[SYCL] fix UT fault cases: count-equal, argsort, pad OPs (#16521)

* fix/refactor OP argsort, pad

* fix count-equal op

* update SYCL OP list

* fix format issue

---------

Co-authored-by: Zhang Jianyu <zhang.jianyu@outlook.com>

.github/workflows/build.yml

@@ -387,6 +387,39 @@ jobs:
           cd build
           ctest -L main --verbose
 
+  ubuntu-24-cmake-vulkan-deb:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-24-cmake-vulkan-deb
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get install -y glslc libvulkan-dev libcurl4-openssl-dev
+
+      - name: Configure
+        id: cmake_configure
+        run: |
+          cmake -B build \
+            -DCMAKE_BUILD_TYPE=RelWithDebInfo \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGGML_VULKAN=ON
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake --build build -j $(nproc)
+
   ubuntu-24-cmake-vulkan:
     runs-on: ubuntu-24.04
 
@@ -444,8 +477,8 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 4200
 
-  ubuntu-22-cmake-webgpu:
-    runs-on: ubuntu-22.04
+  ubuntu-24-cmake-webgpu:
+    runs-on: ubuntu-24.04
 
     steps:
       - name: Clone
@@ -455,16 +488,34 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
-          key: ubuntu-22-cmake-webgpu
+          key: ubuntu-24-cmake-webgpu
           evict-old-files: 1d
 
-      - name: Vulkan SDK Dependencies
-        id: vulkan-depends
+      - name: Dependencies
+        id: depends
         run: |
-          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
-          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+          sudo add-apt-repository -y ppa:kisak/kisak-mesa
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libcurl4-openssl-dev
+
+      - name: Get latest Vulkan SDK version
+        id: vulkan_sdk_version
+        run: |
+          echo "VULKAN_SDK_VERSION=$(curl https://vulkan.lunarg.com/sdk/latest/linux.txt)" >> "$GITHUB_ENV"
+
+      - name: Use Vulkan SDK Cache
+        uses: actions/cache@v4
+        id: cache-sdk
+        with:
+          path: ./vulkan_sdk
+          key: vulkan-sdk-${{ env.VULKAN_SDK_VERSION }}-${{ runner.os }}
+
+      - name: Setup Vulkan SDK
+        if: steps.cache-sdk.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-vulkan
+        with:
+          path: ./vulkan_sdk
+          version: ${{ env.VULKAN_SDK_VERSION }}
 
       - name: Dawn Dependency
         id: dawn-depends
@@ -1497,3 +1548,29 @@ jobs:
         run: |
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
+  ggml-ci-arm64-cpu-kleidiai:
+     runs-on: ubuntu-22.04-arm
+
+     steps:
+       - name: Clone
+         id: checkout
+         uses: actions/checkout@v4
+
+       - name: ccache
+         uses: ggml-org/ccache-action@v1.2.16
+         with:
+           key: ggml-ci-arm64-cpu-kleidiai
+           evict-old-files: 1d
+
+       - name: Dependencies
+         id: depends
+         run: |
+           sudo apt-get update
+           sudo apt-get install -y build-essential libcurl4-openssl-dev
+
+       - name: Test
+         id: ggml-ci
+         run: |
+           GG_BUILD_KLEIDIAI=1 GG_BUILD_EXTRA_TESTS_0=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+

CODEOWNERS

@@ -70,6 +70,7 @@
 /ggml/src/ggml-rpc/                     @rgerganov
 /ggml/src/ggml-threading.*              @ggerganov @slaren
 /ggml/src/ggml-vulkan/                  @0cc4m
+/ggml/src/ggml-webgpu/                  @reeselevine
 /ggml/src/ggml-zdnn/                    @taronaeo @Andreas-Krebbel @AlekseiNikiforovIBM
 /ggml/src/ggml.c                        @ggerganov @slaren
 /ggml/src/ggml.cpp                      @ggerganov @slaren

ci/run.sh

@@ -22,6 +22,9 @@
 # # with MUSA support
 # GG_BUILD_MUSA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
+# # with KLEIDIAI support
+# GG_BUILD_KLEIDIAI=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -115,6 +118,34 @@ if [ ! -z ${GG_BUILD_NO_SVE} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm"
 fi
 
+if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
+    echo ">>===== Enabling KleidiAI support"
+
+    CANDIDATES=("armv9-a+dotprod+i8mm" "armv8.6-a+dotprod+i8mm" "armv8.2-a+dotprod")
+    CPU=""
+
+    for cpu in "${CANDIDATES[@]}"; do
+        if echo 'int main(){}' | ${CXX:-c++} -march="$cpu" -x c++ - -c -o /dev/null >/dev/null 2>&1; then
+            CPU="$cpu"
+            break
+        fi
+    done
+
+    if [ -z "$CPU" ]; then
+        echo "ERROR: None of the required ARM baselines (armv9/armv8.6/armv8.2 + dotprod) are supported by this compiler."
+        exit 1
+    fi
+
+    echo ">>===== Using ARM baseline: ${CPU}"
+
+    CMAKE_EXTRA="${CMAKE_EXTRA:+$CMAKE_EXTRA } \
+        -DGGML_NATIVE=OFF \
+        -DGGML_CPU_KLEIDIAI=ON \
+        -DGGML_CPU_AARCH64=ON \
+        -DGGML_CPU_ARM_ARCH=${CPU} \
+        -DBUILD_SHARED_LIBS=OFF"
+fi
+
 ## helpers
 
 # download a file if it does not exist or if it is outdated

common/arg.cpp

@@ -1935,6 +1935,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_ctx_checkpoints = value;
         }
     ).set_env("LLAMA_ARG_CTX_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
+    add_opt(common_arg(
+        {"--cache-ram", "-cram"}, "N",
+        string_format("set the maximum cache size in MiB (default: %d, -1 - no limit, 0 - disable)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/16391)", params.cache_ram_mib),
+        [](common_params & params, int value) {
+            params.cache_ram_mib = value;
+        }
+    ).set_env("LLAMA_ARG_CACHE_RAM").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
         string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@@ -3350,7 +3358,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--chat-template-kwargs"}, "STRING",
         string_format("sets additional params for the json template parser"),
-        [](common_params & params, const std::string &  value) {
+        [](common_params & params, const std::string & value) {
             auto parsed = json::parse(value);
             for (const auto & item : parsed.items()) {
                 params.default_template_kwargs[item.key()] = item.value().dump();
@@ -3432,7 +3440,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--reasoning-format"}, "FORMAT",
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
         "- none: leaves thoughts unparsed in `message.content`\n"
-        "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
+        "- deepseek: puts thoughts in `message.reasoning_content`\n"
+        "- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`\n"
         "(default: auto)",
         [](common_params & params, const std::string & value) {
             params.reasoning_format = common_reasoning_format_from_name(value);
@@ -3561,21 +3570,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             common_log_set_file(common_log_main(), value.c_str());
         }
     ));
-    add_opt(common_arg({ "--log-colors" }, "[on|off|auto]",
-                       "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
-                       "'auto' enables colors when output is to a terminal",
-                       [](common_params &, const std::string & value) {
-                           if (is_truthy(value)) {
-                               common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
-                           } else if (is_falsey(value)) {
-                               common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
-                           } else if (is_autoy(value)) {
-                               common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
-                           } else {
-                               throw std::invalid_argument(
-                                   string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
-                           }
-                       }).set_env("LLAMA_LOG_COLORS"));
+    add_opt(common_arg(
+        {"--log-colors"}, "[on|off|auto]",
+        "Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
+        "'auto' enables colors when output is to a terminal",
+        [](common_params &, const std::string & value) {
+            if (is_truthy(value)) {
+                common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
+            } else if (is_falsey(value)) {
+                common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
+            } else if (is_autoy(value)) {
+                common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
+            } else {
+                throw std::invalid_argument(
+                    string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
+            }
+        }
+    ).set_env("LLAMA_LOG_COLORS"));
     add_opt(common_arg(
         {"-v", "--verbose", "--log-verbose"},
         "Set verbosity level to infinity (i.e. log all messages, useful for debugging)",
@@ -3841,7 +3852,87 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
-    // model-specific
+    add_opt(common_arg(
+        {"--diffusion-steps"}, "N",
+        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
+        [](common_params & params, int value) { params.diffusion.steps = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-visual"},
+        string_format("enable visual diffusion mode (show progressive generation) (default: %s)", params.diffusion.visual_mode ? "true" : "false"),
+        [](common_params & params) { params.diffusion.visual_mode = true; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-eps"}, "F",
+        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
+        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-algorithm"}, "N",
+        string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)", params.diffusion.algorithm),
+        [](common_params & params, int value) { params.diffusion.algorithm = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-alg-temp"}, "F",
+        string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
+        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-block-length"}, "N",
+        string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
+        [](common_params & params, int value) { params.diffusion.block_length = value; }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-cfg-scale"}, "F",
+        string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
+        [](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        {"--diffusion-add-gumbel-noise"}, "F",
+        string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
+        [](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+    add_opt(common_arg(
+        { "-lr", "--learning-rate" }, "ALPHA",
+        string_format("adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)", (double) params.lr.lr0),
+        [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
+        string_format("(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
+            (double) params.lr.lr_min),
+        [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"-decay-epochs", "--learning-rate-decay-epochs"}, "ALPHA",
+        string_format("(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)", (double) params.lr.decay_epochs),
+        [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"-wd", "--weight-decay"}, "WD",
+        string_format("adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).", (double) params.lr.wd),
+        [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"-val-split", "--val-split"}, "FRACTION",
+        string_format("fraction of data to use as validation set for training (default: %.2g).", (double) params.val_split),
+        [](common_params & params, const std::string & value) { params.val_split = std::stof(value); }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"-epochs", "--epochs"}, "N",
+        string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
+        [](common_params & params, int epochs) { params.lr.epochs = epochs; }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"-opt", "--optimizer"}, "sgd|adamw", "adamw or sgd",
+        [](common_params & params, const std::string & name) {
+            params.optimizer = common_opt_get_optimizer(name.c_str());
+            if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
+                throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
+            }
+        }
+    ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+
+    // presets
     add_opt(common_arg(
         {"--tts-oute-default"},
         string_format("use default OuteTTS models (note: can download weights from the internet)"),
@@ -3854,42 +3945,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_TTS}));
 
     add_opt(common_arg(
-        {"--embd-bge-small-en-default"},
-        string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
+        {"--embd-gemma-default"},
+        string_format("use default EmbeddingGemma model (note: can download weights from the internet)"),
         [](common_params & params) {
-            params.model.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
-            params.model.hf_file = "bge-small-en-v1.5-q8_0.gguf";
-            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
-            params.embd_normalize = 2;
-            params.n_ctx = 512;
-            params.verbose_prompt = true;
-            params.embedding = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
-
-    add_opt(common_arg(
-        {"--embd-e5-small-en-default"},
-        string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
-            params.model.hf_file = "e5-small-v2-q8_0.gguf";
-            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
-            params.embd_normalize = 2;
-            params.n_ctx = 512;
-            params.verbose_prompt = true;
-            params.embedding = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
-
-    add_opt(common_arg(
-        {"--embd-gte-small-default"},
-        string_format("use default gte-small model (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
-            params.model.hf_file = "gte-small-q8_0.gguf";
-            params.pooling_type = LLAMA_POOLING_TYPE_NONE;
-            params.embd_normalize = 2;
-            params.n_ctx = 512;
+            params.model.hf_repo = "ggml-org/embeddinggemma-300M-qat-q4_0-GGUF";
+            params.model.hf_file = "embeddinggemma-300M-qat-Q4_0.gguf";
+            params.port = 8011;
+            params.n_ubatch = 2048;
+            params.n_batch = 2048;
+            params.n_parallel = 32;
+            params.n_ctx = 2048*params.n_parallel;
             params.verbose_prompt = true;
             params.embedding = true;
         }
@@ -3984,96 +4049,65 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     add_opt(common_arg(
-        { "--diffusion-steps" }, "N",
-        string_format("number of diffusion steps (default: %d)", params.diffusion.steps),
-        [](common_params & params, int value) { params.diffusion.steps = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-visual" },
-        string_format("enable visual diffusion mode (show progressive generation) (default: %s)",
-                      params.diffusion.visual_mode ? "true" : "false"),
-        [](common_params & params) { params.diffusion.visual_mode = true; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+        {"--gpt-oss-20b-default"},
+        string_format("use gpt-oss-20b (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/gpt-oss-20b-GGUF";
+            params.model.hf_file = "gpt-oss-20b-mxfp4.gguf";
+            params.port = 8013;
+            params.n_ubatch = 2048;
+            params.n_batch = 32768;
+            params.n_parallel = 2;
+            params.n_ctx = 131072*params.n_parallel;
+            params.sampling.temp = 1.0f;
+            params.sampling.top_p = 1.0f;
+            params.sampling.top_k = 0;
+            params.sampling.min_p = 0.01f;
+            params.use_jinja = true;
+            //params.default_template_kwargs["reasoning_effort"] = "\"high\"";
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     add_opt(common_arg(
-        { "--diffusion-eps" }, "F",
-        string_format("epsilon for timesteps (default: %.6f)", (double) params.diffusion.eps),
-        [](common_params & params, const std::string & value) { params.diffusion.eps = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-algorithm" }, "N",
-        string_format("diffusion algorithm: 0=ORIGIN, 1=ENTROPY_BASED, 2=MARGIN_BASED, 3=RANDOM, 4=LOW_CONFIDENCE (default: %d)",
-                      params.diffusion.algorithm),
-        [](common_params & params, int value) { params.diffusion.algorithm = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-alg-temp" }, "F",
-        string_format("dream algorithm temperature (default: %.3f)", (double) params.diffusion.alg_temp),
-        [](common_params & params, const std::string & value) { params.diffusion.alg_temp = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+        {"--gpt-oss-120b-default"},
+        string_format("use gpt-oss-120b (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/gpt-oss-120b-GGUF";
+            params.port = 8013;
+            params.n_ubatch = 2048;
+            params.n_batch = 32768;
+            params.n_parallel = 2;
+            params.n_ctx = 131072*params.n_parallel;
+            params.sampling.temp = 1.0f;
+            params.sampling.top_p = 1.0f;
+            params.sampling.top_k = 0;
+            params.sampling.min_p = 0.01f;
+            params.use_jinja = true;
+            //params.default_template_kwargs["reasoning_effort"] = "\"high\"";
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     add_opt(common_arg(
-        { "--diffusion-block-length" }, "N",
-        string_format("llada block length for generation (default: %d)", params.diffusion.block_length),
-        [](common_params & params, int value) { params.diffusion.block_length = value; }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-cfg-scale" }, "F",
-        string_format("llada classifier-free guidance scale (default: %.3f)", (double) params.diffusion.cfg_scale),
-        [](common_params & params, const std::string & value) { params.diffusion.cfg_scale = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
-    add_opt(common_arg(
-        { "--diffusion-add-gumbel-noise" }, "F",
-        string_format("add gumbel noise to the logits if temp > 0.0 (default: %s)", params.diffusion.add_gumbel_noise ? "true" : "false"),
-        [](common_params & params, const std::string & value) { params.diffusion.add_gumbel_noise = std::stof(value); }
-    ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
+        {"--vision-gemma-4b-default"},
+        string_format("use Gemma 3 4B QAT (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/gemma-3-4b-it-qat-GGUF";
+            params.port = 8014;
+            params.n_ctx = 0;
+            params.use_jinja = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
-
-    add_opt(
-        common_arg({ "-lr", "--learning-rate" }, "ALPHA",
-                   string_format(
-                       "adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
-                       (double) params.lr.lr0),
-                   [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(
-        common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
-                   string_format(
-                       "(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
-                       (double) params.lr.lr_min),
-                   [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(
-        common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
-                   string_format(
-                       "(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
-                       (double) params.lr.decay_epochs),
-                   [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
-            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
     add_opt(common_arg(
-                { "-wd", "--weight-decay" }, "WD",
-                string_format(
-                    "adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
-                    (double) params.lr.wd),
-                [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
-                       string_format("fraction of data to use as validation set for training (default: %.2g).",
-                                     (double) params.val_split),
-                       [](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-epochs", "--epochs" }, "N",
-                       string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
-                       [](common_params & params, int epochs) { params.lr.epochs = epochs; })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
-    add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
-                       [](common_params & params, const std::string & name) {
-                           params.optimizer = common_opt_get_optimizer(name.c_str());
-                           if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
-                               throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
-                           }
-                       })
-                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+        {"--vision-gemma-12b-default"},
+        string_format("use Gemma 3 12B QAT (note: can download weights from the internet)"),
+        [](common_params & params) {
+            params.model.hf_repo = "ggml-org/gemma-3-12b-it-qat-GGUF";
+            params.port = 8014;
+            params.n_ctx = 0;
+            params.use_jinja = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}));
 
     return ctx_arg;
 }

common/chat-parser.cpp

@@ -3,9 +3,12 @@
 #include "log.h"
 #include "regex-partial.h"
 
+#include <algorithm>
+#include <cctype>
 #include <optional>
 #include <stdexcept>
 #include <string>
+#include <string_view>
 #include <vector>
 
 using json = nlohmann::ordered_json;
@@ -166,6 +169,27 @@ void common_chat_msg_parser::consume_literal(const std::string & literal) {
 }
 
 bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think, const std::string & end_think) {
+    std::string pending_reasoning_prefix;
+
+    if (syntax_.reasoning_format == COMMON_REASONING_FORMAT_NONE) {
+        return false;
+    }
+
+    auto set_reasoning_prefix = [&](size_t prefix_pos) {
+        if (!syntax_.thinking_forced_open || syntax_.reasoning_in_content) {
+            return;
+        }
+        if (prefix_pos + start_think.size() > input_.size()) {
+            pending_reasoning_prefix.clear();
+            return;
+        }
+        // Capture the exact literal that opened the reasoning section so we can
+        // surface it back to callers. This ensures formats that force the
+        // reasoning tag open (e.g. DeepSeek R1) retain their original prefix
+        // instead of dropping it during parsing.
+        pending_reasoning_prefix = input_.substr(prefix_pos, start_think.size());
+    };
+
     auto handle_reasoning = [&](const std::string & reasoning, bool closed) {
         auto stripped_reasoning = string_strip(reasoning);
         if (stripped_reasoning.empty()) {
@@ -178,28 +202,116 @@ bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think
                 add_content(syntax_.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "</think>" : end_think);
             }
         } else {
+            if (!pending_reasoning_prefix.empty()) {
+                add_reasoning_content(pending_reasoning_prefix);
+                pending_reasoning_prefix.clear();
+            }
             add_reasoning_content(stripped_reasoning);
         }
     };
-    if (syntax_.reasoning_format != COMMON_REASONING_FORMAT_NONE) {
-        if (syntax_.thinking_forced_open || try_consume_literal(start_think)) {
-            if (auto res = try_find_literal(end_think)) {
-                handle_reasoning(res->prelude, /* closed */ true);
-                consume_spaces();
-                return true;
-            }
-            auto rest = consume_rest();
+
+    const size_t saved_pos = pos_;
+    const size_t saved_content_size = result_.content.size();
+    const size_t saved_reasoning_size = result_.reasoning_content.size();
+
+    auto restore_state = [&]() {
+        move_to(saved_pos);
+        result_.content.resize(saved_content_size);
+        result_.reasoning_content.resize(saved_reasoning_size);
+    };
+
+    // Allow leading whitespace to be preserved as content when reasoning is present at the start
+    size_t cursor = pos_;
+    size_t whitespace_end = cursor;
+    while (whitespace_end < input_.size() && std::isspace(static_cast<unsigned char>(input_[whitespace_end]))) {
+        ++whitespace_end;
+    }
+
+    if (whitespace_end >= input_.size()) {
+        restore_state();
+        if (syntax_.thinking_forced_open) {
+            auto rest = input_.substr(saved_pos);
             if (!rest.empty()) {
                 handle_reasoning(rest, /* closed */ !is_partial());
             }
-            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
-            // if (!syntax_.thinking_forced_open) {
-            //     throw common_chat_msg_partial_exception(end_think);
-            // }
+            move_to(input_.size());
             return true;
         }
+        return false;
+    }
+
+    cursor = whitespace_end;
+    const size_t remaining = input_.size() - cursor;
+    const size_t start_prefix = std::min(start_think.size(), remaining);
+    const bool has_start_tag = input_.compare(cursor, start_prefix, start_think, 0, start_prefix) == 0;
+
+    if (has_start_tag && start_prefix < start_think.size()) {
+        move_to(input_.size());
+        return true;
+    }
+
+    if (has_start_tag) {
+        if (whitespace_end > pos_) {
+            add_content(input_.substr(pos_, whitespace_end - pos_));
+        }
+        set_reasoning_prefix(cursor);
+        cursor += start_think.size();
+    } else if (syntax_.thinking_forced_open) {
+        cursor = whitespace_end;
+    } else {
+        restore_state();
+        return false;
+    }
+    while (true) {
+        if (cursor >= input_.size()) {
+            move_to(input_.size());
+            return true;
+        }
+
+        size_t end_pos = input_.find(end_think, cursor);
+        if (end_pos == std::string::npos) {
+            std::string_view remaining_view(input_.data() + cursor, input_.size() - cursor);
+            size_t partial_off = string_find_partial_stop(remaining_view, end_think);
+            size_t reasoning_end = partial_off == std::string::npos ? input_.size() : cursor + partial_off;
+            if (reasoning_end > cursor) {
+                handle_reasoning(input_.substr(cursor, reasoning_end - cursor), /* closed */ partial_off == std::string::npos && !is_partial());
+            }
+            move_to(input_.size());
+            return true;
+        }
+
+        if (end_pos > cursor) {
+            handle_reasoning(input_.substr(cursor, end_pos - cursor), /* closed */ true);
+        } else {
+            handle_reasoning("", /* closed */ true);
+        }
+
+        cursor = end_pos + end_think.size();
+
+        while (cursor < input_.size() && std::isspace(static_cast<unsigned char>(input_[cursor]))) {
+            ++cursor;
+        }
+
+        const size_t next_remaining = input_.size() - cursor;
+        if (next_remaining == 0) {
+            move_to(cursor);
+            return true;
+        }
+
+        const size_t next_prefix = std::min(start_think.size(), next_remaining);
+        if (input_.compare(cursor, next_prefix, start_think, 0, next_prefix) == 0) {
+            if (next_prefix < start_think.size()) {
+                move_to(input_.size());
+                return true;
+            }
+            set_reasoning_prefix(cursor);
+            cursor += start_think.size();
+            continue;
+        }
+
+        move_to(cursor);
+        return true;
     }
-    return false;
 }
 
 std::string common_chat_msg_parser::consume_rest() {
@@ -320,7 +432,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
         if (is_arguments_path({})) {
             // Entire JSON is the arguments and was parsed fully.
             return consume_json_result {
-                partial->json.dump(),
+                partial->json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true),
                 /* .is_partial = */ false,
             };
         }
@@ -332,7 +444,7 @@ std::optional<common_chat_msg_parser::consume_json_result> common_chat_msg_parse
     std::vector<std::string> path;
     std::function<json(const json &)> remove_unsupported_healings_and_dump_args = [&](const json & j) -> json {
         if (is_arguments_path(path)) {
-            auto arguments = j.dump();
+            auto arguments = j.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true);
             if (is_partial() && !partial->healing_marker.marker.empty()) {
                 auto idx = arguments.find(partial->healing_marker.json_dump_marker);
                 if (idx != std::string::npos) {

common/chat.cpp

@@ -1408,6 +1408,8 @@ static common_chat_params common_chat_params_init_apertus(const common_chat_temp
     return data;
 }
 static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) {
+    builder.try_parse_reasoning("<think>", "</think>");
+
     if (!builder.syntax().parse_tool_calls) {
         builder.add_content(builder.consume_rest());
         return;
@@ -2862,6 +2864,7 @@ common_chat_params common_chat_templates_apply(
 }
 
 static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
+    builder.try_parse_reasoning("<think>", "</think>");
     builder.add_content(builder.consume_rest());
 }
 

common/chat.h

@@ -33,8 +33,8 @@ struct common_chat_msg_content_part {
 struct common_chat_msg {
     std::string role;
     std::string content;
-    std::vector<common_chat_msg_content_part> content_parts = {};
-    std::vector<common_chat_tool_call> tool_calls = {};
+    std::vector<common_chat_msg_content_part> content_parts;
+    std::vector<common_chat_tool_call> tool_calls;
     std::string reasoning_content;
     std::string tool_name;
     std::string tool_call_id;
@@ -44,7 +44,7 @@ struct common_chat_msg {
     bool empty() const {
         return content.empty() && content_parts.empty() && tool_calls.empty() && reasoning_content.empty() && tool_name.empty() && tool_call_id.empty();
     }
-    void ensure_tool_call_ids_set(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
+    void set_tool_call_ids(std::vector<std::string> & ids_cache, const std::function<std::string()> & gen_tool_call_id) {
         for (auto i = 0u; i < tool_calls.size(); i++) {
             if (ids_cache.size() <= i) {
                 auto id = tool_calls[i].id;

common/common.h

@@ -378,7 +378,7 @@ struct common_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool no_perf           = false; // disable performance metrics
-    bool ctx_shift         = false;  // context shift on infinite text generation
+    bool ctx_shift         = false; // context shift on infinite text generation
     bool swa_full          = false; // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
     bool kv_unified        = false; // enable unified KV cache
 
@@ -425,7 +425,8 @@ struct common_params {
     int32_t timeout_write     = timeout_read; // http write timeout in seconds
     int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
-    int32_t n_ctx_checkpoints = 3;            // max number of context checkpoints per slot
+    int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
+    int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -433,7 +434,7 @@ struct common_params {
     std::string chat_template = "";                                                                         // NOLINT
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
-    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_AUTO;
+    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 

common/json-partial.cpp

@@ -5,6 +5,7 @@
 #include <nlohmann/json.hpp>
 
 #include <string>
+#include <regex>
 
 using json = nlohmann::ordered_json;
 
@@ -168,6 +169,47 @@ bool common_json_parse(
                 }
             }
 
+            // Matches a potentially partial unicode escape sequence, e.g. \u, \uX, \uXX, \uXXX, \uXXXX
+            static const std::regex partial_unicode_regex(R"(\\u(?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F](?:[0-9a-fA-F])?)?)?)?$)");
+
+            auto is_high_surrogate = [&](const std::string & s) {
+                // Check if a partial of a high surrogate (U+D800-U+DBFF)
+                return s.length() >= 4 &&
+                    s[0] == '\\' && s[1] == 'u' &&
+                    std::tolower(s[2]) == 'd' &&
+                    (s[3] == '8' || s[3] == '9' || std::tolower(s[3]) == 'a' || std::tolower(s[3]) == 'b');
+            };
+
+            // Initialize the unicode marker to a low surrogate to handle the edge case
+            // where a high surrogate (U+D800-U+DBFF) is immediately followed by a
+            // backslash (\)
+            std::string unicode_marker_padding = "udc00";
+            std::smatch last_unicode_seq;
+
+            if (std::regex_search(str, last_unicode_seq, partial_unicode_regex)) {
+                std::smatch second_last_seq;
+                std::string prelude = str.substr(0, last_unicode_seq.position());
+
+                // Pad the escape sequence with 0s until it forms a complete sequence of 6 characters
+                unicode_marker_padding = std::string(6 - last_unicode_seq.length(), '0');
+
+                if (is_high_surrogate(last_unicode_seq.str())) {
+                    // If the sequence is a partial match for a high surrogate, add a low surrogate (U+DC00-U+UDFF)
+                    unicode_marker_padding += "\\udc00";
+                } else if (std::regex_search(prelude, second_last_seq, partial_unicode_regex)) {
+                    if (is_high_surrogate(second_last_seq.str())) {
+                        // If this follows a high surrogate, pad it to be a low surrogate
+                        if (last_unicode_seq.length() == 2) {
+                            unicode_marker_padding = "dc00";
+                        } else if (last_unicode_seq.length() == 3) {
+                            unicode_marker_padding = "c00";
+                        } else {
+                            // The original unicode_marker_padding is already padded with 0s
+                        }
+                    }
+                }
+            }
+
             const auto & magic_seed = out.healing_marker.marker = healing_marker;//"$llama.cpp.json$";
 
             if (err_loc.stack.back().type == COMMON_JSON_STACK_ELEMENT_KEY) {
@@ -186,6 +228,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an object value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an object value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else {
                     // find last :
                     auto last_pos = str.find_last_of(':');
@@ -205,6 +250,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\"" + closing)) {
                     // Was inside an array value string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\"" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\"" + closing)) {
+                    // Was inside an array value string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\"" + closing;
                 } else if (!was_maybe_number() && can_parse(str + ", 1" + closing)) {
                     // Had just finished a value
                     str += (out.healing_marker.json_dump_marker = ",\"" + magic_seed) + "\"" + closing;
@@ -230,6 +278,9 @@ bool common_json_parse(
                 } else if (str[str.length() - 1] == '\\' && can_parse(str + "\\\": 1" + closing)) {
                     // Was inside an object key string after an escape
                     str += (out.healing_marker.json_dump_marker = "\\" + magic_seed) + "\": 1" + closing;
+                } else if (can_parse(str + unicode_marker_padding + "\": 1" + closing)) {
+                    // Was inside an object key string after a partial unicode escape
+                    str += (out.healing_marker.json_dump_marker = unicode_marker_padding + magic_seed) + "\": 1" + closing;
                 } else {
                     auto last_pos = str.find_last_of(':');
                     if (last_pos == std::string::npos) {

convert_hf_to_gguf.py

@@ -93,13 +93,15 @@ class ModelBase:
     # Mistral format specifics
     is_mistral_format: bool = False
     disable_mistral_community_chat_template: bool = False
+    sentence_transformers_dense_modules: bool = False
 
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
                  split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
                  small_first_shard: bool = False, hparams: dict[str, Any] | None = None, remote_hf_model_id: str | None = None,
-                 disable_mistral_community_chat_template: bool = False):
+                 disable_mistral_community_chat_template: bool = False,
+                 sentence_transformers_dense_modules: bool = False):
         if type(self) is ModelBase or \
                 type(self) is TextModel or \
                 type(self) is MmprojModel:
@@ -114,6 +116,7 @@ class ModelBase:
         self.lazy = not eager or (remote_hf_model_id is not None)
         self.dry_run = dry_run
         self.remote_hf_model_id = remote_hf_model_id
+        self.sentence_transformers_dense_modules = sentence_transformers_dense_modules
         if remote_hf_model_id is not None:
             self.is_safetensors = True
 
@@ -5269,6 +5272,53 @@ class Gemma3Model(TextModel):
 @ModelBase.register("Gemma3TextModel")
 class EmbeddingGemma(Gemma3Model):
     model_arch = gguf.MODEL_ARCH.GEMMA_EMBEDDING
+    module_paths = []
+    dense_features_dims = {}
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.sentence_transformers_dense_modules:
+            # read modules.json to determine if model has Dense layers
+            modules_file = self.dir_model / "modules.json"
+            if modules_file.is_file():
+                with open(modules_file, encoding="utf-8") as modules_json_file:
+                    mods = json.load(modules_json_file)
+                for mod in mods:
+                    if mod["type"] == "sentence_transformers.models.Dense":
+                        mod_path = mod["path"]
+                        # check if model.safetensors file for Dense layer exists
+                        model_tensors_file = self.dir_model / mod_path / "model.safetensors"
+                        if model_tensors_file.is_file():
+                            self.module_paths.append(mod_path)
+                            # read config.json of the Dense layer to get in/out features
+                            mod_conf_file = self.dir_model / mod_path / "config.json"
+                            if mod_conf_file.is_file():
+                                with open(mod_conf_file, encoding="utf-8") as mod_conf_json_file:
+                                    mod_conf = json.load(mod_conf_json_file)
+                                    # hparams dense_2_feat_out and dense_3_feat_in are required when loading model's dense weights
+                                    prefix = self._get_dense_prefix(mod_path)
+                                    if mod_conf["in_features"] is not None and mod_conf["out_features"] is not None:
+                                        self.dense_features_dims[prefix] = (mod_conf["in_features"], mod_conf["out_features"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        from safetensors.torch import load_file
+        module_paths = list(self.module_paths)
+        for i, module_path in enumerate(module_paths):
+            tensors_file = self.dir_model / module_path / "model.safetensors"
+            local_tensors = load_file(tensors_file)
+            tensor_name = self._get_dense_prefix(module_path)
+            for name, local_tensor in local_tensors.items():
+                if not name.endswith(".weight"):
+                    continue
+                orig_name = name.replace("linear", tensor_name)
+                name = self.map_tensor_name(orig_name)
+                yield name, local_tensor.clone()
+
+    @staticmethod
+    def _get_dense_prefix(module_path) -> str:
+        """Get the tensor name prefix for the Dense layer from module path."""
+        tensor_name = "dense_2" if module_path == "2_Dense" else "dense_3"
+        return tensor_name
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -5285,6 +5335,10 @@ class EmbeddingGemma(Gemma3Model):
             logger.info(f"Using original sliding_window from config: {orig_sliding_window} "
                         f"instead of {self.hparams['sliding_window']}")
             self.gguf_writer.add_sliding_window(orig_sliding_window)
+        if self.sentence_transformers_dense_modules:
+            for dense, dims in self.dense_features_dims.items():
+                logger.info(f"Setting dense layer {dense} in/out features to {dims}")
+                self.gguf_writer.add_dense_features_dims(dense, dims[0], dims[1])
 
         self._try_set_pooling_type()
 
@@ -5912,20 +5966,12 @@ class Mamba2Model(TextModel):
 class JambaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.JAMBA
 
-    def get_vocab_base_pre(self, tokenizer) -> str:
-        del tokenizer  # unused
-
-        return "gpt-2"
-
     def set_vocab(self):
         if (self.dir_model / "tokenizer.model").is_file():
-            # Using Jamba's tokenizer.json causes errors on model load
-            # (something about "byte not found in vocab"),
-            # but there's a working tokenizer.model
             self._set_vocab_sentencepiece()
         else:
-            # Some Jamba models only have a tokenizer.json, which works.
-            self._set_vocab_gpt2()
+            self._set_vocab_llama_hf()
+            self.gguf_writer.add_add_space_prefix(False)
 
     def set_gguf_parameters(self):
         d_model = self.find_hparam(["hidden_size", "mamba_d_model"])
@@ -8836,6 +8882,75 @@ class LFM2Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@ModelBase.register("Lfm2MoeForCausalLM")
+class LFM2MoeModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LFM2MOE
+
+    def set_gguf_parameters(self):
+        # set num_key_value_heads only for attention layers
+        self.hparams["num_key_value_heads"] = [
+            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            for layer_type in self.hparams["layer_types"]
+        ]
+
+        super().set_gguf_parameters()
+
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
+        self.gguf_writer.add_leading_dense_block_count(self.hparams["num_dense_layers"])
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)
+
+        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
+        self.gguf_writer.add_shortconv_l_cache(self.hparams["conv_L_cache"])
+
+    # cache for experts weights for merging
+    _experts_cache: dict[int, dict[str, Tensor]] = {}
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # conv op requires 2d tensor
+        if 'conv.conv' in name:
+            data_torch = data_torch.squeeze(1)
+
+        if name.endswith(".expert_bias"):
+            name = name.replace(".expert_bias", ".expert_bias.bias")
+
+        # merge expert weights
+        if 'experts' in name:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            expert_cache = self._experts_cache.setdefault(bid, {})
+            expert_cache[name] = data_torch
+            expert_weights = ["w1", "w2", "w3"]
+
+            # not enough expert weights to merge
+            if len(expert_cache) < n_experts * len(expert_weights):
+                return []
+
+            tensors: list[tuple[str, Tensor]] = []
+            for w_name in expert_weights:
+                datas: list[Tensor] = []
+
+                for xid in range(n_experts):
+                    ename = f"model.layers.{bid}.feed_forward.experts.{xid}.{w_name}.weight"
+                    datas.append(expert_cache[ename])
+                    del expert_cache[ename]
+
+                data_torch = torch.stack(datas, dim=0)
+                merged_name = f"layers.{bid}.feed_forward.experts.{w_name}.weight"
+                new_name = self.map_tensor_name(merged_name)
+                tensors.append((new_name, data_torch))
+
+            del self._experts_cache[bid]
+            return tensors
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        assert not self._experts_cache
+
+
 @ModelBase.register("Lfm2VlForConditionalGeneration")
 class LFM2VLModel(MmprojModel):
     def __init__(self, *args, **kwargs):
@@ -9266,6 +9381,13 @@ def parse_args() -> argparse.Namespace:
         )
     )
 
+    parser.add_argument(
+        "--sentence-transformers-dense-modules", action="store_true",
+        help=("Whether to include sentence-transformers dense modules."
+              "It can be used for sentence-transformers models, like google/embeddinggemma-300m"
+              "Default these modules are not included.")
+    )
+
     args = parser.parse_args()
     if not args.print_supported_models and args.model is None:
         parser.error("the following arguments are required: model")
@@ -9328,9 +9450,13 @@ def main() -> None:
     if args.remote:
         hf_repo_id = args.model
         from huggingface_hub import snapshot_download
+        allowed_patterns = ["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"]
+        if args.sentence_transformers_dense_modules:
+            # include sentence-transformers dense modules safetensors files
+            allowed_patterns.append("*.safetensors")
         local_dir = snapshot_download(
             repo_id=hf_repo_id,
-            allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
+            allow_patterns=allowed_patterns)
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
     else:
@@ -9398,7 +9524,8 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template
+                                     remote_hf_model_id=hf_repo_id, disable_mistral_community_chat_template=disable_mistral_community_chat_template,
+                                     sentence_transformers_dense_modules=args.sentence_transformers_dense_modules
                                      )
 
         if args.vocab_only:

docs/ops.md

@@ -31,7 +31,7 @@ Legend:
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -51,7 +51,7 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
@@ -65,11 +65,11 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
@@ -92,9 +92,9 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -102,9 +102,11 @@ Legend:
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
 |                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
 |                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                         TOPK_MOE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |
+|                            XIELU | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |

examples/model-conversion/Makefile

@@ -116,20 +116,39 @@ embedding-convert-model:
 	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
 	./scripts/embedding/convert-model.sh
 
+embedding-convert-model-st:
+	$(call validate_embedding_model_path,embedding-convert-model-st)
+	@MODEL_NAME="$(MODEL_NAME)" OUTTYPE="$(OUTTYPE)" MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	METADATA_OVERRIDE="$(METADATA_OVERRIDE)" \
+	./scripts/embedding/convert-model.sh -st
+
 embedding-run-original-model:
 	$(call validate_embedding_model_path,embedding-run-original-model)
 	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" \
+	USE_SENTENCE_TRANSFORMERS="$(USE_SENTENCE_TRANSFORMERS)" \
 	./scripts/embedding/run-original-model.py \
-	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)") \
+	$(if $(USE_SENTENCE_TRANSFORMERS),--use-sentence-transformers)
+
+embedding-run-original-model-st: USE_SENTENCE_TRANSFORMERS=1
+embedding-run-original-model-st: embedding-run-original-model
 
 embedding-run-converted-model:
 	@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
-	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)") \
+	$(if $(USE_POOLING),--pooling)
+
+embedding-run-converted-model-st: USE_POOLING=1
+embedding-run-converted-model-st: embedding-run-converted-model
 
 embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
 	@./scripts/embedding/compare-embeddings-logits.sh \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 
+embedding-verify-logits-st: embedding-run-original-model-st embedding-run-converted-model-st
+	@./scripts/embedding/compare-embeddings-logits.sh \
+	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
+
 embedding-inspect-original-model:
 	$(call validate_embedding_model_path,embedding-inspect-original-model)
 	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}

examples/model-conversion/README.md

@@ -189,6 +189,23 @@ This command will save two files to the `data` directory, one is a binary
 file containing logits which will be used for comparison with the converted
 model, and the other is a text file which allows for manual visual inspection.
 
+#### Using SentenceTransformer with numbered layers
+For models that have numbered SentenceTransformer layers (01_Pooling, 02_Dense,
+03_Dense, 04_Normalize), use the `-st` targets to apply all these layers:
+
+```console
+# Run original model with SentenceTransformer (applies all numbered layers)
+(venv) $ make embedding-run-original-model-st
+
+# Run converted model with pooling enabled
+(venv) $ make embedding-run-converted-model-st
+```
+
+This will use the SentenceTransformer library to load and run the model, which
+automatically applies all the numbered layers in the correct order. This is
+particularly useful when comparing with models that should include these
+additional transformation layers beyond just the base model output.
+
 ### Model conversion
 After updates have been made to [gguf-py](../../gguf-py) to add support for the
 new model the model can be converted to GGUF format using the following command:
@@ -208,6 +225,13 @@ was done manually in the previous steps) and compare the logits:
 (venv) $ make embedding-verify-logits
 ```
 
+For models with SentenceTransformer layers, use the `-st` verification target:
+```console
+(venv) $ make embedding-verify-logits-st
+```
+This convenience target automatically runs both the original model with SentenceTransformer
+and the converted model with pooling enabled, then compares the results.
+
 ### llama-server verification
 To verify that the converted model works with llama-server, the following
 command can be used:

examples/model-conversion/logits.cpp

@@ -1,4 +1,7 @@
 #include "llama.h"
+#include "common.h"
+
+
 #include <cstdio>
 #include <cstring>
 #include <string>
@@ -8,7 +11,10 @@
 
 static void print_usage(int, char ** argv) {
     printf("\nexample usage:\n");
-    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [prompt]\n", argv[0]);
+    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [-pooling] [-embd-norm <norm>] [prompt]\n", argv[0]);
+    printf("\n");
+    printf("  -embd-norm: normalization type for pooled embeddings (default: 2)\n");
+    printf("              -1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm\n");
     printf("\n");
 }
 
@@ -17,6 +23,8 @@ int main(int argc, char ** argv) {
     std::string prompt = "Hello, my name is";
     int ngl = 0;
     bool embedding_mode = false;
+    bool pooling_enabled = false;
+    int32_t embd_norm = 2;  // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
 
     {
         int i = 1;
@@ -41,9 +49,13 @@ int main(int argc, char ** argv) {
                     return 1;
                 }
             } else if (strcmp(argv[i], "-embd-mode") == 0) {
+                embedding_mode = true;
+            } else if (strcmp(argv[i], "-pooling") == 0) {
+                pooling_enabled = true;
+            } else if (strcmp(argv[i], "-embd-norm") == 0) {
                 if (i + 1 < argc) {
                     try {
-                        embedding_mode = true;
+                        embd_norm = std::stoi(argv[++i]);
                     } catch (...) {
                         print_usage(argc, argv);
                         return 1;
@@ -112,7 +124,7 @@ int main(int argc, char ** argv) {
     ctx_params.no_perf = false;
     if (embedding_mode) {
         ctx_params.embeddings = true;
-        ctx_params.pooling_type = LLAMA_POOLING_TYPE_NONE;
+        ctx_params.pooling_type = pooling_enabled ? LLAMA_POOLING_TYPE_MEAN : LLAMA_POOLING_TYPE_NONE;
         ctx_params.n_ubatch = ctx_params.n_batch;
     }
 
@@ -143,17 +155,27 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    float * logits;
-    int n_logits;
+    float * data_ptr;
+    int data_size;
     const char * type;
+    std::vector<float> embd_out;
 
     if (embedding_mode) {
-        logits = llama_get_embeddings(ctx);
-        n_logits = llama_model_n_embd(model) * batch.n_tokens;
+        const int n_embd = llama_model_n_embd(model);
+        const int n_embd_count = pooling_enabled ? 1 : batch.n_tokens;
+        const int n_embeddings = n_embd * n_embd_count;
+        float * embeddings;
         type = "-embeddings";
 
-        const int n_embd = llama_model_n_embd(model);
-        const int n_embd_count = batch.n_tokens;
+        if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_NONE) {
+            embeddings = llama_get_embeddings_seq(ctx, 0);
+            embd_out.resize(n_embeddings);
+            printf("Normalizing embeddings using norm: %d\n", embd_norm);
+            common_embd_normalize(embeddings, embd_out.data(), n_embeddings, embd_norm);
+            embeddings = embd_out.data();
+        } else {
+            embeddings = llama_get_embeddings(ctx);
+        }
 
         printf("Embedding dimension: %d\n", n_embd);
         printf("\n");
@@ -164,7 +186,7 @@ int main(int argc, char ** argv) {
 
             // Print first 3 values
             for (int i = 0; i < 3 && i < n_embd; i++) {
-                printf("%9.6f ", logits[j * n_embd + i]);
+                printf("%9.6f ", embeddings[j * n_embd + i]);
             }
 
             printf(" ... ");
@@ -172,7 +194,7 @@ int main(int argc, char ** argv) {
             // Print last 3 values
             for (int i = n_embd - 3; i < n_embd; i++) {
                 if (i >= 0) {
-                    printf("%9.6f ", logits[j * n_embd + i]);
+                    printf("%9.6f ", embeddings[j * n_embd + i]);
                 }
             }
 
@@ -180,27 +202,33 @@ int main(int argc, char ** argv) {
         }
         printf("\n");
 
-        printf("Embeddings size: %d\n", n_logits);
+        printf("Embeddings size: %d\n", n_embeddings);
+
+        data_ptr = embeddings;
+        data_size = n_embeddings;
     } else {
-        logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-        n_logits = llama_vocab_n_tokens(vocab);
+        float * logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+        const int n_logits = llama_vocab_n_tokens(vocab);
         type = "";
         printf("Vocab size: %d\n", n_logits);
+
+        data_ptr = logits;
+        data_size = n_logits;
     }
 
     std::filesystem::create_directory("data");
 
-    // Save logits to binary file
+    // Save data to binary file
     char bin_filename[512];
     snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
-    printf("Saving logits to %s\n", bin_filename);
+    printf("Saving data to %s\n", bin_filename);
 
     FILE * f = fopen(bin_filename, "wb");
     if (f == NULL) {
         fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
         return 1;
     }
-    fwrite(logits, sizeof(float), n_logits, f);
+    fwrite(data_ptr, sizeof(float), data_size, f);
     fclose(f);
 
     // Also save as text for debugging
@@ -211,27 +239,27 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
         return 1;
     }
-    for (int i = 0; i < n_logits; i++) {
-        fprintf(f, "%d: %.6f\n", i, logits[i]);
+    for (int i = 0; i < data_size; i++) {
+        fprintf(f, "%d: %.6f\n", i, data_ptr[i]);
     }
     fclose(f);
 
     if (!embedding_mode) {
         printf("First 10 logits: ");
-        for (int i = 0; i < 10 && i < n_logits; i++) {
-            printf("%.6f ", logits[i]);
+        for (int i = 0; i < 10 && i < data_size; i++) {
+            printf("%.6f ", data_ptr[i]);
         }
         printf("\n");
 
         printf("Last 10 logits: ");
-        for (int i = n_logits - 10; i < n_logits; i++) {
-            if (i >= 0) printf("%.6f ", logits[i]);
+        for (int i = data_size - 10; i < data_size; i++) {
+            if (i >= 0) printf("%.6f ", data_ptr[i]);
         }
         printf("\n\n");
     }
 
-    printf("Logits saved to %s\n", bin_filename);
-    printf("Logits saved to %s\n", txt_filename);
+    printf("Data saved to %s\n", bin_filename);
+    printf("Data saved to %s\n", txt_filename);
 
     llama_free(ctx);
     llama_model_free(model);

examples/model-conversion/requirements.txt

@@ -4,3 +4,4 @@ torchvision
 transformers
 huggingface-hub
 accelerate
+sentence-transformers

examples/model-conversion/scripts/embedding/convert-model.sh

@@ -2,6 +2,21 @@
 
 set -e
 
+# Parse command line arguments
+SENTENCE_TRANSFORMERS=""
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        -st|--sentence-transformers)
+            SENTENCE_TRANSFORMERS="--sentence-transformers-dense-modules"
+            shift
+            ;;
+        *)
+            echo "Unknown option: $1"
+            exit 1
+            ;;
+    esac
+done
+
 MODEL_NAME="${MODEL_NAME:-$(basename "$EMBEDDING_MODEL_PATH")}"
 OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
 TYPE="${OUTTYPE:-f16}"
@@ -15,7 +30,8 @@ echo "Converted model path:: ${CONVERTED_MODEL}"
 python ../../convert_hf_to_gguf.py --verbose \
     ${EMBEDDING_MODEL_PATH} \
     --outfile ${CONVERTED_MODEL} \
-    --outtype ${TYPE}
+    --outtype ${TYPE} \
+    ${SENTENCE_TRANSFORMERS}
 
 echo ""
 echo "The environment variable CONVERTED_EMBEDDING MODEL can be set to this path using:"

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -5,6 +5,7 @@ set -e
 # Parse command line arguments
 CONVERTED_MODEL=""
 PROMPTS_FILE=""
+USE_POOLING=""
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -12,6 +13,10 @@ while [[ $# -gt 0 ]]; do
             PROMPTS_FILE="$2"
             shift 2
             ;;
+        --pooling)
+            USE_POOLING="1"
+            shift
+            ;;
         *)
             if [ -z "$CONVERTED_MODEL" ]; then
                 CONVERTED_MODEL="$1"
@@ -47,4 +52,8 @@ echo $CONVERTED_MODEL
 
 cmake --build ../../build --target llama-logits -j8
 # TODO: update logits.cpp to accept a --file/-f option for the prompt
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"
+if [ -n "$USE_POOLING" ]; then
+    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode -pooling "$PROMPT"
+else
+    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"
+fi

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -14,6 +14,8 @@ unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 parser = argparse.ArgumentParser(description='Process model with specified path')
 parser.add_argument('--model-path', '-m', help='Path to the model')
 parser.add_argument('--prompts-file', '-p', help='Path to file containing prompts (one per line)')
+parser.add_argument('--use-sentence-transformers', action='store_true',
+                    help='Use SentenceTransformer to apply all numbered layers (01_Pooling, 02_Dense, 03_Dense, 04_Normalize)')
 args = parser.parse_args()
 
 def read_prompt_from_file(file_path):
@@ -31,41 +33,52 @@ model_path = os.environ.get('EMBEDDING_MODEL_PATH', args.model_path)
 if model_path is None:
     parser.error("Model path must be specified either via --model-path argument or EMBEDDING_MODEL_PATH environment variable")
 
-tokenizer = AutoTokenizer.from_pretrained(model_path)
+# Determine if we should use SentenceTransformer
+use_sentence_transformers = args.use_sentence_transformers or os.environ.get('USE_SENTENCE_TRANSFORMERS', '').lower() in ('1', 'true', 'yes')
 
-config = AutoConfig.from_pretrained(model_path)
-
-# This can be used to override the sliding window size for manual testing. This
-# can be useful to verify the sliding window attention mask in the original model
-# and compare it with the converted .gguf model.
-if hasattr(config, 'sliding_window'):
-    original_sliding_window = config.sliding_window
-    #original_sliding_window = 6
-    print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
-
-print(f"Using unreleased model: {unreleased_model_name}")
-if unreleased_model_name:
-    model_name_lower = unreleased_model_name.lower()
-    unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
-    class_name = f"{unreleased_model_name}Model"
-    print(f"Importing unreleased model module: {unreleased_module_path}")
-
-    try:
-        model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
-        model = model_class.from_pretrained(model_path, config=config)
-    except (ImportError, AttributeError) as e:
-        print(f"Failed to import or load model: {e}")
-        exit(1)
+if use_sentence_transformers:
+    from sentence_transformers import SentenceTransformer
+    print("Using SentenceTransformer to apply all numbered layers")
+    model = SentenceTransformer(model_path)
+    tokenizer = model.tokenizer
+    config = model[0].auto_model.config  # type: ignore
 else:
-    model = AutoModel.from_pretrained(model_path, config=config)
-print(f"Model class: {type(model)}")
-print(f"Model file: {type(model).__module__}")
+    tokenizer = AutoTokenizer.from_pretrained(model_path)
+
+    config = AutoConfig.from_pretrained(model_path)
+
+    # This can be used to override the sliding window size for manual testing. This
+    # can be useful to verify the sliding window attention mask in the original model
+    # and compare it with the converted .gguf model.
+    if hasattr(config, 'sliding_window'):
+        original_sliding_window = config.sliding_window
+        #original_sliding_window = 6
+        print(f"Modified sliding window: {original_sliding_window} -> {config.sliding_window}")
+
+    print(f"Using unreleased model: {unreleased_model_name}")
+    if unreleased_model_name:
+        model_name_lower = unreleased_model_name.lower()
+        unreleased_module_path = f"transformers.models.{model_name_lower}.modular_{model_name_lower}"
+        class_name = f"{unreleased_model_name}Model"
+        print(f"Importing unreleased model module: {unreleased_module_path}")
+
+        try:
+            model_class = getattr(importlib.import_module(unreleased_module_path), class_name)
+            model = model_class.from_pretrained(model_path, config=config)
+        except (ImportError, AttributeError) as e:
+            print(f"Failed to import or load model: {e}")
+            exit(1)
+    else:
+        model = AutoModel.from_pretrained(model_path, config=config)
+    print(f"Model class: {type(model)}")
+    print(f"Model file: {type(model).__module__}")
 
 # Verify the model is using the correct sliding window
-if hasattr(model.config, 'sliding_window'):
-    print(f"Model's sliding_window: {model.config.sliding_window}")
-else:
-    print("Model config does not have sliding_window attribute")
+if not use_sentence_transformers:
+    if hasattr(model.config, 'sliding_window'):  # type: ignore
+        print(f"Model's sliding_window: {model.config.sliding_window}")  # type: ignore
+    else:
+        print("Model config does not have sliding_window attribute")
 
 model_name = os.path.basename(model_path)
 
@@ -75,34 +88,56 @@ if args.prompts_file:
 else:
     texts = ["Hello world today"]
 
-encoded = tokenizer(
-    texts,
-    padding=True,
-    truncation=True,
-    return_tensors="pt"
-)
-
-tokens = encoded['input_ids'][0]
-token_strings = tokenizer.convert_ids_to_tokens(tokens)
-for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
-    print(f"{token_id:6d} -> '{token_str}'")
-
 with torch.no_grad():
-    outputs = model(**encoded)
-    hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+    if use_sentence_transformers:
+        embeddings = model.encode(texts, convert_to_numpy=True)
+        all_embeddings = embeddings  # Shape: [batch_size, hidden_size]
 
-    # Extract embeddings for each token (matching LLAMA_POOLING_TYPE_NONE behavior)
-    all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+        encoded = tokenizer(
+            texts,
+            padding=True,
+            truncation=True,
+            return_tensors="pt"
+        )
+        tokens = encoded['input_ids'][0]
+        token_strings = tokenizer.convert_ids_to_tokens(tokens)
+        for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+            print(f"{token_id:6d} -> '{token_str}'")
 
-    print(f"Hidden states shape: {hidden_states.shape}")
-    print(f"All embeddings shape: {all_embeddings.shape}")
-    print(f"Embedding dimension: {all_embeddings.shape[1]}")
+        print(f"Embeddings shape (after all SentenceTransformer layers): {all_embeddings.shape}")
+        print(f"Embedding dimension: {all_embeddings.shape[1] if len(all_embeddings.shape) > 1 else all_embeddings.shape[0]}")  # type: ignore
+    else:
+        # Standard approach: use base model output only
+        encoded = tokenizer(
+            texts,
+            padding=True,
+            truncation=True,
+            return_tensors="pt"
+        )
 
-    # Print embeddings exactly like embedding.cpp does for LLAMA_POOLING_TYPE_NONE
-    n_embd = all_embeddings.shape[1]
-    n_embd_count = all_embeddings.shape[0]
+        tokens = encoded['input_ids'][0]
+        token_strings = tokenizer.convert_ids_to_tokens(tokens)
+        for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
+            print(f"{token_id:6d} -> '{token_str}'")
 
-    print()  # Empty line to match C++ output
+        outputs = model(**encoded)
+        hidden_states = outputs.last_hidden_state  # Shape: [batch_size, seq_len, hidden_size]
+
+        all_embeddings = hidden_states[0].cpu().numpy()  # Shape: [seq_len, hidden_size]
+
+        print(f"Hidden states shape: {hidden_states.shape}")
+        print(f"All embeddings shape: {all_embeddings.shape}")
+        print(f"Embedding dimension: {all_embeddings.shape[1]}")
+
+    if len(all_embeddings.shape) == 1:
+        n_embd = all_embeddings.shape[0]  # type: ignore
+        n_embd_count = 1
+        all_embeddings = all_embeddings.reshape(1, -1)
+    else:
+        n_embd = all_embeddings.shape[1]  # type: ignore
+        n_embd_count = all_embeddings.shape[0]  # type: ignore
+
+    print()
 
     for j in range(n_embd_count):
         embedding = all_embeddings[j]
@@ -120,29 +155,23 @@ with torch.no_grad():
 
         print()  # New line
 
-    print()  # Final empty line to match C++ output
+    print()
 
     data_dir = Path("data")
     data_dir.mkdir(exist_ok=True)
     bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
     txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
 
-    # Save all embeddings flattened (matching what embedding.cpp would save if it did)
     flattened_embeddings = all_embeddings.flatten()
     flattened_embeddings.astype(np.float32).tofile(bin_filename)
 
     with open(txt_filename, "w") as f:
-        f.write(f"# Model class: {model_name}\n")
-        f.write(f"# Tokens: {token_strings}\n")
-        f.write(f"# Shape: {all_embeddings.shape}\n")
-        f.write(f"# n_embd_count: {n_embd_count}, n_embd: {n_embd}\n\n")
-
+        idx = 0
         for j in range(n_embd_count):
-            f.write(f"# Token {j} ({token_strings[j]}):\n")
-            for i, value in enumerate(all_embeddings[j]):
-                f.write(f"{j}_{i}: {value:.6f}\n")
-            f.write("\n")
-    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} tokens × {n_embd} dimensions)")
+            for value in all_embeddings[j]:
+                f.write(f"{idx}: {value:.6f}\n")
+                idx += 1
+    print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} embeddings × {n_embd} dimensions)")
     print("")
     print(f"Saved bin embeddings to: {bin_filename}")
     print(f"Saved txt embeddings to: {txt_filename}")

examples/model-conversion/scripts/utils/semantic_check.py

@@ -35,7 +35,11 @@ def cosine_similarity(a, b=None):
 
 def load_embeddings_from_file(filename, n_tokens, n_embd):
     embeddings = np.fromfile(filename, dtype=np.float32)
-    return embeddings.reshape(n_tokens, n_embd)
+    # Check if this is pooled (single embedding) or per-token embeddings
+    if len(embeddings) == n_embd:
+        return embeddings.reshape(1, n_embd)
+    else:
+        return embeddings.reshape(n_tokens, n_embd)
 
 def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
     np.set_printoptions(suppress=True, precision=6)
@@ -48,58 +52,83 @@ def test_single_prompt_similarity(python_emb, cpp_emb, tokens, prompt):
     print(f"Embeddings shape: Python {python_emb.shape}, llama.cpp {cpp_emb.shape}")
 
     n_tokens = len(tokens)
+    is_pooled = python_emb.shape[0] == 1
 
-    # 1. Direct embedding comparison
-    print(f"\n1. Raw Embedding Magnitude Comparison:")
-    # Check if the distance of each token embedding from the origin and compare
-    # if the vectors are on the same "sphere". This does not tell us about
-    # direction (meaning of the token embedding), just magnitude.
-    for i in range(n_tokens):
-        py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
-        cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+    if is_pooled:
+        print(f"\n[Pooled Embeddings Mode - comparing single sentence embeddings]")
+
+        # 1. Direct embedding comparison for pooled embeddings
+        print(f"\n1. Raw Embedding Magnitude Comparison:")
+        py_mag = np.linalg.norm(python_emb[0])
+        cpp_mag = np.linalg.norm(cpp_emb[0])
         ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
-        print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
+        print(f"   Pooled embedding: Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
 
-    # 2. Cosine similarity between tokens within each model
-    # Here we check the direction of token embeddings to see if the have the
-    # same meaning (similarity). This is done by calculating cosine similarity
-    # of a pair of token embeddings within each model.
-    print(f"\n2. Within-Model Token Similarities:")
-    print("   Python model:")
-    for i in range(n_tokens):
-        for j in range(i+1, n_tokens):
-            sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
-            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+        # 2. Cross-model similarity for pooled embeddings
+        print(f"\n2. Cross-Model Pooled Embedding Similarity:")
+        sim = cosine_similarity([python_emb[0]], [cpp_emb[0]])[0][0]
+        print(f"   Cosine similarity: {sim:.6f}")
 
-    print("   llama.cpp model:")
-    for i in range(n_tokens):
-        for j in range(i+1, n_tokens):
-            sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
-            print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
+        return {
+            'cross_model_similarities': [sim],
+            'similarity_matrix_diff': np.array([[0.0]]),
+            'max_diff': 0.0,
+            'mean_diff': 0.0,
+            'rms_diff': 0.0
+        }
+    else:
+        # Original per-token comparison logic
+        # 1. Direct embedding comparison
+        print(f"\n1. Raw Embedding Magnitude Comparison:")
+        # Check if the distance of each token embedding from the origin and compare
+        # if the vectors are on the same "sphere". This does not tell us about
+        # direction (meaning of the token embedding), just magnitude.
+        for i in range(n_tokens):
+            py_mag = np.linalg.norm(python_emb[i]) # calculate standard euclidean norm for Python embeddings
+            cpp_mag = np.linalg.norm(cpp_emb[i])   # calculate standard euclidean norm for llama.cpp embeddings
+            ratio = py_mag / cpp_mag if cpp_mag > 0 else float('inf')
+            print(f"   Token {i} ({tokens[i]}): Python={py_mag:.3f}, llama.cpp={cpp_mag:.3f}, ratio={ratio:.3f}")
 
-    # 3. Cross-model similarity (same token position)
-    print(f"\n3. Cross-Model Same-Token Similarities:")
-    for i in range(n_tokens):
-        sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
-        print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
+        # 2. Cosine similarity between tokens within each model
+        # Here we check the direction of token embeddings to see if the have the
+        # same meaning (similarity). This is done by calculating cosine similarity
+        # of a pair of token embeddings within each model.
+        print(f"\n2. Within-Model Token Similarities:")
+        print("   Python model:")
+        for i in range(n_tokens):
+            for j in range(i+1, n_tokens):
+                sim = cosine_similarity([python_emb[i]], [python_emb[j]])[0][0]
+                print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
 
-    # 4. Similarity matrix comparison
-    print(f"\n4. Similarity Matrix Differences:")
-    py_sim_matrix = cosine_similarity(python_emb)
-    cpp_sim_matrix = cosine_similarity(cpp_emb)
-    diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+        print("   llama.cpp model:")
+        for i in range(n_tokens):
+            for j in range(i+1, n_tokens):
+                sim = cosine_similarity([cpp_emb[i]], [cpp_emb[j]])[0][0]
+                print(f"     {tokens[i]} ↔ {tokens[j]}: {sim:.4f}")
 
-    print(f"   Max difference: {np.max(diff_matrix):.4f}")
-    print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
-    print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+        # 3. Cross-model similarity (same token position)
+        print(f"\n3. Cross-Model Same-Token Similarities:")
+        for i in range(n_tokens):
+            sim = cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0]
+            print(f"   Token {i} ({tokens[i]}): {sim:.4f}")
 
-    return {
-        'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
-        'similarity_matrix_diff': diff_matrix,
-        'max_diff': np.max(diff_matrix),
-        'mean_diff': np.mean(diff_matrix),
-        'rms_diff': np.sqrt(np.mean(diff_matrix**2))
-    }
+        # 4. Similarity matrix comparison
+        print(f"\n4. Similarity Matrix Differences:")
+        py_sim_matrix = cosine_similarity(python_emb)
+        cpp_sim_matrix = cosine_similarity(cpp_emb)
+        diff_matrix = np.abs(py_sim_matrix - cpp_sim_matrix)
+
+        print(f"   Max difference: {np.max(diff_matrix):.4f}")
+        print(f"   Mean difference: {np.mean(diff_matrix):.4f}")
+        print(f"   RMS difference: {np.sqrt(np.mean(diff_matrix**2)):.4f}")
+
+        return {
+            'cross_model_similarities': [cosine_similarity([python_emb[i]], [cpp_emb[i]])[0][0] for i in range(n_tokens)],
+            'similarity_matrix_diff': diff_matrix,
+            'max_diff': np.max(diff_matrix),
+            'mean_diff': np.mean(diff_matrix),
+            'rms_diff': np.sqrt(np.mean(diff_matrix**2))
+        }
 
 def read_prompt_from_file(file_path):
     try:

ggml/CMakeLists.txt

@@ -222,6 +222,9 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
+option(GGML_WEBGPU_CPU_PROFILE              "ggml: enable WebGPU profiling (CPU)"             OFF)
+option(GGML_WEBGPU_GPU_PROFILE              "ggml: enable WebGPU profiling (GPU)"             OFF)
+
 option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)

ggml/src/CMakeLists.txt

@@ -145,6 +145,9 @@ endif()
 # which was introduced in POSIX.1-2008, forcing us to go higher
 if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
     add_compile_definitions(_XOPEN_SOURCE=700)
+elseif (CMAKE_SYSTEM_NAME MATCHES "AIX")
+    # Don't define _XOPEN_SOURCE.  We need _ALL_SOURCE, which is the default,
+    # in order to define _SC_PHYS_PAGES.
 else()
     add_compile_definitions(_XOPEN_SOURCE=600)
 endif()

ggml/src/ggml-cann/common.h

@@ -341,11 +341,18 @@ private:
 
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
+    // dst tensor
     void * node_address;
-    ggml_op node_op;
     int64_t ne[GGML_MAX_DIMS];
     size_t nb[GGML_MAX_DIMS];
+
+    // src tensor
     void * src_address[GGML_MAX_SRC];
+    int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+    size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
+
+    // op
+    ggml_op node_op;
     int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 };
 

ggml/src/ggml-cann/ggml-cann.cpp

@@ -2186,7 +2186,15 @@ static void add_lru_matched_graph_node_properties(
         std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
 
         for (int src = 0; src < GGML_MAX_SRC; ++src) {
-            prop.src_address[src] = node->src[src] ? node->src[src]->data : nullptr;
+            if (node->src[src]) {
+                prop.src_address[src] = node->src[src]->data;
+                std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
+                std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
+            } else {
+                prop.src_address[src] = nullptr;
+                std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
+                std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
+            }
         }
 
         memcpy(prop.op_params, node->op_params, GGML_MAX_OP_PARAMS);
@@ -2206,14 +2214,18 @@ static void add_lru_matched_graph_node_properties(
  * @param graph_node_properties The stored properties of a CANN graph node.
  * @return true if all fields match (excluding GGML_OP_VIEW); false otherwise.
  */
-static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+static bool ggml_graph_node_has_matching_properties(
+        ggml_tensor * node,
+        ggml_graph_node_properties * graph_node_properties) {
     if (node->data != graph_node_properties->node_address &&
-           node->op != GGML_OP_VIEW) {
+            node->op != GGML_OP_VIEW) {
         return false;
     }
+
     if (node->op != graph_node_properties->node_op) {
         return false;
     }
+
     for (int i = 0; i < GGML_MAX_DIMS; i++) {
         if (node->ne[i] != graph_node_properties->ne[i]) {
             return false;
@@ -2222,17 +2234,31 @@ static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_gra
             return false;
         }
     }
+
     for (int i = 0; i < GGML_MAX_SRC; i++) {
-        if (node->src[i] &&
-            node->src[i]->data != graph_node_properties->src_address[i] &&
-            node->op != GGML_OP_VIEW
-        ) {
-            return false;
+        if (node->src[i]) {
+            if (node->src[i]->data != graph_node_properties->src_address[i] &&
+                node->op != GGML_OP_VIEW) {
+                return false;
+            }
+
+            for (int d = 0; d < GGML_MAX_DIMS; d++) {
+                if (node->src[i]->ne[d] != graph_node_properties->src_ne[i][d]) {
+                    return false;
+                }
+                if (node->src[i]->nb[d] != graph_node_properties->src_nb[i][d]) {
+                    return false;
+                }
+            }
+        } else {
+            if (graph_node_properties->src_address[i] != nullptr) {
+                return false;
+            }
         }
     }
-    if (node->op == GGML_OP_SCALE &&
-        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
-        return false;
+
+    if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU) {
+        return memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
     }
     return true;
 }

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -29,6 +29,108 @@
 
 #define NELEMS(x) sizeof(x) / sizeof(*x)
 
+template<size_t(*Fn)(size_t,size_t,size_t)>
+static inline size_t kernel_offs_fn3(size_t a, size_t b, size_t c) {
+    return Fn(a, b, c);
+}
+
+template<size_t(*Fn)(size_t,size_t)>
+static inline size_t kernel_offs_fn2(size_t a, size_t b, size_t) {
+    return Fn(a, b);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
+static inline void kernel_run_fn11(size_t m, size_t n, size_t k, size_t bl,
+                                     const void* lhs, const void* rhs, void* dst,
+                                     size_t dst_stride_row, size_t dst_stride_col,
+                                     float clamp_min, float clamp_max) {
+    Fn(m, n, k, bl, lhs, rhs, static_cast<float*>(dst), dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,void*,size_t,size_t,float,float)>
+static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
+                                   const void* lhs, const void* rhs, void* dst,
+                                   size_t dst_stride_row, size_t dst_stride_col,
+                                   float clamp_min, float clamp_max) {
+    Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
+static inline size_t lhs_ps_fn6(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
+    return Fn(m, k, bl, mr, kr, sr);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
+static inline size_t lhs_ps_fn5(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
+    return Fn(m, k, mr, kr, sr);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t)>
+static inline size_t lhs_offs_fn6(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr) {
+    return Fn(m_idx, k, bl, mr, kr, sr);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
+static inline size_t lhs_offs_fn5(size_t m_idx, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr) {
+    return Fn(m_idx, k, mr, kr, sr);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const float*,size_t,void*)>
+static inline void lhs_pack_float_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
+                                            size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
+    Fn(m, k, bl, mr, kr, sr, m_idx_start, static_cast<const float*>(lhs), lhs_stride, lhs_packed);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
+static inline void lhs_pack_void_fn10(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
+                                           size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
+    Fn(m, k, bl, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,const void*,size_t,void*)>
+static inline void lhs_pack_void_fn9(size_t m, size_t k, size_t /*bl*/, size_t mr, size_t kr, size_t sr,
+                                             size_t m_idx_start, const void* lhs, size_t lhs_stride, void* lhs_packed) {
+    Fn(m, k, mr, kr, sr, m_idx_start, lhs, lhs_stride, lhs_packed);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t,size_t)>
+static inline size_t rhs_ps_fn5(size_t n, size_t k, size_t nr, size_t kr, size_t bl) {
+    return Fn(n, k, nr, kr, bl);
+}
+
+template<size_t(*Fn)(size_t,size_t)>
+static inline size_t rhs_ps_fn2(size_t n, size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
+    return Fn(n, k);
+}
+
+template<size_t(*Fn)(size_t,size_t,size_t,size_t)>
+static inline size_t rhs_stride_fn4(size_t k, size_t nr, size_t kr, size_t bl) {
+    return Fn(k, nr, kr, bl);
+}
+
+template<size_t(*Fn)(size_t)>
+static inline size_t rhs_stride_fn1(size_t k, size_t /*nr*/, size_t /*kr*/, size_t /*bl*/) {
+    return Fn(k);
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const uint8_t*,const float*,void*,size_t,const struct kai_rhs_pack_qs4cxs1s0_param*)>
+static inline void rhs_pack_fn12(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
+                                      size_t /*rhs_stride*/, const void* rhs, const void* bias, const void* /*scale*/,
+                                      void* rhs_packed, size_t extra_bytes, const void* params) {
+    Fn(num_groups, n, k, nr, kr, sr, bl,
+       static_cast<const uint8_t*>(rhs),
+       static_cast<const float*>(bias),
+       rhs_packed, extra_bytes,
+       static_cast<const kai_rhs_pack_qs4cxs1s0_param*>(params));
+}
+
+template<void(*Fn)(size_t,size_t,size_t,size_t,size_t,size_t,size_t,const void*,const void*,const void*,void*,size_t,const void*)>
+static inline void rhs_pack_fn13(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t /*bl*/,
+                                               size_t rhs_stride, const void* rhs, const void* bias, const void* scale,
+                                               void* rhs_packed, size_t extra_bytes, const void* params) {
+    Fn(num_groups, n, k, nr, kr, sr, rhs_stride, rhs, bias, scale, rhs_packed, extra_bytes, params);
+}
+
 static const size_t INT4_PER_BYTE = 2;
 static const size_t INT4_BITS     = 4;
 static const int Q4_0_ZERO_POINT  = 8;
@@ -122,17 +224,18 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa>,
         },
+
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -142,23 +245,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32_neon>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
-            /* .to_float      = */ dequantize_row_qsi4c32ps1s0scalef16,
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
+            /* .to_float              = */ dequantize_row_qsi4c32ps1s0scalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -174,17 +278,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
+            /* .run_kernel_ex         = */ &kernel_run_fn10<kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
         },
         /* SME GEMV */
         /* .kern_info = */ {
@@ -194,23 +298,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_lhs_offset_ex     = */ nullptr,
+            /* .get_rhs_packed_offset_ex = */ nullptr,
+            /* .run_kernel_ex         = */ nullptr,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme,
-            /* .pack_func             = */ kai_run_lhs_pack_bf16p2vlx2_f32_sme,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
+            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
-            /* .packed_stride = */ NULL,
-            /* .pack_func     = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
-            /* .to_float      = */ NULL,
+            /* .packed_stride         = */ nullptr,
+            /* .to_float              = */ nullptr,
+            /* .packed_size_ex        = */ &rhs_ps_fn2<kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn1<kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
+            /* .pack_func_ex          = */ &rhs_pack_fn13<kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme>,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -229,17 +334,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -249,23 +354,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -283,17 +389,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -303,23 +409,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -338,17 +445,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
         },
         /* i8mm GEMV */
         /* .kern_info = */ {
@@ -358,23 +465,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -392,17 +500,17 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
         },
         /* .gemm_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* DOTPROD GEMV */
         /* .kern_info = */ {
@@ -412,23 +520,24 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_lhs_offset        = */ kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_rhs_packed_offset = */ kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
             /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .run_kernel            = */ kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
         },
         /* .gemv_lhs_info = */ {
             /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset     = */ kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .packed_size           = */ kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32,
-            /* .pack_func             = */ kai_run_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
         },
         /* .rhs_info = */ {
-            /* .packed_size   = */ kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .packed_stride = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .pack_func     = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float      = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
         /* .lhs_type           = */ GGML_TYPE_F32,
@@ -443,6 +552,7 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
     ggml_kleidiai_kernels * kernel = nullptr;
 
     if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
         for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
             if ((cpu_features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
                 gemm_gemv_kernels[i].lhs_type == tensor->src[1]->type &&
@@ -452,6 +562,7 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
                 break;
             }
         }
+#endif
     }
 
     return kernel;
@@ -460,12 +571,14 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {
         if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
             kernels = &gemm_gemv_kernels[i];
             break;
         }
     }
+#endif
 
     return kernels;
 }

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -4,8 +4,6 @@
 
 #pragma once
 
-#include <functional>
-#include <variant>
 #include "ggml.h"
 
 enum cpu_feature {
@@ -15,6 +13,7 @@ enum cpu_feature {
     CPU_FEATURE_SVE     = 4,
     CPU_FEATURE_SME     = 8
 };
+
 inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
     lhs = static_cast<cpu_feature>(lhs | rhs);
     return lhs;
@@ -30,63 +29,52 @@ struct kernel_info {
     size_t (*get_nr)(void);
     size_t (*get_kr)(void);
     size_t (*get_sr)(void);
-    std::variant<
-        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
-        std::function<size_t(size_t m_idx, size_t k)>
-    > get_lhs_offset;
-    std::variant<
-        std::function<size_t(size_t n_idx, size_t k, size_t bl)>,
-        std::function<size_t(size_t n_idx, size_t k)>
-    > get_rhs_packed_offset;
+
     size_t (*get_dst_offset)(size_t m_idx, size_t n_idx, size_t stride);
     size_t (*get_dst_size)(size_t m, size_t n);
-    std::variant<
-        std::function<void(size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed,
-            float* dst, size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max)>,
-        std::function<void(size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row,
-            size_t dst_stride_col, float clamp_min, float clamp_max)>
-    > run_kernel;
+
+    size_t (*get_lhs_offset_ex)(size_t m_idx, size_t k, size_t bl);
+
+    size_t (*get_rhs_packed_offset_ex)(size_t n_idx, size_t k, size_t bl);
+
+    void (*run_kernel_ex)(
+        size_t m, size_t n, size_t k, size_t bl,
+        const void* lhs_packed, const void* rhs_packed,
+        void* dst, size_t dst_stride_row, size_t dst_stride_col,
+        float clamp_min, float clamp_max);
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-    std::variant<
-        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
-        std::function<size_t(size_t m_idx, size_t k, size_t mr, size_t kr, size_t sr)>
-    > get_packed_offset;
-    std::variant<
-        std::function<size_t(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr)>,
-        std::function<size_t(size_t m, size_t k, size_t mr, size_t kr, size_t sr)>
-    > packed_size;
-    std::variant<
-        std::function<void(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const float* lhs,
-            size_t lhs_stride, void* lhs_packed)>,
-        std::function<void(size_t m, size_t k, size_t mr, size_t kr, size_t sr, size_t m_idx_start, const void* lhs, size_t lhs_stride,
-        void* lhs_packed)>
-    > pack_func;
+
+    size_t (*get_packed_offset_ex)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
+
+    size_t (*packed_size_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
+
+    void (*pack_func_ex)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr,
+        size_t m_idx_start, const void * lhs, size_t lhs_stride, void * lhs_packed);
 };
 
 struct rhs_packing_info {
-    std::variant<
-        std::function<size_t(size_t n, size_t k, size_t nr, size_t kr, size_t bl)>,
-        std::function<size_t(size_t n, size_t k)>
-    > packed_size;
     size_t (*packed_stride)(size_t k, size_t nr, size_t kr, size_t bl);
-    std::variant<
-        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl, const uint8_t* rhs,
-            const float* bias, void* rhs_packed, size_t extra_bytes, const struct kai_rhs_pack_qs4cxs1s0_param* params)>,
-        std::function<void(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs,
-            const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params)>
-    > pack_func;
-    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out, size_t nr_pack, size_t packed_row_stride,
-          size_t kr, size_t bl, size_t num_bytes_multiplier);
+
+    void (*to_float)(const void *packed_data, int32_t row_idx, int64_t nc, float *out,
+                     size_t nr_pack, size_t packed_row_stride, size_t kr, size_t bl,
+                     size_t num_bytes_multiplier);
+
+    size_t (*packed_size_ex)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
+
+    size_t (*packed_stride_ex)(size_t k, size_t nr, size_t kr, size_t bl);
+
+    void (*pack_func_ex)(size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t bl,
+        size_t rhs_stride, const void * rhs, const void * bias, const void * scale, void * rhs_packed, size_t extra_bytes, const void * params);
 };
 
 struct ggml_kleidiai_kernels {
-    kernel_info gemm;
+    kernel_info      gemm;
     lhs_packing_info gemm_lhs_info;
 
-    kernel_info gemv;
+    kernel_info      gemv;
     lhs_packing_info gemv_lhs_info;
 
     rhs_packing_info rhs_info;

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -8,6 +8,7 @@
 #include <stdexcept>
 #include <stdint.h>
 #include <string.h>
+#include <string>
 #if defined(__linux__)
 #include <asm/hwcap.h>
 #include <sys/auxv.h>
@@ -87,40 +88,6 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
-template <typename Variant, typename Ret, typename... Args, std::size_t... Is>
-constexpr bool variant_any_invocable_impl(std::index_sequence<Is...>) {
-    using V = std::remove_reference_t<Variant>;
-    return (std::is_invocable_r_v<
-                Ret,
-                std::variant_alternative_t<Is, V>,
-                Args...> || ...);
-}
-
-template <typename Variant, typename Ret, typename... Args>
-constexpr bool variant_any_invocable_v =
-    variant_any_invocable_impl<Variant, Ret, Args...>(
-        std::make_index_sequence<
-            std::variant_size_v<std::remove_reference_t<Variant>>>{});
-
-template<typename Ret, typename Variant, typename... Args>
-static inline Ret variant_call(Variant && var, Args&&... args) {
-    static_assert(variant_any_invocable_v<std::remove_reference_t<Variant>, Ret, Args...>,
-                  "No alternative in Variant is invocable with the provided arguments and return type.");
-
-    return std::visit(
-        [&](auto && f) -> Ret {
-            using F = std::decay_t<decltype(f)>;
-            if constexpr (std::is_invocable_r_v<Ret, F, Args...>) {
-                return std::invoke(std::forward<decltype(f)>(f), std::forward<Args>(args)...);
-            } else {
-                GGML_ABORT("Invalid function type in variant_call");
-                GGML_UNREACHABLE();
-            }
-        },
-        std::forward<Variant>(var)
-    );
-}
-
 namespace ggml::cpu::kleidiai {
 
 static size_t round_down(size_t x, size_t y) {
@@ -145,7 +112,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             return false;
         }
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
         bool is_gemv = op->src[1]->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
@@ -159,16 +128,18 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         size_t sr = kernel->get_sr();
 
         if (kernels->rhs_type == GGML_TYPE_Q4_0) {
-            size = variant_call<size_t>(lhs_info->packed_size, m, k, QK4_0, mr, kr, sr);
+            if (!lhs_info->packed_size_ex) return false;
+            size = lhs_info->packed_size_ex(m, k, QK4_0, mr, kr, sr);
         } else if (kernels->rhs_type == GGML_TYPE_F16) {
+            if (!lhs_info->packed_size_ex || !kernels->rhs_info.packed_size_ex) return false;
             const int64_t lhs_batch_size0 = op->src[1]->ne[2];
             const int64_t rhs_batch_size0 = op->src[0]->ne[2];
             const int64_t r = lhs_batch_size0 / rhs_batch_size0;
-            size = variant_call<size_t>(lhs_info->packed_size, m * r, k, mr, kr, sr) +
-                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
+            size = lhs_info->packed_size_ex(m * r, k, 0, mr, kr, sr) +
+                   kernels->rhs_info.packed_size_ex(n, k, kernel->get_nr(), kernel->get_kr(), 0) +
                    k * n * sizeof(float) + n * sizeof(float);
         } else {
-            GGML_ASSERT(false);
+            return false;
         }
 
         return true;
@@ -196,12 +167,18 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
 
         const bool is_gemv = src1->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
         GGML_ASSERT(kernel);
+        if (!kernels->rhs_info.pack_func_ex ||
+            !kernel->get_lhs_offset_ex || !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex) {
+            return false;
+        }
 
         const int nth = params->nth;
         const int ith = params->ith;
@@ -228,10 +205,10 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const int64_t kr = (int64_t) kernel->get_kr();
         const int64_t sr = (int64_t) kernel->get_sr();
 
-        const size_t lhs_packed_size = variant_call<size_t>(lhs_info->packed_size, (size_t)m, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, (size_t)n, (size_t)k);
-        const size_t kxn_size        = (size_t)k * (size_t)n * sizeof(float);
-        const size_t bias_size       = (size_t)n * sizeof(float);
+        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
+        const size_t rhs_packed_size = kernels->rhs_info.packed_size_ex(n, k, nr, kr, 0);
+        const size_t kxn_size        = k * n * sizeof(float);
+        const size_t bias_size       = n * sizeof(float);
 
         const size_t wsize_required = lhs_packed_size + rhs_packed_size + kxn_size + bias_size;
         GGML_ASSERT(wsize_required <= params->wsize);
@@ -259,10 +236,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                     const int64_t m_count = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
 
                     // Base packed offset (aligned) and per-row stride in bytes
-                    const size_t base_packed_off = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)m_start, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-                    const size_t next_block_off = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)(m_start + mr), (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
+                    const size_t base_packed_off  = lhs_info->get_packed_offset_ex(m_start, k, 0, mr, kr, sr);
+                    const size_t next_block_off   = lhs_info->get_packed_offset_ex(m_start + mr, k, 0, mr, kr, sr);
                     const size_t row_stride_bytes = (next_block_off - base_packed_off) / (size_t)mr;
 
                     int64_t remaining = m_count;
@@ -278,9 +253,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                         const size_t dst_off = base_packed_off + (size_t)(cur - m_start) * row_stride_bytes;
                         void * dst_ptr       = lhs_packed + dst_off;
 
-                        variant_call<void>(lhs_info->pack_func,
-                                        (size_t)take, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr,
-                                        /*m_idx_start*/ 0, src_ptr, lhs_stride, dst_ptr);
+                        lhs_info->pack_func_ex(take, k, 0, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
 
                         cur       += take;
                         remaining -= take;
@@ -296,10 +269,8 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                                         reinterpret_cast<const uint16_t *>(rhs_batch_base),
                                         rhs_stride);
 
-                variant_call<void>(kernels->rhs_info.pack_func,
-                                   /*num_groups*/ 1, (size_t)n, (size_t)k, (size_t)nr, (size_t)kr, (size_t)sr,
-                                   /*rhs_stride (bytes)*/ (size_t)(n * sizeof(float)),
-                                   rhs_kxn, bias, nullptr, rhs_packed, /*extra_bytes*/ 0, /*params*/ nullptr);
+                kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, 0, n * sizeof(float),
+                             rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
@@ -320,20 +291,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                     const int64_t n_to_process = (ith == num_threads_n - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
 
                     // LHS packed base at row 0 (consistent with packing above)
-                    const size_t lhs_packed_offset0 = variant_call<size_t>(
-                        lhs_info->get_packed_offset, (size_t)0, (size_t)k, (size_t)mr, (size_t)kr, (size_t)sr);
-                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, (size_t)n_start, (size_t)k);
-                    const size_t dst_offset        = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
+                    const size_t lhs_packed_offset0 = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+                    const size_t rhs_packed_offset  = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
+                    const size_t dst_offset         = kernel->get_dst_offset((size_t)0, (size_t)n_start, dst_stride);
 
                     const void * lhs_ptr = lhs_packed + lhs_packed_offset0;
                     const void * rhs_ptr = rhs_packed + rhs_packed_offset;
                     float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);
 
-                    variant_call<void>(kernel->run_kernel,
-                                       (size_t)m, (size_t)n_to_process, (size_t)k,
-                                       lhs_ptr, rhs_ptr,
-                                       dst_ptr, dst_stride, sizeof(float),
-                                       -FLT_MAX, FLT_MAX);
+                    kernel->run_kernel_ex(m, n_to_process, k, 0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
                 }
             }
 
@@ -354,13 +320,19 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         GGML_TENSOR_BINARY_OP_LOCALS
 
         ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
-        GGML_ASSERT(kernels);
+        if (!kernels) {
+            return false;
+        }
 
         bool is_gemv = src1->ne[1] == 1;
         kernel_info * kernel = is_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = is_gemv ? &kernels->gemv_lhs_info : &kernels->gemm_lhs_info;
 
         GGML_ASSERT(kernel);
+        if (!lhs_info->get_packed_offset_ex || !lhs_info->pack_func_ex ||
+            !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
+            return false;
+        }
 
         const int ith = params->ith;
         const int nth_raw = params->nth;
@@ -402,25 +374,26 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             // Transform LHS
             const size_t src_stride        = src1->nb[1];
             const float * src_ptr          = reinterpret_cast<const float *>(lhs + lhs_info->get_offset(m_start, dst->src[1]->nb[1]));
-            const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, m_start, k, QK4_0, mr, kr, sr);
+            const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(m_start, k, QK4_0, mr, kr, sr);
             void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
 
-            variant_call<void>(lhs_info->pack_func, m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+            // Pack this thread's chunk with m_idx_start = 0 and per-thread output pointer
+            lhs_info->pack_func_ex(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
         }
 
         ggml_barrier(params->threadpool);
 
         // Perform the operation
         const size_t dst_stride        = dst->nb[1];
-        const size_t lhs_packed_offset = variant_call<size_t>(lhs_info->get_packed_offset, 0, k, QK4_0, mr, kr, sr);
-        const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k, QK4_0);
+        const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, QK4_0, mr, kr, sr);
+        const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, QK4_0);
         const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
         const void * rhs_ptr           = static_cast<const void *>(rhs_packed + rhs_packed_offset);
         const void* lhs_ptr            = (const void*)((const char *)lhs_packed + lhs_packed_offset);
         float *dst_ptr                 = reinterpret_cast<float *>(static_cast<uint8_t *>(dst->data) + dst_offset);
 
         if (n_to_process > 0) {
-            variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+            kernel->run_kernel_ex(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
                                sizeof(float), -FLT_MAX, FLT_MAX);
         }
 
@@ -429,7 +402,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
     bool compute_forward_get_rows(struct ggml_compute_params * params, struct ggml_tensor * dst) {
         GGML_ASSERT(dst->src[0]->type == GGML_TYPE_Q4_0);
-        GGML_ASSERT(ctx.kernels);
+        if (!ctx.kernels) {
+            return false;
+        }
 
         const ggml_tensor * src0 = dst->src[0];
         const ggml_tensor * src1 = dst->src[1];
@@ -438,6 +413,9 @@ class tensor_traits : public ggml::cpu::tensor_traits {
 
         rhs_packing_info * rhs_info = &ctx.kernels->rhs_info;
         kernel_info * kernel        = &ctx.kernels->gemm;
+        if (!rhs_info->to_float || !kernel->get_nr) {
+            return false;
+        }
 
         const int64_t nc     = ne00;
         const int64_t nr     = ggml_nelements(src1);
@@ -480,7 +458,7 @@ public:
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        variant_call<void>(ctx.kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, QK4_0, (const uint8_t*)data, nullptr, tensor->data, 0, &params);
+        ctx.kernels->rhs_info.pack_func_ex(1, n, k, nr, kr, sr, QK4_0, 0, (const uint8_t*)data, nullptr, nullptr, tensor->data, 0, &params);
 
         return 0;
         GGML_UNUSED(data_size);
@@ -548,7 +526,7 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alloc_size(ggml_backend_
     const size_t nr = ctx.kernels->gemm.get_nr();
     const size_t kr = ctx.kernels->gemm.get_kr();
 
-    return variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
+    return ctx.kernels->rhs_info.packed_size_ex(n, k, nr, kr, QK4_0);
 
     GGML_UNUSED(buft);
 }

ggml/src/ggml-cpu/ops.cpp

@@ -3467,31 +3467,27 @@ static void ggml_compute_forward_norm_f32(
 
     GGML_ASSERT(eps >= 0.0f);
 
-    // TODO: optimize
     for (int64_t i03 = 0; i03 < ne03; i03++) {
         for (int64_t i02 = 0; i02 < ne02; i02++) {
             for (int64_t i01 = ith; i01 < ne01; i01 += nth) {
                 const float * x = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
-                ggml_float sum = 0.0;
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    sum += (ggml_float)x[i00];
-                }
-
+                float sum = 0.0;
+                ggml_vec_sum_f32(ne00, &sum, x);
                 float mean = sum/ne00;
 
                 float * y = (float *) ((char *) dst->data + i01*nb1 + i02*nb2 + i03*nb3);
+                float variance = 0;
 
-                ggml_float sum2 = 0.0;
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    float v = x[i00] - mean;
-                    y[i00] = v;
-                    sum2 += (ggml_float)(v*v);
-                }
+#ifdef GGML_USE_ACCELERATE
+                mean = -mean;
+                vDSP_vsadd(x, 1, &mean, y, 1, ne00);
+                vDSP_measqv(y, 1, &variance, ne00);
+#else
+                variance = ggml_vec_cvar_f32(ne00, y, x, mean);
+#endif //GGML_USE_ACCELERATE
 
-                float variance = sum2/ne00;
                 const float scale = 1.0f/sqrtf(variance + eps);
-
                 ggml_vec_scale_f32(ne00, y, scale);
             }
         }

ggml/src/ggml-cpu/vec.cpp

@@ -404,6 +404,72 @@ void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float *
     }
 }
 
+ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
+    int i = 0;
+    ggml_float sum = 0;
+// TODO: optimize to process the remaining elements in groups using the smaller vector sizes from AVX2 and SSE
+// ref: https://github.com/ggml-org/llama.cpp/pull/15953#pullrequestreview-3310928344
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        __m512 val = _mm512_sub_ps(_mm512_loadu_ps(x + i),
+                                   _mm512_set1_ps(mean));
+        _mm512_storeu_ps(y + i, val);
+        sum += (ggml_float)_mm512_reduce_add_ps(_mm512_mul_ps(val, val));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        __m256 val = _mm256_sub_ps(_mm256_loadu_ps(x + i),
+                                   _mm256_set1_ps(mean));
+        _mm256_storeu_ps(y + i, val);
+        val = _mm256_mul_ps(val,val);
+        __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
+                                 _mm256_castps256_ps128(val));
+        val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
+        val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
+        sum += (ggml_float)_mm_cvtss_f32(val2);
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        __m128 val = _mm_sub_ps(_mm_loadu_ps(x + i),
+                                _mm_set1_ps(mean));
+        _mm_storeu_ps(y + i, val);
+        val = _mm_mul_ps(val, val);
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+        val = _mm_add_ps(val, _mm_movehl_ps(val, val));
+        val = _mm_add_ss(val, _mm_movehdup_ps(val));
+#else
+        __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
+        val = _mm_add_ps(val, tmp);
+        tmp = _mm_movehl_ps(tmp, val);
+        val = _mm_add_ss(val, tmp);
+#endif  // __AVX__ || __AVX2__ || __AVX512F__
+        sum += (ggml_float)_mm_cvtss_f32(val);
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        float32x4_t val = vsubq_f32(vld1q_f32(x + i),
+                                    vdupq_n_f32(mean));
+        vst1q_f32(y + i, val);
+        val = vmulq_f32(val, val);
+        sum += (ggml_float)vaddvq_f32(val);
+    }
+#elif defined(__VXE__) || defined(__VXE2__)
+    for (; i + 3 < n; i += 4) {
+        float32x4_t val = vec_sub(vec_xl(0, x + i), vec_splats(mean));
+        vec_xst(val, 0, y + i);
+        val = vec_mul(val, val);
+        sum += (ggml_float)vec_hsum_f32x4(val);
+    }
+#endif
+    for (; i < n; ++i) {
+        float val = x[i] - mean;
+        val *= val;
+        sum += (ggml_float)val;
+        y[i] = val;
+    }
+    return sum/n;
+}
+
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
     int i = 0;
     ggml_float sum = 0;

ggml/src/ggml-cpu/vec.h

@@ -44,6 +44,7 @@ void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t *
 void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
 
 void ggml_vec_silu_f32(const int n, float * y, const float * x);
+ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean); //it will also center y ( y = y - mean )
 ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max);
 ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max);
 
@@ -143,14 +144,14 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         for (int i = 0; i < np; i += ggml_f16_step) {
             ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0); // 8 elements
 
-            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elemnst
+            ax1 = GGML_F16x_VEC_LOAD(x[0] + i + 0*ggml_f16_epr, 0); // 8 elements
             sum_00 = GGML_F16x_VEC_FMA(sum_00, ax1, ay1);     // sum_00 = sum_00+ax1*ay1
             ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 0*ggml_f16_epr, 0); // 8 elements
             sum_10 = GGML_F16x_VEC_FMA(sum_10, ax1, ay1);
 
             ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1); // next 8 elements
 
-            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 ekements
+            ax2 = GGML_F16x_VEC_LOAD(x[0] + i + 1*ggml_f16_epr, 1); // next 8 elements
             sum_01 = GGML_F16x_VEC_FMA(sum_01, ax2, ay2);
             ax2 = GGML_F16x_VEC_LOAD(x[1] + i + 1*ggml_f16_epr, 1);
             sum_11 = GGML_F16x_VEC_FMA(sum_11, ax2, ay2);
@@ -159,7 +160,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
             ax3 = GGML_F16x_VEC_LOAD(x[0] + i + 2*ggml_f16_epr, 2);
             sum_02 = GGML_F16x_VEC_FMA(sum_02, ax3, ay3);
-            ax1 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
+            ax3 = GGML_F16x_VEC_LOAD(x[1] + i + 2*ggml_f16_epr, 2);
             sum_12 = GGML_F16x_VEC_FMA(sum_12, ax3, ay3);
 
             ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
@@ -819,7 +820,8 @@ inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_f
 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] : expm1f(x[i]); }
 inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
+        const float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : expm1f(v));
     }
 }
 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; }

ggml/src/ggml-cuda/CMakeLists.txt

@@ -44,6 +44,8 @@ if (CUDAToolkit_FOUND)
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
 
     file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-tile*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_CUDA ${SRCS})
     file(GLOB   SRCS "template-instances/mmq*.cu")

ggml/src/ggml-cuda/common.cuh

@@ -245,7 +245,8 @@ static bool fp16_available(const int cc) {
 }
 
 static bool fast_fp16_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610);
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
@@ -571,6 +572,10 @@ static __device__ __forceinline__ void ggml_cuda_mad(half2 & acc, const half2 v,
 }
 
 // Aligned memory transfers of 8/16 bytes can be faster than 2 transfers with 4 bytes, especially on AMD.
+// Important: do not use this function if dst and src both point at registers.
+//     Due to the strict aliasing rule the compiler can do incorrect optimizations if src and dst have different types.
+//     The function is intended for copies between registers and SRAM/VRAM to make the compiler emit the right instructions.
+//     If dst and src point at different address spaces then they are guaranteed to not be aliased.
 template <int nbytes, int alignment = 0>
 static __device__ __forceinline__ void ggml_cuda_memcpy_1(void * __restrict__ dst, const void * __restrict__ src) {
     if constexpr (alignment != 0) {

ggml/src/ggml-cuda/fattn-common.cuh

@@ -793,8 +793,6 @@ void launch_fattn(
     GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
         "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
-    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
-
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
     const int id  = ggml_cuda_get_device();
@@ -878,7 +876,7 @@ void launch_fattn(
     // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
     // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
     //     multiple sequences of possibly different lengths.
-    if (mask && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+    if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
         const int s31 = mask->nb[1] / sizeof(half2);
         const int s33 = mask->nb[3] / sizeof(half2);
 
@@ -916,8 +914,7 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        GGML_ASSERT(K->ne[1] % KQ_row_granularity == 0);
-        const int ntiles_KQ = K->ne[1] / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
 
         // parallel_blocks must not be larger than what the tensor size allows:
         parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -946,7 +943,7 @@ void launch_fattn(
 
         blocks_num.x = ntiles_x;
         blocks_num.y = parallel_blocks;
-        blocks_num.z = Q->ne[2]*Q->ne[3];
+        blocks_num.z = (Q->ne[2]/ncols2)*Q->ne[3];
 
         if (parallel_blocks > 1) {
             dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));

ggml/src/ggml-cuda/fattn-tile.cu

@@ -1,756 +1,45 @@
 #include "common.cuh"
-#include "fattn-common.cuh"
 #include "fattn-tile.cuh"
 #include "fattn-wmma-f16.cuh"
 
-// kq_stride == number of KQ rows to process per iteration
-// kq_nbatch == number of K columns to load in parallel for KQ calculation
-
-static int fattn_tile_get_kq_stride_host(const int D, const int ncols, const int cc, const int warp_size) {
-    if (GGML_CUDA_CC_IS_AMD(cc)) {
-        if (GGML_CUDA_CC_IS_RDNA(cc)) {
-            switch (D) {
-                case 64:
-                    return 128;
-                case 128:
-                case 256:
-                    return ncols <= 16 ? 128 : 64;
-                default:
-                    GGML_ABORT("fatal error");
-                    return -1;
-            }
-        }
-        switch (D) {
-            case 64:
-                return ncols == 32 ? 128 : 64;
-            case 128:
-                return ncols == 32 ? 64 : 32;
-            case 256:
-                return 32;
-            default:
-                GGML_ABORT("fatal error");
-                return -1;
-        }
-    }
-    if (fast_fp16_available(cc)) {
-        switch (D) {
-            case 64:
-            case 128:
-            case 256:
-                return ncols <= 16 ? 128 : 64;
-            default:
-                GGML_ABORT("fatal error");
-                return -1;
-        }
-    }
-    switch (D) {
-        case 64:
-            return ncols <= 16 ? 128 : 64;
-        case 128:
-            return ncols <= 16 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            GGML_ABORT("fatal error");
-            return -1;
-    }
-    GGML_UNUSED(warp_size);
-}
-
-static constexpr __device__ int fattn_tile_get_kq_stride_device(int D, int ncols, int warp_size) {
-#ifdef GGML_USE_HIP
-#ifdef RDNA
-    switch (D) {
-        case 64:
-            return 128;
-        case 128:
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return ncols == 32 ? 128 : 64;
-        case 128:
-            return ncols == 32 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            return -1;
-    }
-#endif // RDNA
-#else
-#ifdef FAST_FP16_AVAILABLE
-    switch (D) {
-        case 64:
-        case 128:
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return ncols <= 16 ? 128 : 64;
-        case 128:
-            return ncols <= 16 ? 64 : 32;
-        case 256:
-            return 32;
-        default:
-            return -1;
-    }
-#endif // FAST_FP16_AVAILABLE
-#endif // GGML_USE_HIP
-    GGML_UNUSED_VARS(ncols, warp_size);
-}
-
-static constexpr __device__ int fattn_tile_get_kq_nbatch_device(int D, int ncols, int warp_size) {
-#ifdef GGML_USE_HIP
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-        case 256:
-            return 128;
-        default:
-            return -1;
-    }
-#else
-#ifdef FAST_FP16_AVAILABLE
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-        case 256:
-            return 128;
-        default:
-            return -1;
-    }
-#else
-    switch (D) {
-        case 64:
-            return 64;
-        case 128:
-            return 128;
-        case 256:
-            return ncols <= 16 ? 128 : 64;
-        default:
-            return -1;
-    }
-#endif // FAST_FP16_AVAILABLE
-#endif // GGML_USE_HIP
-    GGML_UNUSED_VARS(ncols, warp_size);
-}
-
-static int fattn_tile_get_nthreads_host(const int cc, const int ncols) {
-    return 256;
-    GGML_UNUSED_VARS(cc, ncols);
-}
-
-static constexpr __device__ int fattn_tile_get_nthreads_device(int ncols) {
-    return 256;
-    GGML_UNUSED(ncols);
-}
-
-static constexpr __device__ int fattn_tile_get_occupancy_device(int ncols) {
-#ifdef RDNA
-    return 3;
-#else
-    return ncols <= 16 ? 3 : 2;
-#endif // RDNA
-    GGML_UNUSED(ncols);
-}
-
-template<int D, int ncols, bool use_logit_softcap> // D == head size
-__launch_bounds__(fattn_tile_get_nthreads_device(ncols), fattn_tile_get_occupancy_device(ncols))
-static __global__ void flash_attn_tile(
-        const char * __restrict__ Q,
-        const char * __restrict__ K,
-        const char * __restrict__ V,
-        const char * __restrict__ mask,
-        const char * __restrict__ sinks,
-        const int  * __restrict__ KV_max,
-        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 float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
-                            const int32_t nb01, const int32_t nb02, const int32_t nb03,
-        const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
-                            const int32_t nb11, const int32_t nb12, const int64_t nb13,
-                            const int32_t nb21, const int32_t nb22, const int64_t nb23,
-                            const int32_t ne31, const int32_t ne32, const int32_t ne33,
-                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#ifdef FLASH_ATTN_AVAILABLE
-
-    // Skip unused kernel variants for faster compilation:
-#ifdef GGML_USE_WMMA_FATTN
-    NO_DEVICE_CODE;
-    return;
-#endif // GGML_USE_WMMA_FATTN
-
-    if (use_logit_softcap && !(D == 128 || D == 256)) {
-        GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-            max_bias, m0, m1, n_head_log2, logit_softcap,
-            ne00, ne01, ne02, ne03,
-                  nb01, nb02, nb03,
-            ne10, ne11, ne12, ne13,
-                  nb11, nb12, nb13,
-                  nb21, nb22, nb23,
-                  ne31, ne32, ne33,
-                  nb31, nb32, nb33);
-        NO_DEVICE_CODE;
-        return;
-    }
-
-    constexpr int warp_size = 32;
-    constexpr int nwarps    = fattn_tile_get_nthreads_device(ncols) / warp_size;
-    constexpr int kq_stride = fattn_tile_get_kq_stride_device(D, ncols, warp_size);
-    static_assert(kq_stride % warp_size == 0, "kq_stride not divisable by warp_size.");
-    constexpr int kq_nbatch = fattn_tile_get_kq_nbatch_device(D, ncols, warp_size);
-    static_assert(kq_nbatch % (2*warp_size) == 0, "bad kq_nbatch");
-
-    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
-
-    const int ic0 = blockIdx.x * ncols; // Index of the Q/QKV column to work on.
-
-    const int sequence = blockIdx.z / ne02;
-    const int head = blockIdx.z - sequence*ne02;
-    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f    = (const float *) (Q    + nb03* sequence         + nb02* head              + nb01*ic0);
-    const half2 * K_h2   = (const half2 *) (K    + nb13* sequence         + nb12*(head / gqa_ratio));
-    const half2 * V_h2   = (const half2 *) (V    + nb13* sequence         + nb12*(head / gqa_ratio)); // K and V have same shape
-    const half  * maskh  = (const half  *) (mask + nb33*(sequence % ne33)                           + nb31*ic0);
-    const float * sinksf = (const float *) (sinks);
-
-    const int stride_KV2 = nb11 / sizeof(half2);
-
-    const float slope = get_alibi_slope(max_bias, head, n_head_log2, m0, m1);
-
-    constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
-    constexpr int cpy_ne = cpy_nb / 4;
-
-    constexpr int cpw = ncols/nwarps; // cols per warp
-
-    // softmax_iter_j == number of KQ columns for which to calculate softmax in parallel.
-    // KQ is originall 2D but uses a Z-shaped memory pattern for larger reads/writes.
-#ifdef FAST_FP16_AVAILABLE
-    constexpr int softmax_iter_j = cpw < 2*cpy_ne ? cpw : 2*cpy_ne;
-
-    __shared__ half  KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
-    __shared__ half2 Q_tmp[ncols][D/2];
-    __shared__ half2 KV_tmp[kq_stride * (kq_nbatch/2 + cpy_ne)]; // Padded to avoid memory bank conflicts.
-    half2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
-#else
-    constexpr int softmax_iter_j = cpw < 1*cpy_ne ? cpw : 1*cpy_ne;
-
-    __shared__ float KQ[ncols/softmax_iter_j][kq_stride][softmax_iter_j];
-    __shared__ float Q_tmp[ncols][D];
-    __shared__ float KV_tmp[kq_stride * (kq_nbatch + cpy_ne)]; // Padded to avoid memory bank conflicts.
-    float2 VKQ[cpw][D/(2*warp_size)] = {{{0.0f, 0.0f}}};
-#endif // FAST_FP16_AVAILABLE
-    static_assert(cpw % softmax_iter_j == 0, "bad softmax_iter_j");
-
-    float KQ_max[cpw];
-#pragma unroll
-    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
-        KQ_max[j0/nwarps] = -FLT_MAX/2.0f;
-    }
-    float KQ_sum[cpw] = {0.0f};
-
-    // Load Q data, convert to FP16 if fast.
-#pragma unroll
-    for (int j0 = 0; j0 < cpw; ++j0) {
-        const int j = j0 + threadIdx.y*cpw;
-
-        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-
-#pragma unroll
-        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-            float tmp_f[cpy_ne_D] = {0.0f};
-            if (ic0 + j < ne01) {
-                ggml_cuda_memcpy_1<sizeof(tmp_f)>(tmp_f, &Q_f[j*(nb01/sizeof(float)) + i0 + threadIdx.x*cpy_ne_D]);
-            }
-
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
-                tmp_f[i1] *= scale;
-            }
-
-#ifdef FAST_FP16_AVAILABLE
-            half2 tmp_h2[cpy_ne_D/2];
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
-                tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
-            }
-            ggml_cuda_memcpy_1<sizeof(tmp_h2)>(&Q_tmp[j][i0/2 + threadIdx.x*(cpy_ne_D/2)], tmp_h2);
-#else
-            ggml_cuda_memcpy_1<sizeof(tmp_f)> (&Q_tmp[j][i0   + threadIdx.x* cpy_ne_D],    tmp_f);
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-    __syncthreads();
-
-    // Main loop over KV cache:
-    const int k_VKQ_max = KV_max ? KV_max[sequence*gridDim.x + blockIdx.x] : ne11;
-    for (int k_VKQ_0 = blockIdx.y*kq_stride; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*kq_stride) {
-        // Calculate KQ tile and keep track of new maximum KQ values:
-
-        float KQ_max_new[cpw];
-#pragma unroll
-        for (int j = 0; j < cpw; ++j) {
-            KQ_max_new[j] = KQ_max[j];
-        }
-
-        float KQ_acc[kq_stride/warp_size][cpw] = {{0.0f}}; // Accumulators for KQ matrix multiplication.
-
-        // KQ = K @ Q matrix multiplication:
-#pragma unroll
-        for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += kq_nbatch) {
-#pragma unroll
-            for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += nwarps) {
-                const int i_KQ = i_KQ_0 + threadIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/(2*warp_size) ? cpy_ne : kq_nbatch/(2*warp_size);
-#pragma unroll
-                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += warp_size*cpy_ne_kqnb) {
-                    ggml_cuda_memcpy_1<cpy_ne_kqnb*4>(
-                        &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb],
-                        &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1 + threadIdx.x*cpy_ne_kqnb]);
-                }
-#else
-                constexpr int cpy_ne_kqnb = cpy_ne < kq_nbatch/warp_size ? cpy_ne : kq_nbatch/warp_size;
-#pragma unroll
-                for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += warp_size*cpy_ne_kqnb) {
-                    half2 tmp_h2[cpy_ne_kqnb/2];
-                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
-                        tmp_h2, &K_h2[int64_t(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + k_KQ_1/2 + threadIdx.x*(cpy_ne_kqnb/2)]);
-
-                    float2 tmp_f2[cpy_ne_kqnb/2];
-#pragma unroll
-                    for (int k_KQ_2 = 0; k_KQ_2 < cpy_ne_kqnb/2; ++k_KQ_2) {
-                        tmp_f2[k_KQ_2] = __half22float2(tmp_h2[k_KQ_2]);
-                    }
-                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
-                        &KV_tmp[i_KQ*(kq_nbatch + cpy_ne) + k_KQ_1 + threadIdx.x*cpy_ne_kqnb], tmp_f2);
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-            __syncthreads();
-
-#ifdef FAST_FP16_AVAILABLE
-#pragma unroll
-            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch/2; k_KQ_1 += cpy_ne) {
-                half2 K_k[kq_stride/warp_size][cpy_ne];
-                half2 Q_k[cpw][cpy_ne];
-#else
-#pragma unroll
-            for (int k_KQ_1 = 0; k_KQ_1 < kq_nbatch; k_KQ_1 += cpy_ne) {
-                float K_k[kq_stride/warp_size][cpy_ne];
-                float Q_k[cpw][cpy_ne];
-#endif // FAST_FP16_AVAILABLE
-
-#pragma unroll
-                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-                    const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#ifdef FAST_FP16_AVAILABLE
-                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch/2 + cpy_ne) + k_KQ_1]);
-#else
-                    ggml_cuda_memcpy_1<cpy_nb>(&K_k[i_KQ_0/warp_size], &KV_tmp[i_KQ*(kq_nbatch   + cpy_ne) + k_KQ_1]);
-#endif // FAST_FP16_AVAILABLE
-                }
-#pragma unroll
-                for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-                    const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
-
-#ifdef FAST_FP16_AVAILABLE
-                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0/2 + k_KQ_1]);
-#else
-                    ggml_cuda_memcpy_1<cpy_nb>(&Q_k[j_KQ_0], &Q_tmp[j_KQ][k_KQ_0   + k_KQ_1]);
-#endif // FAST_FP16_AVAILABLE
-                }
-
-#pragma unroll
-                for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-#pragma unroll
-                    for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-#pragma unroll
-                        for (int k = 0; k < cpy_ne; ++k) {
-                            ggml_cuda_mad(KQ_acc[i_KQ_0/warp_size][j_KQ_0], K_k[i_KQ_0/warp_size][k], Q_k[j_KQ_0][k]);
-                        }
-                    }
-                }
-            }
-
-            if (k_KQ_0 + kq_nbatch < D) {
-                __syncthreads(); // Sync not needed on last iteration.
-            }
-        }
-
-        // Apply logit softcap, mask, update KQ_max:
-#pragma unroll
-        for (int i_KQ_0 = 0; i_KQ_0 < kq_stride; i_KQ_0 += warp_size) {
-            const int i_KQ = i_KQ_0 + threadIdx.x;
-
-#pragma unroll
-            for (int j_KQ_0 = 0; j_KQ_0 < cpw; ++j_KQ_0) {
-                const int j_KQ = j_KQ_0 + threadIdx.y*cpw;
-
-                if (use_logit_softcap) {
-                    KQ_acc[i_KQ_0/warp_size][j_KQ_0] = logit_softcap * tanhf(KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
-                }
-
-                KQ_acc[i_KQ_0/warp_size][j_KQ_0] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
-
-                KQ_max_new[j_KQ_0] = fmaxf(KQ_max_new[j_KQ_0], KQ_acc[i_KQ_0/warp_size][j_KQ_0]);
-            }
-        }
-
-        __syncthreads();
-
-        // Calculate KQ softmax, write to shared KQ buffer, re-scale VKQ accumulators:
-#pragma unroll
-        for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-#ifdef FAST_FP16_AVAILABLE
-            half  tmp[kq_stride/warp_size][softmax_iter_j];
-#else
-            float tmp[kq_stride/warp_size][softmax_iter_j];
-#endif // FAST_FP16_AVAILABLE
-
-#pragma unroll
-            for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
-                KQ_max_new[j0+j1] = warp_reduce_max<warp_size>(KQ_max_new[j0+j1]);
-                const float KQ_max_scale = expf(KQ_max[j0+j1] - KQ_max_new[j0+j1]);
-                KQ_max[j0+j1] = KQ_max_new[j0+j1];
-
-                float KQ_sum_add = 0.0f;
-#pragma unroll
-                for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
-                    const float val = expf(KQ_acc[i0/warp_size][j0+j1] - KQ_max[j0+j1]);
-                    KQ_sum_add += val;
-                    tmp[i0/warp_size][j1] = val;
-                }
-                KQ_sum[j0+j1] = KQ_sum[j0+j1]*KQ_max_scale + KQ_sum_add;
-
-#ifdef FAST_FP16_AVAILABLE
-                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                    VKQ[j0+j1][i0/warp_size] *= KQ_max_scale_h2;
-                }
-#else
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                    VKQ[j0+j1][i0/warp_size].x *= KQ_max_scale;
-                    VKQ[j0+j1][i0/warp_size].y *= KQ_max_scale;
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-#pragma unroll
-            for (int i0 = 0; i0 < kq_stride; i0 += warp_size) {
-                const int i = i0 + threadIdx.x;
-
-                ggml_cuda_memcpy_1<sizeof(tmp[0])>(
-                    KQ[j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j)][i], tmp[i0/warp_size]);
-            }
-        }
-
-        // VKQ = V @ KQ matrix multiplication:
-        constexpr int V_cols_per_iter = kq_stride*kq_nbatch / D; // Number of V columns that fit in SRAM for K.
-        static_assert(kq_stride % V_cols_per_iter == 0, "bad V_cols_per_iter");
-#pragma unroll
-        for (int k0 = 0; k0 < kq_stride; k0 += V_cols_per_iter) {
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; k1 += nwarps) {
-                const int k_tile = k1 + threadIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-                constexpr int cpy_ne_D = cpy_ne < D/(2*warp_size) ? cpy_ne : D/(2*warp_size);
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(
-                        &KV_tmp[k_tile*(D/2) + i0 + threadIdx.x*cpy_ne_D],
-                        &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#else
-                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-                    half2 tmp_h2[cpy_ne_D/2];
-                    ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
-                        tmp_h2, &V_h2[int64_t(k_VKQ_0 + k0 + k_tile)*stride_KV2 + i0/2 + threadIdx.x*(cpy_ne_D/2)]);
-
-                    float2 tmp_f2[cpy_ne_D/2];
-#pragma unroll
-                    for (int i1 = 0; i1 < cpy_ne_D/2; ++i1) {
-                        tmp_f2[i1] = __half22float2(tmp_h2[i1]);
-                    }
-                    ggml_cuda_memcpy_1<sizeof(tmp_f2)>(
-                        &KV_tmp[k_tile*D + i0 + threadIdx.x*cpy_ne_D], tmp_f2);
-                }
-#endif // FAST_FP16_AVAILABLE
-            }
-
-            __syncthreads();
-
-#ifdef FAST_FP16_AVAILABLE
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
-                half2 V_k[(D/2)/warp_size];
-                half2 KQ_k[cpw];
-
-                constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/warp_size], &KV_tmp[k1*(D/2) + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#pragma unroll
-                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
-
-                    half tmp[softmax_iter_j];
-                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(half)>(
-                        &tmp, KQ[j][k0 + k1]);
-#pragma unroll
-                    for (int j1 = 0; j1 < softmax_iter_j; ++j1) {
-                        KQ_k[j0+j1] = __half2half2(tmp[j1]);
-                    }
-                }
-
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-#pragma unroll
-                    for (int j0 = 0; j0 < cpw; ++j0) {
-                        VKQ[j0][i0/warp_size] += V_k[i0/warp_size]*KQ_k[j0];
-                    }
-                }
-            }
-#else
-#pragma unroll
-            for (int k1 = 0; k1 < V_cols_per_iter; ++k1) {
-                float2 V_k[(D/2)/warp_size];
-                float  KQ_k[cpw];
-
-                constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-                for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-                    ggml_cuda_memcpy_1<cpy_ne_D*4>(&V_k[i0/(2*warp_size)], &KV_tmp[k1*D + i0 + threadIdx.x*cpy_ne_D]);
-                }
-#pragma unroll
-                for (int j0 = 0; j0 < cpw; j0 += softmax_iter_j) {
-                    const int j = j0/softmax_iter_j + threadIdx.y*(cpw/softmax_iter_j);
-
-                    ggml_cuda_memcpy_1<softmax_iter_j*sizeof(float)>(
-                        &KQ_k[j0], KQ[j][k0 + k1]);
-                }
-
-#pragma unroll
-                for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-#pragma unroll
-                    for (int j0 = 0; j0 < cpw; ++j0) {
-                        VKQ[j0][i0/warp_size].x += V_k[i0/warp_size].x*KQ_k[j0];
-                        VKQ[j0][i0/warp_size].y += V_k[i0/warp_size].y*KQ_k[j0];
-                    }
-                }
-            }
-#endif // FAST_FP16_AVAILABLE
-
-            __syncthreads();
-        }
-    }
-
-
-    // Attention sink: adjust running max and sum once per head
-    if (sinksf && blockIdx.y == 0) {
-        const float sink = sinksf[head];
-
-#pragma unroll
-        for (int j0 = 0; j0 < cpw; ++j0) {
-            float KQ_max_new_j = fmaxf(KQ_max[j0], sink);
-            KQ_max_new_j = warp_reduce_max<warp_size>(KQ_max_new_j);
-
-            const float KQ_max_scale = expf(KQ_max[j0] - KQ_max_new_j);
-            KQ_max[j0] = KQ_max_new_j;
-
-            const float val = expf(sink - KQ_max[j0]);
-            KQ_sum[j0] = KQ_sum[j0] * KQ_max_scale;
-            if (threadIdx.x == 0) {
-                KQ_sum[j0] += val;
-            }
-
-#ifdef FAST_FP16_AVAILABLE
-            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale, KQ_max_scale);
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                VKQ[j0][i0/warp_size] *= KQ_max_scale_h2;
-            }
-#else
-#pragma unroll
-            for (int i0 = 0; i0 < D/2; i0 += warp_size) {
-                VKQ[j0][i0/warp_size].x *= KQ_max_scale;
-                VKQ[j0][i0/warp_size].y *= KQ_max_scale;
-            }
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-        KQ_sum[j_VKQ_0] = warp_reduce_sum<warp_size>(KQ_sum[j_VKQ_0]);
-    }
-    if (gridDim.y == 1) {
-#pragma unroll
-        for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-#ifdef FAST_FP16_AVAILABLE
-            const half2 KQ_sum_j_inv = make_half2(1.0f/KQ_sum[j_VKQ_0], 1.0f/KQ_sum[j_VKQ_0]);
-#pragma unroll
-            for (int i = 0; i < (D/2)/warp_size; ++i) {
-                VKQ[j_VKQ_0][i] *= KQ_sum_j_inv;
-            }
-#else
-            const float KQ_sum_j_inv = 1.0f/KQ_sum[j_VKQ_0];
-#pragma unroll
-            for (int i = 0; i < (D/2)/warp_size; ++i) {
-                VKQ[j_VKQ_0][i].x *= KQ_sum_j_inv;
-                VKQ[j_VKQ_0][i].y *= KQ_sum_j_inv;
-            }
-#endif // FAST_FP16_AVAILABLE
-        }
-    }
-
-    // Write back results:
-#pragma unroll
-    for (int j_VKQ_0 = 0; j_VKQ_0 < cpw; ++j_VKQ_0) {
-        const int j_VKQ = j_VKQ_0 + threadIdx.y*cpw;
-
-        if (ic0 + j_VKQ >= ne01) {
-            return;
-        }
-
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
-
-#ifdef FAST_FP16_AVAILABLE
-        constexpr int cpy_ne_D = cpy_ne/2 < (D/2)/warp_size ? cpy_ne/2 : (D/2)/warp_size;
-#pragma unroll
-        for (int i0 = 0; i0 < D/2; i0 += warp_size*cpy_ne_D) {
-            float2 tmp[cpy_ne_D];
-#pragma unroll
-            for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
-                tmp[i1] = __half22float2(VKQ[j_VKQ_0][i0/warp_size + i1]);
-            }
-            ggml_cuda_memcpy_1<sizeof(tmp)>(&dst[j_dst_unrolled*D + 2*i0 + threadIdx.x*(2*cpy_ne_D)], tmp);
-        }
-#else
-        constexpr int cpy_ne_D = cpy_ne < D/warp_size ? cpy_ne : D/warp_size;
-#pragma unroll
-        for (int i0 = 0; i0 < D; i0 += warp_size*cpy_ne_D) {
-            ggml_cuda_memcpy_1<cpy_ne_D*4>(
-                &dst[j_dst_unrolled*D + i0 + threadIdx.x*cpy_ne_D], &VKQ[j_VKQ_0][i0/(2*warp_size)]);
-        }
-#endif // FAST_FP16_AVAILABLE
-
-        if (gridDim.y != 1 && threadIdx.x == 0) {
-            dst_meta[j_dst_unrolled] = make_float2(KQ_max[j_VKQ_0], KQ_sum[j_VKQ_0]);
-        }
-    }
-#else
-    GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
-        max_bias, m0, m1, n_head_log2, logit_softcap,
-        ne00, ne01, ne02, ne03,
-              nb01, nb02, nb03,
-        ne10, ne11, ne12, ne13,
-              nb11, nb12, nb13,
-              nb21, nb22, nb23,
-              ne31, ne32, ne33,
-              nb31, nb32, nb33);
-    NO_DEVICE_CODE;
-#endif // FLASH_ATTN_AVAILABLE
-}
-
-template <int D, bool use_logit_softcap>
-static void launch_fattn_tile_switch_ncols(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-
-    const int id        = ggml_cuda_get_device();
-    const int cc        = ggml_cuda_info().devices[id].cc;
-    const int warp_size = 32;
-
-    constexpr size_t nbytes_shared = 0;
-
-#ifdef GGML_USE_HIP
-    if constexpr (D <= 128) {
-        if (Q->ne[1] > 32) {
-            constexpr int cols_per_block = 64;
-            const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-            fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-            const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-            launch_fattn<D, cols_per_block, 1>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-            return;
-        }
-    }
-#endif // GGML_USE_HIP
-
-    if (Q->ne[1] > 16) {
-        constexpr int cols_per_block = 32;
-        const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-        fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-        const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-        launch_fattn<D, cols_per_block, 1>
-            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-        return;
-    }
-
-    constexpr int cols_per_block = 16;
-    const int nwarps = fattn_tile_get_nthreads_host(cc, cols_per_block) / warp_size;
-    fattn_kernel_t fattn_kernel = flash_attn_tile<D, cols_per_block, use_logit_softcap>;
-    const int kq_stride = fattn_tile_get_kq_stride_host(D, cols_per_block, cc, warp_size);
-    launch_fattn<D, cols_per_block, 1>
-        (ctx, dst, fattn_kernel, nwarps, nbytes_shared, kq_stride, true, true, false, warp_size);
-}
-
-template <bool use_logit_softcap>
-static void launch_fattn_tile_switch_head_size(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * Q = dst->src[0];
-    switch (Q->ne[0]) {
+void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+    switch (K->ne[0]) {
+        case  40: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 40,  40>(ctx, dst);
+        } break;
         case  64: {
-            launch_fattn_tile_switch_ncols< 64, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 64,  64>(ctx, dst);
+        } break;
+        case  80: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 80,  80>(ctx, dst);
+        } break;
+        case  96: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case< 96,  96>(ctx, dst);
+        } break;
+        case 112: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<112, 112>(ctx, dst);
         } break;
         case 128: {
-            launch_fattn_tile_switch_ncols<128, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<128, 128>(ctx, dst);
         } break;
         case 256: {
-            launch_fattn_tile_switch_ncols<256, use_logit_softcap>(ctx, dst);
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
+        } break;
+        case 576: {
+            GGML_ASSERT(V->ne[0] == 512);
+            ggml_cuda_flash_attn_ext_tile_case<576, 512>(ctx, dst);
         } break;
         default: {
             GGML_ABORT("Unsupported head size");
         } break;
     }
 }
-
-void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
-
-    float logit_softcap;
-    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
-
-    if (logit_softcap == 0.0f) {
-        constexpr bool use_logit_softcap = false;
-        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
-    } else {
-        constexpr bool use_logit_softcap = true;
-        launch_fattn_tile_switch_head_size<use_logit_softcap>(ctx, dst);
-    }
-}

ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -1,3 +1,5 @@
+#pragma once
+
 #include "common.cuh"
 
 #if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)

ggml/src/ggml-cuda/fattn.cu

@@ -198,6 +198,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     return BEST_FATTN_KERNEL_NONE;
 #endif// FLASH_ATTN_AVAILABLE
 
+    const ggml_tensor * KQV   = dst;
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
@@ -206,31 +207,32 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     const int gqa_ratio = Q->ne[2] / K->ne[2];
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
 
+    float max_bias = 0.0f;
+    memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
+
+    // The effective batch size for the kernel can be increased by gqa_ratio.
+    // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
+    const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+
     const int cc = ggml_cuda_info().devices[device].cc;
 
     switch (K->ne[0]) {
+        case  40:
         case  64:
-        case 128:
-        case 256:
-            if (V->ne[0] != K->ne[0]) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
-            break;
         case  80:
         case  96:
+        case 128:
         case 112:
+        case 256:
             if (V->ne[0] != K->ne[0]) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!ggml_cuda_should_use_wmma_fattn(cc) && !turing_mma_available(cc)) {
-                return BEST_FATTN_KERNEL_NONE;
-            }
             break;
         case 576:
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
             }
-            if (!turing_mma_available(cc) || gqa_ratio % 16 != 0) {
+            if (!gqa_opt_applies || gqa_ratio % 16 != 0) {
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
@@ -264,47 +266,57 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_NONE;
     }
 
-    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0;
-
-    // If Turing tensor cores available, use them except for some cases with batch size 1:
-    if (turing_mma_available(cc)) {
-        best_fattn_kernel best = BEST_FATTN_KERNEL_MMA_F16;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
+    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
+    // If Turing tensor cores available, use them:
+    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40) {
         if (can_use_vector_kernel) {
             if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
-                    best = BEST_FATTN_KERNEL_VEC;
+                    return BEST_FATTN_KERNEL_VEC;
                 }
             } else {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
                     if (Q->ne[1] <= 2) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 } else {
                     if (Q->ne[1] == 1) {
-                        best = BEST_FATTN_KERNEL_VEC;
+                        return BEST_FATTN_KERNEL_VEC;
                     }
                 }
             }
-            if ((gqa_ratio % 2 != 0 || !mask) && Q->ne[1] == 1) {
-                best = BEST_FATTN_KERNEL_VEC; // GQA-specific optimizations in the mma kernel do not apply.
+            if (!gqa_opt_applies && Q->ne[1] == 1) {
+                return BEST_FATTN_KERNEL_VEC;
             }
         }
 
-        return best;
+        return BEST_FATTN_KERNEL_MMA_F16;
     }
 
-    // Use kernels specialized for small batch sizes if possible:
-    if (Q->ne[1] <= 8 && can_use_vector_kernel) {
-        return BEST_FATTN_KERNEL_VEC;
-    }
-
-    // For large batch sizes, use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc)) {
+    // Use the WMMA kernel if possible:
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 576) {
+        if (can_use_vector_kernel && Q->ne[1] <= 2) {
+            return BEST_FATTN_KERNEL_VEC;
+        }
         return BEST_FATTN_KERNEL_WMMA_F16;
     }
 
-    // If there is no suitable kernel for tensor cores or small batch sizes, use the generic kernel for large batch sizes:
+    // If there are no tensor cores available, use the generic tile kernel:
+    if (can_use_vector_kernel) {
+        if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+            if (Q->ne[1] == 1) {
+                if (!gqa_opt_applies) {
+                    return BEST_FATTN_KERNEL_VEC;
+                }
+            }
+        } else {
+            if (Q->ne[1] <= 2) {
+                return BEST_FATTN_KERNEL_VEC;
+            }
+        }
+    }
     return BEST_FATTN_KERNEL_TILE;
 }
 

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -231,7 +231,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-        info.devices[id].integrated = prop.integrated;
+        info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
@@ -3867,7 +3867,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->device = i;
                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
 
-                ggml_cuda_set_device(i);
                 cudaDeviceProp prop;
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq112-dv112.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(112, 112);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq128-dv128.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(128, 128);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq256-dv256.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(256, 256);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq40-dv40.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(40, 40);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq576-dv512.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(576, 512);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq64-dv64.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(64, 64);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq80-dv80.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(80, 80);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq96-dv96.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE(96, 96);

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -3,8 +3,17 @@
 from glob import glob
 import os
 
+HEAD_SIZES_KQ = [40, 64, 80, 96, 112, 128, 256, 576]
+
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0"]
 
+SOURCE_FATTN_TILE = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-tile.cuh"
+
+DECL_FATTN_TILE_CASE({head_size_kq}, {head_size_v});
+"""
+
 SOURCE_FATTN_VEC = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
 
 #include "../fattn-vec.cuh"
@@ -51,6 +60,11 @@ def get_short_name(long_quant_name):
 for filename in glob("*.cu"):
     os.remove(filename)
 
+for head_size_kq in HEAD_SIZES_KQ:
+    head_size_v = head_size_kq if head_size_kq != 576 else 512
+    with open(f"fattn-tile-instance-dkq{head_size_kq}-dv{head_size_v}.cu", "w") as f:
+        f.write(SOURCE_FATTN_TILE.format(head_size_kq=head_size_kq, head_size_v=head_size_v))
+
 for type_k in TYPES_KV:
     for type_v in TYPES_KV:
         with open(f"fattn-vec-instance-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
@@ -64,7 +78,9 @@ for ncols in [8, 16, 32, 64]:
         with open(f"fattn-mma-f16-instance-ncols1_{ncols1}-ncols2_{ncols2}.cu", "w") as f:
             f.write(SOURCE_FATTN_MMA_START)
 
-            for head_size_kq in [64, 80, 96, 112, 128, 256, 576]:
+            for head_size_kq in HEAD_SIZES_KQ:
+                if head_size_kq == 40:
+                    continue
                 if head_size_kq != 576 and ncols2 == 16:
                     continue
                 if head_size_kq == 576 and ncols2 != 16:

ggml/src/ggml-hip/CMakeLists.txt

@@ -53,6 +53,8 @@ file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
 list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
 
 file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-tile*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
 list(APPEND GGML_SOURCES_ROCM ${SRCS})
 file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -924,6 +924,96 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort(ggml_metal_library
     return res;
 }
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        bool    has_mask,
+        int32_t ncpsg) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+    GGML_UNUSED(op);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_%s",
+            "flash_attn_ext_pad");
+
+    snprintf(name, 256, "%s_mask=%d_ncpsg=%d",
+            base,
+            has_mask,
+            ncpsg);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+    ggml_metal_cv_set_bool(cv, has_mask,  FC_FLASH_ATTN_EXT_PAD + 0);
+  //ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_PAD + 1);
+  //ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_PAD + 2);
+  //ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_PAD + 3);
+
+  //ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_PAD + 20);
+  //ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_PAD + 21);
+  //ggml_metal_cv_set_int32(cv, nsg,  FC_FLASH_ATTN_EXT_PAD + 22);
+  //ggml_metal_cv_set_int32(cv, nwg,  FC_FLASH_ATTN_EXT_PAD + 23);
+  //ggml_metal_cv_set_int32(cv, nqptg, FC_FLASH_ATTN_EXT_PAD + 24);
+    ggml_metal_cv_set_int32(cv, ncpsg, FC_FLASH_ATTN_EXT_PAD + 25);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
+
+    return res;
+}
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_blk(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        int32_t nqptg,
+        int32_t ncpsg) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+    GGML_UNUSED(op);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_%s",
+            "flash_attn_ext_blk");
+
+    snprintf(name, 256, "%s_nqptg=%d_ncpsg=%d",
+            base,
+            nqptg,
+            ncpsg);
+
+    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
+    if (res) {
+        return res;
+    }
+
+    ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+  //ggml_metal_cv_set_bool(cv, has_mask,  FC_FLASH_ATTN_EXT_BLK + 0);
+  //ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_BLK + 1);
+  //ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_BLK + 2);
+  //ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_BLK + 3);
+
+  //ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_BLK + 20);
+  //ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_BLK + 21);
+  //ggml_metal_cv_set_int32(cv, nsg,  FC_FLASH_ATTN_EXT_BLK + 22);
+  //ggml_metal_cv_set_int32(cv, nwg,  FC_FLASH_ATTN_EXT_BLK + 23);
+    ggml_metal_cv_set_int32(cv, nqptg, FC_FLASH_ATTN_EXT_BLK + 24);
+    ggml_metal_cv_set_int32(cv, ncpsg, FC_FLASH_ATTN_EXT_BLK + 25);
+
+    res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+    ggml_metal_cv_free(cv);
+
+    return res;
+}
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         ggml_metal_library_t lib,
         const ggml_tensor * op,
@@ -931,6 +1021,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
@@ -943,18 +1034,23 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
     const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
     const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
 
+    // do bounds checks for the mask?
+    const bool bc_mask = op->src[3] && (op->src[3]->ne[1] % 8 != 0);
+
     snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
             "flash_attn_ext",
             ggml_type_name(op->src[1]->type),
             dk,
             dv);
 
-    snprintf(name, 256, "%s_mask=%d_sinks=%d_bias=%d_scap=%d_ns10=%d_ns20=%d_nsg=%d",
+    snprintf(name, 256, "%s_mask=%d_sinks=%d_bias=%d_scap=%d_kvpad=%d_bcm=%d_ns10=%d_ns20=%d_nsg=%d",
             base,
             has_mask,
             has_sinks,
             has_bias,
             has_scap,
+            has_kvpad,
+            bc_mask,
             ns10,
             ns20,
             nsg);
@@ -970,6 +1066,9 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
     ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT + 1);
     ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT + 2);
     ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT + 3);
+    ggml_metal_cv_set_bool(cv, has_kvpad, FC_FLASH_ATTN_EXT + 4);
+
+    ggml_metal_cv_set_bool(cv, bc_mask, FC_FLASH_ATTN_EXT + 10);
 
     ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT + 20);
     ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT + 21);
@@ -989,6 +1088,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg,
         int32_t nwg) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
@@ -1008,12 +1108,13 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
             dk,
             dv);
 
-    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_softcap=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
+    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
             base,
             has_mask,
             has_sinks,
             has_bias,
             has_scap,
+            has_kvpad,
             ns10,
             ns20,
             nsg, nwg);
@@ -1029,6 +1130,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
     ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_VEC + 1);
     ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_VEC + 2);
     ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_VEC + 3);
+    ggml_metal_cv_set_bool(cv, has_kvpad, FC_FLASH_ATTN_EXT_VEC + 4);
 
     ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_VEC + 20);
     ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_VEC + 21);

ggml/src/ggml-metal/ggml-metal-device.h

@@ -135,6 +135,18 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_me
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
 
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        bool    has_mask,
+        int32_t ncpsg);
+
+ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_blk(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op,
+        int32_t nqptg,
+        int32_t ncpsg);
+
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
@@ -142,6 +154,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg);
 
 ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
@@ -151,6 +164,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         bool    has_sinks,
         bool    has_bias,
         bool    has_scap,
+        bool    has_kvpad,
         int32_t nsg,
         int32_t nwg);
 

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -69,11 +69,20 @@
 #define N_SG_IQ4_XS 2
 
 // function constants offsets
-#define FC_FLASH_ATTN_EXT              100
-#define FC_FLASH_ATTN_EXT_VEC          200
-#define FC_FLASH_ATTN_EXT_VEC_REDUCE   300
-#define FC_MUL_MV                      400
-#define FC_MUL_MM                      500
+#define FC_FLASH_ATTN_EXT_PAD          100
+#define FC_FLASH_ATTN_EXT_BLK          200
+#define FC_FLASH_ATTN_EXT              300
+#define FC_FLASH_ATTN_EXT_VEC          400
+#define FC_FLASH_ATTN_EXT_VEC_REDUCE   500
+#define FC_MUL_MV                      600
+#define FC_MUL_MM                      700
+
+// op-specific constants
+#define OP_FLASH_ATTN_EXT_NQPTG 8
+#define OP_FLASH_ATTN_EXT_NCPSG 64
+
+#define OP_FLASH_ATTN_EXT_VEC_NQPTG 1
+#define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
 // kernel argument structs
 //
@@ -244,6 +253,35 @@ typedef struct {
     int32_t  sect_3;
 } ggml_metal_kargs_rope;
 
+typedef struct {
+    int32_t  ne11;
+    int32_t  ne_12_2; // assume K and V are same shape
+    int32_t  ne_12_3;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ne31;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+} ggml_metal_kargs_flash_attn_ext_pad;
+
+typedef struct {
+    int32_t  ne01;
+    int32_t  ne30;
+    int32_t  ne31;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+} ggml_metal_kargs_flash_attn_ext_blk;
+
 typedef struct {
     int32_t  ne01;
     int32_t  ne02;
@@ -262,6 +300,7 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne31;
     int32_t  ne32;
     int32_t  ne33;
     uint64_t nb31;
@@ -296,6 +335,7 @@ typedef struct {
     uint64_t nb21;
     uint64_t nb22;
     uint64_t nb23;
+    int32_t  ne31;
     int32_t  ne32;
     int32_t  ne33;
     uint64_t nb31;

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -226,6 +226,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             GGML_TENSOR_LOCALS(uint64_t, nb0, node->src[0], nb);
             GGML_TENSOR_LOCALS( int64_t, ne1, node->src[1], ne);
             GGML_TENSOR_LOCALS(uint64_t, nb1, node->src[1], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne2, node->src[2], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb2, node->src[2], nb);
+            GGML_TENSOR_LOCALS( int64_t, ne3, node->src[3], ne);
+            GGML_TENSOR_LOCALS(uint64_t, nb3, node->src[3], nb);
             GGML_TENSOR_LOCALS( int64_t, ne,  node,         ne);
             GGML_TENSOR_LOCALS(uint64_t, nb,  node,         nb);
 
@@ -237,6 +241,14 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
                 GGML_LOG_DEBUG("%s: src1 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[1]->type), ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
                         ggml_is_contiguous(node->src[1]), node->src[1]->name);
             }
+            if (node->src[2]) {
+                GGML_LOG_DEBUG("%s: src2 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[2]->type), ne20, ne21, ne22, ne23, nb20, nb21, nb22, nb23,
+                        ggml_is_contiguous(node->src[2]), node->src[2]->name);
+            }
+            if (node->src[3]) {
+                GGML_LOG_DEBUG("%s: src3 - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(node->src[3]->type), ne30, ne31, ne32, ne33, nb30, nb31, nb32, nb33,
+                        ggml_is_contiguous(node->src[3]), node->src[3]->name);
+            }
             if (node) {
                 GGML_LOG_DEBUG("%s: node  - %4s [%5lld, %5lld, %5lld, %5lld] [%5lld, %5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(node->type), ne0, ne1, ne2, ne3, nb0, nb1, nb2, nb3,
                         node->name);
@@ -1534,9 +1546,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         !ggml_is_transposed(op->src[1]) &&
         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-        props_dev->has_simdgroup_mm && ne00 >= 64 &&
-        (ne11 > ne11_mm_min || (ggml_is_quantized(op->src[0]->type) && ne12 > 1))) {
-        //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+        props_dev->has_simdgroup_mm && ne00 >= 64 && ne11 > ne11_mm_min) {
+        //GGML_LOG_INFO("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
 
         // some Metal matrix data types require aligned pointers
         // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
@@ -1889,20 +1900,107 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
     return (ne01 < 20) && (ne00 % 32 == 0);
 }
 
+size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    } else {
+        const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_NCPSG != 0;
+
+        if (has_kvpad) {
+            res += OP_FLASH_ATTN_EXT_NCPSG*(
+                nb11*ne12*ne13 +
+                nb21*ne22*ne23 +
+                (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
+        }
+    }
+
+    return res;
+}
+
+size_t ggml_metal_op_flash_attn_ext_extra_blk(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
+
+    size_t res = 0;
+
+    const bool has_mask = op->src[3] != nullptr;
+
+    if (!has_mask) {
+        return res;
+    }
+
+    const bool is_vec = ggml_metal_op_flash_attn_ext_use_vec(op);
+
+    // this optimization is not useful for the vector kernels
+    if (is_vec) {
+        return res;
+    }
+
+    const int nqptg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NQPTG : OP_FLASH_ATTN_EXT_NQPTG;
+    const int ncpsg = is_vec ? OP_FLASH_ATTN_EXT_VEC_NCPSG : OP_FLASH_ATTN_EXT_NCPSG;
+
+    const int64_t ne1 = (ne01 + nqptg - 1)/nqptg;
+    const int64_t ne0 = (ne30 + ncpsg - 1)/ncpsg;
+
+    res += GGML_PAD(ggml_type_size(GGML_TYPE_I8)*ne0*ne1*ne32*ne33, 32);
+
+    return res;
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
-    const int64_t nwg = 32;
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+  //GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+  //GGML_TENSOR_LOCALS(uint64_t, nb3, op->src[3], nb);
 
-    const int64_t ne01 = op->src[0]->ne[1];
-    const int64_t ne02 = op->src[0]->ne[2];
-    const int64_t ne03 = op->src[0]->ne[3];
-    const int64_t ne20 = op->src[2]->ne[0];
+    size_t res = 0;
 
-    // temp buffer for writing the results from each workgroup
-    // - ne20: the size of the Value head
-    // -  + 2: the S and M values for each intermediate result
-    return ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+    if (ggml_metal_op_flash_attn_ext_use_vec(op)) {
+        const int64_t nwg = 32;
+
+        // temp buffer for writing the results from each workgroup
+        // - ne20: the size of the Value head
+        // -  + 2: the S and M values for each intermediate result
+        res += ggml_type_size(GGML_TYPE_F32)*(ne01*ne02*ne03*nwg*(ne20 + 2));
+    }
+
+    return res;
 }
 
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
@@ -1924,8 +2022,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS( int32_t, nb,  op,         nb);
 
-    GGML_ASSERT(ne00 % 4  == 0);
-    GGML_ASSERT(ne11 % 32 == 0);
+    GGML_ASSERT(ne00 % 4 == 0);
 
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == op->src[2]->type);
@@ -1935,8 +2032,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     GGML_ASSERT(ne12 == ne22);
 
     GGML_ASSERT(!op->src[3] || op->src[3]->type == GGML_TYPE_F16);
-    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= GGML_PAD(op->src[0]->ne[1], 8) &&
-            "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+    GGML_ASSERT(!op->src[3] || op->src[3]->ne[1] >= op->src[0]->ne[1] &&
+            "the Flash-Attention Metal kernel requires the mask to be at least n_queries big");
 
     float scale;
     float max_bias;
@@ -1963,15 +2060,111 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
     GGML_ASSERT(ne01 < 65536);
 
+    ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
+    ggml_metal_buffer_id bid_src1 = ggml_metal_get_buffer_id(op->src[1]);
+    ggml_metal_buffer_id bid_src2 = ggml_metal_get_buffer_id(op->src[2]);
+    ggml_metal_buffer_id bid_src3 = has_mask  ? ggml_metal_get_buffer_id(op->src[3]) : bid_src0;
+    ggml_metal_buffer_id bid_src4 = has_sinks ? ggml_metal_get_buffer_id(op->src[4]) : bid_src0;
+
+    ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
+
+    ggml_metal_buffer_id bid_pad = bid_dst;
+    bid_pad.offs += ggml_nbytes(op);
+
+    ggml_metal_buffer_id bid_blk = bid_pad;
+    bid_blk.offs += ggml_metal_op_flash_attn_ext_extra_pad(op);
+
+    ggml_metal_buffer_id bid_tmp = bid_blk;
+    bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);
+
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
-        const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 64; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nqptg = OP_FLASH_ATTN_EXT_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 8  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (has_mask) {
+            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_blk args0 = {
+                /*.ne01 =*/ ne01,
+                /*.ne30 =*/ ne30,
+                /*.ne31 =*/ ne31,
+                /*.ne32 =*/ ne32,
+                /*.ne33 =*/ ne33,
+                /*.nb31 =*/ nb31,
+                /*.nb32 =*/ nb32,
+                /*.nb33 =*/ nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_blk(lib, op, nqptg, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_blk,  2);
+
+            const int32_t nblk1 = ((ne01 + nqptg - 1)/nqptg);
+            const int32_t nblk0 = ((ne30 + ncpsg - 1)/ncpsg);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, nblk0, nblk1, ne32*ne33, 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_blk(op) == 0);
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
         const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
 
         // 2*(2*ncpsg)
@@ -2021,6 +2214,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2037,24 +2231,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg);
+        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         6);
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+        ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
+        ggml_metal_encoder_set_buffer  (enc, bid_blk,  7);
+        ggml_metal_encoder_set_buffer  (enc, bid_dst,  8);
 
         ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
@@ -2062,14 +2250,62 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
-        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
-        const int64_t nkpsg = 1*ncpsg;
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPTG; // queries per threadgroup
+        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
+        const int nkpsg = 1*ncpsg;
 
         GGML_ASSERT(nqptg <= 32);
         GGML_ASSERT(nqptg  % 1  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
+        bool need_sync = false;
+
+        const bool has_kvpad = ne11 % ncpsg != 0;
+
+        if (has_kvpad) {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
+
+            ggml_metal_kargs_flash_attn_ext_pad args0 = {
+                /*.ne11    =*/ne11,
+                /*.ne_12_2 =*/ne12,
+                /*.ne_12_3 =*/ne13,
+                /*.nb11    =*/nb11,
+                /*.nb12    =*/nb12,
+                /*.nb13    =*/nb13,
+                /*.nb21    =*/nb21,
+                /*.nb22    =*/nb22,
+                /*.nb23    =*/nb23,
+                /*.ne31    =*/ne31,
+                /*.ne32    =*/ne32,
+                /*.ne33    =*/ne33,
+                /*.nb31    =*/nb31,
+                /*.nb32    =*/nb32,
+                /*.nb33    =*/nb33,
+            };
+
+            ggml_metal_pipeline_t pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_pad(lib, op, has_mask, ncpsg);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
+            ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
+
+            assert(ne12 == ne22);
+            assert(ne13 == ne23);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ncpsg, std::max(ne12, ne32), std::max(ne13, ne33), 32, 1, 1);
+
+            need_sync = true;
+        } else {
+            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) == 0);
+        }
+
+        if (need_sync) {
+            ggml_metal_op_concurrency_reset(ctx);
+        }
+
         // ne00 + 2*ncpsg*(nsg)
         // for each query, we load it as f16 in shared memory (ne00)
         // and store the soft_max values and the mask
@@ -2134,6 +2370,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.nb21          =*/ nb21,
             /*.nb22          =*/ nb22,
             /*.nb23          =*/ nb23,
+            /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
             /*.nb31          =*/ nb31,
@@ -2150,25 +2387,17 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, nsg, nwg);
+        ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        if (op->src[3]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 4);
-        }
-        if (op->src[4]) {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        } else {
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op->src[0]), 5);
-        }
+        ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
+        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
+        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
 
         const size_t smem = FATTN_SMEM(nsg);
 
@@ -2176,23 +2405,25 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);
 
         if (nwg == 1) {
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) == 0);
+
             // using 1 workgroup -> write the result directly into dst
-            ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_dst, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
             ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);
         } else {
             // sanity checks
+            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) != 0);
+
             GGML_ASSERT(ne01*ne02*ne03 == ne1*ne2*ne3);
             GGML_ASSERT((uint64_t)ne1*ne2*ne3 <= (1u << 31));
 
-            ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
-
             // write the results from each workgroup into a temp buffer
-            ggml_metal_buffer_id bid_tmp = bid_dst;
-            bid_tmp.offs += ggml_nbytes(op);
-            ggml_metal_encoder_set_buffer(enc, bid_tmp, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
+            ggml_metal_encoder_set_buffer(enc, bid_tmp, 7);
 
             ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
             ggml_metal_encoder_dispatch_threadgroups(enc, (ne01 + nqptg - 1)/nqptg, ne02, ne03*nwg, 32, nsg, 1);

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -39,6 +39,8 @@ size_t ggml_metal_op_mul_mat_id_extra_ids(const struct ggml_tensor * op);
 // return true if we should use the FA vector kernel for this op
 bool ggml_metal_op_flash_attn_ext_use_vec(const struct ggml_tensor * op);
 
+size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
 
 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal.cpp

@@ -193,9 +193,9 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
             } break;
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                if (ggml_metal_op_flash_attn_ext_use_vec(tensor)) {
-                    res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
-                }
+                res += ggml_metal_op_flash_attn_ext_extra_pad(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
             } break;
         default:
             break;

ggml/src/ggml-metal/ggml-metal.metal

@@ -4349,10 +4349,142 @@ kernel void kernel_leaky_relu_f32_4(
     dst[tpig] = float4(x > 0.0f)*x + float4(x <= 0.0f)*(x * args.slope);
 }
 
+constant bool FC_flash_attn_ext_pad_has_mask [[function_constant(FC_FLASH_ATTN_EXT_PAD + 0)]];
+
+constant int32_t FC_flash_attn_ext_pad_ncpsg [[function_constant(FC_FLASH_ATTN_EXT_PAD + 25)]];
+
+// pad the last chunk of C elements of k and v into a an extra pad buffer
+kernel void kernel_flash_attn_ext_pad(
+        constant ggml_metal_kargs_flash_attn_ext_pad & args,
+        device const char * k,
+        device const char * v,
+        device const char * mask,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int32_t C = FC_flash_attn_ext_pad_ncpsg;
+
+    device char * k_pad    = dst;
+    device char * v_pad    = k_pad + args.nb11*C*args.ne_12_2*args.ne_12_3;
+    device char * mask_pad = v_pad + args.nb21*C*args.ne_12_2*args.ne_12_3;
+
+    const int32_t icp = args.ne11 % C;
+    const int32_t ic0 = args.ne11 - icp;
+
+    const int32_t i1 = tgpig[0];
+    const int32_t i2 = tgpig[1];
+    const int32_t i3 = tgpig[2];
+
+    if (i2 < args.ne_12_2 && i3 < args.ne_12_3) {
+        device const char * k_src = k + args.nb11*(ic0 + i1) + args.nb12*i2 + args.nb13*i3;
+        device const char * v_src = v + args.nb21*(ic0 + i1) + args.nb22*i2 + args.nb23*i3;
+
+        device char * k_dst = k_pad + args.nb11*i1 + args.nb11*C*i2 + args.nb11*C*args.ne_12_2*i3;
+        device char * v_dst = v_pad + args.nb21*i1 + args.nb21*C*i2 + args.nb21*C*args.ne_12_2*i3;
+
+        if (i1 >= icp) {
+            // here it is not important the exact value that will be used as we rely on masking out the scores in the attention
+            for (uint64_t i = tiitg; i < args.nb11; i += ntg.x) {
+                k_dst[i] = 0;
+            }
+            for (uint64_t i = tiitg; i < args.nb21; i += ntg.x) {
+                v_dst[i] = 0;
+            }
+        } else {
+            for (uint64_t i = tiitg; i < args.nb11; i += ntg.x) {
+                k_dst[i] = k_src[i];
+            }
+            for (uint64_t i = tiitg; i < args.nb21; i += ntg.x) {
+                v_dst[i] = v_src[i];
+            }
+        }
+    }
+
+    if (FC_flash_attn_ext_pad_has_mask) {
+        if (i2 < args.ne32 && i3 < args.ne33) {
+            for (int ib = i1; ib < args.ne31; ib += C) {
+                device const half * mask_src = (device const half *)(mask      + args.nb31*ib + args.nb32*i2 + args.nb33*i3) + ic0;
+                device       half * mask_dst = (device       half *)(mask_pad) + C*ib + C*args.ne31*i2 + C*args.ne31*args.ne32*i3;
+
+                for (int i = tiitg; i < C; i += ntg.x) {
+                    if (i >= icp) {
+                        mask_dst[i] = -MAXHALF;
+                    } else {
+                        mask_dst[i] = mask_src[i];
+                    }
+                }
+            }
+        }
+    }
+}
+
+constant int32_t FC_flash_attn_ext_blk_nqptg [[function_constant(FC_FLASH_ATTN_EXT_BLK + 24)]];
+constant int32_t FC_flash_attn_ext_blk_ncpsg [[function_constant(FC_FLASH_ATTN_EXT_BLK + 25)]];
+
+// scan the blocks of the mask that are not masked
+// 0 -     masked (i.e. full of -INF, skip)
+// 1 - not masked (i.e. at least one element of the mask is not -INF)
+kernel void kernel_flash_attn_ext_blk(
+        constant ggml_metal_kargs_flash_attn_ext_blk & args,
+        device const char * mask,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
+    // block size C x Q
+    const int32_t Q = FC_flash_attn_ext_blk_nqptg;
+    const int32_t C = FC_flash_attn_ext_blk_ncpsg;
+
+    constexpr short NW  = N_SIMDWIDTH;
+
+    const int32_t i3 = tgpig[2]/args.ne32;
+    const int32_t i2 = tgpig[2]%args.ne32;
+    const int32_t i1 = tgpig[1];
+    const int32_t i0 = tgpig[0];
+
+    char res = i0*C + C > args.ne30 ? 1 : 0;
+
+    device const half * mask_src = (device const half *) (mask + (i1*Q)*args.nb31 + i2*args.nb32 + i3*args.nb33) + i0*C + tiisg;
+
+    // fast route
+    if (res == 0) {
+        if (simd_max(*mask_src) > -MAXHALF/2) {
+            res = 1;
+        }
+    }
+
+    // detailed check of the elements of the block
+    if ((C > NW || Q > 1) && res == 0) {
+        half m = -MAXHALF;
+
+        FOR_UNROLL (short j = 0; j < Q; ++j) {
+            FOR_UNROLL (short ii = 0; ii < C/NW; ++ii) {
+                m = max(m, mask_src[ii*NW]);
+            }
+
+            mask_src += args.nb31/2;
+        }
+
+        if (simd_max(m) > -MAXHALF/2) {
+            res = 1;
+        }
+    }
+
+    const int32_t nblk1 = ((args.ne01 + Q - 1)/Q);
+    const int32_t nblk0 = ((args.ne30 + C - 1)/C);
+
+    if (tiisg == 0) {
+        dst[((i3*args.ne32 + i2)*nblk1 + i1)*nblk0 + i0] = res;
+    }
+}
+
 constant bool FC_flash_attn_ext_has_mask  [[function_constant(FC_FLASH_ATTN_EXT + 0)]];
 constant bool FC_flash_attn_ext_has_sinks [[function_constant(FC_FLASH_ATTN_EXT + 1)]];
 constant bool FC_flash_attn_ext_has_bias  [[function_constant(FC_FLASH_ATTN_EXT + 2)]];
 constant bool FC_flash_attn_ext_has_scap  [[function_constant(FC_FLASH_ATTN_EXT + 3)]];
+constant bool FC_flash_attn_ext_has_kvpad [[function_constant(FC_FLASH_ATTN_EXT + 4)]];
+
+constant bool FC_flash_attn_ext_bc_mask [[function_constant(FC_FLASH_ATTN_EXT + 10)]];
 
 //constant float FC_flash_attn_ext_scale         [[function_constant(FC_FLASH_ATTN_EXT + 10)]];
 //constant float FC_flash_attn_ext_max_bias      [[function_constant(FC_FLASH_ATTN_EXT + 11)]];
@@ -4399,6 +4531,8 @@ void kernel_flash_attn_ext_impl(
         device const char * v,
         device const char * mask,
         device const char * sinks,
+        device const char * pad,
+        device const char * blk,
         device       char * dst,
         threadgroup  half * shmem_f16,
         uint3   tgpig,
@@ -4464,6 +4598,13 @@ void kernel_flash_attn_ext_impl(
         pm2[jj] = (device const half2 *) ((device const char *) mask + (iq1 + j)*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);
     }
 
+    {
+        const int32_t nblk1 = ((args.ne01 + Q - 1)/Q);
+        const int32_t nblk0 = ((args.ne11 + C - 1)/C);
+
+        blk += (((iq3%args.ne33)*args.ne32 + (iq2%args.ne32))*nblk1 + iq1/Q)*nblk0;
+    }
+
     {
         q += iq1*args.nb01 + iq2*args.nb02 + iq3*args.nb03;
 
@@ -4523,16 +4664,75 @@ void kernel_flash_attn_ext_impl(
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic = 0; ic < args.ne11; ic += C) {
+        for (int ic0 = 0; ; ++ic0) {
+            int ic = ic0*C;
+            if (ic >= args.ne11) {
+                break;
+            }
+
+            // the last partial chunk uses the pad buffer as source
+            if (FC_flash_attn_ext_has_kvpad && ic + C > args.ne11) {
+                k    = pad;
+                v    = k + args.nb11*C*args.ne_12_2*args.ne_12_3;
+                mask = v + args.nb21*C*args.ne_12_2*args.ne_12_3;
+
+                const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+                const short ikv3 = iq3/(args.ne03/args.ne_12_3);
+
+                k += (ikv2 + ikv3*args.ne_12_2)*args.nb11*C;
+                v += (ikv2 + ikv3*args.ne_12_2)*args.nb21*C;
+
+                if (!FC_flash_attn_ext_has_mask) {
+                    threadgroup half * sm = (threadgroup half *) (sm2);
+
+                    FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
+                        const short j = jj*NSG + sgitg;
+
+                        for (short i = tiisg; i < C; i += NW) {
+                            if (ic + i >= args.ne11) {
+                                sm[2*j*SH + i] = -MAXHALF;
+                            }
+                        }
+                    }
+                } else {
+                    FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
+                        const short j = jj*NSG + sgitg;
+
+                        pm2[jj] = (device const half2 *) ((device const half *) mask +
+                                (iq1 + j)*C +
+                                (iq2%args.ne32)*(C*args.ne31) +
+                                (iq3%args.ne33)*(C*args.ne31*args.ne32));
+                    }
+                }
+
+                ic = 0;
+            }
+
             // read the mask into shared mem
             if (FC_flash_attn_ext_has_mask) {
+                if (blk[ic0] == 0) {
+                    FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
+                        pm2[jj] += NW;
+                    }
+
+                    continue;
+                }
+
                 FOR_UNROLL (short jj = 0; jj < NQ; ++jj) {
                     const short j = jj*NSG + sgitg;
 
-                    sm2[j*SH + tiisg] = pm2[jj][tiisg];
+                    if (FC_flash_attn_ext_bc_mask) {
+                        sm2[j*SH + tiisg] = (iq1 + j) < args.ne31 ? pm2[jj][tiisg] : half2(-MAXHALF, -MAXHALF);
+                    } else {
+                        sm2[j*SH + tiisg] = pm2[jj][tiisg];
+                    }
+
                     pm2[jj] += NW;
                 }
 
+#if 0
+                // note: old -INF block optimization - obsoleted by pre-computing non-masked blocks
+
                 threadgroup_barrier(mem_flags::mem_threadgroup);
 
                 // used to detect blocks full of -INF
@@ -4551,13 +4751,14 @@ void kernel_flash_attn_ext_impl(
 
                     continue;
                 }
+#endif
             }
 
             // Q*K^T
             // this is compile-time check, so it does not have runtime overhead
             if (is_same<kd4x4_t, k4x4_t>::value) {
                 // we can read directly from global memory
-                device      const k_t * pk = (device const k_t *) ((device const char *) k + ic*args.nb11);
+                device      const k_t * pk = (device const k_t *) (k + ic*args.nb11);
                 threadgroup const q_t * pq = sq;
                 threadgroup       s_t * ps = ss;
 
@@ -4568,26 +4769,24 @@ void kernel_flash_attn_ext_impl(
 
                 constexpr short NC = (C/8)/NSG;
 
-                // TODO: not good to unroll for large contexts - not sure why?
+                // note: do not unroll for large heads
+                #pragma unroll (DK <= 64 ? NC : 1)
                 for (short cc = 0; cc < NC; ++cc) {
                     qk8x8_t mqk = make_filled_simdgroup_matrix<qk_t, 8>((qk_t) 0.0f);
 
-                    if (DK8 % 16 != 0) {
+                    if (DK % 16 != 0) {
                         k8x8_t mk;
                         q8x8_t mq;
 
                         FOR_UNROLL (short i = 0; i < DK8; ++i) {
                             simdgroup_barrier(mem_flags::mem_none);
 
-                            simdgroup_load(mk, pk, NS10, 0, true);
-                            simdgroup_load(mq, pq, DK);
+                            simdgroup_load(mk, pk + 8*i, NS10, 0, true);
+                            simdgroup_load(mq, pq + 8*i, DK);
 
                             simdgroup_barrier(mem_flags::mem_none);
 
                             simdgroup_multiply_accumulate(mqk, mq, mk, mqk);
-
-                            pk += 8;
-                            pq += 8;
                         }
                     } else {
                         k8x8_t mk[2];
@@ -4596,26 +4795,22 @@ void kernel_flash_attn_ext_impl(
                         FOR_UNROLL (short i = 0; i < DK8/2; ++i) {
                             simdgroup_barrier(mem_flags::mem_none);
 
-                            simdgroup_load(mk[0], pk + 0*8, NS10, 0, true);
-                            simdgroup_load(mk[1], pk + 1*8, NS10, 0, true);
+                            simdgroup_load(mq[0], pq + 0*8 + 16*i, DK);
+                            simdgroup_load(mq[1], pq + 1*8 + 16*i, DK);
 
-                            simdgroup_load(mq[0], pq + 0*8, DK);
-                            simdgroup_load(mq[1], pq + 1*8, DK);
+                            simdgroup_load(mk[0], pk + 0*8 + 16*i, NS10, 0, true);
+                            simdgroup_load(mk[1], pk + 1*8 + 16*i, NS10, 0, true);
 
                             simdgroup_barrier(mem_flags::mem_none);
 
                             simdgroup_multiply_accumulate(mqk, mq[0], mk[0], mqk);
                             simdgroup_multiply_accumulate(mqk, mq[1], mk[1], mqk);
-
-                            pk += 16;
-                            pq += 16;
                         }
                     }
 
                     simdgroup_store(mqk, ps, SH, 0, false);
 
-                    pk += 8*(NSG*NS10 - DK8);
-                    pq += 8*(NSG*0    - DK8);
+                    pk += 8*(NSG*NS10);
                     ps += 8*(NSG);
                 }
             } else {
@@ -4629,7 +4824,7 @@ void kernel_flash_attn_ext_impl(
                     qk8x8_t mqk = make_filled_simdgroup_matrix<qk_t, 8>((qk_t) 0.0f);
 
                     for (short ii = 0; ii < DK16; ii += 4) {
-                        device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb11));
+                        device const kd4x4_t * pk4x4 = (device const kd4x4_t *) (k + ((ic + 8*cc + ty)*args.nb11));
 
                         if (DK16%4 == 0) {
                             // the head is evenly divisible by 4*16 = 64, so no need for bound checks
@@ -4749,27 +4944,50 @@ void kernel_flash_attn_ext_impl(
                     }
 
                     {
-                        auto sst = ss;
-
-                        device const v_t * pv = (device const v_t *) ((device const char *) v + ic*args.nb21);
+                        device const v_t * pv = (device const v_t *) (v + ic*args.nb21);
 
                         pv += 8*sgitg;
 
-                        FOR_UNROLL (short cc = 0; cc < C/8; ++cc) {
-                            s8x8_t vs;
-                            simdgroup_load(vs, sst, SH, 0, false);
+                        if (DV <= 64) {
+                            FOR_UNROLL (short cc = 0; cc < C/8; ++cc) {
+                                s8x8_t vs;
+                                simdgroup_load(vs, ss + 8*cc, SH, 0, false);
 
-                            FOR_UNROLL (short ii = 0; ii < NO; ++ii) {
-                                v8x8_t mv;
+                                FOR_UNROLL (short ii = 0; ii < NO/2; ++ii) {
+                                    v8x8_t mv[2];
 
-                                simdgroup_load(mv, pv, NS20, 0, false);
-                                simdgroup_multiply_accumulate(lo[ii], vs, mv, lo[ii]);
+                                    simdgroup_load(mv[0], pv + 0*NSG + 16*ii*NSG, NS20, 0, false);
+                                    simdgroup_load(mv[1], pv + 8*NSG + 16*ii*NSG, NS20, 0, false);
 
-                                pv += 8*NSG;
+                                    simdgroup_multiply_accumulate(lo[2*ii + 0], vs, mv[0], lo[2*ii + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*ii + 1], vs, mv[1], lo[2*ii + 1]);
+                                }
+
+                                pv  += 8*NS20;
                             }
+                        } else {
+                            FOR_UNROLL (short cc = 0; cc < (C/8)/2; ++cc) {
+                                s8x8_t vs[2];
 
-                            pv  += 8*(NS20 - NO*NSG);
-                            sst += 8;
+                                simdgroup_load(vs[0], ss + 16*cc + 0, SH, 0, false);
+                                simdgroup_load(vs[1], ss + 16*cc + 8, SH, 0, false);
+
+                                FOR_UNROLL (short ii = 0; ii < NO/2; ++ii) {
+                                    v8x8_t mv[4];
+
+                                    simdgroup_load(mv[0], pv + 0*NSG + 16*ii*NSG + 0*8*NS20, NS20, 0, false);
+                                    simdgroup_load(mv[1], pv + 8*NSG + 16*ii*NSG + 0*8*NS20, NS20, 0, false);
+                                    simdgroup_load(mv[2], pv + 0*NSG + 16*ii*NSG + 1*8*NS20, NS20, 0, false);
+                                    simdgroup_load(mv[3], pv + 8*NSG + 16*ii*NSG + 1*8*NS20, NS20, 0, false);
+
+                                    simdgroup_multiply_accumulate(lo[2*ii + 0], vs[0], mv[0], lo[2*ii + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*ii + 1], vs[0], mv[1], lo[2*ii + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*ii + 0], vs[1], mv[2], lo[2*ii + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*ii + 1], vs[1], mv[3], lo[2*ii + 1]);
+                                }
+
+                                pv  += 2*8*NS20;
+                            }
                         }
                     }
 
@@ -4793,7 +5011,7 @@ void kernel_flash_attn_ext_impl(
                         simdgroup_load(vs, ss + 8*cc, SH, 0, false);
 
                         for (short ii = 4*sgitg; ii < DV16; ii += 4*NSG) {
-                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb21));
+                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) (v + ((ic + 8*cc + ty)*args.nb21));
 
                             if (DV16%4 == 0) {
                                 // no need for bound checks
@@ -4883,7 +5101,7 @@ void kernel_flash_attn_ext_impl(
 
         device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
 
-        const float scale = 1.0f/S[jj];
+        const float scale = S[jj] == 0.0 ? 0.0f : 1.0f/S[jj];
 
         if (DV4 % NW == 0) {
             FOR_UNROLL (short ii = 0; ii < DV4/NW; ++ii) {
@@ -4928,8 +5146,8 @@ template<
     void (*deq_v)(device const vd4x4_t *, short, thread v4x4_t &),
     short DK,         // K head size
     short DV,         // V head size
-    short Q  = 8,     // queries per threadgroup
-    short C  = 64>    // cache items per threadgroup
+    short Q  = OP_FLASH_ATTN_EXT_NQPTG, // queries per threadgroup
+    short C  = OP_FLASH_ATTN_EXT_NCPSG> // cache items per threadgroup
 kernel void kernel_flash_attn_ext(
         constant ggml_metal_kargs_flash_attn_ext & args,
         device const char * q,
@@ -4937,13 +5155,15 @@ kernel void kernel_flash_attn_ext(
         device const char * v,
         device const char * mask,
         device const char * sinks,
+        device const char * pad,
+        device const char * blk,
         device       char * dst,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
         ushort  tiisg[[thread_index_in_simdgroup]],
         ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
 #define FWD_TMPL q_t, q4_t, q8x8_t, k_t, k4x4_t, k8x8_t, v_t, v4x4_t, v8x8_t, qk_t, qk8x8_t, s_t, s2_t, s8x8_t, o_t, o4_t, o8x8_t, kd4x4_t, nl_k, deq_k, vd4x4_t, nl_v, deq_v, DK, DV, Q, C
-#define FWD_ARGS args, q, k, v, mask, sinks, dst, shmem_f16, tgpig, tiisg, sgitg
+#define FWD_ARGS args, q, k, v, mask, sinks, pad, blk, dst, shmem_f16, tgpig, tiisg, sgitg
     switch (FC_flash_attn_ext_nsg) {
       // note: disabled cases to reduce library load time
       //case 1: kernel_flash_attn_ext_impl<FWD_TMPL, 1>(FWD_ARGS); break;
@@ -5063,6 +5283,7 @@ constant bool FC_flash_attn_ext_vec_has_mask  [[function_constant(FC_FLASH_ATTN_
 constant bool FC_flash_attn_ext_vec_has_sinks [[function_constant(FC_FLASH_ATTN_EXT_VEC + 1)]];
 constant bool FC_flash_attn_ext_vec_has_bias  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 2)]];
 constant bool FC_flash_attn_ext_vec_has_scap  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 3)]];
+constant bool FC_flash_attn_ext_vec_has_kvpad [[function_constant(FC_FLASH_ATTN_EXT_VEC + 4)]];
 
 //constant float FC_flash_attn_ext_vec_scale         [[function_constant(FC_FLASH_ATTN_EXT_VEC + 10)]];
 //constant float FC_flash_attn_ext_vec_max_bias      [[function_constant(FC_FLASH_ATTN_EXT_VEC + 11)]];
@@ -5089,9 +5310,9 @@ template<
     void (*deq_v_t4)(device const vd4_t *, short, thread v4_t &),
     short DK,       // K head size
     short DV,       // V head size
-    short NE = 4,   // head elements per thread
-    short Q  = 1,   // queries per threadgroup
-    short C  = 32,  // cache items per threadgroup
+    short NE,       // head elements per thread
+    short Q,        // queries per threadgroup
+    short C,        // cache items per threadgroup
     short NSG>      // number of simd groups
 void kernel_flash_attn_ext_vec_impl(
         constant ggml_metal_kargs_flash_attn_ext_vec & args,
@@ -5100,6 +5321,7 @@ void kernel_flash_attn_ext_vec_impl(
         device const char * v,
         device const char * mask,
         device const char * sinks,
+        device const char * pad,
         device       char * dst,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
@@ -5205,12 +5427,38 @@ void kernel_flash_attn_ext_vec_impl(
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = (int) iwg*C*NSG; ic0 < args.ne11; ic0 += (int) NWG*C*NSG) {
-            const int ic = ic0 + C*sgitg;
+        for (int ic0 = iwg*NSG + sgitg; ; ic0 += NWG*NSG) {
+            int ic = ic0*C;
             if (ic >= args.ne11) {
                 break;
             }
 
+            // the last partial chunk uses the pad buffer as source
+            if (FC_flash_attn_ext_vec_has_kvpad && ic + C > args.ne11) {
+                k    = pad;
+                v    = k + args.nb11*C*args.ne_12_2*args.ne_12_3;
+                mask = v + args.nb21*C*args.ne_12_2*args.ne_12_3;
+
+                const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+                const short ikv3 = iq3/(args.ne03/args.ne_12_3);
+
+                k += (ikv2 + ikv3*args.ne_12_2)*args.nb11*C;
+                v += (ikv2 + ikv3*args.ne_12_2)*args.nb21*C;
+
+                if (!FC_flash_attn_ext_vec_has_mask) {
+                    if (ic + tiisg >= args.ne11) {
+                        sm[tiisg] = -MAXHALF;
+                    }
+                } else {
+                    pm = (device const half *) (mask) +
+                        iq1*C +
+                        (iq2%args.ne32)*(C*args.ne31) +
+                        (iq3%args.ne33)*(C*args.ne31*args.ne32);
+                }
+
+                ic = 0;
+            }
+
             if (FC_flash_attn_ext_vec_has_mask) {
                 sm[tiisg] = pm[ic + tiisg];
             }
@@ -5222,7 +5470,7 @@ void kernel_flash_attn_ext_vec_impl(
 
             // Q*K^T
             {
-                device      const k4_t * pk4 = (device const k4_t *) ((device const char *) k + ic*args.nb11);
+                device      const k4_t * pk4 = (device const k4_t *) (k + ic*args.nb11);
                 threadgroup const q4_t * pq4 = sq4;
 
                 pk4 += ty*NS10/4 + tx;
@@ -5237,7 +5485,7 @@ void kernel_flash_attn_ext_vec_impl(
                             mqk[cc] += dot((float4) pk4[cc*NE*NS10/4 +  ii*NL], (float4) pq4[ii*NL]);
                         }
                     } else {
-                        device const kd4_t * pk = (device const kd4_t *) ((device const char *) k + ((ic + NE*cc + ty)*args.nb11));
+                        device const kd4_t * pk = (device const kd4_t *) (k + ((ic + NE*cc + ty)*args.nb11));
 
                         k4_t mk;
 
@@ -5335,7 +5583,7 @@ void kernel_flash_attn_ext_vec_impl(
                 }
 
                 if (is_same<vd4_t, v4_t>::value) {
-                    device const v4_t * pv4 = (device const v4_t *) ((device const char *) v + ic*args.nb21);
+                    device const v4_t * pv4 = (device const v4_t *) (v + ic*args.nb21);
 
                     pv4 += ty*NS20/4 + tx;
 
@@ -5348,7 +5596,7 @@ void kernel_flash_attn_ext_vec_impl(
                     }
                 } else {
                     FOR_UNROLL (short cc = 0; cc < C/NE; ++cc) {
-                        device const vd4_t * pv4 = (device const vd4_t *) ((device const char *) v + ((ic + NE*cc + ty)*args.nb21));
+                        device const vd4_t * pv4 = (device const vd4_t *) (v + ((ic + NE*cc + ty)*args.nb21));
 
                         FOR_UNROLL (short ii = 0; ii < DV4/NL; ++ii) {
                             const short i = ii*NL + tx;
@@ -5473,7 +5721,7 @@ void kernel_flash_attn_ext_vec_impl(
         device float4 * dst4 = (device float4 *) dst;
         device float  * dst1 = (device float  *) dst + nrows*DV*NWG; // the S and M are stored after the results
 
-        const float S = NWG == 1 ? 1.0f/ss[0] : 1.0f;
+        const float S = NWG == 1 ? (ss[0] == 0.0f ? 0.0f : 1.0f/ss[0]) : 1.0f;
 
         // interleave the workgroup data
         for (short i = tiisg; i < DV4; i += NW) {
@@ -5511,8 +5759,8 @@ template<
     short DK,       // K head size
     short DV,       // V head size
     short NE = 4,   // head elements per thread
-    short Q  = 1,   // queries per threadgroup
-    short C  = 32>  // cache items per threadgroup
+    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPTG,  // queries per threadgroup
+    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
 kernel void kernel_flash_attn_ext_vec(
         constant ggml_metal_kargs_flash_attn_ext_vec & args,
         device const char * q,
@@ -5520,13 +5768,14 @@ kernel void kernel_flash_attn_ext_vec(
         device const char * v,
         device const char * mask,
         device const char * sinks,
+        device const char * pad,
         device       char * dst,
         threadgroup  half * shmem_f16 [[threadgroup(0)]],
         uint3   tgpig[[threadgroup_position_in_grid]],
         ushort  tiisg[[thread_index_in_simdgroup]],
         ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
 #define FWD_TMPL q4_t, k4_t, v4_t, qk_t, s_t, s4_t, o4_t, kd4_t, nl_k, deq_k_t4, vd4_t, nl_v, deq_v_t4, DK, DV, NE, Q, C
-#define FWD_ARGS args, q, k, v, mask, sinks, dst, shmem_f16, tgpig, tiisg, sgitg
+#define FWD_ARGS args, q, k, v, mask, sinks, pad, dst, shmem_f16, tgpig, tiisg, sgitg
     switch (FC_flash_attn_ext_vec_nsg) {
       // note: disabled cases to reduce library load time
         case 1:  kernel_flash_attn_ext_vec_impl<FWD_TMPL,  1>(FWD_ARGS); break;
@@ -5650,7 +5899,8 @@ kernel void kernel_flash_attn_ext_vec_reduce(
     const float m  = simd_max(M);
     const float ms = exp(M - m);
 
-    S = 1.0f/simd_sum(S*ms);
+    S = simd_sum(S*ms);
+    S = S == 0.0f ? 0.0f : 1.0f/S;
 
     const short DV4 = DV/4;
 
@@ -7237,7 +7487,7 @@ kernel void kernel_mul_mv_iq1_m_f32(
     kernel_mul_mv_iq1_m_f32_impl<N_R0_IQ1_M, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
-template<int nr0, typename args_t>
+template<int NR0, typename args_t>
 void kernel_mul_mv_iq4_nl_f32_impl(
         args_t args,
         device const char * src0,
@@ -7250,13 +7500,12 @@ void kernel_mul_mv_iq4_nl_f32_impl(
     const short NSG = FC_mul_mv_nsg;
 
     threadgroup float * shmem_f32 = (threadgroup float *) shmem;
-    const int nb = args.ne00/QK4_NL;
 
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
-    const int first_row = (r0 * NSG + sgitg) * nr0;
+    const int first_row = (r0 * NSG + sgitg) * NR0;
 
     const uint i12 = im%args.ne12;
     const uint i13 = im/args.ne12;
@@ -7267,6 +7516,9 @@ void kernel_mul_mv_iq4_nl_f32_impl(
     device const block_iq4_nl * x = (device const block_iq4_nl *) (src0 + offset0);
     device const float        * y = (device const float        *) (src1 + offset1);
 
+    const int nb   = args.ne00/QK4_NL;
+    const int ns01 = args.nb01/args.nb00;
+
     const short ix = tiisg/2;  // 0...15
     const short it = tiisg%2;  // 0 or 1
 
@@ -7274,24 +7526,25 @@ void kernel_mul_mv_iq4_nl_f32_impl(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
-    float sumf[nr0]={0.f};
+    float sumf[NR0]={0.f};
 
-    device const float * yb = y + ix * QK4_NL + it * 8;
+    device const float * yb = y + ix*QK4_NL + it*8;
 
     uint32_t aux32[2];
     thread const uint8_t * q8 = (thread const uint8_t *)aux32;
 
     float4 qf1, qf2;
 
-    for (int ib = ix; ib < nb; ib += 16) {
+    // [TAG_MUL_MV_WEIRD]
+    for (int ib = ix; ib < nb && ib < ns01; ib += 16) {
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0];
         yl[1] = y4[4];
         yl[2] = y4[1];
         yl[3] = y4[5];
 
-        for (short row = 0; row < nr0; row++) {
-            device const block_iq4_nl & xb = x[row*nb + ib];
+        for (short row = 0; row < NR0; row++) {
+            device const block_iq4_nl & xb = x[row*ns01 + ib];
             device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
 
             float4 acc1 = {0.f}, acc2 = {0.f};
@@ -7322,7 +7575,7 @@ void kernel_mul_mv_iq4_nl_f32_impl(
 
     device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
 
-    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
+    for (int row = 0; row < NR0 && first_row + row < args.ne0; ++row) {
         float sum_all = simd_sum(sumf[row]);
         if (tiisg == 0) {
             dst_f32[first_row + row] = sum_all;
@@ -7344,7 +7597,7 @@ kernel void kernel_mul_mv_iq4_nl_f32(
     kernel_mul_mv_iq4_nl_f32_impl<N_R0_IQ4_NL, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
-template<int nr0, typename args_t>
+template<int NR0, typename args_t>
 void kernel_mul_mv_iq4_xs_f32_impl(
         args_t args,
         device const char * src0,
@@ -7357,12 +7610,11 @@ void kernel_mul_mv_iq4_xs_f32_impl(
     const short NSG = FC_mul_mv_nsg;
 
     threadgroup float * shmem_f32 = (threadgroup float *) shmem;
-    const int nb = args.ne00/QK_K;
 
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
-    const int first_row = (r0 * NSG + sgitg) * nr0;
+    const int first_row = (r0 * NSG + sgitg) * NR0;
 
     const uint i12 = im%args.ne12;
     const uint i13 = im/args.ne12;
@@ -7373,6 +7625,9 @@ void kernel_mul_mv_iq4_xs_f32_impl(
     device const block_iq4_xs * x = (device const block_iq4_xs *) (src0 + offset0);
     device const float        * y = (device const float        *) (src1 + offset1);
 
+    const int nb   = args.ne00/QK_K;
+    const int ns01 = args.nb01/args.nb00;
+
     const short ix = tiisg/16;  // 0 or 1
     const short it = tiisg%16;  // 0...15
     const short ib = it/2;
@@ -7382,7 +7637,7 @@ void kernel_mul_mv_iq4_xs_f32_impl(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
-    float sumf[nr0]={0.f};
+    float sumf[NR0]={0.f};
 
     device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
 
@@ -7391,15 +7646,16 @@ void kernel_mul_mv_iq4_xs_f32_impl(
 
     float4 qf1, qf2;
 
-    for (int ibl = ix; ibl < nb; ibl += 2) {
+    // [TAG_MUL_MV_WEIRD]
+    for (int ibl = ix; ibl < nb && ibl < ns01; ibl += 2) {
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0];
         yl[1] = y4[4];
         yl[2] = y4[1];
         yl[3] = y4[5];
 
-        for (short row = 0; row < nr0; ++row) {
-            device const block_iq4_xs & xb = x[row*nb + ibl];
+        for (short row = 0; row < NR0; ++row) {
+            device const block_iq4_xs & xb = x[row*ns01 + ibl];
             device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
 
             float4 acc1 = {0.f}, acc2 = {0.f};
@@ -7429,7 +7685,7 @@ void kernel_mul_mv_iq4_xs_f32_impl(
 
     device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
 
-    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
+    for (int row = 0; row < NR0 && first_row + row < args.ne0; ++row) {
         float sum_all = simd_sum(sumf[row]);
         if (tiisg == 0) {
             dst_f32[first_row + row] = sum_all;
@@ -7451,7 +7707,7 @@ kernel void kernel_mul_mv_iq4_xs_f32(
     kernel_mul_mv_iq4_xs_f32_impl<N_R0_IQ4_XS, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
-template<int nr0, typename args_t>
+template<int NR0, typename args_t>
 void kernel_mul_mv_mxfp4_f32_impl(
         args_t args,
         device const char * src0,
@@ -7464,13 +7720,12 @@ void kernel_mul_mv_mxfp4_f32_impl(
     const short NSG = FC_mul_mv_nsg;
 
     threadgroup float * shmem_f32 = (threadgroup float *) shmem;
-    const int nb = args.ne00/QK_MXFP4;
 
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
-    const int first_row = (r0 * NSG + sgitg) * nr0;
+    const int first_row = (r0 * NSG + sgitg) * NR0;
 
     const uint i12 = im%args.ne12;
     const uint i13 = im/args.ne12;
@@ -7481,6 +7736,9 @@ void kernel_mul_mv_mxfp4_f32_impl(
     device const block_mxfp4 * x = (device const block_mxfp4 *) (src0 + offset0);
     device const float       * y = (device const float       *) (src1 + offset1);
 
+    const int nb   = args.ne00/QK_MXFP4;
+    const int ns01 = args.nb01/args.nb00; // this can be larger than nb for permuted src0 tensors
+
     const short ix = tiisg/2;  // 0...15
     const short it = tiisg%2;  // 0 or 1
 
@@ -7488,20 +7746,22 @@ void kernel_mul_mv_mxfp4_f32_impl(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
-    float sumf[nr0]={0.f};
+    float sumf[NR0]={0.f};
 
-    device const float * yb = y + ix * QK_MXFP4 + it * 8;
+    device const float * yb = y + ix*QK_MXFP4 + it*8;
+
+    // note: just the check `ib < nb` is enough, but adding the redundant `&& ib < ns01` check makes the kernel a bit faster
+    //       no idea why that is - needs some deeper investigation [TAG_MUL_MV_WEIRD]
+    for (int ib = ix; ib < nb && ib < ns01; ib += 16) {
+        device const float4 * y4 = (device const float4 *) yb;
 
-    for (int ib = ix; ib < nb; ib += 16) {
-        device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0];
         yl[1] = y4[4];
         yl[2] = y4[1];
         yl[3] = y4[5];
 
-#pragma unroll(nr0)
-        for (short row = 0; row < nr0; row++) {
-            device const block_mxfp4 & xb = x[row*nb + ib];
+        FOR_UNROLL (short row = 0; row < NR0; row++) {
+            device const block_mxfp4 & xb = x[row*ns01 + ib];
             device const uint8_t     * q2 = (device const uint8_t *)(xb.qs + 8*it);
 
             float4 acc1 = yl[0]*float4(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
@@ -7519,7 +7779,7 @@ void kernel_mul_mv_mxfp4_f32_impl(
 
     device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
 
-    for (int row = 0; row < nr0 && first_row + row < args.ne0; ++row) {
+    for (int row = 0; row < NR0 && first_row + row < args.ne0; ++row) {
         float sum_all = simd_sum(sumf[row]);
         if (tiisg == 0) {
             dst_f32[first_row + row] = sum_all;

ggml/src/ggml-musa/CMakeLists.txt

@@ -30,6 +30,8 @@ if (MUSAToolkit_FOUND)
     list(APPEND GGML_HEADERS_MUSA "../ggml-musa/mudnn.cuh")
 
     file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-tile*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
     file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-mma*.cu")
     list(APPEND GGML_SOURCES_MUSA ${SRCS})
     file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")

ggml/src/ggml-sycl/backend.hpp

@@ -18,6 +18,7 @@
 #include "concat.hpp"
 #include "conv.hpp"
 #include "convert.hpp"
+#include "count-equal.hpp"
 #include "cpy.hpp"
 #include "dequantize.hpp"
 #include "dmmv.hpp"
@@ -28,6 +29,7 @@
 #include "mmvq.hpp"
 #include "norm.hpp"
 #include "outprod.hpp"
+#include "pad.hpp"
 #include "quantize.hpp"
 #include "quants.hpp"
 #include "rope.hpp"

ggml/src/ggml-sycl/binbcast.cpp

@@ -303,10 +303,6 @@ inline void ggml_sycl_op_sub(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_sub>>(ctx, dst->src[0], dst->src[1], dst);
 }
 
-inline void ggml_sycl_op_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_count_equal>>(ctx, dst->src[0], dst->src[1], dst);
-}
-
 inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
 
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, dst->src[0], dst->src[1], dst);
@@ -332,11 +328,6 @@ void ggml_sycl_sub(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_sub(ctx, dst);
 }
 
-void ggml_sycl_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
-    ggml_sycl_op_count_equal(ctx, dst);
-}
-
 void ggml_sycl_mul(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
     ggml_sycl_op_mul(ctx, dst);

ggml/src/ggml-sycl/binbcast.hpp

@@ -16,12 +16,6 @@ static __dpct_inline__ float op_sub(const float a, const float b) {
     return a - b;
 }
 
-static __dpct_inline__ float op_count_equal(const float a, const float b) {
-    return (a == b) ? 1.0f : 0.0f;
-}
-
-void ggml_sycl_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
 static __dpct_inline__ float op_mul(const float a, const float b) {
     return a * b;
 }

ggml/src/ggml-sycl/common.hpp

@@ -195,8 +195,10 @@ struct optimize_feature {
 
 struct sycl_device_info {
     int     cc;                 // compute capability
-    // int     nsm;                // number of streaming multiprocessors
+    int nsm; // number of streaming multiprocessors (CUDA) maps to the maximum
+             // number of compute units on a SYCL device.
     // size_t  smpb;               // max. shared memory per block
+    size_t  smpbo;              // max. shared memory per block (with opt-in)
     bool    vmm;                // virtual memory support
     size_t  total_vram;
     //sycl_hw_info hw_info;     \\ device id and aarch, currently not used
@@ -416,13 +418,6 @@ static __dpct_inline__ float warp_reduce_sum(float x,
     const sycl::nd_item<3>& item_ct1) {
 #pragma unroll
     for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
-        /*
-        DPCT1096:98: The right-most dimension of the work-group used in the SYCL
-        kernel that calls this function may be less than "32". The function
-        "dpct::permute_sub_group_by_xor" may return an unexpected result on the
-        CPU device. Modify the size of the work-group to ensure that the value
-        of the right-most dimension is a multiple of "32".
-        */
         x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask);
     }
     return x;
@@ -440,17 +435,67 @@ warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3>& item_ct1) {
     return a;
 }
 
+template <int width = WARP_SIZE>
+static __dpct_inline__ int warp_reduce_sum(int x) {
+  return sycl::reduce_over_group(
+      sycl::ext::oneapi::this_work_item::get_sub_group(), x, sycl::plus<>());
+}
+
+template <int width = WARP_SIZE>
+static __dpct_inline__ float warp_reduce_sum(float x) {
+#pragma unroll
+  for (int offset = width / 2; offset > 0; offset >>= 1) {
+    x += dpct::permute_sub_group_by_xor(
+        sycl::ext::oneapi::this_work_item::get_sub_group(), x, offset, width);
+  }
+  return x;
+}
+
+template <int width = WARP_SIZE>
+static __dpct_inline__ sycl::float2 warp_reduce_sum(sycl::float2 a) {
+#pragma unroll
+  for (int offset = width / 2; offset > 0; offset >>= 1) {
+    a.x() += dpct::permute_sub_group_by_xor(
+        sycl::ext::oneapi::this_work_item::get_sub_group(), a.x(), offset,
+        width);
+    a.y() += dpct::permute_sub_group_by_xor(
+        sycl::ext::oneapi::this_work_item::get_sub_group(), a.y(), offset,
+        width);
+  }
+  return a;
+}
+
+template <int width = WARP_SIZE>
+static __dpct_inline__ sycl::half2 warp_reduce_sum(sycl::half2 a) {
+#pragma unroll
+  for (int offset = width / 2; offset > 0; offset >>= 1) {
+    a = a + dpct::permute_sub_group_by_xor(
+                sycl::ext::oneapi::this_work_item::get_sub_group(), a, offset,
+                width);
+  }
+  return a;
+}
+
+static constexpr int ggml_sycl_get_physical_warp_size() {
+  // todo: for old iGPU + dGPU case, need to be changed.
+  return WARP_SIZE;
+}
+
+template <int width = WARP_SIZE>
+static __dpct_inline__ float warp_reduce_max(float x) {
+#pragma unroll
+  for (int offset = width / 2; offset > 0; offset >>= 1) {
+    x = sycl::fmax(x, dpct::permute_sub_group_by_xor(
+                          sycl::ext::oneapi::this_work_item::get_sub_group(), x,
+                          offset, width));
+  }
+  return x;
+}
+
 static __dpct_inline__ float warp_reduce_max(float x,
     const sycl::nd_item<3>& item_ct1) {
 #pragma unroll
     for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
-        /*
-        DPCT1096:97: The right-most dimension of the work-group used in the SYCL
-        kernel that calls this function may be less than "32". The function
-        "dpct::permute_sub_group_by_xor" may return an unexpected result on the
-        CPU device. Modify the size of the work-group to ensure that the value
-        of the right-most dimension is a multiple of "32".
-        */
         x = sycl::fmax(x, dpct::permute_sub_group_by_xor(
             item_ct1.get_sub_group(), x, mask));
     }
@@ -558,4 +603,18 @@ struct scope_op_debug_print {
     std::string_view func_suffix;
 };
 
+static __dpct_inline__ float get_alibi_slope(const float    max_bias,
+                                             const uint32_t h,
+                                             const uint32_t n_head_log2,
+                                             const float    m0,
+                                             const float    m1) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return dpct::pow(base, exph);
+}
+
 #endif // GGML_SYCL_COMMON_HPP

ggml/src/ggml-sycl/count-equal.cpp

@@ -0,0 +1,79 @@
+#include "count-equal.hpp"
+
+#include <cstdint>
+
+template <typename T>
+static void count_equal(const T *__restrict__ x, const T *__restrict__ y,
+                        int64_t *__restrict__ dst, const int64_t dk,
+                        const int64_t k) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int64_t i0 = (int64_t)item_ct1.get_group(2) * dk;
+    const int64_t i1 = sycl::min(i0 + dk, k);
+
+    int nequal = 0;
+
+    for (int64_t i = i0 + item_ct1.get_local_id(2); i < i1; i += WARP_SIZE) {
+        const T xi = x[i];
+        const T yi = y[i];
+        nequal += xi == yi;
+    }
+
+    nequal = warp_reduce_sum(nequal);
+
+    if (item_ct1.get_local_id(2) != 0) {
+        return;
+    }
+
+    dpct::atomic_fetch_add<sycl::access::address_space::generic_space>(
+        (int *)dst, nequal);
+}
+
+void ggml_sycl_count_equal(ggml_backend_sycl_context &ctx, ggml_tensor *dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT( dst->type == GGML_TYPE_I64);
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    int64_t * dst_d  = (int64_t *) dst->data;
+
+    dpct::queue_ptr stream = ctx.stream();
+    const int id       = get_current_device_id();
+    const int nsm = ggml_sycl_info().devices[id].nsm;
+
+    const int64_t ne = ggml_nelements(src0);
+    GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int");
+    const int64_t dne =
+        GGML_PAD((ne + 4 * nsm - 1) / (4 * nsm), SYCL_COUNT_EQUAL_CHUNK_SIZE);
+
+    SYCL_CHECK(CHECK_TRY_ERROR(stream->memset(dst_d, 0, ggml_nbytes(dst))));
+
+    const dpct::dim3 block_dims(WARP_SIZE, 1, 1);
+    const dpct::dim3 block_nums(
+        std::min((int64_t)4 * nsm, (ne + SYCL_COUNT_EQUAL_CHUNK_SIZE - 1) /
+                                       SYCL_COUNT_EQUAL_CHUNK_SIZE),
+        1, 1);
+
+    switch (src0->type) {
+    case GGML_TYPE_I32: {
+        const int *src0_d = (const int *)src0->data;
+        const int *src1_d = (const int *)src1->data;
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                count_equal(src0_d, src1_d, dst_d, dne, ne);
+                GGML_UNUSED(item_ct1);
+            });
+
+    } break;
+    default:
+        GGML_ASSERT(false);
+        break;
+    }
+}

ggml/src/ggml-sycl/count-equal.hpp

@@ -0,0 +1,9 @@
+#ifndef GGML_SYCL_COUNT_EQUAL_HPP
+#define GGML_SYCL_COUNT_EQUAL_HPP
+#include "common.hpp"
+
+#define SYCL_COUNT_EQUAL_CHUNK_SIZE 128
+
+void ggml_sycl_count_equal(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif //GGML_SYCL_COUNT_EQUAL_HPP

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -277,6 +277,26 @@ namespace dpct
 
     } // namespace detail
 
+    // COPY from DPCT head files
+    /// dim3 is used to store 3 component dimensions.
+    class dim3 {
+        public:
+        unsigned x, y, z;
+
+        constexpr dim3(unsigned x = 1, unsigned y = 1, unsigned z = 1)
+            : x(x), y(y), z(z) {}
+
+        dim3(const sycl::id<3> &r) : dim3(r[2], r[1], r[0]) {}
+
+        operator sycl::range<3>() const { return sycl::range<3>(z, y, x); }
+    }; // namespace dim3
+
+    inline dim3 operator*(const dim3 &a, const dim3 &b) {
+    return dim3{a.x * b.x, a.y * b.y, a.z * b.z};
+    }
+    // COPY from DPCT head files
+
+
     /// Pitched 2D/3D memory data.
     class pitched_data
     {

ggml/src/ggml-sycl/element_wise.cpp

@@ -328,26 +328,6 @@ static void upscale(const T  *x, T *dst, const int nb00, const int nb01,
     dst[index] = *(const T *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
 
-template <typename T>
-static void pad(const T  *x, T *dst, const int ne0, const int ne00, const int ne01, const int ne02,
-                    const sycl::nd_item<3> &item_ct1) {
-    int nidx = SYCL_LOCAL_ID_CALC(item_ct1, 2);
-    if (nidx >= ne0) {
-        return;
-    }
-
-    // operation
-    int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
-                     item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-    if (nidx < ne00 && item_ct1.get_group(1) < (size_t) ne01 && item_ct1.get_group(0) < (size_t) ne02) {
-        int offset_src = nidx + item_ct1.get_group(1) * ne00 +
-                         item_ct1.get_group(0) * ne00 * ne01;
-            dst[offset_dst] = x[offset_src];
-    } else {
-        dst[offset_dst] = static_cast<T>(0.0f);
-    }
-}
-
 template<typename T>
 static void clamp(const T * x, T * dst, const float min, const float max, const int k,
                       const sycl::nd_item<1> &item_ct1) {
@@ -431,18 +411,6 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
         });
 }
 
-template<typename T>
-static void pad_sycl(const T *x, T *dst, const int ne00,
-                         const int ne01, const int ne02, const int ne0,
-                         const int ne1, const int ne2, queue_ptr stream) {
-    int num_blocks = ceil_div(ne0, SYCL_PAD_BLOCK_SIZE);
-    sycl::range<3> gridDim(ne2, ne1, num_blocks);
-    stream->parallel_for(
-                      sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
-                                        sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
-                      [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
-}
-
 template<typename KernelInvoker, typename... Args>
 static inline void dispatch_ggml_sycl_op_unary(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
 #if defined (GGML_SYCL_F16)
@@ -596,40 +564,6 @@ static inline void dispatch_ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx
     }
 }
 
-template<typename KernelInvoker, typename... Args>
-static inline void dispatch_ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst, KernelInvoker kernel_invoker, Args&&... args) {
-#if defined (GGML_SYCL_F16)
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
-#else
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-#endif
-    GGML_ASSERT(dst->src[0]->type == dst->type);
-    GGML_ASSERT(dst->src[0]->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-    dpct::queue_ptr main_stream = ctx.stream();
-    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
-    switch (dst->type) {
-#if defined (GGML_SYCL_F16)
-        case GGML_TYPE_F16:
-            {
-                auto data_pts = cast_data<sycl::half>(dst);
-                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
-                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
-                break;
-            }
-#endif
-        case GGML_TYPE_F32:
-            {
-                auto data_pts = cast_data<float>(dst);
-                kernel_invoker(data_pts.src, data_pts.dst, (int)dst->src[0]->ne[0], (int)dst->src[0]->ne[1], (int)dst->src[0]->ne[2], (int)dst->ne[0],
-                               (int)dst->ne[1], (int)dst->ne[2], main_stream, std::forward<Args>(args)...);
-                break;
-            }
-        default:
-            GGML_ABORT("GGML tensor type not supported!\n");
-    }
-}
 
 } // namespace ggml_sycl_detail
 
@@ -919,14 +853,6 @@ static inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_te
         });
 }
 
-static inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    ggml_sycl_detail::dispatch_ggml_sycl_op_pad(ctx, dst,
-        [](const auto* src, auto* dst_ptr, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2,
-           queue_ptr stream) {
-            ggml_sycl_detail::pad_sycl(src, dst_ptr, ne00, ne01, ne02, ne0, ne1, ne2, stream);
-        });
-}
-
 static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     float min_val;
     float max_val;
@@ -1119,10 +1045,6 @@ void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     ggml_sycl_op_upscale(ctx, dst);
 }
 
-void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
-    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
-    ggml_sycl_op_pad(ctx, dst);
-}
 
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);

ggml/src/ggml-sycl/element_wise.hpp

@@ -67,8 +67,6 @@ void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
-void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
 void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
 
 void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -85,7 +85,10 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
         info.devices[i].cc =
             100 * prop.get_major_version() + 10 * prop.get_minor_version();
+        info.devices[i].nsm = prop.get_max_compute_units();
         info.devices[i].opt_feature.reorder = device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
+        info.devices[i].smpbo = prop.get_local_mem_size();
+
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
 
@@ -1511,60 +1514,70 @@ static inline void ggml_sycl_swap(T & a, T & b) {
 template <ggml_sort_order order>
 __dpct_inline__ static void
 k_argsort_f32_i32(const float *x, int *dst, const int ncols, int ncols_pad,
-                  const sycl::nd_item<3> &item_ct1, uint8_t *dpct_local) {
+                  const int tasks_per_thread, const sycl::nd_item<3> &item_ct1,
+                  uint8_t *dpct_local) {
     // bitonic sort
-    int col = item_ct1.get_local_id(2);
+    int col_index =  item_ct1.get_local_id(2);
     int row = item_ct1.get_group(1);
 
-    if (col >= ncols_pad) {
-        return;
+    for (int i = 0; i < tasks_per_thread; i++) {
+        int col = col_index * tasks_per_thread + i;
+        if (col >= ncols_pad) {
+            return;
+        }
     }
 
     const float * x_row = x + row * ncols;
     auto dst_row = (int *)dpct_local;
 
     // initialize indices
-    dst_row[col] = col;
+    for (int i=0;i<tasks_per_thread;i++){
+        int col = col_index*tasks_per_thread+i;
+        dst_row[col] = col;
+    }
 
     item_ct1.barrier(sycl::access::fence_space::local_space);
 
     for (int k = 2; k <= ncols_pad; k *= 2) {
         for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (dst_row[col] >= ncols ||
-                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (dst_row[ixj] >= ncols ||
-                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
-                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
-                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
-                    ) {
-                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+            for (int i = 0; i < tasks_per_thread; i++) {
+                int col = col_index * tasks_per_thread + i;
+                int ixj = col ^ j;
+                if (ixj > col) {
+                    if ((col & k) == 0) {
+                        if (dst_row[col] >= ncols ||
+                            (dst_row[ixj] < ncols &&
+                             (order == GGML_SORT_ORDER_ASC
+                                  ? x_row[dst_row[col]] > x_row[dst_row[ixj]]
+                                  : x_row[dst_row[col]] <
+                                        x_row[dst_row[ixj]]))) {
+                            ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                        }
+                    } else {
+                        if (dst_row[ixj] >= ncols ||
+                            (dst_row[col] < ncols &&
+                             (order == GGML_SORT_ORDER_ASC
+                                  ? x_row[dst_row[col]] < x_row[dst_row[ixj]]
+                                  : x_row[dst_row[col]] >
+                                        x_row[dst_row[ixj]]))) {
+                            ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                        }
                     }
                 }
+                item_ct1.barrier(sycl::access::fence_space::local_space);
             }
-            /*
-            DPCT1118:1: SYCL group functions and algorithms must be encountered
-            in converged control flow. You may need to adjust the code.
-            */
-            item_ct1.barrier(sycl::access::fence_space::local_space);
         }
     }
 
     // copy the result to dst without the padding
-    if (col < ncols) {
-        dst[row * ncols + col] = dst_row[col];
+    for (int i = 0; i < tasks_per_thread; i++) {
+        int col = col_index * tasks_per_thread + i;
+        if (col < ncols) {
+            dst[row * ncols + col] = dst_row[col];
+        }
     }
 }
 
-
 static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past,
                               const sycl::nd_item<3> &item_ct1) {
     const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
@@ -1737,11 +1750,20 @@ static int next_power_of_2(int x) {
 
 static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                                  const int nrows, ggml_sort_order order,
-                                 queue_ptr stream) {
+                                 queue_ptr stream, int device) {
     // bitonic sort requires ncols to be power of 2
     const int ncols_pad = next_power_of_2(ncols);
 
-    const sycl::range<3> block_dims(1, 1, ncols_pad);
+    int nth = 1;
+    int max_block_size = ggml_sycl_info().max_work_group_sizes[device];
+    while (nth < ncols_pad && nth < max_block_size)
+        nth *= 2;
+    if (nth > max_block_size)
+        nth = max_block_size;
+
+    const int tasks_per_thread = ncols_pad / nth;
+
+    const sycl::range<3> block_dims(1, 1, nth);
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = ncols_pad * sizeof(int);
 
@@ -1754,8 +1776,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
-                        x, dst, ncols, ncols_pad, item_ct1,
-                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                        x, dst, ncols, ncols_pad, tasks_per_thread, item_ct1,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
                             .get());
                 });
         });
@@ -1768,8 +1791,9 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
-                        x, dst, ncols, ncols_pad, item_ct1,
-                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                        x, dst, ncols, ncols_pad, tasks_per_thread, item_ct1,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
                             .get());
                 });
         });
@@ -2141,7 +2165,8 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor *
 
     enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
 
-    argsort_f32_i32_sycl(src0_dd, (int *) dst_dd, ncols, nrows, order, main_stream);
+    argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order,
+                         main_stream, ctx.device);
 }
 
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
@@ -3741,6 +3766,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
         case GGML_OP_SOFT_MAX:
             ggml_sycl_op_soft_max(ctx, dst);
             break;
+        case GGML_OP_SOFT_MAX_BACK:
+            ggml_sycl_op_soft_max_back(ctx, dst);
+            break;
         case GGML_OP_ROPE:
             ggml_sycl_rope(ctx, dst);
             break;
@@ -3778,6 +3806,7 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
     return true;
 } catch (sycl::exception & e) {
     std::cerr << e.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+    std::cerr << "Error OP "<<ggml_op_name(dst->op)<< std::endl;
     std::exit(1);
 }
 
@@ -4386,19 +4415,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
             return true;
         case GGML_OP_CONT:
             return op->src[0]->type != GGML_TYPE_BF16;
-        case GGML_OP_SOFT_MAX:
-            // TODO: support batching
-            if (op->src[0]->ne[3] != 1) {
-                return false;
-            }
-            // TODO: support attention sinks [TAG_ATTN_SINKS]
-            if (op->src[2]) {
-                return false;
-            }
-            // TODO: support broadcast
-            // ref: https://github.com/ggml-org/llama.cpp/pull/14435
-            return !op->src[1] || (op->src[1]->ne[2] == 1 && op->src[1]->ne[3] == 1);
         case GGML_OP_DIAG_MASK_INF:
+            return true;
+        case GGML_OP_SOFT_MAX:
+            return true;
+        case GGML_OP_SOFT_MAX_BACK: {
+            float max_bias = 0.0f;
+            memcpy(&max_bias, (const float *) op->op_params + 1, sizeof(float));
+            return max_bias == 0.0f;
+        }
         case GGML_OP_ROPE:
         case GGML_OP_IM2COL:
             return true;
@@ -4412,8 +4437,7 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_ACC:
             return true;
         case GGML_OP_PAD:
-            return (ggml_get_op_params_i32(op, 0) == 0) && (ggml_get_op_params_i32(op, 2) == 0) &&
-                   (ggml_get_op_params_i32(op, 4) == 0) && (ggml_get_op_params_i32(op, 6) == 0);
+            return ggml_is_contiguous(op->src[0]);
         case GGML_OP_LEAKY_RELU:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_RWKV_WKV6:

ggml/src/ggml-sycl/pad.cpp

@@ -0,0 +1,97 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+//#include "common.hpp"
+#include "pad.hpp"
+
+static void pad_f32(const float * src, float * dst,
+                               const int lp0, const int rp0, const int lp1, const int rp1,
+                               const int lp2, const int rp2, const int lp3, const int rp3,
+                               const int ne0, const int ne1, const int ne2, const int ne3) {
+    auto item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    int i0 = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    int i1 = item_ct1.get_group(1);
+    int i2 = item_ct1.get_group(0) % ne2;
+    int i3 = item_ct1.get_group(0) / ne2;
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    // operation
+    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+    if ((i0 >= lp0 && i0 < ne0 - rp0) &&
+        (i1 >= lp1 && i1 < ne1 - rp1) &&
+        (i2 >= lp2 && i2 < ne2 - rp2) &&
+        (i3 >= lp3 && i3 < ne3 - rp3)) {
+        const int64_t i00 = i0 - lp0;
+        const int64_t i01 = i1 - lp1;
+        const int64_t i02 = i2 - lp2;
+        const int64_t i03 = i3 - lp3;
+        const int64_t ne02 = ne2 - lp2 - rp2;
+        const int64_t ne01 = ne1 - lp1 - rp1;
+        const int64_t ne00 = ne0 - lp0 - rp0;
+
+        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) +
+                                i02 * (ne00 * ne01) + i01 * ne00 + i00;
+
+        dst[dst_idx] = src[src_idx];
+    } else {
+        dst[dst_idx] = 0.0f;
+    }
+}
+
+static void pad_f32_sycl(const float *src, float *dst, const int lp0,
+                         const int rp0, const int lp1, const int rp1,
+                         const int lp2, const int rp2, const int lp3,
+                         const int rp3, const int ne0, const int ne1,
+                         const int ne2, const int ne3,
+                         dpct::queue_ptr stream) {
+    int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
+    dpct::dim3 gridDim(num_blocks, ne1, ne2 * ne3);
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pad_f32(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1,
+                    ne2, ne3);
+        });
+}
+
+void ggml_sycl_op_pad(ggml_backend_sycl_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;
+    dpct::queue_ptr     stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int32_t lp0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t rp0 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t lp1 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t rp1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t lp2 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t rp2 = ((const int32_t*)(dst->op_params))[5];
+    const int32_t lp3 = ((const int32_t*)(dst->op_params))[6];
+    const int32_t rp3 = ((const int32_t*)(dst->op_params))[7];
+
+    pad_f32_sycl(src0_d, dst_d,
+                 lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
+                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
+
+void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);
+    ggml_sycl_op_pad(ctx, dst);
+}

ggml/src/ggml-sycl/pad.hpp

@@ -0,0 +1,24 @@
+//
+// MIT license
+// Copyright (C) 2025 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_PAD_HPP
+#define GGML_SYCL_PAD_HPP
+
+#include "common.hpp"
+
+#define SYCL_PAD_BLOCK_SIZE 256
+
+void ggml_sycl_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
+#endif // GGML_SYCL_PAD_HPP

ggml/src/ggml-sycl/softmax.cpp

@@ -1,37 +1,94 @@
 #include "softmax.hpp"
+#include <cstdint>
+#include <utility>
+#include <cmath>
 
-template <bool vals_smem, int ncols_template, int block_size_template, typename T>
-static 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 sycl::nd_item<3> &item_ct1, float *buf) {
-    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
 
-    const int tid = item_ct1.get_local_id(2);
-    const int rowx = item_ct1.get_group(2);
-    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
+template <typename T> static __dpct_inline__ float t2f32(T val) {
+    return (float) val;
+}
 
-    const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template;
+template <> float __dpct_inline__ t2f32<sycl::half>(sycl::half val) {
+  return sycl::vec<sycl::half, 1>(val)
+      .convert<float, sycl::rounding_mode::automatic>()[0];
+}
 
-    const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
-    const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+struct soft_max_params {
+
+    int64_t nheads;
+    uint32_t n_head_log2;
+    int64_t ncols;
+    int64_t nrows_x;
+    int64_t nrows_y;
+    int64_t ne00;
+    int64_t ne01;
+    int64_t ne02;
+    int64_t ne03;
+    int64_t nb11;
+    int64_t nb12;
+    int64_t nb13;
+
+    int64_t ne12;
+    int64_t ne13;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+};
+
+// When ncols_template == 0 the bounds for the loops in this function are not known and can't be unrolled.
+// As we want to keep pragma unroll for all other cases we supress the clang transformation warning here.
+#ifdef __clang__
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpass-failed"
+#endif // __clang__
+template <bool use_shared, int ncols_template, int block_size_template, typename T>
+static void soft_max_f32(const float *         x,
+                         const T *             mask,
+                         const float *         sinks,
+                         float *               dst,
+                         const soft_max_params p,
+                         uint8_t *             dpct_local) {
+    auto      item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int ncols    = ncols_template == 0 ? p.ncols : ncols_template;
+    const int block_size = block_size_template == 0
+                               ? item_ct1.get_local_range(2)
+                               : block_size_template;
     const int nthreads = block_size;
     const int nwarps = nthreads / WARP_SIZE;
     size_t nreduce = nwarps / WARP_SIZE;
-    float slope = 1.0f;
 
-    // ALiBi
-    if (max_bias > 0.0f) {
-        const uint32_t h = rowx/nrows_y; // head index
+    const int tid = item_ct1.get_local_id(2);
 
-        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 int64_t i03 = item_ct1.get_group(0);
+    const int64_t i02 = item_ct1.get_group(1);
+    const int64_t i01 = item_ct1.get_group(2);
 
-        slope = sycl::pow(base, float(exp));
-    }
+    //TODO: noncontigous inputs/outputs
+    const int rowx = item_ct1.get_group(2) +
+                     item_ct1.get_group(1) * item_ct1.get_group_range(2) +
+                     item_ct1.get_group(0) * item_ct1.get_group_range(2) *
+                         item_ct1.get_group_range(1);
 
-    float *vals = vals_smem ? buf + sycl::max(nwarps, WARP_SIZE) : dst + rowx * ncols;
-    float max_val = -INFINITY;
+    const int64_t i11 = i01;
+    const int64_t i12 = i02 % p.ne12;
+    const int64_t i13 = i03 % p.ne13;
 
+    x    += int64_t(rowx)*ncols;
+    mask += (i11*p.nb11 + i12*p.nb12 + i13*p.nb13) / sizeof(T) * (mask != nullptr);
+    dst  += int64_t(rowx)*ncols;
+
+    const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+    const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+
+    const float slope = get_alibi_slope(p.max_bias, i02, p.n_head_log2, p.m0, p.m1);
+
+    float * buf_iw = (float *) dpct_local;
+
+    // shared memory buffer to cache values between iterations:
+    float *vals = use_shared ? buf_iw + sycl::max(nwarps, WARP_SIZE) : dst;
+    float max_val = sinks ? sinks[i02] : -INFINITY;
+#pragma unroll
     for (int col0 = 0; col0 < ncols; col0 += block_size) {
         const int col = col0 + tid;
 
@@ -39,42 +96,35 @@ static void soft_max_f32(const float * x, const T * mask, float * dst, const int
             break;
         }
 
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-
-        const float val = x[ix]*scale + (mask ? slope*static_cast<float>(mask[iy]) : 0.0f);
+        const float val = x[col]*p.scale + (mask ? slope*t2f32(mask[col]) : 0.0f);
 
         vals[col] = val;
-        max_val = sycl::max(max_val, val);
+        max_val   = sycl::max(max_val, val);
     }
-
     // find the max value in the block
-    max_val = warp_reduce_max(max_val, item_ct1);
+    max_val = warp_reduce_max(max_val);
+
     if (block_size > WARP_SIZE) {
         if (warp_id == 0) {
-            buf[lane_id] = -INFINITY;
-            for (size_t i = 1; i < nreduce; i += 1) {
-                buf[lane_id + i * WARP_SIZE] = -INFINITY;
-            }
+            buf_iw[lane_id] = -INFINITY;
         }
-        item_ct1.barrier(sycl::access::fence_space::local_space);
+        item_ct1.barrier();
 
         if (lane_id == 0) {
-            buf[warp_id] = max_val;
+            buf_iw[warp_id] = max_val;
         }
-        item_ct1.barrier(sycl::access::fence_space::local_space);
-        max_val = buf[lane_id];
-        for (size_t i = 1; i < nreduce; i += 1) {
-            max_val = sycl::max(max_val, buf[lane_id + i * WARP_SIZE]);
-        }
-        max_val = warp_reduce_max(max_val, item_ct1);
-    }
+        item_ct1.barrier();
+
+        max_val = buf_iw[lane_id];
+        max_val = warp_reduce_max(max_val);
+    }
+    float tmp = 0.0f; // partial sum
 
-    float tmp = 0.f;
 #pragma unroll
     for (int col0 = 0; col0 < ncols; col0 += block_size) {
         const int col = col0 + tid;
-                if (ncols_template == 0 && col >= ncols) {
+
+        if (ncols_template == 0 && col >= ncols) {
             break;
         }
 
@@ -82,32 +132,33 @@ static void soft_max_f32(const float * x, const T * mask, float * dst, const int
         tmp += val;
         vals[col] = val;
     }
-
     // find the sum of exps in the block
-    tmp = warp_reduce_sum(tmp, item_ct1);
+    tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
-        item_ct1.barrier(sycl::access::fence_space::local_space);
+        item_ct1.barrier();
         if (warp_id == 0) {
-            buf[lane_id] = 0.f;
+            buf_iw[lane_id] = 0.0f;
             for (size_t i = 1; i < nreduce; i += 1) {
-                buf[lane_id + i * WARP_SIZE] = 0.f;
+                buf_iw[lane_id + i * WARP_SIZE] = 0.f;
             }
         }
-        item_ct1.barrier(sycl::access::fence_space::local_space);
+        item_ct1.barrier();
 
         if (lane_id == 0) {
-            buf[warp_id] = tmp;
+            buf_iw[warp_id] = tmp;
         }
-        item_ct1.barrier(sycl::access::fence_space::local_space);
+        item_ct1.barrier();
 
-        tmp = buf[lane_id];
+        tmp = buf_iw[lane_id];
         for (size_t i = 1; i < nreduce; i += 1) {
-            tmp += buf[lane_id + i * WARP_SIZE];
+            tmp += buf_iw[lane_id + i * WARP_SIZE];
         }
-        tmp = warp_reduce_sum(tmp, item_ct1);
+        tmp = warp_reduce_sum(tmp);
     }
-
-    const float inv_sum = 1.f / tmp;
+    if (sinks) {
+        tmp += sycl::native::exp(sinks[i02] - max_val);
+    }
+    const float inv_sum = 1.0f / tmp;
 
 #pragma unroll
     for (int col0 = 0; col0 < ncols; col0 += block_size) {
@@ -117,145 +168,259 @@ static void soft_max_f32(const float * x, const T * mask, float * dst, const int
             return;
         }
 
-        const int idst = rowx*ncols + col;
-        dst[idst] = vals[col] * inv_sum;
+        dst[col] = vals[col] * inv_sum;
+    }
+}
+#ifdef __clang__
+#pragma clang diagnostic pop
+#endif // __clang__
+
+static void soft_max_back_f32(const float *grad, const float *dstf, float *dst,
+                              const int ncols, const float scale) {
+    auto      item_ct1 = sycl::ext::oneapi::this_work_item::get_nd_item<3>();
+    const int tid      = item_ct1.get_local_id(2);
+    const int rowx     = item_ct1.get_group(2);
+
+    grad += int64_t(rowx)*ncols;
+    dstf += int64_t(rowx)*ncols;
+    dst  += int64_t(rowx)*ncols;
+
+    float dgf_dot = 0.0f; // dot product of dst from forward pass and gradients
+
+    for (int col = tid; col < ncols; col += WARP_SIZE) {
+        dgf_dot += dstf[col]*grad[col];
+    }
+
+    dgf_dot = warp_reduce_sum(dgf_dot);
+
+    for (int col = tid; col < ncols; col += WARP_SIZE) {
+        dst[col] = scale * (grad[col] - dgf_dot) * dstf[col];
     }
 }
 
-template <bool vals_smem, int ncols_template, int block_size_template, typename T>
-static void soft_max_f32_submitter(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, sycl::range<3> block_nums, sycl::range<3> block_dims,
-                                   const size_t n_local_scratch, queue_ptr stream) {
+template <int... Ns, typename T>
+static void launch_soft_max_kernels(const float *           x,
+                                    const T *               mask,
+                                    const float *           sinks,
+                                    float *                 dst,
+                                    const soft_max_params & p,
+                                    dpct::queue_ptr         stream,
+                                    dpct::dim3              block_dims,
+                                    dpct::dim3              block_nums,
+                                    size_t                  nbytes_shared)
+{
+    auto launch_kernel = [=](auto I) -> bool {
+        constexpr int ncols = decltype(I)::value;
+        constexpr int block = (ncols > 1024 ? 1024 : ncols);
+        if (p.ncols == ncols) {
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
+                    sycl::range<1>(nbytes_shared), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(
+                        WARP_SIZE)]] {
+                        soft_max_f32<true, ncols, block>(
+                            x, mask, sinks, dst, p,
+                            dpct_local_acc_ct1
+                                .get_multi_ptr<sycl::access::decorated::no>()
+                                .get());
+                        GGML_UNUSED(item_ct1);
+                    });
+            });
+            return true;
+        }
+        return false;
+    };
+
+    // unary fold over launch_kernel
+    if ((launch_kernel(std::integral_constant<int, Ns>{}) || ...)) {
+        return;
+    }
+
     stream->submit([&](sycl::handler &cgh) {
-        sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
+        sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
+            sycl::range<1>(nbytes_shared), cgh);
 
         cgh.parallel_for(
             sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                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,
-                                                                             get_pointer(local_buf_acc));
-            });
+            [=](sycl::nd_item<3> item_ct1)
+                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                    soft_max_f32<true, 0, 0>(
+                        x, mask, sinks, dst, p,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
+                            .get());
+                    GGML_UNUSED(item_ct1);
+                });
     });
 }
 
-template<typename T>
-static void soft_max_f32_sycl(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,
-                              queue_ptr stream, int device) {
+template <typename T>
+static void soft_max_f32_sycl(const float *x, const T *mask,
+                              const float *sinks, float *dst,
+                              const soft_max_params &params,
+                              dpct::queue_ptr stream, int device) {
     int nth = WARP_SIZE;
     int max_block_size = ggml_sycl_info().max_work_group_sizes[device];
+    const int64_t ncols_x = params.ncols;
+
     while (nth < ncols_x && nth < max_block_size) nth *= 2;
     if (nth>max_block_size) nth = max_block_size;
 
-    const sycl::range<3> block_dims(1, 1, nth);
-    const sycl::range<3> block_nums(1, 1, nrows_x);
-    const size_t n_val_tmp = nth / WARP_SIZE;
-    const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + n_val_tmp);
+    const dpct::dim3 block_dims(nth, 1, 1);
+    const dpct::dim3 block_nums(params.ne01, params.ne02, params.ne03);
+    const size_t nbytes_shared =
+        (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE) * sizeof(float);
 
-    const uint32_t n_head_kv   = nrows_x/nrows_y;
-    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+    const int id       = get_current_device_id();
+    const size_t smpbo = ggml_sycl_info().devices[id].smpbo;
 
-    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
-
-    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, 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, 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, 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, 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, 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, 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, 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, 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, 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, dst, ncols_x, nrows_y, scale,
-                                                   max_bias, m0, m1, n_head_log2, block_nums,
-                                                   block_dims, n_local_scratch, stream);
-                break;
-        }
+    if (nbytes_shared <= smpbo) {
+        launch_soft_max_kernels<32, 64, 128, 256, 512, 1024, 2048, 4096>(
+            x, mask, sinks, dst, params, stream, block_dims, block_nums,
+            nbytes_shared);
     } else {
-        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);
+        const size_t nbytes_shared_low = WARP_SIZE * sizeof(float);
+
+        stream->submit([&](sycl::handler &cgh) {
+            sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
+                sycl::range<1>(nbytes_shared_low), cgh);
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
+                    soft_max_f32<false, 0, 0>(
+                        x, mask, sinks, dst, params,
+                        dpct_local_acc_ct1
+                            .get_multi_ptr<sycl::access::decorated::no>()
+                            .get());
+                    GGML_UNUSED(item_ct1);
+                });
+        });
     }
 }
 
+static void soft_max_back_f32_sycl(const float *   grad,
+                                   const float *   dstf,
+                                   float *         dst,
+                                   const int       ncols,
+                                   const int       nrows,
+                                   const float     scale,
+                                   dpct::queue_ptr stream) {
+    const dpct::dim3 block_dims(WARP_SIZE, 1, 1);
+    const dpct::dim3 block_nums(nrows, 1, 1);
+
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             soft_max_back_f32(grad, dstf, dst, ncols, scale);
+                             GGML_UNUSED(item_ct1);
+                         });
+}
+
 void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
     scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
-    GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
+
+    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;
+    const void  * src2_d = src2 ? (const void *) src2->data : nullptr;
+    float       *  dst_d = (float *) dst->data;
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    GGML_ASSERT(!dst->src[1] || dst->src[1]->type == GGML_TYPE_F16 || dst->src[1]->type == GGML_TYPE_F32); // src1 contains mask and it is optional
+    // src1 contains mask and it is optional
+    GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = dst->src[0]->ne[0];
-    const int64_t nrows_x = ggml_nrows(dst->src[0]);
-    const int64_t nrows_y = dst->src[0]->ne[1];
+    const int64_t nrows_x = ggml_nrows(src0);
+    const int64_t nrows_y = src0->ne[1];
 
-    float scale = 1.0f;
+    const int64_t ne00 = src0->ne[0];
+
+    float scale    = 1.0f;
     float max_bias = 0.0f;
 
-    memcpy(&scale, dst->op_params + 0, sizeof(float));
-    memcpy(&max_bias, dst->op_params + 1, sizeof(float));
+    memcpy(&scale,    (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
 
-    const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
-    float * dst_dd = static_cast<float *>(dst->data);
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
 
-    ggml_sycl_set_device(ctx.device);
-    dpct::queue_ptr main_stream = ctx.stream();
+    const int64_t nb11 = src1 ? src1->nb[1] : 1;
+    const int64_t nb12 = src1 ? src1->nb[2] : 1;
+    const int64_t nb13 = src1 ? src1->nb[3] : 1;
 
-    if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F16) {
-        const sycl::half * src1_dd = static_cast<sycl::half *>(dst->src[1]->data);
-        soft_max_f32_sycl<sycl::half>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias,
-                          main_stream, ctx.device);
-    } else if (dst->src[1] && dst->src[1]->type == GGML_TYPE_F32) {
-        const float * src1_dd = static_cast<const float *>(dst->src[1]->data);
-        soft_max_f32_sycl<float>(src0_dd, src1_dd, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
+    const int64_t ne12 = src1 ? src1->ne[2] : 1;
+    const int64_t ne13 = src1 ? src1->ne[3] : 1;
+
+    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);
+
+
+    soft_max_params params = {};
+    params.nheads = src0->ne[2];
+    params.n_head_log2 = n_head_log2;
+    params.ncols = ne00;
+    params.nrows_x = nrows_x;
+    params.nrows_y = nrows_y;
+    params.ne00 = src0->ne[0];
+    params.ne01 = src0->ne[1];
+    params.ne02 = src0->ne[2];
+    params.ne03 = src0->ne[3];
+    params.nb11 = nb11;
+    params.nb12 = nb12;
+    params.nb13 = nb13;
+    params.ne12 = ne12;
+    params.ne13 = ne13;
+    params.scale = scale;
+    params.max_bias = max_bias;
+    params.m0 = m0;
+    params.m1 = m1;
+
+    if (use_f16) {
+        soft_max_f32_sycl(src0_d, (const sycl::half *)src1_d,
+                          (const float *)src2_d, dst_d, params, stream,
+                          ctx.device);
     } else {
-        /* mask unavailable */
-        soft_max_f32_sycl<float>(src0_dd, nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
+        soft_max_f32_sycl(src0_d, (const float *)src1_d, (const float *)src2_d,
+                          dst_d, params, stream, ctx.device);
     }
 }
+
+void ggml_sycl_op_soft_max_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
+    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+    const ggml_tensor * src0 = dst->src[0]; // grad
+    const ggml_tensor * src1 = dst->src[1]; // forward pass output
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (const float *) dst->op_params + 1, sizeof(float));
+
+    GGML_ASSERT(max_bias == 0.0f);
+
+    soft_max_back_f32_sycl(src0_d, src1_d, dst_d, ncols, nrows, scale, stream);
+}

ggml/src/ggml-sycl/softmax.hpp

@@ -15,6 +15,10 @@
 
 #include "common.hpp"
 
+#define SYCL_SOFT_MAX_BLOCK_SIZE 1024
+
 void ggml_sycl_op_soft_max(ggml_backend_sycl_context &ctx, ggml_tensor *dst);
 
+void ggml_sycl_op_soft_max_back(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
+
 #endif // GGML_SYCL_SOFTMAX_HPP

ggml/src/ggml-webgpu/CMakeLists.txt

@@ -50,5 +50,13 @@ if (GGML_WEBGPU_DEBUG)
     target_compile_definitions(ggml-webgpu PRIVATE GGML_WEBGPU_DEBUG=1)
 endif()
 
+if (GGML_WEBGPU_CPU_PROFILE)
+    target_compile_definitions(ggml-webgpu PRIVATE GGML_WEBGPU_CPU_PROFILE=1)
+endif()
+
+if (GGML_WEBGPU_GPU_PROFILE)
+    target_compile_definitions(ggml-webgpu PRIVATE GGML_WEBGPU_GPU_PROFILE=1)
+endif()
+
 target_include_directories(ggml-webgpu PRIVATE ${SHADER_OUTPUT_DIR})
 target_link_libraries(ggml-webgpu PRIVATE ${DawnWebGPU_TARGET})

ggml/src/ggml-webgpu/wgsl-shaders/mul_mat.tmpl.wgsl

@@ -870,7 +870,7 @@ struct MulMatParams {
 
 @group(0) @binding(3) var<uniform> params: MulMatParams;
 
-@compute @workgroup_size(64)
+@compute @workgroup_size(256)
 fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
     let total = params.m * params.n * params.bs02 * params.broadcast2 * params.bs03 * params.broadcast3;
     if (global_id.x >= total) {

gguf-py/gguf/constants.py

@@ -128,6 +128,8 @@ class Keys:
         ALTUP_ACTIVE_IDX                  = "{arch}.altup.active_idx"
         ALTUP_NUM_INPUTS                  = "{arch}.altup.num_inputs"
         EMBD_LENGTH_PER_LAYER_INP         = "{arch}.embedding_length_per_layer_input"
+        DENSE_FEAT_IN_SIZE                = "{arch}.{dense}_feat_in"
+        DENSE_FEAT_OUT_SIZE               = "{arch}.{dense}_feat_out"
 
     class Attention:
         HEAD_COUNT                   = "{arch}.attention.head_count"
@@ -407,6 +409,7 @@ class MODEL_ARCH(IntEnum):
     SMOLLM3          = auto()
     GPT_OSS          = auto()
     LFM2             = auto()
+    LFM2MOE          = auto()
     DREAM            = auto()
     SMALLTHINKER     = auto()
     LLADA            = auto()
@@ -432,6 +435,8 @@ class MODEL_TENSOR(IntEnum):
     TOKEN_TYPES          = auto()
     POS_EMBD             = auto()
     OUTPUT               = auto()
+    DENSE_2_OUT          = auto() # embeddinggemma 2_Dense
+    DENSE_3_OUT          = auto() # embeddinggemma 3_Dense
     OUTPUT_NORM          = auto()
     ROPE_FREQS           = auto()
     ROPE_FACTORS_LONG    = auto()
@@ -749,6 +754,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.SMOLLM3:          "smollm3",
     MODEL_ARCH.GPT_OSS:          "gpt-oss",
     MODEL_ARCH.LFM2:             "lfm2",
+    MODEL_ARCH.LFM2MOE:          "lfm2moe",
     MODEL_ARCH.DREAM:            "dream",
     MODEL_ARCH.SMALLTHINKER:     "smallthinker",
     MODEL_ARCH.LLADA:            "llada",
@@ -775,6 +781,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.POS_EMBD:                  "position_embd",
     MODEL_TENSOR.OUTPUT_NORM:               "output_norm",
     MODEL_TENSOR.OUTPUT:                    "output",
+    MODEL_TENSOR.DENSE_2_OUT:                "dense_2", # embeddinggemma 2_Dense
+    MODEL_TENSOR.DENSE_3_OUT:                "dense_3", # embeddinggemma 2_Dense
     MODEL_TENSOR.ROPE_FREQS:                "rope_freqs",
     MODEL_TENSOR.ROPE_FACTORS_LONG:         "rope_factors_long",
     MODEL_TENSOR.ROPE_FACTORS_SHORT:        "rope_factors_short",
@@ -1757,6 +1765,8 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     MODEL_ARCH.GEMMA_EMBEDDING: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.DENSE_2_OUT,
+        MODEL_TENSOR.DENSE_3_OUT,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_Q_NORM,
@@ -2698,6 +2708,29 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.OUTPUT,
     ],
+    MODEL_ARCH.LFM2MOE: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.SHORTCONV_CONV,
+        MODEL_TENSOR.SHORTCONV_INPROJ,
+        MODEL_TENSOR.SHORTCONV_OUTPROJ,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.ATTN_NORM, # operator_norm
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_EXP_PROBS_B,
+    ],
     MODEL_ARCH.SMALLTHINKER: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,

gguf-py/gguf/gguf_writer.py

@@ -730,6 +730,10 @@ class GGUFWriter:
     def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
         self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value)
 
+    def add_dense_features_dims(self, dense:str, in_f:int, out_f:int) -> None:
+        self.add_uint32(Keys.LLM.DENSE_FEAT_IN_SIZE.format(arch=self.arch, dense=dense), in_f)
+        self.add_uint32(Keys.LLM.DENSE_FEAT_OUT_SIZE.format(arch=self.arch, dense=dense), out_f)
+
     def add_logit_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
 

gguf-py/gguf/tensor_mapping.py

@@ -76,7 +76,12 @@ class TensorNameMap:
             "lm_head",                   # llama4
             "model.transformer.ff_out",  # llada
         ),
-
+        MODEL_TENSOR.DENSE_2_OUT: (
+            "dense_2_out",  # embeddinggemma
+        ),
+        MODEL_TENSOR.DENSE_3_OUT: (
+            "dense_3_out",  # embeddinggemma
+        ),
         # Output norm
         MODEL_TENSOR.OUTPUT_NORM: (
             "gpt_neox.final_layer_norm",               # gptneox
@@ -358,6 +363,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.router",                    # openai-moe
             "model.layers.{bid}.mlp.gate.wg",                   # hunyuan
             "model.layers.{bid}.block_sparse_moe.primary_router", # smallthinker
+            "model.layers.{bid}.feed_forward.gate",               # lfm2moe
         ),
 
         MODEL_TENSOR.FFN_GATE_INP_SHEXP: (
@@ -367,6 +373,7 @@ class TensorNameMap:
         MODEL_TENSOR.FFN_EXP_PROBS_B: (
             "model.layers.{bid}.mlp.gate.e_score_correction",               # deepseek-v3 dots1
             "model.layers.{bid}.mlp.moe_statics.e_score_correction",        # ernie4.5-moe
+            "model.layers.{bid}.feed_forward.expert_bias",                  # lfm2moe
         ),
 
         # Feed-forward up

requirements/requirements-all.txt

@@ -14,3 +14,5 @@
 -r ./requirements-tool_bench.txt
 
 -r ./requirements-gguf_editor_gui.txt
+
+-r ../examples/model-conversion/requirements.txt

src/llama-arch.cpp

@@ -93,6 +93,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_SMOLLM3,          "smollm3"          },
     { LLM_ARCH_OPENAI_MOE,       "gpt-oss"          },
     { LLM_ARCH_LFM2,             "lfm2"             },
+    { LLM_ARCH_LFM2MOE,          "lfm2moe"          },
     { LLM_ARCH_DREAM,            "dream"            },
     { LLM_ARCH_SMALLTHINKER,     "smallthinker"     },
     { LLM_ARCH_LLADA,            "llada"            },
@@ -218,6 +219,11 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_CLASSIFIER_OUTPUT_LABELS, "%s.classifier.output_labels" },
 
     { LLM_KV_SHORTCONV_L_CACHE, "%s.shortconv.l_cache" },
+    // sentence-transformers dense modules feature dims
+    { LLM_KV_DENSE_2_FEAT_IN,        "%s.dense_2_feat_in"  },
+    { LLM_KV_DENSE_2_FEAT_OUT,       "%s.dense_2_feat_out"  },
+    { LLM_KV_DENSE_3_FEAT_IN,        "%s.dense_3_feat_in"   },
+    { LLM_KV_DENSE_3_FEAT_OUT,       "%s.dense_3_feat_out"  },
 
     { LLM_KV_TOKENIZER_MODEL,                "tokenizer.ggml.model"                    },
     { LLM_KV_TOKENIZER_PRE,                  "tokenizer.ggml.pre"                      },
@@ -1070,6 +1076,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
             { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_DENSE_2_OUT,     "dense_2" },
+            { LLM_TENSOR_DENSE_3_OUT,     "dense_3" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
@@ -2104,6 +2112,32 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_OUTPUT,            "output" },
         }
     },
+    {
+        LLM_ARCH_LFM2MOE,
+        {
+            { LLM_TENSOR_ATTN_NORM,         "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,            "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,            "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,            "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,          "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_K_NORM,       "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,       "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_FFN_DOWN,          "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_GATE,          "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_NORM,          "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,            "blk.%d.ffn_up" },
+            { LLM_TENSOR_SHORTCONV_CONV,    "blk.%d.shortconv.conv" },
+            { LLM_TENSOR_SHORTCONV_INPROJ,  "blk.%d.shortconv.in_proj" },
+            { LLM_TENSOR_SHORTCONV_OUTPROJ, "blk.%d.shortconv.out_proj" },
+            { LLM_TENSOR_TOKEN_EMBD,        "token_embd" },
+            { LLM_TENSOR_TOKEN_EMBD_NORM,   "token_embd_norm" },
+            { LLM_TENSOR_FFN_GATE_INP,      "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_GATE_EXPS,     "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,     "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,       "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_EXP_PROBS_B,   "blk.%d.exp_probs_b" },
+        }
+    },
     {
         LLM_ARCH_SMALLTHINKER,
         {
@@ -2254,6 +2288,8 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_DENSE_2_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
+    {LLM_TENSOR_DENSE_3_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
     {LLM_TENSOR_OUTPUT_NORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_DEC_OUTPUT_NORM,            {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_ENC_OUTPUT_NORM,            {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
@@ -2493,6 +2529,7 @@ bool llm_arch_is_hybrid(const llm_arch & arch) {
         case LLM_ARCH_PLAMO2:
         case LLM_ARCH_GRANITE_HYBRID:
         case LLM_ARCH_LFM2:
+        case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_NEMOTRON_H:
             return true;
         default:

src/llama-arch.h

@@ -97,6 +97,7 @@ enum llm_arch {
     LLM_ARCH_SMOLLM3,
     LLM_ARCH_OPENAI_MOE,
     LLM_ARCH_LFM2,
+    LLM_ARCH_LFM2MOE,
     LLM_ARCH_DREAM,
     LLM_ARCH_SMALLTHINKER,
     LLM_ARCH_LLADA,
@@ -270,6 +271,12 @@ enum llm_kv {
     LLM_KV_TOKENIZER_PREFIX_ID,
     LLM_KV_TOKENIZER_SUFFIX_ID,
     LLM_KV_TOKENIZER_MIDDLE_ID,
+
+    // sentence-transformers dense layers in and out features
+    LLM_KV_DENSE_2_FEAT_IN,
+    LLM_KV_DENSE_2_FEAT_OUT,
+    LLM_KV_DENSE_3_FEAT_IN,
+    LLM_KV_DENSE_3_FEAT_OUT,
 };
 
 enum llm_tensor {
@@ -277,6 +284,8 @@ enum llm_tensor {
     LLM_TENSOR_TOKEN_EMBD_NORM,
     LLM_TENSOR_TOKEN_TYPES,
     LLM_TENSOR_POS_EMBD,
+    LLM_TENSOR_DENSE_2_OUT,
+    LLM_TENSOR_DENSE_3_OUT,
     LLM_TENSOR_OUTPUT,
     LLM_TENSOR_OUTPUT_NORM,
     LLM_TENSOR_ROPE_FREQS,

src/llama-context.cpp

@@ -2346,6 +2346,12 @@ llama_context * llama_init_from_model(
         return nullptr;
     }
 
+    if (params.pooling_type != model->hparams.pooling_type) {
+        //user-specified pooling-type is different from the model default
+        LLAMA_LOG_WARN("%s: model default pooling_type is [%d], but [%d] was specified\n", __func__,
+                       model->hparams.pooling_type, params.pooling_type);
+    }
+
     try {
         auto * ctx = new llama_context(*model, params);
         return ctx;

src/llama-graph.cpp

@@ -1853,6 +1853,23 @@ llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const {
     return (llm_graph_input_mem_hybrid *) res->add_input(std::move(inp));
 }
 
+void llm_graph_context::build_dense_out(
+    ggml_tensor * dense_2,
+    ggml_tensor * dense_3) const {
+    if (!cparams.embeddings || dense_2 == nullptr || dense_3 == nullptr) {
+        return;
+    }
+    ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
+    GGML_ASSERT(cur != nullptr && "missing t_embd_pooled/t_embd");
+
+    cur = ggml_mul_mat(ctx0, dense_2, cur);
+    cur = ggml_mul_mat(ctx0, dense_3, cur);
+    cb(cur, "result_embd_pooled", -1);
+    res->t_embd_pooled = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+
+
 void llm_graph_context::build_pooling(
         ggml_tensor * cls,
         ggml_tensor * cls_b,

src/llama-graph.h

@@ -814,6 +814,14 @@ struct llm_graph_context {
             ggml_tensor * cls_b,
             ggml_tensor * cls_out,
             ggml_tensor * cls_out_b) const;
+
+    //
+    // dense (out)
+    //
+
+    void build_dense_out(
+            ggml_tensor * dense_2,
+            ggml_tensor * dense_3) const;
 };
 
 // TODO: better name

src/llama-hparams.cpp

@@ -140,7 +140,11 @@ uint32_t llama_hparams::n_embd_s() const {
 }
 
 bool llama_hparams::is_recurrent(uint32_t il) const {
-    return recurrent_layer_arr[il];
+    if (il < n_layer) {
+        return recurrent_layer_arr[il];
+    }
+
+    GGML_ABORT("%s: il (%u) out of bounds (n_layer: %u)\n", __func__, il, n_layer);
 }
 
 uint32_t llama_hparams::n_pos_per_embd() const {

src/llama-hparams.h

@@ -169,6 +169,12 @@ struct llama_hparams {
     uint32_t laurel_rank  = 64;
     uint32_t n_embd_altup = 256;
 
+    // needed for sentence-transformers dense layers
+    uint32_t dense_2_feat_in  = 0;  // in_features of the 2_Dense
+    uint32_t dense_2_feat_out = 0;  // out_features of the 2_Dense
+    uint32_t dense_3_feat_in  = 0;  // in_features of the 3_Dense
+    uint32_t dense_3_feat_out = 0;  // out_features of the 3_Dense
+
     // xIELU
     std::array<float, LLAMA_MAX_LAYERS> xielu_alpha_n;
     std::array<float, LLAMA_MAX_LAYERS> xielu_alpha_p;

src/llama-kv-cache.cpp

@@ -123,11 +123,8 @@ llama_kv_cache::llama_kv_cache(
             throw std::runtime_error("failed to create ggml context for kv cache");
         }
 
-        ggml_tensor * k;
-        ggml_tensor * v;
-
-        k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
-        v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
+        ggml_tensor * k = ggml_new_tensor_3d(ctx, type_k, n_embd_k_gqa, kv_size, n_stream);
+        ggml_tensor * v = ggml_new_tensor_3d(ctx, type_v, n_embd_v_gqa, kv_size, n_stream);
 
         ggml_format_name(k, "cache_k_l%d", il);
         ggml_format_name(v, "cache_v_l%d", il);

src/llama-memory-hybrid.cpp

@@ -73,7 +73,9 @@ llama_memory_context_ptr llama_memory_hybrid::init_batch(llama_batch_allocr & ba
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, false);
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                ubatch = balloc.split_equal(n_ubatch, true);
             }
 
             if (ubatch.n_tokens == 0) {

src/llama-memory-recurrent.cpp

@@ -382,7 +382,9 @@ llama_memory_context_ptr llama_memory_recurrent::init_batch(llama_batch_allocr &
                 // if all tokens are output, split by sequence
                 ubatch = balloc.split_seq(n_ubatch);
             } else {
-                ubatch = balloc.split_equal(n_ubatch, false);
+                // TODO: non-sequential equal split can be done if using unified KV cache
+                //       for simplicity, we always use sequential equal split for now
+                ubatch = balloc.split_equal(n_ubatch, true);
             }
 
             if (ubatch.n_tokens == 0) {
@@ -859,9 +861,12 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
 bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell_count, llama_seq_id dest_seq_id) {
     if (dest_seq_id != -1) {
         // single sequence
-
         seq_rm(dest_seq_id, -1, -1);
 
+        if (cell_count == 0) {
+            return true;
+        }
+
         llama_batch_allocr balloc(hparams.n_pos_per_embd());
 
         llama_ubatch ubatch = balloc.ubatch_reserve(cell_count, 1);

src/llama-model.cpp

@@ -114,6 +114,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_17B_16E:       return "17Bx16E (Scout)";
         case LLM_TYPE_17B_128E:      return "17Bx128E (Maverick)";
         case LLM_TYPE_A13B:          return "A13B";
+        case LLM_TYPE_8B_A1B:        return "8B.A1B";
         case LLM_TYPE_21B_A3B:       return "21B.A3B";
         case LLM_TYPE_30B_A3B:       return "30B.A3B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
@@ -1217,12 +1218,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 hparams.set_swa_pattern(6);
 
                 hparams.causal_attn = false; // embeddings do not use causal attention
-                hparams.rope_freq_base_train_swa  = 10000.0f;
+                hparams.rope_freq_base_train_swa = 10000.0f;
                 hparams.rope_freq_scale_train_swa = 1.0f;
 
-                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-                ml.get_key(LLM_KV_POOLING_TYPE,                hparams.pooling_type);
+                ml.get_key(LLM_KV_POOLING_TYPE, hparams.pooling_type);
+
+                //applied only if model converted with --sentence-transformers-dense-modules
+                ml.get_key(LLM_KV_DENSE_2_FEAT_IN, hparams.dense_2_feat_in, false);
+                ml.get_key(LLM_KV_DENSE_2_FEAT_OUT, hparams.dense_2_feat_out, false);
+                ml.get_key(LLM_KV_DENSE_3_FEAT_IN, hparams.dense_3_feat_in, false);
+                ml.get_key(LLM_KV_DENSE_3_FEAT_OUT, hparams.dense_3_feat_out, false);
+
+                GGML_ASSERT((hparams.dense_2_feat_in == 0 || hparams.dense_2_feat_in == hparams.n_embd) && "dense_2_feat_in must be equal to n_embd");
+                GGML_ASSERT((hparams.dense_3_feat_out == 0 || hparams.dense_3_feat_out == hparams.n_embd) && "dense_3_feat_out must be equal to n_embd");
 
                 switch (hparams.n_layer) {
                     case 24: type = LLM_TYPE_0_3B; break;
@@ -1995,14 +2005,29 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 for (uint32_t il = 0; il < hparams.n_layer; ++il) {
                     hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
                 }
+                hparams.n_layer_dense_lead = hparams.n_layer;
                 switch (hparams.n_ff()) {
                     case  4608: type = LLM_TYPE_350M; break;
                     case  6912: type = LLM_TYPE_700M; break;
                     case  8192: type = LLM_TYPE_1_2B; break;
                     case 10752: type = LLM_TYPE_2_6B; break;
-                    default:   type = LLM_TYPE_UNKNOWN;
+                    default:    type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_LFM2MOE:
+            {
+                ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func);
+
+                for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+                    hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
+                }
+
+                type = LLM_TYPE_8B_A1B;
+            } break;
         case LLM_ARCH_SMALLTHINKER:
             {
                 const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
@@ -3670,6 +3695,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
                     }
 
+                    // Dense linear weights
+                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);
+                    dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);
+
+
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
 
@@ -5814,6 +5844,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     }
                 } break;
             case LLM_ARCH_LFM2:
+            case LLM_ARCH_LFM2MOE:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
                     tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, 0);
@@ -5825,11 +5856,23 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
 
                     for (int i = 0; i < n_layer; ++i) {
                         auto & layer = layers[i];
-                        // ffn is same for transformer and conv layers
+
+                        const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);
+
+                        // ffn/moe is same for transformer and conv layers
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
-                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
-                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        if (is_moe_layer) {
+                            GGML_ASSERT(n_expert && n_expert_used);
+                            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);
+                            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);
+                            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);
+                            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+                        } else {  // dense
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        }
 
                         // for operator_norm
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
@@ -6310,7 +6353,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: expert_weights_norm  = %d\n",     __func__, hparams.expert_weights_norm);
     }
 
-    if (arch == LLM_ARCH_SMALLTHINKER) {
+    if (arch == LLM_ARCH_SMALLTHINKER || arch == LLM_ARCH_LFM2MOE) {
         LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
         LLAMA_LOG_INFO("%s: expert_gating_func   = %s\n",     __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func));
     }
@@ -16270,10 +16313,10 @@ struct llm_build_granite_hybrid : public llm_graph_context_mamba {
     }
 
     ggml_tensor * build_layer_ffn(
-              ggml_tensor       * cur,
-              ggml_tensor       * inpSA,
-        const llama_model       & model,
-        const int                 il) {
+              ggml_tensor * cur,
+              ggml_tensor * inpSA,
+        const llama_model & model,
+        const int           il) {
 
         // For Granite architectures - scale residual
         if (hparams.f_residual_scale) {
@@ -18602,6 +18645,8 @@ struct llm_build_lfm2 : public llm_graph_context {
         ggml_tensor * inp_out_ids = build_inp_out_ids();
 
         for (int il = 0; il < n_layer; ++il) {
+            const bool is_moe_layer = il >= static_cast<int>(hparams.n_layer_dense_lead);
+
             auto * prev_cur = cur;
             cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
             cb(cur, "model.layers.{}.operator_norm", il);
@@ -18616,7 +18661,16 @@ struct llm_build_lfm2 : public llm_graph_context {
             }
 
             cur = ggml_add(ctx0, prev_cur, cur);
-            cur = ggml_add(ctx0, cur, build_feed_forward(cur, il));
+
+            auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
+            cb(ffn_norm_out, "model.layers.{}.ffn_norm", il);
+
+            ggml_tensor * ffn_out = is_moe_layer ?
+                build_moe_feed_forward(ffn_norm_out, il) :
+                build_dense_feed_forward(ffn_norm_out, il);
+            cb(ffn_norm_out, "model.layers.{}.ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_out);
         }
 
         cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1);
@@ -18631,23 +18685,32 @@ struct llm_build_lfm2 : public llm_graph_context {
         ggml_build_forward_expand(gf, cur);
     }
 
-    ggml_tensor * build_feed_forward(ggml_tensor * cur,
-                                     int           il) const {
-        cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
-        cb(cur, "model.layers.{}.ffn_norm", il);
+    ggml_tensor * build_moe_feed_forward(ggml_tensor * cur,
+                                         int           il) const {
+        return build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    static_cast<llama_expert_gating_func_type>(hparams.expert_gating_func),
+                    il);
+    }
 
+    ggml_tensor * build_dense_feed_forward(ggml_tensor * cur,
+                                           int           il) const {
         GGML_ASSERT(!model.layers[il].ffn_up_b);
         GGML_ASSERT(!model.layers[il].ffn_gate_b);
         GGML_ASSERT(!model.layers[il].ffn_down_b);
-        cur = build_ffn(cur,
+        return build_ffn(cur,
                 model.layers[il].ffn_up,   NULL, NULL,
                 model.layers[il].ffn_gate, NULL, NULL,
                 model.layers[il].ffn_down, NULL, NULL,
                 NULL,
                 LLM_FFN_SILU, LLM_FFN_PAR, il);
-        cb(cur, "model.layers.{}.feed_forward.w2", il);
-
-        return cur;
     }
 
     ggml_tensor * build_attn_block(ggml_tensor             * cur,
@@ -19817,6 +19880,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
                 llm = std::make_unique<llm_build_falcon_h1>(*this, params);
             } break;
         case LLM_ARCH_LFM2:
+        case LLM_ARCH_LFM2MOE:
             {
                 llm = std::make_unique<llm_build_lfm2>(*this, params);
             } break;
@@ -19843,6 +19907,12 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
     // add on pooling layer
     llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
 
+    // if the gguf model was converted with --sentence-transformers-dense-modules
+    // there will be two additional dense projection layers
+    // dense linear projections are applied after pooling
+    // TODO: move reranking logic here and generalize
+    llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
+
     return llm->res->get_gf();
 }
 
@@ -20039,6 +20109,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_OPENAI_MOE:
         case LLM_ARCH_HUNYUAN_DENSE:
         case LLM_ARCH_LFM2:
+        case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_SMALLTHINKER:
         case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_SEED_OSS:

src/llama-model.h

@@ -107,6 +107,7 @@ enum llm_type {
     LLM_TYPE_17B_16E, // llama4 Scout
     LLM_TYPE_17B_128E, // llama4 Maverick
     LLM_TYPE_A13B,
+    LLM_TYPE_8B_A1B, // lfm2moe
     LLM_TYPE_21B_A3B, // Ernie MoE small
     LLM_TYPE_30B_A3B,
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
@@ -437,6 +438,12 @@ struct llama_model {
 
     std::vector<llama_layer> layers;
 
+    //Dense linear projections for SentenceTransformers models like embeddinggemma
+    // For Sentence Transformers models structure see
+    // https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
+    struct ggml_tensor * dense_2_out_layers = nullptr;
+    struct ggml_tensor * dense_3_out_layers = nullptr;
+
     llama_model_params params;
 
     // gguf metadata

src/llama-sampling.cpp

@@ -2541,8 +2541,13 @@ static void llama_sampler_infill_apply(struct llama_sampler * smpl, llama_token_
     if (n_non_eog == 0) {
         cur_p->size = 1;
         cur_p->data[0].id = ctx->vocab->token_eot();
+        if (cur_p->data[0].id == LLAMA_TOKEN_NULL) {
+            cur_p->data[0].id = ctx->vocab->token_eos();
+        }
         cur_p->data[0].logit = 1.0f;
 
+        GGML_ASSERT(cur_p->data[0].id != LLAMA_TOKEN_NULL);
+
         return;
     }
 

src/llama-vocab.cpp

@@ -2171,6 +2171,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                         || t.first == "<|end|>"
                         || t.first == "<end_of_turn>"
                         || t.first == "<|endoftext|>"
+                        || t.first == "<|end_of_text|>" // granite
                         || t.first == "<EOT>"
                         || t.first == "_<EOT>"
                         || t.first == "<｜end▁of▁sentence｜>" // DeepSeek

tests/test-backend-ops.cpp

@@ -6773,7 +6773,8 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                                     if (hsk > 64 && nr3 > 1) continue; // skip broadcast for large head sizes
                                     for (int nr2 : { 1, 4, 16 }) {
                                         if (nr2 == 16 && hsk != 128) continue;
-                                        for (int kv : { 512, 1024, }) {
+                                        //for (int kv : { 1, 17, 31, 33, 61, 113, 65, 127, 129, 130, 255, 260, 371, 380, 407, 512, 1024, }) {
+                                        for (int kv : { 113, 512, 1024, }) {
                                             if (nr2 != 1 && kv != 512) continue;
                                             for (int nb : { 1, 3, 32, 35, }) {
                                                 for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {

tests/test-chat-parser.cpp

@@ -106,6 +106,34 @@ static void test_reasoning() {
     assert_equals("<think>Cogito</think>", builder.result().content);
     assert_equals("Ergo sum", builder.consume_rest());
   }
+  {
+    const std::string variant("content_only_inline_think");
+    common_chat_syntax syntax = {
+        /* .format = */ COMMON_CHAT_FORMAT_CONTENT_ONLY,
+        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+        /* .reasoning_in_content = */ false,
+        /* .thinking_forced_open = */ false,
+        /* .parse_tool_calls = */ false,
+    };
+    const std::string input = "<think>Pense</think>Bonjour";
+    auto msg = common_chat_parse(input, false, syntax);
+    assert_equals(variant, std::string("Pense"), msg.reasoning_content);
+    assert_equals(variant, std::string("Bonjour"), msg.content);
+  }
+  {
+    const std::string variant("llama_3_inline_think");
+    common_chat_syntax syntax = {
+        /* .format = */ COMMON_CHAT_FORMAT_LLAMA_3_X,
+        /* .reasoning_format = */ COMMON_REASONING_FORMAT_DEEPSEEK,
+        /* .reasoning_in_content = */ false,
+        /* .thinking_forced_open = */ false,
+        /* .parse_tool_calls = */ false,
+    };
+    const std::string input = "<think>Plan</think>Réponse";
+    auto msg = common_chat_parse(input, false, syntax);
+    assert_equals(variant, std::string("Plan"), msg.reasoning_content);
+    assert_equals(variant, std::string("Réponse"), msg.content);
+  }
   // Test DeepSeek V3.1 parsing - reasoning content followed by "</think>" and then regular content
   {
     common_chat_syntax syntax = {
@@ -496,6 +524,64 @@ static void test_json_with_dumped_args() {
     R"({"foo": "bar", "args": {"arg1": [)",
     R"({"foo":"bar","args":"{\"arg1\":["})"
   );
+
+  // Unicode tests
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\u)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\u"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\u0)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\u0"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\u00)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\u00"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\u000)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\u000"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\u0000)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\u0000"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud8)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud8"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud80)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud80"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\u)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\u"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\ud)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\ud"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\udc)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\udc"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\udc0)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\udc0"})"
+  );
+  test_with_args(
+    R"({"foo": "bar", "args": {"arg1": "\ud800\udc00)",
+    R"({"foo":"bar","args":"{\"arg1\":\"\\ud800\\udc00"})"
+  );
 }
 
 static void test_positions() {

tests/test-json-partial.cpp

@@ -58,7 +58,7 @@ static void test_json_healing() {
       for (const auto & input : inputs) {
         common_json out;
         assert_equals(true, common_json_parse(input, "$foo", out));
-        assert_equals<std::string>(expected, out.json.dump());
+        assert_equals<std::string>(expected, out.json.dump(/* indent */ -1, /* indent_char */ ' ', /* ensure_ascii */ true));
         assert_equals<std::string>(expected_marker, out.healing_marker.json_dump_marker);
       }
   };
@@ -228,6 +228,56 @@ static void test_json_healing() {
     R"({"key":"$foo"})",
     R"(:"$foo)"
   );
+  // Test unicode escape sequences
+  test(
+    {
+      R"({"a":"\u)",
+    },
+    R"({"a":"\u0000$foo"})",
+    R"(0000$foo)"
+  );
+  test(
+    {
+      R"({"a":"\u00)",
+    },
+    R"({"a":"\u0000$foo"})",
+    R"(00$foo)"
+  );
+  test(
+    {
+      R"({"a":"\ud300)",
+    },
+    R"({"a":"\ud300$foo"})",
+    R"($foo)"
+  );
+  test(
+    {
+      R"({"a":"\ud800)",
+    },
+    R"({"a":"\ud800\udc00$foo"})",
+    R"(\udc00$foo)"
+  );
+  test(
+    {
+      R"({"a":"\ud800\)",
+    },
+    R"({"a":"\ud800\udc00$foo"})",
+    R"(udc00$foo)"
+  );
+  test(
+    {
+      R"({"a":"\ud800\u)",
+    },
+    R"({"a":"\ud800\udc00$foo"})",
+    R"(dc00$foo)"
+  );
+  test(
+    {
+      R"({"a":"\ud800\udc00)",
+    },
+    R"({"a":"\ud800\udc00$foo"})",
+    R"($foo)"
+  );
 }
 
 int main() {

tools/rpc/README.md

@@ -4,7 +4,7 @@
 > This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
 > insecure. **Never run the RPC server on an open network or in a sensitive environment!**
 
-The `rpc-server` allows  running `ggml` backend on a remote host.
+The `rpc-server` allows exposing `ggml` devices on a remote host.
 The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
 This can be used for distributed LLM inference with `llama.cpp` in the following way:
 
@@ -14,28 +14,34 @@ flowchart TD
     rpcb<-->|TCP|srvb
     rpcb<-.->|TCP|srvn
     subgraph hostn[Host N]
-    srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"]
+    srvn[rpc-server]<-.->dev4["CUDA0"]
+    srvn[rpc-server]<-.->dev5["CPU"]
     end
     subgraph hostb[Host B]
-    srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"]
+    srvb[rpc-server]<-->dev3["Metal"]
     end
     subgraph hosta[Host A]
-    srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"]
+    srva[rpc-server]<-->dev["CUDA0"]
+    srva[rpc-server]<-->dev2["CUDA1"]
     end
     subgraph host[Main Host]
-    local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli]
+    local["Local devices"]<-->ggml[llama-cli]
     ggml[llama-cli]<-->rpcb[RPC backend]
     end
     style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
+    classDef devcls fill:#5B9BD5
+    class local,dev,dev2,dev3,dev4,dev5 devcls
 ```
 
-Each host can run a different backend, e.g. one with CUDA and another with Metal.
-You can also run multiple `rpc-server` instances on the same host, each with a different backend.
+By default, `rpc-server` exposes all available accelerator devices on the host.
+If there are no accelerators, it exposes a single `CPU` device.
 
 ## Usage
 
-On each host, build the corresponding backend with `cmake` and add `-DGGML_RPC=ON` to the build options.
-For example, to build the CUDA backend with RPC support:
+### Remote hosts
+
+On each remote host, build the backends for each accelerator by adding `-DGGML_RPC=ON` to the build options.
+For example, to build the `rpc-server` with support for CUDA accelerators:
 
 ```bash
 mkdir build-rpc-cuda
@@ -44,33 +50,38 @@ cmake .. -DGGML_CUDA=ON -DGGML_RPC=ON
 cmake --build . --config Release
 ```
 
-Then, start the `rpc-server` with the backend:
+When started, the `rpc-server` will detect and expose all available `CUDA` devices:
 
 ```bash
-$ bin/rpc-server -p 50052
-create_backend: using CUDA backend
-ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
-ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
+$ bin/rpc-server
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
+ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
 ggml_cuda_init: found 1 CUDA devices:
-  Device 0: NVIDIA T1200 Laptop GPU, compute capability 7.5, VMM: yes
-Starting RPC server on 0.0.0.0:50052
+  Device 0: NVIDIA GeForce RTX 5090, compute capability 12.0, VMM: yes
+Starting RPC server v3.0.0
+  endpoint       : 127.0.0.1:50052
+  local cache    : n/a
+Devices:
+  CUDA0: NVIDIA GeForce RTX 5090 (32109 MiB, 31588 MiB free)
 ```
 
-When using the CUDA backend, you can specify the device with the `CUDA_VISIBLE_DEVICES` environment variable, e.g.:
+You can control the set of exposed CUDA devices with the `CUDA_VISIBLE_DEVICES` environment variable or the `--device` command line option. The following two commands have the same effect:
 ```bash
 $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
+$ bin/rpc-server --device CUDA0 -p 50052
 ```
-This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
 
+### Main host
 
-On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options.
-Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`:
+On the main host build `llama.cpp` with the backends for the local devices and add `-DGGML_RPC=ON` to the build options.
+Finally, when running `llama-cli` or `llama-server`, use the `--rpc` option to specify the host and port of each `rpc-server`:
 
 ```bash
-$ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
+$ llama-cli -hf ggml-org/gemma-3-1b-it-GGUF -ngl 99 --rpc 192.168.88.10:50052,192.168.88.11:50052
 ```
 
-This way you can offload model layers to both local and remote devices.
+By default, llama.cpp distributes model weights and the KV cache across all available devices -- both local and remote -- in proportion to each device's available memory.
+You can override this behavior with the `--tensor-split` option and set custom proportions when splitting tensor data across devices.
 
 ### Local cache
 
@@ -83,3 +94,11 @@ $ bin/rpc-server -c
 ```
 
 By default, the cache is stored in the `$HOME/.cache/llama.cpp/rpc` directory and can be controlled via the `LLAMA_CACHE` environment variable.
+
+### Troubleshooting
+
+Use the `GGML_RPC_DEBUG` environment variable to enable debug messages from `rpc-server`:
+```bash
+$ GGML_RPC_DEBUG=1 bin/rpc-server
+```
+

tools/server/README.md

@@ -190,7 +190,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
 | `--jinja` | use jinja template for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
-| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: deepseek)<br/>(env: LLAMA_ARG_THINK) |
 | `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, exaone4, falcon3, gemma, gigachat, glmedge, gpt-oss, granite, hunyuan-dense, hunyuan-moe, kimi-k2, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, seed_oss, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
@@ -393,7 +393,7 @@ node index.js
 
 ### GET `/health`: Returns health check result
 
-This endpoint is public (no API key check).
+This endpoint is public (no API key check). `/v1/health` also works.
 
 **Response format**
 

tools/server/tests/unit/test_basic.py

@@ -66,8 +66,7 @@ def test_server_slots():
     assert len(res.body) == server.n_slots
     assert server.n_ctx is not None and server.n_slots is not None
     assert res.body[0]["n_ctx"] == server.n_ctx / server.n_slots
-    assert "params" in res.body[0]
-    assert res.body[0]["params"]["seed"] == server.seed
+    assert "params" not in res.body[0]
 
 
 def test_load_split_model():

tools/server/tests/unit/test_chat_completion.py

@@ -19,8 +19,8 @@ def create_server():
         (None, "Book", "What is the best book", 8, "(Suddenly)+|\\{ \" Sarax.", 77, 8, "length", True,  None),
         (None, "Book", "What is the best book", 8, "(Suddenly)+|\\{ \" Sarax.", 77, 8, "length", True, 'chatml'),
         (None, "Book", "What is the best book", 8, "^ blue",                    23, 8, "length", True, "This is not a chat template, it is"),
-        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False, None),
-        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True, None),
+        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 128, "length", False, None),
+        ("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 128, "length", True, None),
         (None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", False, None),
         (None, "Book", [{"type": "text", "text": "What is"}, {"type": "text", "text": "the best book"}], 8, "Whillicter", 79, 8, "length", True, None),
     ]
@@ -54,7 +54,7 @@ def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_conte
     "system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason",
     [
         ("Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"),
-        ("You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"),
+        ("You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 128, "length"),
     ]
 )
 def test_chat_completion_stream(system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason):
@@ -408,6 +408,28 @@ def test_context_size_exceeded():
     assert res.body["error"]["n_ctx"] == server.n_ctx // server.n_slots
 
 
+def test_context_size_exceeded_stream():
+    global server
+    server.start()
+    try:
+        for _ in server.make_stream_request("POST", "/chat/completions", data={
+            "messages": [
+                {"role": "system", "content": "Book"},
+                {"role": "user", "content": "What is the best book"},
+            ] * 100, # make the prompt too long
+            "stream": True}):
+                pass
+        assert False, "Should have failed"
+    except ServerError as e:
+        assert e.code == 400
+        assert "error" in e.body
+        assert e.body["error"]["type"] == "exceed_context_size_error"
+        assert e.body["error"]["n_prompt_tokens"] > 0
+        assert server.n_ctx is not None
+        assert server.n_slots is not None
+        assert e.body["error"]["n_ctx"] == server.n_ctx // server.n_slots
+
+
 @pytest.mark.parametrize(
     "n_batch,batch_count,reuse_cache",
     [