-        return size * self.dtype.itemsize
-
-    def tofile(self, *args, **kwargs) -> None:
-        data = self.data()
-        for f in self.functions:
-            data = f(data)
-        assert data.shape == self.shape
-        assert data.dtype == self.dtype
-        assert data.nbytes == self.nbytes
-        self.functions = []
-        self.data = lambda: data
-        data.tofile(*args, **kwargs)
-
-    def byteswap(self, *args, **kwargs) -> LazyTensor:
-        self.functions.append(lambda n: n.byteswap(*args, **kwargs))
-        return self
-
-
 class WriterState(Enum):
     EMPTY   = auto()
     HEADER  = auto()
@@ -79,7 +38,7 @@ class WriterState(Enum):
 class GGUFWriter:
     fout: BufferedWriter
     temp_file: tempfile.SpooledTemporaryFile[bytes] | None
-    tensors: list[np.ndarray[Any, Any] | LazyTensor]
+    tensors: list[np.ndarray[Any, Any]]
     _simple_value_packing = {
         GGUFValueType.UINT8:   "B",
         GGUFValueType.INT8:    "b",
@@ -278,7 +237,7 @@ class GGUFWriter:
         self.ti_data_count += 1
 
     def add_tensor(
-        self, name: str, tensor: np.ndarray[Any, Any] | LazyTensor, raw_shape: Sequence[int] | None = None,
+        self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
         raw_dtype: GGMLQuantizationType | None = None,
     ) -> None:
         if self.endianess == GGUFEndian.BIG:
@@ -303,7 +262,7 @@ class GGUFWriter:
         if pad != 0:
             fp.write(bytes([0] * pad))
 
-    def write_tensor_data(self, tensor: np.ndarray[Any, Any] | LazyTensor) -> None:
+    def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
         if self.state is not WriterState.TI_DATA:
             raise ValueError(f'Expected output file to contain tensor info, got {self.state}')
 
@@ -391,7 +350,7 @@ class GGUFWriter:
     def add_name(self, name: str) -> None:
         self.add_string(Keys.General.NAME, name)
 
-    def add_quantization_version(self, quantization_version: GGMLQuantizationType) -> None:
+    def add_quantization_version(self, quantization_version: int) -> None:
         self.add_uint32(
             Keys.General.QUANTIZATION_VERSION, quantization_version)
 

gguf-py/gguf/lazy.py

@@ -0,0 +1,225 @@
+from __future__ import annotations
+from abc import ABC, ABCMeta, abstractmethod
+
+import logging
+from typing import Any, Callable
+from collections import deque
+
+import numpy as np
+from numpy.typing import DTypeLike
+
+
+logger = logging.getLogger(__name__)
+
+
+class LazyMeta(ABCMeta):
+
+    def __new__(cls, name: str, bases: tuple[type, ...], namespace: dict[str, Any], **kwargs):
+        def __getattr__(self, __name: str) -> Any:
+            meta_attr = getattr(self._meta, __name)
+            if callable(meta_attr):
+                return type(self)._wrap_fn(
+                    (lambda s, *args, **kwargs: getattr(s, __name)(*args, **kwargs)),
+                    use_self=self,
+                )
+            elif isinstance(meta_attr, self._tensor_type):
+                # e.g. self.T with torch.Tensor should still be wrapped
+                return type(self)._wrap_fn(lambda s: getattr(s, __name))(self)
+            else:
+                # no need to wrap non-tensor properties,
+                # and they likely don't depend on the actual contents of the tensor
+                return meta_attr
+
+        namespace["__getattr__"] = __getattr__
+
+        # need to make a builder for the wrapped wrapper to copy the name,
+        # or else it fails with very cryptic error messages,
+        # because somehow the same string would end up in every closures
+        def mk_wrap(op_name: str, *, meta_noop: bool = False):
+            # need to wrap the wrapper to get self
+            def wrapped_special_op(self, *args, **kwargs):
+                return type(self)._wrap_fn(
+                    getattr(type(self)._tensor_type, op_name),
+                    meta_noop=meta_noop,
+                )(self, *args, **kwargs)
+            return wrapped_special_op
+
+        # special methods bypass __getattr__, so they need to be added manually
+        # ref: https://docs.python.org/3/reference/datamodel.html#special-lookup
+        # NOTE: doing this from a metaclass is very convenient
+        # TODO: make this even more comprehensive
+        for binary_op in (
+            "lt", "le", "eq", "ne", "ge", "gt", "not"
+            "abs", "add", "and", "floordiv", "invert", "lshift", "mod", "mul", "matmul",
+            "neg", "or", "pos", "pow", "rshift", "sub", "truediv", "xor",
+            "iadd", "iand", "ifloordiv", "ilshift", "imod", "imul", "ior", "irshift", "isub", "ixor",
+            "radd", "rand", "rfloordiv", "rmul", "ror", "rpow", "rsub", "rtruediv", "rxor",
+        ):
+            attr_name = f"__{binary_op}__"
+            # the result of these operators usually has the same shape and dtype as the input,
+            # so evaluation on the meta tensor can be skipped.
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=True)
+
+        for special_op in (
+            "getitem", "setitem", "len",
+        ):
+            attr_name = f"__{special_op}__"
+            namespace[attr_name] = mk_wrap(attr_name, meta_noop=False)
+
+        return super().__new__(cls, name, bases, namespace, **kwargs)
+
+
+# Tree of lazy tensors
+class LazyBase(ABC, metaclass=LazyMeta):
+    _tensor_type: type
+    _meta: Any
+    _data: Any | None
+    _lazy: deque[LazyBase]  # shared within a graph, to avoid deep recursion when making eager
+    _args: tuple
+    _func: Callable[[tuple], Any] | None
+
+    def __init__(self, *, meta: Any, data: Any | None = None, lazy: deque[LazyBase] | None = None, args: tuple = (), func: Callable[[tuple], Any] | None = None):
+        super().__init__()
+        self._meta = meta
+        self._data = data
+        self._lazy = lazy if lazy is not None else deque()
+        self._args = args
+        self._func = func
+        assert self._func is not None or self._data is not None
+        if self._data is None:
+            self._lazy.append(self)
+
+    def __init_subclass__(cls) -> None:
+        if "_tensor_type" not in cls.__dict__:
+            raise TypeError(f"property '_tensor_type' must be defined for {cls!r}")
+        return super().__init_subclass__()
+
+    @staticmethod
+    def _recurse_apply(o: Any, fn: Callable[[Any], Any]) -> Any:
+        # TODO: dict and set
+        if isinstance(o, (list, tuple)):
+            L = []
+            for item in o:
+                L.append(LazyBase._recurse_apply(item, fn))
+            if isinstance(o, tuple):
+                L = tuple(L)
+            return L
+        elif isinstance(o, LazyBase):
+            return fn(o)
+        else:
+            return o
+
+    @classmethod
+    def _wrap_fn(cls, fn: Callable, *, use_self: LazyBase | None = None, meta_noop: bool | DTypeLike = False) -> Callable[[Any], Any]:
+        def wrapped_fn(*args, **kwargs):
+            if kwargs is None:
+                kwargs = {}
+            args = ((use_self,) if use_self is not None else ()) + args
+
+            meta_args = LazyBase._recurse_apply(args, lambda t: t._meta)
+
+            if isinstance(meta_noop, bool) and not meta_noop:
+                try:
+                    res = fn(*meta_args, **kwargs)
+                except NotImplementedError:
+                    # running some operations on PyTorch's Meta tensors can cause this exception
+                    res = None
+            else:
+                # some operators don't need to actually run on the meta tensors
+                assert len(args) > 0
+                res = args[0]
+                assert isinstance(res, cls)
+                res = res._meta
+                # allow operations to override the dtype
+                if meta_noop is not True:
+                    res = cls.meta_with_dtype(res, meta_noop)
+
+            if isinstance(res, cls._tensor_type):
+                def collect_replace(t: LazyBase):
+                    if collect_replace.shared_lazy is None:
+                        collect_replace.shared_lazy = t._lazy
+                    else:
+                        collect_replace.shared_lazy.extend(t._lazy)
+                        t._lazy = collect_replace.shared_lazy
+
+                # emulating a static variable
+                collect_replace.shared_lazy = None
+
+                LazyBase._recurse_apply(args, collect_replace)
+
+                shared_lazy = collect_replace.shared_lazy
+
+                return cls(meta=cls.eager_to_meta(res), lazy=shared_lazy, args=args, func=lambda a: fn(*a, **kwargs))
+            else:
+                del res  # not needed
+                # non-tensor return likely relies on the contents of the args
+                # (e.g. the result of torch.equal)
+                eager_args = cls.to_eager(args)
+                return fn(*eager_args, **kwargs)
+        return wrapped_fn
+
+    @classmethod
+    def to_eager(cls, t: Any) -> Any:
+        def simple_to_eager(_t: LazyBase) -> Any:
+            def already_eager_to_eager(_t: LazyBase) -> Any:
+                assert _t._data is not None
+                return _t._data
+
+            while _t._data is None:
+                lt = _t._lazy.popleft()
+                if lt._data is not None:
+                    raise ValueError(f"{lt} did not belong in the lazy queue")
+                assert lt._func is not None
+                lt._args = cls._recurse_apply(lt._args, already_eager_to_eager)
+                lt._data = lt._func(lt._args)
+                # sanity check
+                assert lt._data.dtype == lt._meta.dtype
+                assert lt._data.shape == lt._meta.shape
+
+            return _t._data
+
+        # recurse into lists and/or tuples, keeping their structure
+        return cls._recurse_apply(t, simple_to_eager)
+
+    @classmethod
+    def eager_to_meta(cls, t: Any) -> Any:
+        return cls.meta_with_dtype(t, t.dtype)
+
+    # must be overridden, meta tensor init is backend-specific
+    @classmethod
+    @abstractmethod
+    def meta_with_dtype(cls, m: Any, dtype: Any) -> Any: pass
+
+    @classmethod
+    def from_eager(cls, t: Any) -> Any:
+        if type(t) is cls:
+            # already eager
+            return t
+        elif isinstance(t, cls._tensor_type):
+            return cls(meta=cls.eager_to_meta(t), data=t)
+        else:
+            return TypeError(f"{type(t)!r} is not compatible with {cls._tensor_type!r}")
+
+
+class LazyNumpyTensor(LazyBase):
+    _tensor_type = np.ndarray
+
+    @classmethod
+    def meta_with_dtype(cls, m: np.ndarray[Any, Any], dtype: DTypeLike) -> np.ndarray[Any, Any]:
+        # The initial idea was to use np.nan as the fill value,
+        # but non-float types like np.int16 can't use that.
+        # So zero it is.
+        cheat = np.zeros(1, dtype)
+        return np.lib.stride_tricks.as_strided(cheat, m.shape, (0 for _ in m.shape))
+
+    def astype(self, dtype, *args, **kwargs):
+        meta = type(self).meta_with_dtype(self._meta, dtype)
+        full_args = (self, dtype,) + args
+        # very important to pass the shared _lazy deque, or else there's an infinite loop somewhere.
+        return type(self)(meta=meta, args=full_args, lazy=self._lazy, func=(lambda a: a[0].astype(*a[1:], **kwargs)))
+
+    def tofile(self, *args, **kwargs):
+        eager = LazyNumpyTensor.to_eager(self)
+        return eager.tofile(*args, **kwargs)
+
+    # TODO: __array_function__

gguf-py/gguf/tensor_mapping.py

@@ -137,6 +137,7 @@ class TensorNameMap:
             "layers.{bid}.attention.wk",                               # llama-pth
             "encoder.layer.{bid}.attention.self.key",                  # bert
             "transformer.h.{bid}.attn.k_proj",                         # gpt-j
+            "transformer.h.{bid}.attn.k",                              # refact
             "model.layers.layers.{bid}.self_attn.k_proj",              # plamo
             "model.layers.{bid}.attention.wk",                         # internlm2
             "transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok
@@ -148,6 +149,7 @@ class TensorNameMap:
             "layers.{bid}.attention.wv",                                 # llama-pth
             "encoder.layer.{bid}.attention.self.value",                  # bert
             "transformer.h.{bid}.attn.v_proj",                           # gpt-j
+            "transformer.h.{bid}.attn.v",                                # refact
             "model.layers.layers.{bid}.self_attn.v_proj",                # plamo
             "model.layers.{bid}.attention.wv",                           # internlm2
             "transformer.decoder_layer.{bid}.multi_head_attention.value" # Grok
@@ -229,6 +231,7 @@ class TensorNameMap:
             "layers.{bid}.feed_forward.w3",                           # llama-pth
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
+            "transformer.h.{bid}.mlp.linear_3",                       # refact
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
             "model.layers.{bid}.mlp.dense_h_to_4h",                   # persimmon
             "transformer.h.{bid}.mlp.w1",                             # qwen
@@ -240,6 +243,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
             "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
+            "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -266,6 +270,8 @@ class TensorNameMap:
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
             "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
+            "encoder.layer.{bid}.mlp.gated_layers_w",     # jina-bert-v2
+            "transformer.h.{bid}.mlp.linear_1",           # refact
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -299,6 +305,7 @@ class TensorNameMap:
             "model.layers.{bid}.feed_forward.w2",                     # internlm2
             "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
+            "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -317,6 +324,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
             "model.layers.{bid}.self_attn.q_norm",                            # cohere
             "transformer.blocks.{bid}.attn.q_ln",                             # sea-lion
+            "encoder.layer.{bid}.attention.self.layer_norm_q"                 # jina-bert-v2
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
@@ -324,6 +332,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
             "model.layers.{bid}.self_attn.k_norm",                            # cohere
             "transformer.blocks.{bid}.attn.k_ln",                             # sea-lion
+            "encoder.layer.{bid}.attention.self.layer_norm_k"                 # jina-bert-v2
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (
@@ -334,6 +343,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.output.LayerNorm",         # bert
             "encoder.layers.{bid}.norm2",                   # nomic-bert
             "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
+            "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
         ),
 
         MODEL_TENSOR.SSM_IN: (

gguf-py/scripts/gguf-new-metadata.py

@@ -7,7 +7,8 @@ import json
 from pathlib import Path
 
 import numpy as np
-from typing import Any, Sequence
+from tqdm import tqdm
+from typing import Any, Sequence, NamedTuple
 
 # Necessary to load the local gguf package
 if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
@@ -18,6 +19,12 @@ import gguf
 logger = logging.getLogger("gguf-new-metadata")
 
 
+class MetadataDetails(NamedTuple):
+    type: gguf.GGUFValueType
+    value: Any
+    description: str = ''
+
+
 def get_byteorder(reader: gguf.GGUFReader) -> gguf.GGUFEndian:
     if np.uint32(1) == np.uint32(1).newbyteorder("<"):
         # Host is little endian
@@ -59,7 +66,16 @@ def get_field_data(reader: gguf.GGUFReader, key: str) -> Any:
     return decode_field(field)
 
 
-def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: dict[str, str], remove_metadata: Sequence[str]) -> None:
+def find_token(token_list: Sequence[int], token: str) -> Sequence[int]:
+    token_ids = [index for index, value in enumerate(token_list) if value == token]
+
+    if len(token_ids) == 0:
+        raise LookupError(f'Unable to find "{token}" in token list!')
+
+    return token_ids
+
+
+def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new_metadata: dict[str, MetadataDetails], remove_metadata: Sequence[str]) -> None:
     for field in reader.fields.values():
         # Suppress virtual fields and fields written by GGUFWriter
         if field.name == gguf.Keys.General.ARCHITECTURE or field.name.startswith('GGUF.'):
@@ -75,54 +91,64 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
             logger.debug(f'Removing {field.name}')
             continue
 
-        old_val = decode_field(field)
+        old_val = MetadataDetails(field.types[0], decode_field(field))
         val = new_metadata.get(field.name, old_val)
 
         if field.name in new_metadata:
-            logger.debug(f'Modifying {field.name}: "{old_val}" -> "{val}"')
+            logger.debug(f'Modifying {field.name}: "{old_val.value}" -> "{val.value}" {val.description}')
             del new_metadata[field.name]
-        elif val is not None:
+        elif val.value is not None:
             logger.debug(f'Copying {field.name}')
 
-        if val is not None:
+        if val.value is not None:
             writer.add_key(field.name)
-            writer.add_val(val, field.types[0])
+            writer.add_val(val.value, val.type)
 
     if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
         logger.debug('Adding chat template(s)')
-        writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE])
+        writer.add_chat_template(new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE].value)
         del new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE]
 
-    # TODO: Support other types than string?
     for key, val in new_metadata.items():
-        logger.debug(f'Adding {key}: {val}')
+        logger.debug(f'Adding {key}: "{val.value}" {val.description}')
         writer.add_key(key)
-        writer.add_val(val, gguf.GGUFValueType.STRING)
+        writer.add_val(val.value, val.type)
+
+    total_bytes = 0
 
     for tensor in reader.tensors:
+        total_bytes += tensor.n_bytes
         # Dimensions are written in reverse order, so flip them first
         shape = np.flipud(tensor.shape).tolist()
         writer.add_tensor_info(tensor.name, shape, tensor.data.dtype, tensor.data.nbytes, tensor.tensor_type)
 
+    bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
+
     writer.write_header_to_file()
     writer.write_kv_data_to_file()
     writer.write_ti_data_to_file()
 
     for tensor in reader.tensors:
         writer.write_tensor_data(tensor.data)
+        bar.update(tensor.n_bytes)
 
     writer.close()
 
 
 def main() -> None:
+    tokenizer_metadata = (getattr(gguf.Keys.Tokenizer, n) for n in gguf.Keys.Tokenizer.__dict__.keys() if not n.startswith('_'))
+    token_names = dict((n.split('.')[-1][:-len('_token_id')], n) for n in tokenizer_metadata if n.endswith('_token_id'))
+
     parser = argparse.ArgumentParser(description="Make a copy of a GGUF file with new metadata")
     parser.add_argument("input",                                       type=Path, help="GGUF format model input filename")
     parser.add_argument("output",                                      type=Path, help="GGUF format model output filename")
-    parser.add_argument("--general-name",                              type=str,  help="The models general.name")
+    parser.add_argument("--general-name",                              type=str,  help="The models general.name", metavar='"name"')
-    parser.add_argument("--general-description",                       type=str,  help="The models general.description")
+    parser.add_argument("--general-description",                       type=str,  help="The models general.description", metavar='"Description ..."')
-    parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)")
+    parser.add_argument("--chat-template",                             type=str,  help="Chat template string (or JSON string containing templates)", metavar='"{% ... %} ..."')
-    parser.add_argument("--chat-template-config",                      type=Path, help="Config file (tokenizer_config.json) containing chat template(s)")
+    parser.add_argument("--chat-template-config",                      type=Path, help="Config file containing chat template(s)", metavar='tokenizer_config.json')
-    parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model")
+    parser.add_argument("--remove-metadata",      action="append",     type=str,  help="Remove metadata (by key name) from output model", metavar='general.url')
+    parser.add_argument("--special-token",        action="append",     type=str,  help="Special token by value", nargs=2, metavar=(' | '.join(token_names.keys()), '"<token>"'))
+    parser.add_argument("--special-token-by-id",  action="append",     type=str,  help="Special token by id", nargs=2, metavar=(' | '.join(token_names.keys()), '0'))
     parser.add_argument("--force",                action="store_true",            help="Bypass warnings without confirmation")
     parser.add_argument("--verbose",              action="store_true",            help="Increase output verbosity")
     args = parser.parse_args(None if len(sys.argv) > 2 else ["--help"])
@@ -133,20 +159,20 @@ def main() -> None:
     remove_metadata = args.remove_metadata or []
 
     if args.general_name:
-        new_metadata[gguf.Keys.General.NAME] = args.general_name
+        new_metadata[gguf.Keys.General.NAME] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_name)
 
     if args.general_description:
-        new_metadata[gguf.Keys.General.DESCRIPTION] = args.general_description
+        new_metadata[gguf.Keys.General.DESCRIPTION] = MetadataDetails(gguf.GGUFValueType.STRING, args.general_description)
 
     if args.chat_template:
-        new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template
+        new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, json.loads(args.chat_template) if args.chat_template.startswith('[') else args.chat_template)
 
     if args.chat_template_config:
         with open(args.chat_template_config, 'r') as fp:
             config = json.load(fp)
             template = config.get('chat_template')
             if template:
-                new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = template
+                new_metadata[gguf.Keys.Tokenizer.CHAT_TEMPLATE] = MetadataDetails(gguf.GGUFValueType.STRING, template)
 
     if remove_metadata:
         logger.warning('*** Warning *** Warning *** Warning **')
@@ -166,6 +192,32 @@ def main() -> None:
     arch = get_field_data(reader, gguf.Keys.General.ARCHITECTURE)
     endianess = get_byteorder(reader)
 
+    token_list = get_field_data(reader, gguf.Keys.Tokenizer.LIST) or []
+
+    for name, token in args.special_token or []:
+        if name not in token_names:
+            logger.warning(f'Unknown special token "{name}", ignoring...')
+        else:
+            ids = find_token(token_list, token)
+            new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, ids[0], f'= {token}')
+
+            if len(ids) > 1:
+                logger.warning(f'Multiple "{token}" tokens found, choosing ID {ids[0]}, use --special-token-by-id if you want another:')
+                logger.warning(', '.join(str(i) for i in ids))
+
+    for name, id_string in args.special_token_by_id or []:
+        if name not in token_names:
+            logger.warning(f'Unknown special token "{name}", ignoring...')
+        elif not id_string.isdecimal():
+            raise LookupError(f'Token ID "{id_string}" is not a valid ID!')
+        else:
+            id_int = int(id_string)
+
+            if id_int >= 0 and id_int < len(token_list):
+                new_metadata[token_names[name]] = MetadataDetails(gguf.GGUFValueType.UINT32, id_int, f'= {token_list[id_int]}')
+            else:
+                raise LookupError(f'Token ID {id_int} is not within token list!')
+
     if os.path.isfile(args.output) and not args.force:
         logger.warning('*** Warning *** Warning *** Warning **')
         logger.warning(f'* The "{args.output}" GGUF file already exists, it will be overwritten!')

llama.cpp

@@ -205,6 +205,7 @@ enum llm_arch {
     LLM_ARCH_REFACT,
     LLM_ARCH_BERT,
     LLM_ARCH_NOMIC_BERT,
+    LLM_ARCH_JINA_BERT_V2,
     LLM_ARCH_BLOOM,
     LLM_ARCH_STABLELM,
     LLM_ARCH_QWEN,
@@ -241,6 +242,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_REFACT,          "refact"       },
     { LLM_ARCH_BERT,            "bert"         },
     { LLM_ARCH_NOMIC_BERT,      "nomic-bert"   },
+    { LLM_ARCH_JINA_BERT_V2,    "jina-bert-v2" },
     { LLM_ARCH_BLOOM,           "bloom"        },
     { LLM_ARCH_STABLELM,        "stablelm"     },
     { LLM_ARCH_QWEN,            "qwen"         },
@@ -691,6 +693,25 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_JINA_BERT_V2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
+            { LLM_TENSOR_TOKEN_TYPES,     "token_types" },
+            { LLM_TENSOR_ATTN_OUT_NORM,   "blk.%d.attn_output_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_LAYER_OUT_NORM,  "blk.%d.layer_output_norm" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_BLOOM,
         {
@@ -1845,7 +1866,7 @@ struct llama_hparams {
     float f_logit_scale    = 0.0f;
 
     bool causal_attn = true;
-    bool use_alibi   = false; // currently, we need KQ_pos data for ALiBi-based models
+    bool use_alibi   = false;
 
     enum llama_pooling_type      pooling_type            = LLAMA_POOLING_TYPE_NONE;
     enum llama_rope_type         rope_type               = LLAMA_ROPE_TYPE_NONE;
@@ -2317,7 +2338,6 @@ struct llama_context {
     struct ggml_tensor * inp_pos;       // I32 [n_batch]
     struct ggml_tensor * inp_out_ids;   // I32 [n_outputs]
     struct ggml_tensor * inp_KQ_mask;   // F32 [kv_size, n_batch]
-    struct ggml_tensor * inp_KQ_pos;    // F32 [n_kv]
     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]
@@ -3779,6 +3799,12 @@ static void llm_load_hparams(
 
     // get hparams kv
     ml.get_key(LLM_KV_VOCAB_SIZE,           hparams.n_vocab,       false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, hparams.n_vocab);
+
+    // everything past this point is not vocab-related
+    if (hparams.vocab_only) {
+        return;
+    }
+
     ml.get_key(LLM_KV_CONTEXT_LENGTH,       hparams.n_ctx_train);
     ml.get_key(LLM_KV_EMBEDDING_LENGTH,     hparams.n_embd);
     ml.get_key(LLM_KV_FEED_FORWARD_LENGTH,  hparams.n_ff);
@@ -3860,7 +3886,7 @@ static void llm_load_hparams(
                     switch (hparams.n_layer) {
                         case 22: model.type = e_model::MODEL_1B; break;
                         case 26: model.type = e_model::MODEL_3B; break;
-                        case 32: model.type = hparams.n_head == hparams.n_head_kv ? e_model::MODEL_7B : e_model::MODEL_8B; break; // LLaMa 8B v3 uses GQA
+                        case 32: model.type = hparams.n_vocab < 40000 ? e_model::MODEL_7B : e_model::MODEL_8B; break;
                         case 40: model.type = e_model::MODEL_13B; break;
                         case 48: model.type = e_model::MODEL_34B; break;
                         case 60: model.type = e_model::MODEL_30B; break;
@@ -3962,6 +3988,19 @@ static void llm_load_hparams(
                         model.type = e_model::MODEL_335M; break; // bge-large
                 }
             } break;
+        case LLM_ARCH_JINA_BERT_V2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
+                ml.get_key(LLM_KV_ATTENTION_CAUSAL,           hparams.causal_attn);
+                ml.get_key(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT, hparams.n_vocab_type);
+                ml.get_key(LLM_KV_POOLING_TYPE,               hparams.pooling_type);
+                hparams.f_max_alibi_bias = 8.0f;
+
+                switch (hparams.n_layer) {
+                    case 4: model.type = e_model::MODEL_33M; break; // jina-embeddings-small
+                    case 12: model.type = e_model::MODEL_137M; break; // jina-embeddings-base
+                }
+            } break;
         case LLM_ARCH_NOMIC_BERT:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
@@ -4383,7 +4422,9 @@ static void llm_load_vocab(
                     tokenizer_pre == "starcoder") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_STARCODER;
             } else if (
-                    tokenizer_pre == "gpt-2") {
+                    tokenizer_pre == "gpt-2"   ||
+                    tokenizer_pre == "jina-es" ||
+                    tokenizer_pre == "jina-de") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
             } else if (
                     tokenizer_pre == "refact") {
@@ -5245,6 +5286,50 @@ static bool llm_load_tensors(
                         layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd});
                     }
                 } break;
+            case LLM_ARCH_JINA_BERT_V2:
+                {
+                    model.tok_embd     = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}); // word_embeddings
+                    model.type_embd    = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_vocab_type}); //token_type_embeddings
+                    model.tok_norm   = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}); // LayerNorm
+                    model.tok_norm_b = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"),   {n_embd}); //LayerNorm bias
+
+                    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]; // JinaBertLayer
+
+                        layer.wq   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd});
+                        layer.bq   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q,   "bias", i),   {n_embd});
+
+                        layer.attn_q_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, false);
+                        layer.attn_q_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {n_embd}, false);
+
+                        layer.wk   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.bk   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K,   "bias", i),   {n_embd_gqa});
+
+                        layer.attn_k_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, false);
+                        layer.attn_k_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {n_embd}, false);
+
+                        layer.wv   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa});
+                        layer.bv   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V,   "bias", i),   {n_embd_gqa});
+
+                        layer.wo              = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {n_embd, n_embd}); //output_dens
+                        layer.bo              = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT,      "bias", i), {n_embd}); //output_dens
+
+                        layer.attn_out_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
+                        layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd});
+
+                        layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,        "weight", i), {n_embd, n_ff});
+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE,    "weight", i), {n_embd, n_ff});
+
+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN,        "weight", i), {n_ff, n_embd});
+                        layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN,      "bias", i), {n_embd});
+
+                        layer.layer_out_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_LAYER_OUT_NORM,        "weight", i), {n_embd});
+                        layer.layer_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_LAYER_OUT_NORM,        "bias", i), {n_embd});
+                    }
+                } break;
             case LLM_ARCH_BLOOM:
                 {
                     model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
@@ -6321,7 +6406,7 @@ static struct ggml_tensor * llm_build_ffn(
           llm_ffn_gate_type   type_gate,
          const llm_build_cb & cb,
                         int   il) {
-    struct ggml_tensor * tmp = ggml_mul_mat(ctx, up, cur);
+    struct ggml_tensor * tmp = up ? ggml_mul_mat(ctx, up, cur) : cur;
     cb(tmp, "ffn_up", il);
 
     if (up_b) {
@@ -6503,7 +6588,6 @@ static struct ggml_tensor * llm_build_kqv(
          struct ggml_tensor * wo_b,
          struct ggml_tensor * q_cur,
          struct ggml_tensor * kq_mask,
-         struct ggml_tensor * kq_pos,
                     int32_t   n_tokens,
                     int32_t   n_kv,
                     float     kq_scale,
@@ -6533,10 +6617,6 @@ static struct ggml_tensor * llm_build_kqv(
         GGML_UNUSED(model);
         GGML_UNUSED(n_ctx);
 
-        // note: if this assert triggers, then some check has failed earlier
-        //       the idea is to detect during context creation that ALiBi would be used and disable Flash Attention
-        GGML_ASSERT(kq_pos == nullptr && "ALiBi is not yet supported with Flash Attention");
-
         // split cached v into n_head heads (not transposed)
         struct ggml_tensor * v =
             ggml_view_3d(ctx, kv.v_l[il],
@@ -6546,7 +6626,7 @@ static struct ggml_tensor * llm_build_kqv(
                     0);
         cb(v, "v", il);
 
-        cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale);
+        cur = ggml_flash_attn_ext(ctx, q, k, v, kq_mask, kq_scale, hparams.f_max_alibi_bias);
 
         if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3) {
             ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
@@ -6577,28 +6657,8 @@ static struct ggml_tensor * llm_build_kqv(
             kq = ggml_scale(ctx, kq, 30);
         }
 
-#if defined(GGML_USE_KOMPUTE)
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, hparams.f_max_alibi_bias);
-#pragma message("TODO: ALiBi support in ggml_soft_max_ext is not implemented for Kompute")
-#pragma message("      Falling back to ggml_alibi(). Will become an error in Mar 2024")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5488")
-        if (hparams.use_alibi) {
-            kq = ggml_scale(ctx, kq, kq_scale);
-            cb(kq, "kq_scaled", il);
-
-            kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, hparams.f_max_alibi_bias);
-            cb(kq, "kq_scaled_alibi", il);
-
-            kq = ggml_add(ctx, kq, kq_mask);
-            cb(kq, "kq_masked", il);
-
-            kq = ggml_soft_max(ctx, kq);
-            cb(kq, "kq_soft_max", il);
-        } else
-#endif
-        {
-            kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_pos, kq_scale, hparams.f_max_alibi_bias);
         cb(kq, "kq_soft_max_ext", il);
-        }
 
         GGML_ASSERT(kv.size == n_ctx);
 
@@ -6648,7 +6708,6 @@ static struct ggml_tensor * llm_build_kv(
          struct ggml_tensor * v_cur,
          struct ggml_tensor * q_cur,
          struct ggml_tensor * kq_mask,
-         struct ggml_tensor * kq_pos,
                     int32_t   n_tokens,
                     int32_t   kv_head,
                     int32_t   n_kv,
@@ -6667,7 +6726,7 @@ static struct ggml_tensor * llm_build_kv(
     struct ggml_tensor * cur;
 
     cur  = llm_build_kqv(ctx, model, hparams, cparams, kv, graph, wo, wo_b,
-            q_cur, kq_mask, kq_pos, n_tokens, n_kv, kq_scale, cb, il);
+            q_cur, kq_mask, n_tokens, n_kv, kq_scale, cb, il);
     cb(cur, "kqv_out", il);
 
     return cur;
@@ -6779,7 +6838,6 @@ struct llm_build_context {
         lctx.inp_pos     = nullptr;
         lctx.inp_out_ids = nullptr;
         lctx.inp_KQ_mask = nullptr;
-        lctx.inp_KQ_pos = nullptr;
         lctx.inp_K_shift = nullptr;
         lctx.inp_mean    = nullptr;
         lctx.inp_cls     = nullptr;
@@ -6935,19 +6993,6 @@ struct llm_build_context {
         return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_mask, GGML_TYPE_F16) : lctx.inp_KQ_mask;
     }
 
-    struct ggml_tensor * build_inp_KQ_pos(bool causal = true) {
-        if (causal) {
-            lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_kv);
-        } else {
-            // TODO: this will be needed for ALiBi-based BERT models
-            //       https://github.com/ggerganov/llama.cpp/pull/6826
-            lctx.inp_KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_tokens);
-        }
-        cb(lctx.inp_KQ_pos, "KQ_pos", -1);
-        ggml_set_input(lctx.inp_KQ_pos);
-        return flash_attn ? ggml_cast(ctx0, lctx.inp_KQ_pos, GGML_TYPE_F16) : lctx.inp_KQ_pos;
-    }
-
     struct ggml_tensor * build_inp_mean() {
         lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
         cb(lctx.inp_mean, "inp_mean", -1);
@@ -7053,7 +7098,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7146,9 +7191,6 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
-
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -7193,7 +7235,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7263,9 +7305,6 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
-
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -7300,7 +7339,7 @@ struct llm_build_context {
                 cb(Kcur, "Kcur", il);
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7420,7 +7459,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7545,7 +7584,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7697,7 +7736,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -7809,7 +7848,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8013,7 +8052,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Q, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Q, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8079,9 +8118,6 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
-
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
 
@@ -8109,7 +8145,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8171,8 +8207,11 @@ struct llm_build_context {
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
+        struct ggml_tensor * inp_pos = nullptr;
 
-        struct ggml_tensor * inp_pos  = build_inp_pos();
+        if (model.arch != LLM_ARCH_JINA_BERT_V2) {
+            inp_pos = build_inp_pos();
+        }
         struct ggml_tensor * inp_mean = build_inp_mean();
         struct ggml_tensor * inp_cls  = build_inp_cls();
 
@@ -8203,13 +8242,26 @@ struct llm_build_context {
             struct ggml_tensor * Vcur;
 
             // self-attention
-            if (model.arch == LLM_ARCH_BERT) {
+            if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_JINA_BERT_V2) {
                 Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), model.layers[il].bq);
                 cb(Qcur, "Qcur", il);
 
+                if (model.layers[il].attn_q_norm) {
+                    Qcur = llm_build_norm(ctx0, Qcur, hparams,
+                            model.layers[il].attn_q_norm,
+                            model.layers[il].attn_q_norm_b,
+                            LLM_NORM, cb, il);
+                }
+
                 Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), model.layers[il].bk);
                 cb(Kcur, "Kcur", il);
 
+                if (model.layers[il].attn_k_norm) {
+                    Kcur = llm_build_norm(ctx0, Kcur, hparams,
+                            model.layers[il].attn_k_norm,
+                            model.layers[il].attn_k_norm_b,
+                            LLM_NORM, cb, il);
+                }
                 Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
 
@@ -8249,7 +8301,7 @@ struct llm_build_context {
             struct ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
             cb(kq, "kq", il);
 
-            kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, nullptr, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
+            kq = ggml_soft_max_ext(ctx0, kq, KQ_mask, 1.0f/sqrtf(float(n_embd_head)), hparams.f_max_alibi_bias);
             cb(kq, "kq_soft_max_ext", il);
 
             struct ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_tokens)));
@@ -8300,6 +8352,13 @@ struct llm_build_context {
                         model.layers[il].ffn_down, model.layers[il].ffn_down_b,
                         NULL,
                         LLM_FFN_GELU, LLM_FFN_SEQ, cb, il);
+            } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   NULL,
+                        model.layers[il].ffn_gate, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+                        NULL,
+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
             } else {
                 cur = llm_build_ffn(ctx0, cur,
                         model.layers[il].ffn_up,   NULL,
@@ -8366,9 +8425,6 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
-
         inpL = llm_build_norm(ctx0, inpL, hparams,
                 model.tok_norm,
                 model.tok_norm_b,
@@ -8402,7 +8458,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8467,9 +8523,6 @@ struct llm_build_context {
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
 
-        // positions of the tokens in the KV cache
-        struct ggml_tensor * KQ_pos = build_inp_KQ_pos();
-
         if (model.pos_embd) {
             // inp_pos - contains the positions
             struct ggml_tensor * inp_pos = build_inp_pos();
@@ -8533,13 +8586,13 @@ struct llm_build_context {
 
                     cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                             model.layers[il].wo, model.layers[il].bo,
-                            Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 } else {
                     Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
 
                     cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                             model.layers[il].wo, model.layers[il].bo,
-                            Kcur, Vcur, Qcur, KQ_mask, KQ_pos, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                            Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
                 }
             }
 
@@ -8683,7 +8736,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8801,7 +8854,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -8914,7 +8967,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9028,7 +9081,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9183,7 +9236,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9300,7 +9353,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9413,7 +9466,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
             struct ggml_tensor * sa_out = cur;
 
@@ -9516,7 +9569,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9623,7 +9676,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9739,7 +9792,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9856,7 +9909,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -9986,7 +10039,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10107,7 +10160,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, NULL,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f, cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10226,7 +10279,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10516,7 +10569,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10647,7 +10700,7 @@ struct llm_build_context {
 
                 cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
                         model.layers[il].wo, nullptr,
-                        Kcur, Vcur, Qcur, KQ_mask, nullptr, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
             }
 
             if (il == n_layer - 1) {
@@ -10828,6 +10881,7 @@ static struct ggml_cgraph * llama_build_graph(
                 result = llm.build_refact();
             } break;
         case LLM_ARCH_BERT:
+        case LLM_ARCH_JINA_BERT_V2:
         case LLM_ARCH_NOMIC_BERT:
             {
                 result = llm.build_bert();
@@ -11034,12 +11088,22 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
                         float f;
                         if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) {
                             f = -INFINITY;
+                        } else {
+                            if (hparams.use_alibi) {
+                                f = -fabs(lctx.kv_self.cells[i].pos - pos);
                             } else {
                                 f = 0.0f;
                             }
+                        }
                         data[h*(n_kv*n_tokens) + j*n_kv + i] = f;
                     }
                 }
+
+                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                    for (int j = 0; j < n_kv; ++j) {
+                        data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                    }
+                }
             }
         } else {
             // when using kv cache, the mask needs to match the kv cache size
@@ -11058,7 +11122,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
                         float f = -INFINITY;
                         for (int s = 0; s < batch.n_seq_id[i]; ++s) {
                             if (batch.seq_id[i][s] == seq_id) {
+                                if (hparams.use_alibi) {
+                                    f = -fabs(batch.pos[i] - batch.pos[j]);
+                                } else {
                                     f = 0.0f;
+                                }
                                 break;
                             }
                         }
@@ -11074,21 +11142,6 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
         }
     }
 
-    // ALiBi requires the KQ_pos tensor to provide the sequence position of each token in the batch
-    // this allows to process multiple sequences in parallel with ALiBi-based models
-    if (hparams.use_alibi) {
-        const int64_t n_kv = kv_self.n;
-
-        GGML_ASSERT(lctx.inp_KQ_pos);
-        GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_KQ_pos->buffer));
-
-        float * data = (float *) lctx.inp_KQ_pos->data;
-
-        for (int i = 0; i < n_kv; ++i) {
-            data[i] = float(lctx.kv_self.cells[i].pos);
-        }
-    }
-
     if (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
         const int64_t n_tokens = batch.n_tokens;
 
@@ -12203,13 +12256,14 @@ struct llm_tokenizer_bpe {
 
     void tokenize(const std::string & text, std::vector<llama_vocab::id> & output) {
         int final_prev_index = -1;
+        bool ignore_merges = false;
 
         std::vector<std::string> word_collection;
         switch (vocab.type) {
             case LLAMA_VOCAB_TYPE_BPE:
                 switch (vocab.type_pre) {
                     case LLAMA_VOCAB_PRE_TYPE_LLAMA3:
-                    case LLAMA_VOCAB_PRE_TYPE_DBRX:
+                        ignore_merges = true;
                         word_collection = unicode_regex_split(text, {
                             // original regex from tokenizer.json
                             //"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
@@ -12218,6 +12272,12 @@ struct llm_tokenizer_bpe {
                             "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                         });
                         break;
+                    case LLAMA_VOCAB_PRE_TYPE_DBRX:
+                        word_collection = unicode_regex_split(text, {
+                            // same as llama3
+                            "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                        });
+                        break;
                     case LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM:
                         word_collection = unicode_regex_split(text, {
                             "[\r\n]",
@@ -12307,6 +12367,11 @@ struct llm_tokenizer_bpe {
             int index = 0;
             size_t offset = 0;
 
+            if (ignore_merges && vocab.token_to_id.find(word) != vocab.token_to_id.end()) {
+                symbols.emplace_back(llm_symbol{-1, -1, word.c_str(), word.size()});
+                offset = word.size();
+            }
+
             while (offset < word.size()) {
                 llm_symbol sym;
                 size_t char_len = std::min(word.size() - offset, (size_t) ::utf8_len(word[offset]));
@@ -12497,7 +12562,7 @@ struct llm_tokenizer_wpm {
                 continue;
             }
             code = unicode_tolower(code);
-            if (type == CODEPOINT_TYPE_WHITESPACE) {
+            if (type == CODEPOINT_TYPE_SEPARATOR) {
                 code = ' ';
             }
             std::string s = unicode_cpt_to_utf8(code);
@@ -12761,7 +12826,10 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                     }
                 }
 
-                GGML_ASSERT(vocab.special_add_eos != 1);
+                if (add_special && vocab.special_add_eos == 1) {
+                    GGML_ASSERT(vocab.special_add_eos != -1);
+                    output.push_back(vocab.special_eos_id);
+                }
             } break;
         case LLAMA_VOCAB_TYPE_WPM:
             {
@@ -15518,23 +15586,11 @@ struct llama_context * llama_new_context_with_model(
         }
     }
 
-    if (cparams.flash_attn && hparams.use_alibi) {
-        LLAMA_LOG_WARN("%s: flash_attn is not yet compatible with ALiBi - forcing off\n", __func__);
-        cparams.flash_attn = false;
-    }
-
     if (cparams.flash_attn && model->arch == LLM_ARCH_GROK) {
         LLAMA_LOG_WARN("%s: flash_attn is not compatible with Grok - forcing off\n", __func__);
         cparams.flash_attn = false;
     }
 
-#ifdef GGML_USE_HIPBLAS
-    if (cparams.flash_attn) {
-        LLAMA_LOG_WARN("%s: flash_attn is not yet compatible with HIPBLAS builds - forcing off\n", __func__);
-        cparams.flash_attn = false;
-    }
-#endif
-
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
     }
@@ -15824,6 +15880,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_REFACT:
         case LLM_ARCH_BLOOM:
         case LLM_ARCH_MAMBA:
+        case LLM_ARCH_JINA_BERT_V2:
             return LLAMA_ROPE_TYPE_NONE;
 
         // use what we call a normal RoPE, operating on pairs of consecutive head values
@@ -17888,7 +17945,7 @@ struct llama_timings llama_get_timings(struct llama_context * ctx) {
         /*.t_eval_ms   =*/ 1e-3 * ctx->t_eval_us,
 
         /*.n_sample =*/ std::max(1, ctx->n_sample),
-        /*.n_p_eval =*/ std::max(1, ctx->n_p_eval),
+        /*.n_p_eval =*/ std::max(0, ctx->n_p_eval),
         /*.n_eval   =*/ std::max(1, ctx->n_eval),
     };
 

models/ggml-vocab-llama-bpe.gguf.inp

@@ -104,3 +104,5 @@ __ggml_vocab_test__
      
 🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
 __ggml_vocab_test__
+ Việt
+__ggml_vocab_test__

models/ggml-vocab-llama-bpe.gguf.out

@@ -41,3 +41,4 @@
  8765 8765 1644
  8765 8765 8765
  198 4815 15073 66597 8004 1602 2355 79772 11187 9468 248 222 320 8416 8 27623 114 102470 9468 234 104 31643 320 36773 100166 98634 8 26602 227 11410 99 247 9468 99 247 220 18 220 1644 220 8765 220 8765 18 220 8765 1644 220 8765 8765 220 8765 8765 18 220 8765 8765 1644 220 18 13 18 220 18 497 18 220 18 1131 18 220 21549 222 98629 241 45358 233 21549 237 45358 224 21549 244 21549 115 21549 253 45358 223 21549 253 21549 95 98629 227 76460 223 949 37046 101067 19000 23182 102301 9263 18136 16 36827 21909 56560 54337 19175 102118 13373 64571 34694 3114 112203 80112 3436 106451 14196 14196 74694 3089 3089 29249 17523 3001 27708 7801 358 3077 1027 364 83 820 568 596 1070 11 364 793 499 2771 30 364 44 539 2771 358 3358 1304 433 11 364 35 499 1093 1063 15600 30 1226 6 43712 264 64966 43
+ 101798

requirements.txt

@@ -9,5 +9,4 @@
 -r ./requirements/requirements-convert-hf-to-gguf.txt
 -r ./requirements/requirements-convert-hf-to-gguf-update.txt
 -r ./requirements/requirements-convert-llama-ggml-to-gguf.txt
--r ./requirements/requirements-convert-lora-to-ggml.txt
 -r ./requirements/requirements-convert-persimmon-to-gguf.txt

requirements/requirements-convert-lora-to-ggml.txt

@@ -1,2 +0,0 @@
--r ./requirements-convert.txt
-torch~=2.1.1

scripts/compare-llama-bench.py

@@ -325,8 +325,12 @@ table = []
 for row in rows_show:
     n_prompt = int(row[-4])
     n_gen    = int(row[-3])
-    assert n_prompt == 0 or n_gen == 0
+    if n_prompt != 0 and n_gen == 0:
-    test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}"
+        test_name = f"pp{n_prompt}"
+    elif n_prompt == 0 and n_gen != 0:
+        test_name = f"tg{n_gen}"
+    else:
+        test_name = f"pp{n_prompt}+tg{n_gen}"
     #           Regular columns    test name    avg t/s values              Speedup
     #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
     table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])

scripts/debug-test.sh

@@ -0,0 +1,117 @@
+#!/bin/bash
+test_suite=${1:-}
+test_number=${2:-}
+
+PROG=${0##*/}
+build_dir="build-ci-debug"
+
+if [ x"$1" = x"-h" ] || [ x"$1" = x"--help" ]; then
+    echo "Usage: $PROG [OPTION]... <test_regex> (test_number)"
+    echo "Debug specific ctest program."
+    echo
+    echo "Options:"
+    echo "  -h, --help       Display this help and exit"
+    echo
+    echo "Arguments:"
+    echo "  <test_regex>     (Mandatory) Supply one regex to the script to filter tests"
+    echo "  (test_number)    (Optional) Test number to run a specific test"
+    echo
+    echo "Example:"
+    echo "  $PROG test-tokenizer"
+    echo "  $PROG test-tokenizer 3"
+    echo
+    exit 0
+fi
+
+# Function to select and debug a test
+function select_test() {
+    test_suite=${1:-test}
+    test_number=${2:-}
+
+    # Sanity Check If Tests Is Detected
+    printf "\n\nGathering tests that fit REGEX: ${test_suite} ...\n"
+    tests=($(ctest -R ${test_suite} -V -N | grep -E " +Test +#[0-9]+*" | cut -d':' -f2 | awk '{$1=$1};1'))
+    if [ ${#tests[@]} -eq 0 ]
+    then
+        echo "No tests avaliable... check your compliation process..."
+        echo "Exiting."
+        exit 1
+    fi
+
+    if [ -z $test_number ]
+    then
+        # List out avaliable tests
+        printf "Which test would you like to debug?\n"
+        id=0
+        for s in "${tests[@]}"
+        do
+            echo "Test# ${id}"
+            echo "  $s"
+            ((id++))
+        done
+
+        # Prompt user which test they wanted to run
+        printf "\nRun test#? "
+        read test_number
+    else
+        printf "\nUser Already Requested #${test_number}"
+    fi
+
+    # Start GDB with the requested test binary and arguments
+    printf "Debugging(GDB) test: ${tests[test_number]}\n"
+    # Change IFS (Internal Field Separator)
+    sIFS=$IFS
+    IFS=$'\n'
+
+    # Get test args
+    gdb_args=($(ctest -R ${test_suite} -V -N | grep "Test command" | cut -d':' -f3 | awk '{$1=$1};1' ))
+    IFS=$sIFS
+    printf "Debug arguments: ${gdb_args[test_number]}\n\n"
+
+    # Expand paths if needed
+    args=()
+    for x in $(echo ${gdb_args[test_number]} | sed -e 's/"\/\<//' -e 's/\>"//')
+    do
+        args+=($(echo $x | sed -e 's/.*\/..\//..\//'))
+    done
+
+    # Execute debugger
+    echo "gdb args: ${args[@]}"
+    gdb --args ${args[@]}
+}
+
+# Step 0: Check the args
+if [ -z "$test_suite" ]
+then
+    echo "Usage: $PROG [OPTION]... <test_regex> (test_number)"
+    echo "Supply one regex to the script to filter tests,"
+    echo "and optionally a test number to run a specific test."
+    echo "Use --help flag for full instructions"
+    exit 1
+fi
+
+# Step 1: Reset and Setup folder context
+## Sanity check that we are actually in a git repo
+repo_root=$(git rev-parse --show-toplevel)
+if [ ! -d "$repo_root" ]; then
+    echo "Error: Not in a Git repository."
+    exit 1
+fi
+
+## Reset folder to root context of git repo
+pushd "$repo_root" || exit 1
+
+## Create and enter build directory
+rm -rf "$build_dir" && mkdir "$build_dir" || exit 1
+
+# Step 2: Setup Build Environment and Compile Test Binaries
+cmake -B "./$build_dir" -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON || exit 1
+pushd "$build_dir" && make -j || exit 1
+
+# Step 3: Debug the Test
+select_test "$test_suite" "$test_number"
+
+# Step 4: Return to the directory from which the user ran the command.
+popd || exit 1
+popd || exit 1
+popd || exit 1

scripts/gen-unicode-data.py

@@ -1,31 +1,14 @@
 import regex
 
 
-def cpt_to_utf8_str(cpt):
-    if cpt <= 0xFF:
-        return bytes([cpt, 0, 0, 0])
-    elif cpt <= 0xFFFF:
-        return bytes([cpt & 0xFF, cpt >> 8, 0, 0])
-    elif cpt <= 0xFFFFFF:
-        return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, 0])
-    else:
-        return bytes([cpt & 0xFF, (cpt >> 8) & 0xFF, (cpt >> 16) & 0xFF, cpt >> 24])
-
-
-def is_match(codepoint, regex_expr):
-    try:
-        res = regex.match(regex_expr, cpt_to_utf8_str(codepoint).decode('utf-32'))
-        return res is not None
-    except Exception:
-        return False
-
-
 def get_matches(regex_expr):
+    regex_expr_compiled = regex.compile(regex_expr)
     unicode_ranges = []
     current_range = None
 
     for codepoint in range(0x110000):
-        if is_match(codepoint, regex_expr):
+        char = chr(codepoint)
+        if regex_expr_compiled.match(char):
             if current_range is None:
                 current_range = [codepoint, codepoint]
             else:
@@ -40,27 +23,42 @@ def get_matches(regex_expr):
     return unicode_ranges
 
 
-def print_cat(cat, ranges):
+def print_cat(mode, cat, ranges):
+    if mode == "range":
         print("const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_{} = {{".format(cat)) # noqa: NP100
-    cnt = 0
+    if mode == "map":
-    for start, end in ranges:
+        print("const std::map<uint32_t, uint32_t> unicode_map_{} = {{".format(cat)) # noqa: NP100
-        if cnt % 4 != 0:
+    for i, values in enumerate(ranges):
-            print(" ", end="") # noqa: NP100
+        end = ",\n" if (i % 4 == 3 or i + 1 == len(ranges)) else ", "
-        print("{{0x{:08X}, 0x{:08X}}},".format(start, end), end="") # noqa: NP100
+        values = ["0x%08X" % value for value in values]
-        if cnt % 4 == 3:
+        print("{" + ", ".join(values) + "}", end=end) # noqa: NP100
-            print("") # noqa: NP100
-        cnt += 1
-
-    if cnt % 4 != 0:
-        print("") # noqa: NP100
     print("};") # noqa: NP100
     print("") # noqa: NP100
 
 
-print_cat("number",      get_matches(r'\p{N}'))
+print_cat("range", "number",      get_matches(r'\p{N}'))
-print_cat("letter",      get_matches(r'\p{L}'))
+print_cat("range", "letter",      get_matches(r'\p{L}'))
-print_cat("whitespace",  get_matches(r'\p{Z}'))
+print_cat("range", "separator",   get_matches(r'\p{Z}'))
-print_cat("accent_mark", get_matches(r'\p{M}'))
+print_cat("range", "accent_mark", get_matches(r'\p{M}'))
-print_cat("punctuation", get_matches(r'\p{P}'))
+print_cat("range", "punctuation", get_matches(r'\p{P}'))
-print_cat("symbol",      get_matches(r'\p{S}'))
+print_cat("range", "symbol",      get_matches(r'\p{S}'))
-print_cat("control",     get_matches(r'\p{C}'))
+print_cat("range", "control",     get_matches(r'\p{C}'))
+
+print_cat("range", "whitespace",  get_matches(r'\s'))
+
+
+map_lowercase = []
+map_uppercase = []
+for codepoint in range(0x110000):
+    char = chr(codepoint)
+    lower = ord(char.lower()[0])
+    upper = ord(char.upper()[0])
+    if codepoint != lower:
+        map_lowercase.append((codepoint, lower))
+    if codepoint != upper:
+        map_uppercase.append((codepoint, upper))
+print_cat("map", "lowercase", map_lowercase)
+print_cat("map", "uppercase", map_uppercase)
+
+
+# TODO: generate unicode_map_nfd

scripts/sync-ggml.last

@@ -1 +1 @@
-98875cdb7e9ceeb726d1c196d2fecb3cbb59b93a
+30f54cbb3ada3e4c5bc6924de3e5918e5be4ff11

tests/CMakeLists.txt

@@ -93,7 +93,7 @@ target_link_libraries(test-tokenizer-1-bpe PRIVATE common)
 install(TARGETS test-tokenizer-1-bpe RUNTIME)
 
 # TODO: disabled due to slowness
-#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
+#llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-llama-bpe ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf --ignore-merges)
 #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-falcon    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf)
 #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
 #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-mpt       ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf)

tests/test-backend-ops.cpp

@@ -2,6 +2,7 @@
 #include <ggml-alloc.h>
 #include <ggml-backend.h>
 #include <ggml-backend-impl.h>
+
 #include <algorithm>
 #include <array>
 #include <cfloat>
@@ -1111,11 +1112,7 @@ struct test_soft_max : public test_case {
         if (this->mask) {
             mask = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, ne[0], ne[1]);
         }
-        ggml_tensor * pos = nullptr;
+        ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, scale, max_bias);
-        if (max_bias > 0.0f) {
-            pos = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ne[0]);
-        }
-        ggml_tensor * out = ggml_soft_max_ext(ctx, a, mask, pos, scale, max_bias);
         return out;
     }
 };
@@ -1490,23 +1487,25 @@ struct test_flash_attn_ext : public test_case {
     const int64_t kv; // kv size
     const int64_t nb; // batch size
 
+    const float max_bias; // ALiBi
+
     std::string vars() override {
-        return VARS_TO_STR4(hs, nh, kv, nb);
+        return VARS_TO_STR5(hs, nh, kv, nb, max_bias);
     }
 
     double max_nmse_err() override {
         return 5e-4;
     }
 
-    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8)
+    test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t kv = 96, int64_t nb = 8, float max_bias = 0.0f)
-        : hs(hs), nh(nh), kv(kv), nb(nb) {}
+        : hs(hs), nh(nh), kv(kv), nb(nb), max_bias(max_bias) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
         ggml_tensor * k = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
         ggml_tensor * v = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, hs, kv, nh, 1);
         ggml_tensor * mask = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1);
-        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, mask, 1.0f/sqrtf(hs));
+        ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, mask, 1.0f/sqrtf(hs), max_bias);
         return out;
     }
 };
@@ -1611,7 +1610,7 @@ public:
 
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
 
-        kq = ggml_soft_max_ext(ctx, kq, kq_mask, nullptr, kq_scale, 0.0f);
+        kq = ggml_soft_max_ext(ctx, kq, kq_mask, kq_scale, 0.0f);
 
         // split cached v into n_head heads
         struct ggml_tensor * v =
@@ -2128,6 +2127,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 #endif
     for (bool mask : {false, true}) {
         for (float max_bias : {0.0f, 8.0f}) {
+            if (!mask && max_bias > 0.0f) continue;
             for (float scale : {1.0f, 0.1f}) {
                 for (int64_t ne0 : {16, 1024}) {
                     for (int64_t ne1 : {16, 1024}) {
@@ -2141,7 +2141,6 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
 
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 0.0f));
-    test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 8.0f));
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 8.0f));
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
@@ -2176,10 +2175,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_leaky_relu());
 
     for (int hs : { 64, 80, 128, 256, }) {
+        for (float max_bias : {0.0f, 8.0f}) {
             for (int nh : { 32, }) {
                 for (int kv : { 512, 1024, }) {
                     for (int nb : { 1, 2, 4, 8, }) {
-                    test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb));
+                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, kv, nb, max_bias));
+                    }
                 }
             }
         }

tests/test-tokenizer-1-bpe.cpp

@@ -13,15 +13,27 @@
 #include <vector>
 
 int main(int argc, char **argv) {
-    if (argc < 2) {
+    if (argc < 2 || argc > 3) {
-        fprintf(stderr, "Usage: %s <vocab-file>\n", argv[0]);
+        fprintf(stderr, "Usage: %s <vocab-file> [--ignore-merges]\n", argv[0]);
         return 1;
     }
 
     const std::string fname = argv[1];
+    bool ignore_merges = false;
+    if (argc == 3) {
+        if (std::strcmp(argv[2], "--ignore-merges") != 0) {
+            fprintf(stderr, "Usage: %s <vocab-file> [--ignore-merges]\n", argv[0]);
+            return 1;
+        }
+        ignore_merges = true;
+    }
 
     fprintf(stderr, "%s : reading vocab from: '%s'\n", __func__, fname.c_str());
 
+    if (ignore_merges) {
+        fprintf(stderr, "%s : ignoring merges for tokens inside vocab\n", __func__);
+    }
+
     llama_model * model;
     llama_context * ctx;
 
@@ -65,7 +77,19 @@ int main(int argc, char **argv) {
         std::string str = llama_detokenize_bpe(ctx, std::vector<int>(1, i));
         try {
             auto cps = unicode_cpts_from_utf8(str);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false, true);
+            if (ignore_merges && tokens.size() > 1) {
+                fprintf(stderr,
+                        "%s : error: token %d detokenizes to '%s'(%zu) but "
+                        "tokenization of this to multiple tokens: [",
+                        __func__, i, str.c_str(), str.length());
+                fprintf(stderr, "%d", tokens[0]);
+                for (size_t i = 1; i < tokens.size(); i++) {
+                    fprintf(stderr, ", %d", tokens[i]);
+                }
+                fprintf(stderr, "]\n");
+                return 2;
+            }
             std::string check = llama_detokenize_bpe(ctx, tokens);
             if (check != str) {
                 fprintf(stderr, "%s : error: token %d detokenizes to '%s'(%zu) but tokenization of this detokenizes to '%s'(%zu)\n",

tests/test-tokenizer-random.py

@@ -0,0 +1,295 @@
+# Test libllama tokenizer == AutoTokenizer.
+# Brute force random tokens/text generation.
+#
+# Sample usage:
+#
+#   python3 tests/test-tokenizer-random.py ./models/ggml-vocab-llama-bpe.gguf ./models/tokenizers/llama-bpe
+#
+
+import time
+import logging
+import argparse
+import subprocess
+import random
+
+from typing import Iterator
+
+import cffi
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
+
+logger = logging.getLogger("test-tokenizer-random-bpe")
+
+
+class LibLlama:
+
+    DEFAULT_PATH_LLAMA_H = "./llama.h"
+    DEFAULT_PATH_LIBLLAMA = "./build/libllama.so"  # CMakeLists.txt: BUILD_SHARED_LIBS ON
+
+    def __init__(self, path_llama_h: str = None, path_libllama: str = None):
+        path_llama_h = path_llama_h or self.DEFAULT_PATH_LLAMA_H
+        path_libllama = path_libllama or self.DEFAULT_PATH_LIBLLAMA
+        (self.ffi, self.lib) = self._load_libllama_cffi(path_llama_h, path_libllama)
+        self.lib.llama_backend_init()
+
+    def _load_libllama_cffi(self, path_llama_h: str, path_libllama: str):
+        cmd = ["gcc", "-E", "-P", "-D__restrict=", "-D__attribute__(x)=", "-D__asm__(x)=", path_llama_h]
+        res = subprocess.run(cmd, stdout=subprocess.PIPE)
+        assert (res.returncode == 0)
+        source = res.stdout.decode()
+        ffi = cffi.FFI()
+        if True:  # workarounds for pycparser
+            source = "typedef struct { } __builtin_va_list;" + "\n" + source
+            source = source.replace("sizeof (int)",    str(ffi.sizeof("int")))
+            source = source.replace("sizeof (void *)", str(ffi.sizeof("void*")))
+            source = source.replace("sizeof (size_t)", str(ffi.sizeof("size_t")))
+            source = source.replace("sizeof(int32_t)", str(ffi.sizeof("int32_t")))
+        ffi.cdef(source, override=True)
+        lib = ffi.dlopen(path_libllama)
+        return (ffi, lib)
+
+    def model_default_params(self, **kwargs):
+        mparams = self.lib.llama_model_default_params()
+        for k, v in kwargs.items():
+            setattr(mparams, k, v)
+        return mparams
+
+    def context_default_params(self, **kwargs):
+        cparams = self.lib.llama_context_default_params()
+        for k, v in kwargs.items():
+            setattr(cparams, k, v)
+        return cparams
+
+
+class LibLlamaModel:
+
+    def __init__(self, libllama: LibLlama, path_model: str, mparams={}, cparams={}):
+        self.lib = libllama.lib
+        self.ffi = libllama.ffi
+        if isinstance(mparams, dict):
+            mparams = libllama.model_default_params(**mparams)
+        self.model = self.lib.llama_load_model_from_file(path_model.encode(), mparams)
+        if not self.model:
+            raise RuntimeError("error: failed to load model '%s'" % path_model)
+        if isinstance(cparams, dict):
+            cparams = libllama.context_default_params(**cparams)
+        self.ctx = self.lib.llama_new_context_with_model(self.model, cparams)
+        if not self.ctx:
+            raise RuntimeError("error: failed to create context for model '%s'" % path_model)
+        n_tokens_max = self.lib.llama_n_ctx(self.ctx)
+        self.token_ids = self.ffi.new("llama_token[]", n_tokens_max)
+
+    def free(self):
+        if self.ctx:
+            self.lib.llama_free(self.ctx)
+        if self.model:
+            self.lib.llama_free_model(self.model)
+        self.ctx = None
+        self.model = None
+        self.lib = None
+
+    def tokenize(self, text: str, n_tokens_max: int = 0, add_special: bool = False, parse_special: bool = False) -> list[int]:
+        n_tokens_max = n_tokens_max if n_tokens_max > 0 else len(self.token_ids)
+        text = text.encode("utf-8")
+        num = self.lib.llama_tokenize(self.model, text, len(text), self.token_ids, n_tokens_max, add_special, parse_special)
+        if num < 0:
+            return []
+        return list(self.token_ids[0:num])
+
+
+def generator_custom_text() -> Iterator[str]:
+    """General tests"""
+    yield from [
+        "",
+        " ",
+        "  ",
+        "   ",
+        "\t",
+        "\n",
+        "\n\n",
+        "\n\n\n",
+        "\t\n",
+        "Hello world",
+        " Hello world",
+        "Hello World",
+        " Hello World",
+        " Hello World!",
+        "Hello, world!",
+        " Hello, world!",
+        " this is 🦙.cpp",
+        "w048 7tuijk dsdfhu",
+        "нещо на Български",
+        "កាន់តែពិសេសអាចខលចេញ",
+        "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+        "Hello",
+        " Hello",
+        "  Hello",
+        "   Hello",
+        "    Hello",
+        "    Hello\n    Hello",
+        " (",
+        "\n =",
+        "' era",
+        "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
+        "3",
+        "33",
+        "333",
+        "3333",
+        "33333",
+        "333333",
+        "3333333",
+        "33333333",
+        "333333333",
+    ]
+
+
+def generator_custom_text_edge_cases() -> Iterator[str]:
+    """Edge cases found while debugging"""
+    yield from [
+        '\x1f-a',   # unicode_ranges_control, {0x00001C, 0x00001F}
+        '¼-a',      # unicode_ranges_digit, 0x00BC
+        '½-a',      # unicode_ranges_digit, 0x00BD
+        '¾-a',      # unicode_ranges_digit, 0x00BE
+        'a 〇b',    # unicode_ranges_digit, 0x3007
+        'Ⅵ-a',     # unicode_ranges_digit, {0x00002150, 0x0000218F} // Number Forms
+        '\uFEFF//', # unicode_ranges_control, 0xFEFF (BOM)
+        '<s>a'      # TODO: Phi-3 fail
+    ]
+
+
+def generator_random_chars(iterations = 100) -> Iterator[str]:
+    """Brute force random text with simple characters"""
+
+    WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
+    CHARS = list(set("""
+        ABCDEFGHIJKLMNOPQRSTUVWXYZ
+        abcdefghijklmnopqrstuvwxyz
+        ÁÉÍÓÚÀÈÌÒÙÂÊÎÔÛÄËÏÖÜ
+        áéíóúàèìòùâêîôûäëïöü
+        .-,*/-+ª!"·$%&/()=?¿[]{}<>\\|@#~½¬~;:_
+    """))
+
+    rand = random.Random()
+    for m in range(iterations):
+        rand.seed(m)
+        text = []
+        num_words = rand.randint(300, 400)
+        for i in range(num_words):
+            k = rand.randint(1, 7)
+            word = rand.choices(CHARS, k=k)
+            space = rand.choice(WHITESPACES)
+            text.append("".join(word) + space)
+        yield "".join(text)
+
+
+def generator_random_vocab_chars(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
+    """Brute force random text with vocab characters"""
+
+    vocab_ids = list(tokenizer.vocab.values())
+    vocab_text = tokenizer.decode(vocab_ids, skip_special_tokens=True)
+    vocab_chars = list(set(vocab_text))
+    del vocab_ids, vocab_text
+
+    rand = random.Random()
+    for m in range(iterations):
+        rand.seed(m)
+        text = rand.choices(vocab_chars, k=1024)
+        yield "".join(text)
+
+
+def generator_random_vocab_tokens(tokenizer: PreTrainedTokenizerBase, iterations = 100) -> Iterator[str]:
+    """Brute force random text from vocab tokens"""
+
+    space_id = tokenizer.encode(" ", add_special_tokens=False)[0]
+    vocab_ids = list(tokenizer.vocab.values())
+    vocab_ids = list(sorted(vocab_ids + vocab_ids))
+    for i in range(1, len(vocab_ids), 2):
+        vocab_ids[i] = space_id
+    vocab_tokens = tokenizer.decode(vocab_ids, skip_special_tokens=True)
+    vocab_tokens = vocab_tokens.split(" ")
+    del vocab_ids
+
+    yield from vocab_tokens
+
+    rand = random.Random()
+    for m in range(iterations):
+        rand.seed(m)
+        text = []
+        num_words = rand.randint(300, 400)
+        for i in range(num_words):
+            k = rand.randint(1, 3)
+            tokens = rand.choices(vocab_tokens, k=k)
+            tokens = [t.strip(" \n\r\t") for t in tokens]
+            sep = rand.choice("     \n\r\t")
+            text.append("".join(tokens) + sep)
+        yield "".join(text)
+
+
+def generator_random_bytes(iterations = 100) -> Iterator[str]:
+    """Brute force random bytes"""
+
+    WHITESPACES = list(" " * 20 + "\n" * 5 + "\r\n" * 5 + "\t" * 5)
+
+    rand = random.Random()
+    for m in range(iterations):
+        rand.seed(m)
+        text = []
+        num_words = rand.randint(300, 400)
+        for i in range(num_words):
+            k = rand.randint(1, 8)
+            word = [chr(r) for r in rand.randbytes(k) if r]
+            word.append(rand.choice(WHITESPACES))
+            text.append("".join(word))
+        yield "".join(text)
+
+
+def test_compare_tokenizer(model: LibLlamaModel, tokenizer: PreTrainedTokenizerBase, generator: Iterator[str]):
+
+    def find_first_mismatch(ids1: list[int], ids2: list[int]):
+        for i, (a,b) in enumerate(zip(ids1, ids2)):
+            if a != b:
+                return i
+        if len(ids1) == len(ids2):
+            return -1
+        return min(len(ids1), len(ids2))
+
+    t0 = time.perf_counter()
+    logger.info("%s: %s" % (generator.__name__, "ini"))
+    for text in generator:
+        ids1 = model.tokenize(text, add_special=False, parse_special=False)
+        ids2 = tokenizer.encode(text, add_special_tokens=False)
+        if ids1 != ids2:
+            i = find_first_mismatch(ids1, ids2)
+            ids1 = list(ids1)[max(0, i - 2) : i + 2 + 1]
+            ids2 = list(ids2)[max(0, i - 2) : i + 2 + 1]
+            text2 = tokenizer.decode(ids2, skip_special_tokens=True)
+            assert (text2 in text)
+            logger.info(" Text:     " + repr(text2))
+            logger.info(" TokenIDs: " + str(ids1))
+            logger.info(" Expected: " + str(ids2))
+            raise Exception()
+    t1 = time.perf_counter()
+    logger.info("%s: end, time: %.3f secs" % (generator.__name__, t1 - t0))
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("vocab_file", help="path to vocab 'gguf' file")
+    parser.add_argument("dir_tokenizer", help="directory containing 'tokenizer.model' file")
+    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+
+    model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=2048))
+
+    tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
+
+    test_compare_tokenizer(model, tokenizer, generator_custom_text())
+    test_compare_tokenizer(model, tokenizer, generator_custom_text_edge_cases())
+    test_compare_tokenizer(model, tokenizer, generator_random_chars(10_000))
+    test_compare_tokenizer(model, tokenizer, generator_random_vocab_chars(tokenizer, 10_000))
+    test_compare_tokenizer(model, tokenizer, generator_random_vocab_tokens(tokenizer, 10_000))
+    # test_compare_tokenizer(model, tokenizer, generator_random_bytes(10_000)) # FAIL
+
+    model.free()

unicode-data.h

@@ -7,6 +7,7 @@
 
 extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_number;
 extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_letter;
+extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_separator;
 extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_whitespace;
 extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_accent_mark;
 extern const std::vector<std::pair<uint32_t, uint32_t>> unicode_ranges_punctuation;

unicode.cpp

@@ -9,6 +9,7 @@
 #include <stdexcept>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <utility>
 #include <vector>
 #include <locale>
@@ -120,9 +121,9 @@ static std::unordered_map<uint32_t, int> unicode_cpt_type_map() {
             cpt_types[i] = CODEPOINT_TYPE_LETTER;
         }
     }
-    for (auto p : unicode_ranges_whitespace) {
+    for (auto p : unicode_ranges_separator) {
         for (auto i = p.first; i <= p.second; ++i) {
-            cpt_types[i] = CODEPOINT_TYPE_WHITESPACE;
+            cpt_types[i] = CODEPOINT_TYPE_SEPARATOR;
         }
     }
     for (auto p : unicode_ranges_accent_mark) {
@@ -224,138 +225,256 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
 
+    const auto cpts = unicode_cpts_from_utf8(text);
+
     size_t start = 0;
+    for (auto offset : offsets) {
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
+
+        auto _get_cpt = [&] (const size_t pos) -> char32_t {
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
+        };
+
+        auto _get_cpt_type = [&] (const size_t pos) -> int {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        };
+
+        size_t _prev_end = offset_ini;
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
+            }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
+
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
+            const char32_t cpt = _get_cpt(pos);
+            const int cpt_type = _get_cpt_type(pos);
+
+            // regex: 's|'t|'re|'ve|'m|'ll|'d
+            if (cpt == '\'' && pos+1 < offset_end) {
+                char32_t cpt_next = _get_cpt(pos+1);
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
+                    pos += _add_token(pos+2);
+                    continue;
+                }
+                if (pos+2 < offset_end) {
+                    char32_t cpt_next_next = _get_cpt(pos+2);
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
+                        pos += _add_token(pos+3);
+                        continue;
+                    }
+                }
+            }
+
+            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
+            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            // regex: <space>?\p{L}+
+            if (cpt2_type == CODEPOINT_TYPE_LETTER) {
+                pos += (cpt == ' ');
+                while (cpt2_type == CODEPOINT_TYPE_LETTER) {
+                    cpt2_type = _get_cpt_type(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: <space>?\p{N}+
+            if (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+                pos += (cpt == ' ');
+                while (cpt2_type == CODEPOINT_TYPE_NUMBER) {
+                    cpt2_type = _get_cpt_type(++pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+            // regex: <space>?[^\s\p{L}\p{N}]+
+            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                pos += (cpt == ' ');
+                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                    cpt2_type = _get_cpt_type(++pos);
+                    cpt2 = _get_cpt(pos);
+                }
+                _add_token(pos);
+                continue;
+            }
+
+            size_t num_whitespaces = 0;
+            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
+                num_whitespaces++;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
+    }
+
+    return bpe_offsets;
+}
+
+// LLAMA3 system regex: "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"
+static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string & text, const std::vector<size_t> & offsets) {
+    std::vector<size_t> bpe_offsets; // store the offset of each word
+    bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
 
     const auto cpts = unicode_cpts_from_utf8(text);
 
+    size_t start = 0;
     for (auto offset : offsets) {
-        std::string token;
+        const size_t offset_ini = start;
+        const size_t offset_end = start + offset;
+        assert(offset_end <= cpts.size());
+        start = offset_end;
 
-        bool collecting_numeric = false;
+        auto _get_cpt = [&] (const size_t pos) -> char32_t {
-        bool collecting_letter = false;
+            return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : 0;
-        bool collecting_special = false;
+        };
-        bool collecting_whitespace_lookahead = false;
-        bool collecting = false;
 
-        std::vector<std::string> text_utf;
+        auto _get_cpt_type = [&] (const size_t pos) -> int {
-        text_utf.reserve(offset);
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_type(cpts[pos]) : CODEPOINT_TYPE_UNIDENTIFIED;
+        };
 
-        for (size_t i = start; i < start + offset; ++i) {
+        size_t _prev_end = offset_ini;
-            text_utf.emplace_back(unicode_cpt_to_utf8(cpts[i]));
+        auto _add_token = [&] (const size_t end) -> size_t {
+            assert(_prev_end <= end && end <= offset_end);
+            size_t len = end - _prev_end;
+            if (len > 0) {
+                bpe_offsets.push_back(len);
             }
+            _prev_end = end;
+            //if (len > 0) {
+            //    std::string s = "";
+            //    for(size_t p = end-len; p < end; p++)
+            //        s += unicode_cpt_to_utf8(cpts[p]);
+            //    printf(">>> '%s'\n", s.c_str());
+            //}
+            return len;
+        };
 
-        for (int i = 0; i < (int)text_utf.size(); i++) {
+        for (size_t pos = offset_ini; pos < offset_end; /*pos++*/ ) {
-            const std::string & utf_char = text_utf[i];
+            const char32_t cpt = _get_cpt(pos);
-            bool split_condition = false;
+            const int cpt_type = _get_cpt_type(pos);
-            int bytes_remain = text_utf.size() - i;
 
-            // forward backward lookups
+            // regex: (?i:'s|'t|'re|'ve|'m|'ll|'d) // case insensitive
-            const std::string & utf_char_next      = (i + 1 < (int)text_utf.size()) ? text_utf[i + 1] : "";
+            if (cpt == '\'' && pos+1 < offset_end) {
-            const std::string & utf_char_next_next = (i + 2 < (int)text_utf.size()) ? text_utf[i + 2] : "";
+                char32_t cpt_next = unicode_tolower(_get_cpt(pos+1));
-
+                if (cpt_next == 's' || cpt_next == 't' || cpt_next == 'm' || cpt_next == 'd') {
-            // handling contractions
+                    pos += _add_token(pos+2);
-            if (!split_condition && bytes_remain >= 2) {
+                    continue;
-                // 's|'t|'m|'d
-                if (utf_char == "\'" && (utf_char_next == "s" || utf_char_next == "t" || utf_char_next == "m" || utf_char_next == "d")) {
-                    split_condition = true;
                 }
-                if (split_condition) {
+                if (pos+2 < offset_end) {
-                    if (token.size()) {
+                    char32_t cpt_next_next = unicode_tolower(_get_cpt(pos+2));
-                        bpe_offsets.emplace_back(unicode_cpts_from_utf8(token).size());
+                    if ((cpt_next == 'r' && cpt_next_next == 'e') ||
-                    }
+                        (cpt_next == 'v' && cpt_next_next == 'e') ||
-                    token = utf_char + utf_char_next;
+                        (cpt_next == 'l' && cpt_next_next == 'l')) {
-                    bpe_offsets.emplace_back(unicode_cpts_from_utf8(token).size());
+                        pos += _add_token(pos+3);
-                    token = "";
-                    i++;
                         continue;
                     }
                 }
-            if (!split_condition && bytes_remain >= 3) {
-                // 're|'ve|'ll
-                if (utf_char == "\'" && (
-                    (utf_char_next == "r" && utf_char_next_next == "e") ||
-                    (utf_char_next == "v" && utf_char_next_next == "e") ||
-                    (utf_char_next == "l" && utf_char_next_next == "l"))
-                    ) {
-                    split_condition = true;
             }
-                if (split_condition) {
-                    // current token + next token can be defined
-                    if (token.size()) {
-                        bpe_offsets.emplace_back(unicode_cpts_from_utf8(token).size());
-                    }
-                    token =  utf_char;
-                    token += utf_char_next;
-                    token += utf_char_next_next;
 
-                    bpe_offsets.emplace_back(unicode_cpts_from_utf8(token).size());
+            // regex: [^\r\n\p{L}\p{N}]?\p{L}+  //####FIXME: the first \p{L} is correct?
-                    token = "";
+            if (cpt != '\r' && cpt != '\n' && /*cpt_type != CODEPOINT_TYPE_LETTER &&*/ cpt_type != CODEPOINT_TYPE_NUMBER) {
-                    i += 2;
+                if (cpt_type == CODEPOINT_TYPE_LETTER || _get_cpt_type(pos+1) == CODEPOINT_TYPE_LETTER) {  // one or more letters
+                    pos++;
+                    while (_get_cpt_type(pos) == CODEPOINT_TYPE_LETTER) {
+                        pos++;
+                    }
+                    _add_token(pos);
                     continue;
                 }
             }
 
-            if (!split_condition && !collecting) {
+            // regex: \p{N}{1,3}
-                if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || (token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER)) {
+            if (cpt_type == CODEPOINT_TYPE_NUMBER) {
-                    collecting_letter = true;
+                size_t ini = pos;
-                    collecting = true;
+                while (_get_cpt_type(pos) == CODEPOINT_TYPE_NUMBER) {
-                }
+                    if (++pos - ini >= 3 ) {
-                else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_NUMBER || (token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_NUMBER)) {
+                        _add_token(pos);
-                    collecting_numeric = true;
+                        ini = pos;
-                    collecting = true;
-                }
-                else if (
-                    ((unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_NUMBER) && (unicode_cpt_type(utf_char) != CODEPOINT_TYPE_WHITESPACE)) ||
-                    (token.empty() && utf_char == " " && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_LETTER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_NUMBER && unicode_cpt_type(utf_char_next) != CODEPOINT_TYPE_WHITESPACE)
-                    ) {
-                    collecting_special = true;
-                    collecting = true;
-                }
-                else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE && unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_WHITESPACE) {
-                    collecting_whitespace_lookahead = true;
-                    collecting = true;
-                }
-                else if (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE) {
-                    split_condition = true;
                     }
                 }
-            else if (!split_condition && collecting) {
+                _add_token(pos);
-                if (collecting_letter && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_LETTER) {
+                continue;
-                    split_condition = true;
-                }
-                else if (collecting_numeric && unicode_cpt_type(utf_char) != CODEPOINT_TYPE_NUMBER) {
-                    split_condition = true;
-                }
-                else if (collecting_special && (unicode_cpt_type(utf_char) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_NUMBER || unicode_cpt_type(utf_char) == CODEPOINT_TYPE_WHITESPACE)) {
-                    split_condition = true;
-                }
-                else if (collecting_whitespace_lookahead && (unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_LETTER || unicode_cpt_type(utf_char_next) == CODEPOINT_TYPE_NUMBER)) {
-                    split_condition = true;
-                }
             }
 
-            if (utf_char_next == "") {
+            // regex: <space>?[^\s\p{L}\p{N}]+[\r\n]*
-                split_condition = true; // final
+            char32_t cpt2 = (cpt == ' ' ? _get_cpt(pos+1) : cpt);
-                token += utf_char;
+            int cpt2_type = (cpt == ' ' ? _get_cpt_type(pos+1) : cpt_type);
+            if (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                pos += (cpt == ' ');
+                while (!unicode_cpt_is_whitespace(cpt2) && cpt2_type != CODEPOINT_TYPE_LETTER && cpt2_type != CODEPOINT_TYPE_NUMBER && cpt2_type != CODEPOINT_TYPE_UNIDENTIFIED) {
+                    cpt2_type = _get_cpt_type(++pos);
+                    cpt2 = _get_cpt(pos);
+                }
+                while (cpt2 == '\r' || cpt2 == '\n') {
+                    cpt2 = _get_cpt(++pos);
+                }
+                _add_token(pos);
+                continue;
             }
 
-            if (split_condition) {
+            size_t num_whitespaces = 0;
-                if (token.size()) {
+            size_t last_end_r_or_n = 0;
-                    bpe_offsets.emplace_back(unicode_cpts_from_utf8(token).size());
+            while (unicode_cpt_is_whitespace(_get_cpt(pos+num_whitespaces))) {
-                }
+                char32_t cpt2 = _get_cpt(pos+num_whitespaces);
-                token = utf_char;
+                if (cpt2 == '\r' || cpt2 == '\n') {
-                collecting = false;
+                    last_end_r_or_n = pos + num_whitespaces + 1;
-                collecting_letter = false;
-                collecting_numeric = false;
-                collecting_special = false;
-                collecting_whitespace_lookahead = false;
-            }
-            else {
-                token += utf_char;
                 }
+                num_whitespaces++;
             }
 
-        start += offset;
+            // regex: \s*[\r\n]+
+            if (last_end_r_or_n > 0) {
+                pos = last_end_r_or_n;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+(?!\S)
+            if (num_whitespaces > 1 && _get_cpt(pos+num_whitespaces) != 0) {
+                pos += num_whitespaces - 1;
+                _add_token(pos);
+                continue;
+            }
+
+            // regex: \s+
+            if (num_whitespaces > 0) {
+                pos += num_whitespaces;
+                _add_token(pos);
+                continue;
+            }
+
+            // no matches
+            _add_token(++pos);
+        }
     }
 
     return bpe_offsets;
@@ -424,14 +543,14 @@ static std::vector<size_t> unicode_regex_split_stl(const std::string & text, con
 static std::vector<size_t> unicode_regex_split_custom(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
     std::vector<size_t> bpe_offsets;
 
-    (void)(text);
+    if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
-    (void)(regex_expr);
+        bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
-    (void)(offsets);
+    } else if (
-    // TODO: this implementation is actually wrong, uncomment and run:
+            regex_expr == "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" ||
-    //       make -j && ./bin/test-tokenizer-0 ../models/ggml-vocab-gpt-2.gguf
+            regex_expr == "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+") {
-    //if (regex_expr == "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)") {
+
-    //    bpe_offsets = unicode_regex_split_custom_gpt2(text, offsets);
+        bpe_offsets = unicode_regex_split_custom_llama3(text, offsets);
-    //}
+    }
 
     return bpe_offsets;
 }
@@ -506,6 +625,19 @@ int unicode_cpt_type(const std::string & utf8) {
     return unicode_cpt_type(unicode_cpt_from_utf8(utf8, offset));
 }
 
+bool unicode_cpt_is_whitespace(uint32_t cp) {
+    static const std::unordered_set<uint32_t> is_whitespace = [] {
+        std::unordered_set<uint32_t> is_whitespace;
+        for (auto p : unicode_ranges_whitespace) {
+            for (auto i = p.first; i <= p.second; ++i) {
+                is_whitespace.insert(i);
+            }
+        }
+        return is_whitespace;
+    }();
+    return (bool)is_whitespace.count(cp);
+}
+
 std::string unicode_byte_to_utf8(uint8_t byte) {
     static std::unordered_map<uint8_t, std::string> map = unicode_byte_to_utf8_map();
     return map.at(byte);

unicode.h

@@ -7,7 +7,7 @@
 #define CODEPOINT_TYPE_UNIDENTIFIED 0
 #define CODEPOINT_TYPE_NUMBER       1
 #define CODEPOINT_TYPE_LETTER       2
-#define CODEPOINT_TYPE_WHITESPACE   3
+#define CODEPOINT_TYPE_SEPARATOR    3
 #define CODEPOINT_TYPE_ACCENT_MARK  4
 #define CODEPOINT_TYPE_PUNCTUATION  5
 #define CODEPOINT_TYPE_SYMBOL       6
@@ -21,6 +21,8 @@ std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & c
 int unicode_cpt_type(uint32_t cp);
 int unicode_cpt_type(const std::string & utf8);
 
+bool unicode_cpt_is_whitespace(uint32_t cp);
+
 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);