vulkan: move common FA code to flash_attn_base.comp (#13556)

* vulkan: move common FA code to flash_attn_base.comp

* vulkan: move common FA index/stride setup code to flash_attn_base.comp

* build fix

.github/actions/windows-setup-curl/action.yml

@@ -5,6 +5,10 @@ inputs:
     description: 'CURL version'
     required: false
     default: '8.6.0_6'
+  architecture:
+    description: 'Architecture of the libcurl to download'
+    required: false
+    default: 'win64'
 outputs:
   curl_path:
     description: "Path to the downloaded libcurl"
@@ -18,8 +22,9 @@ runs:
       shell: powershell
       env:
         CURL_VERSION: ${{ inputs.curl_version }}
+        ARCHITECTURE: ${{ inputs.architecture }}
       run: |
-        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
+        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-${env:ARCHITECTURE}-mingw.zip"
         mkdir $env:RUNNER_TEMP/libcurl
         tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
         echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

.github/workflows/build-linux-cross.yml

@@ -140,3 +140,94 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  ubuntu-24-ppc64el-cpu-cross:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu \
+                  libcurl4-openssl-dev:ppc64el
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  ubuntu-24-ppc64el-vulkan-cross:
+    runs-on: ubuntu-24.04
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup PowerPC64le
+        run: |
+          sudo dpkg --add-architecture ppc64el
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | sudo tee /etc/apt/sources.list.d/ppc64el-ports.list
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-updates main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-security main universe
+          deb [arch=ppc64el] http://ports.ubuntu.com/ubuntu-ports/ noble-backports main universe
+          EOF
+
+          sudo apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-powerpc64le-linux-gnu \
+                  g++-14-powerpc64le-linux-gnu \
+                  libvulkan-dev:ppc64el \
+                  libcurl4-openssl-dev:ppc64el
+
+      - name: Build
+        run: |
+          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=ppc64 \
+                         -DCMAKE_C_COMPILER=powerpc64le-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=powerpc64le-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/powerpc64le-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/docker.yml

@@ -42,8 +42,7 @@ jobs:
           - { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "cuda", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           - { tag: "musa", dockerfile: ".devops/musa.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
-          # Note: the intel images are failing due to an out of disk space error
-          # - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
+          - { tag: "intel", dockerfile: ".devops/intel.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true }
           - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false }
           # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
           #- {tag: "rocm", dockerfile: ".devops/rocm.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: true }

.github/workflows/release.yml

@@ -238,14 +238,19 @@ jobs:
       matrix:
         include:
           - build: 'cpu-x64'
+            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
           #- build: 'openblas-x64'
+          #  arch: 'x64'
           #  defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
+            arch: 'x64'
             defines: '-DGGML_NATIVE=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'cpu-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF'
           - build: 'opencl-adreno-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
 
     steps:
@@ -312,6 +317,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -339,7 +346,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-x64.dll .\build\bin\Release\libcurl-x64.dll
+          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts

README.md

@@ -572,4 +572,11 @@ automatically. For example:
 $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 ```
 
-## References
+## Dependencies
+
+- [yhirose/cpp-httplib](https://github.com/yhirose/cpp-httplib) - Single-header HTTP server, used by `llama-server` - MIT license
+- [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
+- [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
+- [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
+- [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
+- [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)

build-xcframework.sh

@@ -117,6 +117,7 @@ setup_framework_structure() {
     # Copy all required headers (common for all platforms)
     cp include/llama.h             ${header_path}
     cp ggml/include/ggml.h         ${header_path}
+    cp ggml/include/ggml-opt.h     ${header_path}
     cp ggml/include/ggml-alloc.h   ${header_path}
     cp ggml/include/ggml-backend.h ${header_path}
     cp ggml/include/ggml-metal.h   ${header_path}

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET} STATIC
     minja/minja.hpp
     ngram-cache.cpp
     ngram-cache.h
+    regex-partial.cpp
+    regex-partial.h
     sampling.cpp
     sampling.h
     speculative.cpp
@@ -119,8 +121,8 @@ if (LLAMA_LLGUIDANCE)
 
     ExternalProject_Add(llguidance_ext
         GIT_REPOSITORY https://github.com/guidance-ai/llguidance
-        # v0.7.19 (+ fancy-regex build fix):
-        GIT_TAG b59f98f85269892a7de3d3641ad155366f13daa6
+        # v0.7.20 (+ fix to build on GCC 15):
+        GIT_TAG b5b8b64dba11c4e4ee6b1d1450d3a3ae279891e8
         PREFIX ${CMAKE_BINARY_DIR}/llguidance
         SOURCE_DIR ${LLGUIDANCE_SRC}
         BUILD_IN_SOURCE TRUE

common/arg.cpp

@@ -2437,6 +2437,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ));
+    add_opt(common_arg(
+        {"--no-op-offload"},
+        string_format("disable offloading host tensor operations to device (default: %s)", params.no_op_offload ? "true" : "false"),
+        [](common_params & params) {
+            params.no_op_offload = true;
+        }
+    ));
     add_opt(common_arg(
         {"--lora"}, "FNAME",
         "path to LoRA adapter (can be repeated to use multiple adapters)",

common/chat.cpp

@@ -6,6 +6,15 @@
 
 #include <optional>
 
+static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
+    auto time = std::chrono::system_clock::to_time_t(now);
+    auto local_time = *std::localtime(&time);
+    std::ostringstream ss;
+    ss << std::put_time(&local_time, format.c_str());
+    auto res = ss.str();
+    return res;
+}
+
 typedef minja::chat_template common_chat_template;
 
 struct common_chat_templates {
@@ -24,6 +33,7 @@ struct templates_params {
     std::string grammar;
     bool add_generation_prompt = true;
     bool extract_reasoning     = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -939,78 +949,83 @@ static void expect_tool_parameters(const std::string & name, const json & parame
     }
 }
 
-static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
+static common_chat_params common_chat_params_init_llama_3_x(const common_chat_template & tmpl, const struct templates_params & inputs, bool allow_python_tag_builtin_tools) {
     auto builtin_tools = json::array();
     common_chat_params data;
-    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-        std::vector<std::string> tool_rules;
+    if (!inputs.tools.is_null()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
 
-        auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
-            if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
-                expect_tool_parameters(name, parameters, {"query"});
-            } else if (name == "python" || name == "code_interpreter") {
-                // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
-                expect_tool_parameters(name, parameters, {"code"});
-            } else {
-                return false;
+            auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
+                if (name == "wolfram_alpha" || name == "web_search" || name == "brave_search") {
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
+                    expect_tool_parameters(name, parameters, {"query"});
+                } else if (name == "python" || name == "code_interpreter") {
+                    // https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
+                    expect_tool_parameters(name, parameters, {"code"});
+                } else {
+                    return false;
+                }
+
+                std::vector<std::string> kvs;
+                for (const auto & [key, value] : parameters.at("properties").items()) {
+                    kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
+                }
+
+                tool_rules.push_back(
+                    builder.add_rule(
+                        name + "-call",
+                        "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
+                builtin_tools.push_back(name);
+
+                return true;
+            };
+
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
+                if (allow_python_tag_builtin_tools) {
+                    handle_builtin_tool(name, parameters);
+                }
+                tool_rules.push_back(
+                    builder.add_rule(
+                        name + "-call",
+                        "\"{\" space "
+                        "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
+                        "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
+                        "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
+                        "\"}\" space"));
+            });
+            // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
+                "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
+            });
+            if (!builtin_tools.empty()) {
+                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+                data.preserved_tokens.push_back("<|python_tag|>");
             }
-
-            std::vector<std::string> kvs;
-            for (const auto & [key, value] : parameters.at("properties").items()) {
-                kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value)); // NOLINT
-            }
-
-            tool_rules.push_back(
-                builder.add_rule(
-                    name + "-call",
-                    "\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
-            builtin_tools.push_back(name);
-
-            return true;
-        };
-
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            std::string name = function.at("name");
-            auto parameters = function.at("parameters");
-            builder.resolve_refs(parameters);
-
-            // https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
-            if (allow_python_tag_builtin_tools) {
-                handle_builtin_tool(name, parameters);
-            }
-            tool_rules.push_back(
-                builder.add_rule(
-                    name + "-call",
-                    "\"{\" space "
-                    "( \"\\\"type\\\"\"       space \":\" space \"\\\"function\\\"\"     space \",\" space )? "
-                    "  \"\\\"name\\\"\"       space \":\" space \"\\\"" + name + "\\\"\" space \",\" space "
-                    "  \"\\\"parameters\\\"\" space \":\" space " + builder.add_schema(name + "-args", parameters) + " "
-                    "\"}\" space"));
+            // Allow a few empty lines on top of the usual constrained json schema space rule.
+            builder.add_rule("root", string_join(tool_rules, " | "));
+            data.additional_stops.push_back("<|eom_id|>");
         });
-        // Small models may hallucinate function names so we match anything (*at the start*) that looks like the JSON of a function call, regardless of the name.
-        data.grammar_triggers.push_back({
-            COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_START,
-            "\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"", // + name + "\"[\\s\\S]*",
-        });
-        if (!builtin_tools.empty()) {
-            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-            data.preserved_tokens.push_back("<|python_tag|>");
-        }
-        // Allow a few empty lines on top of the usual constrained json schema space rule.
-        builder.add_rule("root", string_join(tool_rules, " | "));
-    });
-    data.additional_stops.push_back("<|eom_id|>");
+        data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
+            ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
+            : COMMON_CHAT_FORMAT_LLAMA_3_X;
+    } else {
+        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    }
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+        {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
     });
-    data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
-        ? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
-        : COMMON_CHAT_FORMAT_LLAMA_3_X;
     return data;
 }
 static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {
@@ -1150,7 +1165,7 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
     data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
-        {"datetime", "Jan 29 2025 13:00:00 GMT"},
+        {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
@@ -1285,55 +1300,59 @@ static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & in
 static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     // https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt
     common_chat_params data;
-    json tools = inputs.tools.is_null() ? inputs.tools : json::array();
-    std::string python_code_argument_name;
-    auto has_raw_python = false;
 
-    data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-    data.grammar = build_grammar([&](const common_grammar_builder & builder) {
-        std::vector<std::string> tool_rules;
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const auto & function = tool.at("function");
-            const auto & parameters = function.at("parameters");
-            std::string name = function.at("name");
-            if (name == "python" || name == "ipython") {
-                if (!parameters.contains("type")) {
-                    throw std::runtime_error("Missing type in python tool");
-                }
-                has_raw_python = true;
-                const auto & type = parameters.at("type");
-                if (type == "object") {
-                    auto properties = parameters.at("properties");
-                    for (auto it = properties.begin(); it != properties.end(); ++it) {
-                        if (it.value().at("type") == "string") {
-                            if (!python_code_argument_name.empty()) {
-                                throw std::runtime_error("Multiple string arguments found in python tool");
+    if (!inputs.tools.is_null()) {
+        std::string python_code_argument_name;
+        auto has_raw_python = false;
+
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                const auto & parameters = function.at("parameters");
+                std::string name = function.at("name");
+                if (name == "python" || name == "ipython") {
+                    if (!parameters.contains("type")) {
+                        throw std::runtime_error("Missing type in python tool");
+                    }
+                    has_raw_python = true;
+                    const auto & type = parameters.at("type");
+                    if (type == "object") {
+                        auto properties = parameters.at("properties");
+                        for (auto it = properties.begin(); it != properties.end(); ++it) {
+                            if (it.value().at("type") == "string") {
+                                if (!python_code_argument_name.empty()) {
+                                    throw std::runtime_error("Multiple string arguments found in python tool");
+                                }
+                                python_code_argument_name = it.key();
                             }
-                            python_code_argument_name = it.key();
                         }
+                        if (python_code_argument_name.empty()) {
+                            throw std::runtime_error("No string argument found in python tool");
+                        }
+                    } else if (type != "string") {
+                        throw std::runtime_error("Invalid type in python tool: " + type.dump());
                     }
-                    if (python_code_argument_name.empty()) {
-                        throw std::runtime_error("No string argument found in python tool");
-                    }
-                } else if (type != "string") {
-                    throw std::runtime_error("Invalid type in python tool: " + type.dump());
                 }
+                tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
+            });
+            if (has_raw_python) {
+                tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
+                data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
+                data.preserved_tokens.push_back("<|python_tag|>");
             }
-            tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
+            auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
+            builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
+            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
         });
-        if (has_raw_python) {
-            tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
-            data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<|python_tag|>"});
-            data.preserved_tokens.push_back("<|python_tag|>");
-        }
-        auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
-        builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
-        data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<function="});
-    });
+        data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
+    } else {
+        data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
+    }
 
     data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
     // TODO: if (has_raw_python)
-    data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
     return data;
 }
 static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) {
@@ -1593,6 +1612,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.extract_reasoning = inputs.extract_reasoning;
     params.tool_choice = inputs.tool_choice;
     params.grammar = inputs.grammar;
+    params.now = inputs.now;
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }
@@ -1644,21 +1664,21 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_firefunction_v2(tmpl, params);
     }
 
-    // Plain handler (no tools)
-    if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
-        return common_chat_params_init_without_tools(tmpl, params);
-    }
-
     // Functionary v3.1 (w/ tools)
     if (src.find("<|start_header_id|>") != std::string::npos
         && src.find("<function=") != std::string::npos) {
         return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, params);
     }
 
-    // Llama 3.1, 3.2, 3.3 (w/ tools)
+    // Llama 3.1, 3.2, 3.3 (also requires date_string so using it even w/o tools)
     if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
         auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
-        return common_chat_params_init_llama_3_1_tool_calls(tmpl, params, allow_python_tag_builtin_tools);
+        return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
+    }
+
+    // Plain handler (no tools)
+    if (params.tools.is_null() || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+        return common_chat_params_init_without_tools(tmpl, params);
     }
 
     // Mistral Nemo (w/ tools)

common/chat.h

@@ -3,6 +3,7 @@
 #pragma once
 
 #include "common.h"
+#include <chrono>
 #include <string>
 #include <vector>
 
@@ -71,6 +72,7 @@ struct common_chat_templates_inputs {
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
     bool extract_reasoning     = true;
+    std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
 struct common_chat_params {

common/common.cpp

@@ -443,6 +443,25 @@ void string_replace_all(std::string & s, const std::string & search, const std::
     s = std::move(builder);
 }
 
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix) {
+    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
+}
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop) {
+    if (!str.empty() && !stop.empty()) {
+        const char text_last_char = str.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
+            if (stop[char_index] == text_last_char) {
+                const auto current_partial = stop.substr(0, char_index + 1);
+                if (string_ends_with(str, current_partial)) {
+                    return str.size() - char_index - 1;
+                }
+            }
+        }
+    }
+
+    return std::string::npos;
+}
+
 std::string regex_escape(const std::string & s) {
     static const std::regex special_chars("[.^$|()*+?\\[\\]{}\\\\]");
     return std::regex_replace(s, special_chars, "\\$0");
@@ -1113,6 +1132,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.offload_kqv       = !params.no_kv_offload;
     cparams.flash_attn        = params.flash_attn;
     cparams.no_perf           = params.no_perf;
+    cparams.op_offload        = !params.no_op_offload;
 
     if (params.reranking) {
         cparams.embeddings    = true;
@@ -1564,3 +1584,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
 
     return result;
 }
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
+    const int64_t ne_datapoint = llama_n_ctx(ctx);
+    const int64_t ndata        = (tokens.size() - ne_datapoint - 1) / stride;
+    ggml_opt_dataset_t result = ggml_opt_dataset_init(
+        GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
+
+    llama_token * data   = (llama_token *) ggml_opt_dataset_data(result)->data;
+    llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
+
+    for (int64_t idata = 0; idata < ndata; ++idata) {
+        memcpy(data   + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
+        memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
+    }
+
+    return result;
+}

common/common.h

@@ -6,6 +6,7 @@
 
 #include <set>
 #include <string>
+#include <string_view>
 #include <vector>
 #include <sstream>
 
@@ -332,6 +333,7 @@ struct common_params {
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
     bool check_tensors     = false; // validate tensor data
+    bool no_op_offload     = false; // globally disable offload host tensor operations to device
 
     bool single_turn       = false; // single turn chat conversation
 
@@ -502,10 +504,9 @@ static bool string_starts_with(const std::string & str,
     return str.rfind(prefix, 0) == 0;
 }
 
-static bool string_ends_with(const std::string & str,
-                               const std::string & suffix) {  // While we wait for C++20's std::string::ends_with...
-    return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
-}
+// While we wait for C++20's std::string::ends_with...
+bool string_ends_with(const std::string_view & str, const std::string_view & suffix);
+size_t string_find_partial_stop(const std::string_view & str, const std::string_view & stop);
 
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
@@ -665,3 +666,9 @@ const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 
 }
+
+//
+// training utils
+//
+
+ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);

common/llguidance.cpp

@@ -189,6 +189,7 @@ static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab)
         /* .tokenize_fn                        = */ llama_sampler_llg_tokenize_fn,
         /* .use_approximate_greedy_tokenize_fn = */ false,
         /* .tokenize_user_data                 = */ vocab,
+        /* .slices                             = */ nullptr,
     };
 
     char           error_buffer[1024];

common/minja/chat-template.hpp

@@ -13,10 +13,12 @@
 #include <chrono>
 #include <cstddef>
 #include <cstdio>
+#include <ctime>
 #include <exception>
 #include <iomanip>
 #include <memory>
 #include <sstream>
+#include <stdexcept>
 #include <string>
 #include <vector>
 
@@ -393,8 +395,8 @@ class chat_template {
 
             for (const auto & message_ : adjusted_messages) {
                 auto message = message_;
-                if (!message.contains("role") || !message.contains("content")) {
-                    throw std::runtime_error("message must have 'role' and 'content' fields: " + message.dump());
+                if (!message.contains("role") || (!message.contains("content") && !message.contains("tool_calls"))) {
+                    throw std::runtime_error("message must have 'role' and one of 'content' or 'tool_calls' fields: " + message.dump());
                 }
                 std::string role = message.at("role");
 
@@ -415,7 +417,6 @@ class chat_template {
                         }
                     }
                     if (polyfill_tool_calls) {
-                        auto content = message.at("content");
                         auto tool_calls = json::array();
                         for (const auto & tool_call : message.at("tool_calls")) {
                             if (tool_call.at("type") != "function") {
@@ -434,8 +435,11 @@ class chat_template {
                         auto obj = json {
                             {"tool_calls", tool_calls},
                         };
-                        if (!content.is_null() && !content.empty()) {
-                            obj["content"] = content;
+                        if (message.contains("content")) {
+                            auto content = message.at("content");
+                            if (!content.is_null() && !content.empty()) {
+                                obj["content"] = content;
+                            }
                         }
                         message["content"] = obj.dump(2);
                         message.erase("tool_calls");

common/minja/minja.hpp

@@ -11,6 +11,7 @@
 #include <algorithm>
 #include <cctype>
 #include <cstddef>
+#include <cstdint>
 #include <cmath>
 #include <exception>
 #include <functional>
@@ -233,7 +234,7 @@ public:
       }
     } else if (is_object()) {
       if (!index.is_hashable())
-        throw std::runtime_error("Unashable type: " + index.dump());
+        throw std::runtime_error("Unhashable type: " + index.dump());
       auto it = object_->find(index.primitive_);
       if (it == object_->end())
         throw std::runtime_error("Key not found: " + index.dump());
@@ -252,7 +253,7 @@ public:
       auto index = key.get<int>();
       return array_->at(index < 0 ? array_->size() + index : index);
     } else if (object_) {
-      if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+      if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
       auto it = object_->find(key.primitive_);
       if (it == object_->end()) return Value();
       return it->second;
@@ -261,7 +262,7 @@ public:
   }
   void set(const Value& key, const Value& value) {
     if (!object_) throw std::runtime_error("Value is not an object: " + dump());
-    if (!key.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!key.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     (*object_)[key.primitive_] = value;
   }
   Value call(const std::shared_ptr<Context> & context, ArgumentsValue & args) const {
@@ -398,7 +399,7 @@ public:
       }
       return false;
     } else if (object_) {
-      if (!value.is_hashable()) throw std::runtime_error("Unashable type: " + value.dump());
+      if (!value.is_hashable()) throw std::runtime_error("Unhashable type: " + value.dump());
       return object_->find(value.primitive_) != object_->end();
     } else {
       throw std::runtime_error("contains can only be called on arrays and objects: " + dump());
@@ -416,7 +417,7 @@ public:
     return const_cast<Value*>(this)->at(index);
   }
   Value& at(const Value & index) {
-    if (!index.is_hashable()) throw std::runtime_error("Unashable type: " + dump());
+    if (!index.is_hashable()) throw std::runtime_error("Unhashable type: " + dump());
     if (is_array()) return array_->at(index.get<int>());
     if (is_object()) return object_->at(index.primitive_);
     throw std::runtime_error("Value is not an array or object: " + dump());
@@ -676,8 +677,8 @@ public:
 class VariableExpr : public Expression {
     std::string name;
 public:
-    VariableExpr(const Location & location, const std::string& n)
-      : Expression(location), name(n) {}
+    VariableExpr(const Location & loc, const std::string& n)
+      : Expression(loc), name(n) {}
     std::string get_name() const { return name; }
     Value do_evaluate(const std::shared_ptr<Context> & context) const override {
         if (!context->contains(name)) {
@@ -1200,9 +1201,9 @@ public:
 
 class SliceExpr : public Expression {
 public:
-    std::shared_ptr<Expression> start, end;
-    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e)
-      : Expression(loc), start(std::move(s)), end(std::move(e)) {}
+    std::shared_ptr<Expression> start, end, step;
+    SliceExpr(const Location & loc, std::shared_ptr<Expression> && s, std::shared_ptr<Expression> && e, std::shared_ptr<Expression> && st = nullptr)
+      : Expression(loc), start(std::move(s)), end(std::move(e)), step(std::move(st)) {}
     Value do_evaluate(const std::shared_ptr<Context> &) const override {
         throw std::runtime_error("SliceExpr not implemented");
     }
@@ -1219,18 +1220,35 @@ public:
         if (!index) throw std::runtime_error("SubscriptExpr.index is null");
         auto target_value = base->evaluate(context);
         if (auto slice = dynamic_cast<SliceExpr*>(index.get())) {
-          auto start = slice->start ? slice->start->evaluate(context).get<int64_t>() : 0;
-          auto end = slice->end ? slice->end->evaluate(context).get<int64_t>() : (int64_t) target_value.size();
+          auto len = target_value.size();
+          auto wrap = [len](int64_t i) -> int64_t {
+            if (i < 0) {
+              return i + len;
+            }
+            return i;
+          };
+          int64_t step = slice->step ? slice->step->evaluate(context).get<int64_t>() : 1;
+          if (!step) {
+            throw std::runtime_error("slice step cannot be zero");
+          }
+          int64_t start = slice->start ? wrap(slice->start->evaluate(context).get<int64_t>()) : (step < 0 ? len - 1 : 0);
+          int64_t end = slice->end ? wrap(slice->end->evaluate(context).get<int64_t>()) : (step < 0 ? -1 : len);
           if (target_value.is_string()) {
             std::string s = target_value.get<std::string>();
-            if (start < 0) start = s.size() + start;
-            if (end < 0) end = s.size() + end;
-            return s.substr(start, end - start);
+
+            std::string result;
+            if (start < end && step == 1) {
+              result = s.substr(start, end - start);
+            } else {
+              for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
+                result += s[i];
+              }
+            }
+            return result;
+
           } else if (target_value.is_array()) {
-            if (start < 0) start = target_value.size() + start;
-            if (end < 0) end = target_value.size() + end;
             auto result = Value::array();
-            for (auto i = start; i < end; ++i) {
+            for (int64_t i = start; step > 0 ? i < end : i > end; i += step) {
               result.push_back(target_value.at(i));
             }
             return result;
@@ -1305,6 +1323,8 @@ public:
               if (name == "iterable") return l.is_iterable();
               if (name == "sequence") return l.is_array();
               if (name == "defined") return !l.is_null();
+              if (name == "true") return l.to_bool();
+              if (name == "false") return !l.to_bool();
               throw std::runtime_error("Unknown type for 'is' operator: " + name);
             };
             auto value = eval();
@@ -1520,6 +1540,10 @@ public:
             vargs.expectArgs("endswith method", {1, 1}, {0, 0});
             auto suffix = vargs.args[0].get<std::string>();
             return suffix.length() <= str.length() && std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
+          } else if (method->get_name() == "startswith") {
+            vargs.expectArgs("startswith method", {1, 1}, {0, 0});
+            auto prefix = vargs.args[0].get<std::string>();
+            return prefix.length() <= str.length() && std::equal(prefix.begin(), prefix.end(), str.begin());
           } else if (method->get_name() == "title") {
             vargs.expectArgs("title method", {0, 0}, {0, 0});
             auto res = str;
@@ -2082,28 +2106,37 @@ private:
 
       while (it != end && consumeSpaces() && peekSymbols({ "[", "." })) {
         if (!consumeToken("[").empty()) {
-            std::shared_ptr<Expression> index;
+          std::shared_ptr<Expression> index;
+          auto slice_loc = get_location();
+          std::shared_ptr<Expression> start, end, step;
+          bool has_first_colon = false, has_second_colon = false;
+
+          if (!peekSymbols({ ":" })) {
+            start = parseExpression();
+          }
+
+          if (!consumeToken(":").empty()) {
+            has_first_colon = true;
+            if (!peekSymbols({ ":", "]" })) {
+              end = parseExpression();
+            }
             if (!consumeToken(":").empty()) {
-              auto slice_end = parseExpression();
-              index = std::make_shared<SliceExpr>(slice_end->location, nullptr, std::move(slice_end));
-            } else {
-              auto slice_start = parseExpression();
-              if (!consumeToken(":").empty()) {
-                consumeSpaces();
-                if (peekSymbols({ "]" })) {
-                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), nullptr);
-                } else {
-                  auto slice_end = parseExpression();
-                  index = std::make_shared<SliceExpr>(slice_start->location, std::move(slice_start), std::move(slice_end));
-                }
-              } else {
-                index = std::move(slice_start);
+              has_second_colon = true;
+              if (!peekSymbols({ "]" })) {
+                step = parseExpression();
               }
             }
-            if (!index) throw std::runtime_error("Empty index in subscript");
-            if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");
+          }
 
-            value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
+          if ((has_first_colon || has_second_colon) && (start || end || step)) {
+            index = std::make_shared<SliceExpr>(slice_loc, std::move(start), std::move(end), std::move(step));
+          } else {
+            index = std::move(start);
+          }
+          if (!index) throw std::runtime_error("Empty index in subscript");
+          if (consumeToken("]").empty()) throw std::runtime_error("Expected closing bracket in subscript");
+
+          value = std::make_shared<SubscriptExpr>(value->location, std::move(value), std::move(index));
         } else if (!consumeToken(".").empty()) {
             auto identifier = parseIdentifier();
             if (!identifier) throw std::runtime_error("Expected identifier in subscript");

common/regex-partial.cpp

@@ -0,0 +1,204 @@
+#include "regex-partial.h"
+#include "common.h"
+#include <functional>
+#include <optional>
+
+common_regex::common_regex(const std::string & pattern) :
+    pattern(pattern),
+    rx(pattern),
+    rx_reversed_partial(regex_to_reversed_partial_regex(pattern)) {}
+
+common_regex_match common_regex::search(const std::string & input, size_t pos, bool as_match) const {
+    std::smatch match;
+    if (pos > input.size()) {
+        throw std::runtime_error("Position out of bounds");
+    }
+    auto start = input.begin() + pos;
+    auto found = as_match
+        ? std::regex_match(start, input.end(), match, rx)
+        : std::regex_search(start, input.end(), match, rx);
+    if (found) {
+        common_regex_match res;
+        res.type = COMMON_REGEX_MATCH_TYPE_FULL;
+        for (size_t i = 0; i < match.size(); ++i) {
+            auto begin = pos + match.position(i);
+            res.groups.emplace_back(begin, begin + match.length(i));
+        }
+        return res;
+    }
+    std::match_results<std::string::const_reverse_iterator> srmatch;
+    if (std::regex_match(input.rbegin(), input.rend() - pos, srmatch, rx_reversed_partial)) {
+        auto group = srmatch[1].str();
+        if (group.length() != 0) {
+            auto it = srmatch[1].second.base();
+            // auto position = static_cast<size_t>(std::distance(input.begin(), it));
+            if ((!as_match) || it == input.begin()) {
+                common_regex_match res;
+                res.type = COMMON_REGEX_MATCH_TYPE_PARTIAL;
+                const size_t begin = std::distance(input.begin(), it);
+                const size_t end = input.size();
+                if (begin == std::string::npos || end == std::string::npos || begin > end) {
+                    throw std::runtime_error("Invalid range");
+                }
+                res.groups.push_back({begin, end});
+                return res;
+            }
+        }
+    }
+    return {};
+}
+
+/*
+  Transforms a regex pattern to a partial match pattern that operates on a reversed input string to find partial final matches of the original pattern.
+
+  Ideally we'd like to use boost::match_partial (https://beta.boost.org/doc/libs/1_59_0/libs/regex/doc/html/boost_regex/partial_matches.html)
+  to see if a string ends with a partial regex match, but but it's not in std::regex yet.
+  Instead, we'll the regex into a partial match regex operating as a full match on the reverse iterators of the input.
+
+  - /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:(?:d)?c)?b)?a).*
+  - /a|b/ -> (a|b).*
+  - /a*?/ -> error, could match ""
+  - /a*b/ -> ((?:b)?a*+).* (final repetitions become eager)
+  - /.*?ab/ -> ((?:b)?a).* (merge .*)
+  - /a.*?b/ -> ((?:b)?.*?a).* (keep reluctant matches)
+  - /a(bc)d/ -> ((?:(?:d)?(?:(?:c)?b))?a).*
+  - /a(bc|de)/ -> ((?:(?:(?:e)?d)?|(?:(?:c)?b)?)?a).*
+  - /ab{2,4}c/ -> abbb?b?c -> ((?:(?:(?:(?:(?:c)?b)?b)?b?)?b?)?a).*
+
+  The regex will match a reversed string fully, and the end of the first (And only) capturing group will indicate the reversed start of the original partial pattern
+  (i.e. just where the final .* starts in the inverted pattern; all other groups are turned into non-capturing groups, and reluctant quantifiers are ignored)
+*/
+std::string regex_to_reversed_partial_regex(const std::string & pattern) {
+    auto it = pattern.begin();
+    const auto end = pattern.end();
+
+    std::function<std::string()> process = [&]() {
+        std::vector<std::vector<std::string>> alternatives(1);
+        std::vector<std::string> * sequence = &alternatives.back();
+
+        while (it != end) {
+            if (*it == '[') {
+                auto start = it;
+                ++it;
+                while (it != end) {
+                    if ((*it == '\\') && (++it != end)) {
+                        ++it;
+                    } else if ((it != end) && (*it == ']')) {
+                        break;
+                    } else {
+                        ++it;
+                    }
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '[' in pattern");
+                }
+                ++it;
+                sequence->push_back(std::string(start, it));
+            } else if (*it == '*' || *it == '?' || *it == '+') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Quantifier without preceding element");
+                }
+                sequence->back() += *it;
+                auto is_star = *it == '*';
+                ++it;
+                if (is_star) {
+                    if (*it == '?') {
+                        ++it;
+                    }
+                }
+            } else if (*it == '{') {
+                if (sequence->empty()) {
+                    throw std::runtime_error("Repetition without preceding element");
+                }
+                ++it;
+                auto start = it;
+                while (it != end && *it != '}') {
+                    ++it;
+                }
+                if (it == end) {
+                    throw std::runtime_error("Unmatched '{' in pattern");
+                }
+                auto parts = string_split(std::string(start, it), ",");
+                ++it;
+                if (parts.size() > 2) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+
+                auto parseOptInt = [&](const std::string & s, const std::optional<int> & def = std::nullopt) -> std::optional<int> {
+                    if (s.empty()) {
+                        return def;
+                    }
+                    return std::stoi(s);
+                };
+                auto min = parseOptInt(parts[0], 0);
+                auto max = parts.size() == 1 ? min : parseOptInt(parts[1]);
+                if (min && max && *max < *min) {
+                    throw std::runtime_error("Invalid repetition range in pattern");
+                }
+                // Brutal but... let's repeat at least min times, then ? for the delta between min & max (or * for unbounded)
+                auto part = sequence->back();
+                sequence->pop_back();
+                for (int i = 0; i < *min; i++) {
+                    sequence->push_back(part);
+                }
+                if (max) {
+                    for (int i = *min; i < *max; i++) {
+                        sequence->push_back(part + "?");
+                    }
+                } else {
+                    sequence->push_back(part + "*");
+                }
+            } else if (*it == '(') {
+                ++it;
+                if (it != end && *it == '?' && (it + 1 != end) && *(it + 1) == ':') {
+                    it += 2;
+                }
+                auto sub = process();
+                if (*it != ')') {
+                    throw std::runtime_error("Unmatched '(' in pattern");
+                }
+                ++it;
+                auto & part = sequence->emplace_back("(?:");
+                part += sub;
+                part += ")";
+            } else if (*it == ')') {
+                break;
+            } else if (*it == '|') {
+                ++it;
+                alternatives.emplace_back();
+                sequence = &alternatives.back();
+            } else if (*it == '\\' && (++it != end)) {
+                auto str = std::string("\\") + *it;
+                sequence->push_back(str);
+                ++it;
+            } else if (it != end) {
+                sequence->push_back(std::string(1, *it));
+                ++it;
+            }
+        }
+
+        // /abcd/ -> (dcba|cba|ba|a).* -> ((?:(?:(?:d)?c)?b)?a).*
+        // if n(=4) parts, opening n-1(=3) non-capturing groups after the 1 capturing group
+        // We'll do the outermost capturing group and final .* in the enclosing function.
+        std::vector<std::string> res_alts;
+        for (const auto & parts : alternatives) {
+            auto & res = res_alts.emplace_back();
+            for (size_t i = 0; i < parts.size() - 1; i++) {
+                res += "(?:";
+            }
+            for (auto it = parts.rbegin(); it != parts.rend(); ++it) {
+                res += *it;
+                if (it != parts.rend() - 1) {
+                    res += ")?";
+                }
+            }
+        }
+        return string_join(res_alts, "|");
+    };
+    auto res = process();
+    if (it != end) {
+        throw std::runtime_error("Unmatched '(' in pattern");
+    }
+
+    return "(" + res + ")[\\s\\S]*";
+}

common/regex-partial.h

@@ -0,0 +1,56 @@
+#pragma once
+
+#include <regex>
+#include <string>
+
+enum common_regex_match_type {
+    COMMON_REGEX_MATCH_TYPE_NONE,
+    COMMON_REGEX_MATCH_TYPE_PARTIAL,
+    COMMON_REGEX_MATCH_TYPE_FULL,
+};
+
+struct common_string_range {
+    size_t begin;
+    size_t end;
+    common_string_range(size_t begin, size_t end) : begin(begin), end(end) {
+        if (begin > end) {
+            throw std::runtime_error("Invalid range");
+        }
+    }
+    // prevent default ctor
+    common_string_range() = delete;
+    bool empty() const {
+        return begin == end;
+    }
+    bool operator==(const common_string_range & other) const {
+        return begin == other.begin && end == other.end;
+    }
+};
+
+struct common_regex_match {
+    common_regex_match_type type = COMMON_REGEX_MATCH_TYPE_NONE;
+    std::vector<common_string_range> groups;
+
+    bool operator==(const common_regex_match & other) const {
+        return type == other.type && groups == other.groups;
+    }
+    bool operator!=(const common_regex_match & other) const {
+        return !(*this == other);
+    }
+};
+
+class common_regex {
+    std::string pattern;
+    std::regex rx;
+    std::regex rx_reversed_partial;
+
+  public:
+    explicit common_regex(const std::string & pattern);
+
+    common_regex_match search(const std::string & input, size_t pos, bool as_match = false) const;
+
+    const std::string & str() const { return pattern; }
+};
+
+// For testing only (pretty print of failures).
+std::string regex_to_reversed_partial_regex(const std::string & pattern);

convert_hf_to_gguf.py

@@ -798,6 +798,9 @@ class TextModel(ModelBase):
         if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
             # ref: https://huggingface.co/mistral-community/pixtral-12b
             res = "pixtral"
+        if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
+            # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
+            res = "seed-coder"
 
         if res is None:
             logger.warning("\n")
@@ -2066,6 +2069,9 @@ class Llama4Model(LlamaModel):
         self.gguf_writer.add_expert_feed_forward_length(self.hparams["intermediate_size_moe"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None):
+        if name.startswith("language_model."):
+            name = name.replace("language_model.", "")
+
         # split the gate_up into gate and up
         if "gate_up_proj" in name:
             name_up = name.replace("gate_up_proj", "up_proj.weight")
@@ -5743,11 +5749,20 @@ class GraniteModel(LlamaModel):
             logger.info("gguf: (granite) logits_scale = %s", logits_scale)
 
 
-@ModelBase.register("GraniteMoeForCausalLM")
+@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
 class GraniteMoeModel(GraniteModel):
     """Conversion for IBM's GraniteMoeForCausalLM"""
     model_arch = gguf.MODEL_ARCH.GRANITE_MOE
 
+    def set_gguf_parameters(self):
+        """GraniteMoeShared uses GraniteMoe parameters plus the following:
+        - shared_intermediate_size
+        """
+        super().set_gguf_parameters()
+        if shared_feed_forward_length := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_feed_forward_length)
+            logger.info("gguf: (granitemoeshared) shared_feed_forward_length = %s", shared_feed_forward_length)
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         """In modeling_granitemoe, the JetMoe implementation of parallel experts
         is used. This essentially merges w1 and w3 into a single tensor with 2x
@@ -5758,12 +5773,21 @@ class GraniteMoeModel(GraniteModel):
         if name.endswith("block_sparse_moe.input_linear.weight"):
             ffn_dim = self.hparams["intermediate_size"]
             assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
-            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
+            gate, up = data_torch.split(ffn_dim, dim=-2)
             return [
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
                 (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
             ]
 
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams["shared_intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            gate, up = data_torch.split(ffn_dim, dim=-2)
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_SHEXP, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), up),
+            ]
+
         return super().modify_tensors(data_torch, name, bid)
 
 

convert_hf_to_gguf_update.py

@@ -116,6 +116,7 @@ models = [
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
+    {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]
 
 

docs/backend/SYCL.md

@@ -731,6 +731,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DEVICE_ARCH | Optional (except for AMD)             | Set the SYCL device architecture, optional except for AMD. Setting the device architecture can improve the performance. See the table [--offload-arch](https://github.com/intel/llvm/blob/sycl/sycl/doc/design/OffloadDesign.md#--offload-arch) for a list of valid architectures. |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
 | GGML_SYCL_GRAPH    | ON *(default)* \|OFF *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
+| GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
@@ -741,6 +742,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
 | GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
+| GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
 

docs/multimodal.md

@@ -6,7 +6,7 @@ llama.cpp supports multimodal input via `libmtmd`. Currently, there are 2 tools
 
 To enable it, can use use one of the 2 methods below:
 
-- Use `-hf` option with a [supported model](../../docs/multimodal.md)
+- Use `-hf` option with a supported model (see a list of pre-quantized model below)
     - To load a model using `-hf` while disabling multimodal, use `--no-mmproj`
     - To load a model using `-hf` while using a custom mmproj file, use `--mmproj local_file.gguf`
 - Use `-m model.gguf` option with `--mmproj file.gguf` to specify text and multimodal projector respectively
@@ -31,7 +31,7 @@ llama-server -hf ggml-org/gemma-3-4b-it-GGUF --no-mmproj-offload
 
 ## Pre-quantized models
 
-These are ready-to-use models, most of them come with `Q4_K_M` quantization by default.
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default. They can be found at the Hugging Face page of the ggml-org: https://huggingface.co/ggml-org
 
 Replaces the `(tool_name)` with the name of binary you want to use. For example, `llama-mtmd-cli` or `llama-server`
 
@@ -69,9 +69,9 @@ NOTE: some models may require large context window, for example: `-c 8192`
 
 # InternVL 2.5 and 3
 (tool_name) -hf ggml-org/InternVL2_5-1B-GGUF
-(tool_name) -hf ggml-org/InternVL2_5-2B-GGUF
+(tool_name) -hf ggml-org/InternVL2_5-4B-GGUF
 (tool_name) -hf ggml-org/InternVL3-1B-Instruct-GGUF
 (tool_name) -hf ggml-org/InternVL3-2B-Instruct-GGUF
-(tool_name) -hf ggml-org/InternVL3-4B-Instruct-GGUF
+(tool_name) -hf ggml-org/InternVL3-8B-Instruct-GGUF
 (tool_name) -hf ggml-org/InternVL3-14B-Instruct-GGUF
 ```

examples/CMakeLists.txt

@@ -32,6 +32,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(speculative-simple)
     add_subdirectory(gen-docs)
+    add_subdirectory(training)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/training/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-finetune)
+add_executable(${TARGET} finetune.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/training/README.md

@@ -0,0 +1,17 @@
+# llama.cpp/examples/training
+
+This directory contains examples related to language model training using llama.cpp/GGML.
+So far finetuning is technically functional (for FP32 models and limited hardware setups) but the code is very much WIP.
+Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
+**For CPU training, compile llama.cpp without any additional backends such as CUDA.**
+**For CUDA training, use the maximum number of GPU layers.**
+
+Proof of concept:
+
+``` sh
+export model_name=llama_3.2-1b && export quantization=f32
+./build/bin/finetune --file wikitext-2-raw/wiki.test.raw -ngl 999 --model models/${model_name}-${quantization}.gguf -c 512 -b 512 -ub 512
+./build/bin/perplexity --file wikitext-2-raw/wiki.test.raw -ngl 999 --model finetuned-model.gguf
+```
+
+The perplexity value of the finetuned model should be lower after training on the test set for 2 epochs.

examples/training/finetune.cpp

@@ -0,0 +1,96 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    params.escape = false;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+        return 1;
+    }
+
+    if (params.use_mmap) {
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        params.use_mmap = false;
+    }
+    if (params.cache_type_k != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing k cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_k = GGML_TYPE_F32;
+    }
+    if (params.cache_type_v != GGML_TYPE_F32) {
+        LOG_INF("%s: force changing v cache type to f32 due to a lack of f16 support for OUT_PROD\n", __func__);
+        params.cache_type_v = GGML_TYPE_F32;
+    }
+
+    common_init();
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model and apply lora adapter, if any
+    common_init_result llama_init = common_init_from_params(params);
+    llama_model_ptr   & model = llama_init.model;
+    llama_context_ptr & ctx   = llama_init.context;
+
+    if (model == NULL) {
+        LOG_ERR("%s: unable to load model\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+    }
+
+    constexpr float val_split = 0.05f;
+
+    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
+
+    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
+    optimizer_params.adamw.alpha = 1e-7f; // learning rate
+
+    struct llama_opt_params lopt_params {
+        /*n_ctx_train     =*/ 0,
+        /*param_filter    =*/ llama_opt_param_filter_all,
+        /*param_filter_ud =*/ nullptr,
+        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
+        /*get_opt_pars_ud =*/ &optimizer_params,
+    };
+    llama_opt_init(ctx.get(), model.get(), lopt_params);
+
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_eval  = ggml_opt_result_init();
+
+    for (int epoch = 0; epoch < 2; ++epoch) {
+        llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
+            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+        fprintf(stderr, "\n");
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_eval);
+    }
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_eval);
+
+    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+
+    llama_backend_free();
+
+    return 0;
+}

ggml/CMakeLists.txt

@@ -193,6 +193,7 @@ option(GGML_RPC                             "ggml: use RPC"
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
+option(GGML_SYCL_DNN                        "ggml: enable oneDNN in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING

ggml/include/ggml-backend.h

@@ -248,7 +248,7 @@ extern "C" {
         // preferrably to run on the same backend as the buffer
         ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         // initialize buffers from a max size graph (optional)
         reserve_graph = build_graph(sched, max_batch_size);
@@ -289,7 +289,7 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler, backends with low index are given priority over backends with high index
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel, bool op_offload);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
     // Initialize backend buffers from a measure graph

ggml/include/ggml-opt.h

@@ -37,13 +37,16 @@ extern "C" {
     // ====== Dataset ======
 
     GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
-            int64_t ne_datapoint, // number of elements per datapoint
-            int64_t ne_label,     // number of elements per label
-            int64_t ndata,        // total number of datapoints/labels
-            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+            enum ggml_type type_data,    // the type for the internal data tensor
+            enum ggml_type type_label,   // the type for the internal labels tensor
+            int64_t        ne_datapoint, // number of elements per datapoint
+            int64_t        ne_label,     // number of elements per label
+            int64_t        ndata,        // total number of datapoints/labels
+            int64_t        ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
     GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
 
     // get underlying tensors that store the data
+    GGML_API int64_t              ggml_opt_dataset_ndata (ggml_opt_dataset_t dataset);
     GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
     GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
 
@@ -56,13 +59,19 @@ extern "C" {
             struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
             struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
             int64_t              ibatch);
+    GGML_API void ggml_opt_dataset_get_batch_host(
+            ggml_opt_dataset_t   dataset,
+            void               * data_batch,
+            size_t               nb_data_batch,
+            void               * labels_batch,
+            int64_t              ibatch);
 
     // ====== Model / Context ======
 
     enum ggml_opt_build_type {
-        GGML_OPT_BUILD_TYPE_FORWARD,
-        GGML_OPT_BUILD_TYPE_GRAD,
-        GGML_OPT_BUILD_TYPE_OPT,
+        GGML_OPT_BUILD_TYPE_FORWARD = 10,
+        GGML_OPT_BUILD_TYPE_GRAD    = 20,
+        GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
@@ -81,20 +90,22 @@ extern "C" {
     // userdata can be used to pass arbitrary data
     typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
 
-    // returns the default optimizer params (constant)
+    // returns the default optimizer params (constant, hard-coded values)
     // userdata is not used
     GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
 
+    // casts userdata to ggml_opt_optimizer_params and returns it
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata);
+
     // parameters for initializing a new optimization context
     struct ggml_opt_params {
         ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
 
-        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
-
-        // the forward graph is defined by inputs and outputs
-        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
-        struct ggml_tensor * inputs;
-        struct ggml_tensor * outputs;
+        // by default the forward graph needs to be reconstructed for each eval
+        // if ctx_compute, inputs, and outputs are set the graphs are instead allocated statically
+        struct ggml_context * ctx_compute;
+        struct ggml_tensor  * inputs;
+        struct ggml_tensor  * outputs;
 
         enum ggml_opt_loss_type  loss_type;
         enum ggml_opt_build_type build_type;
@@ -107,12 +118,9 @@ extern "C" {
 
     // get parameters for an optimization context with defaults set where possible
     // parameters for which no sensible defaults exist are supplied as arguments to this function
-    GGML_API ggml_opt_params ggml_opt_default_params(
-            ggml_backend_sched_t      backend_sched,
-            struct ggml_context     * ctx_compute,
-            struct ggml_tensor      * inputs,
-            struct ggml_tensor      * outputs,
-            enum ggml_opt_loss_type   loss_type);
+    GGML_API struct ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t    backend_sched,
+            enum ggml_opt_loss_type loss_type);
 
     GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
     GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
@@ -121,6 +129,7 @@ extern "C" {
     GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
 
     // get underlying tensors that store data
+    // if not using static graphs these pointers become invalid with the next call to ggml_opt_alloc
     GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
     GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
     GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
@@ -128,11 +137,12 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
     GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
 
+    // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
     // ====== Optimization Result ======
 
-    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API ggml_opt_result_t ggml_opt_result_init(void);
     GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
     GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
 
@@ -144,11 +154,20 @@ extern "C" {
 
     // ====== Computation ======
 
-    // do forward pass, increment result if not NULL
-    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // if not using static graphs, this function must be called prior to ggml_opt_alloc
+    GGML_API void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs);
 
-    // do forward pass, increment result if not NULL, do backward pass
-    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+    // allocate the next graph for evaluation, either forward or forward + backward
+    // must be called exactly once prior to calling ggml_opt_eval
+    GGML_API void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward);
+
+    // do forward pass, increment result if not NULL, do backward pass if allocated
+    GGML_API void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
 
     // ############################################################################
     // ## The high-level functions start here. They do not depend on any private ##
@@ -200,9 +219,9 @@ extern "C" {
     // fit model defined by inputs and outputs to dataset
     GGML_API void ggml_opt_fit(
             ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
-            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
-            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
-            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            struct ggml_context           * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            struct ggml_tensor            * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)

ggml/include/ggml.h

@@ -768,7 +768,7 @@ extern "C" {
     // Tensor flags
     GGML_API void ggml_set_input(struct ggml_tensor * tensor);
     GGML_API void ggml_set_output(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_tensor * tensor);
     GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     //
@@ -938,7 +938,7 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_repeat_back(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
 
     // concat a and b along dim
     // used in stable-diffusion
@@ -2049,15 +2049,14 @@ extern "C" {
 
     GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
-        struct ggml_context * ctx_compute, // context for gradient computation
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
+        struct ggml_context *  ctx,        // context for gradient computation
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs);
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
     GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads);
     GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
     GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
     GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);

ggml/src/ggml-backend.cpp

@@ -674,6 +674,8 @@ struct ggml_backend_sched {
     char * context_buffer;
     size_t context_buffer_size;
 
+    bool op_offload;
+
     int debug;
 };
 
@@ -766,7 +768,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
+            if (sched->op_offload && src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");
@@ -1109,7 +1111,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             const int node_backend_id = tensor_backend_id(node);
 
-            assert(node_backend_id != -1); // all nodes should be assigned by now
+            assert(node_backend_id != -1); // all nodes should be assigned by now, this can happen if there is no CPU fallback
 
             // check if we should start a new split based on the sources of the current node
             bool need_new_split = false;
@@ -1452,7 +1454,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
         ggml_backend_buffer_type_t * bufts,
         int n_backends,
         size_t graph_size,
-        bool parallel) {
+        bool parallel,
+        bool op_offload) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
@@ -1497,6 +1500,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+    sched->op_offload = op_offload;
 
     ggml_backend_sched_reset(sched);
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -385,9 +385,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.5.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "ea22e1aefb800e9bc8c74d91633cc58e")
+        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)
@@ -428,6 +428,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${KLEIDIAI_SRC}/kai/ukernels/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
             ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
 
         set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
@@ -438,17 +439,19 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
         string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
 
-        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS})
+        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
 
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c)
-        list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
+        list(APPEND GGML_KLEIDIAI_SOURCES
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c)
 
         if (NOT DOTPROD_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c)
         endif()
 
         if (NOT I8MM_ENABLED MATCHES -1)
@@ -456,9 +459,13 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (NOT SME_ENABLED MATCHES -1)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c)
-            list(APPEND GGML_KLEIDIAI_SOURCES ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c)
-            set(PRIVATE_ARCH_FLAGS "${PRIVATE_ARCH_FLAGS}+sve+sve2")
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c)
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}+sve+sve2")
         endif()
 
         set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -8519,7 +8519,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
     assert(nrc == 1);
+#endif
     UNUSED(nrc);
     UNUSED(bx);
     UNUSED(by);
@@ -8530,6 +8534,197 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     const int nb = n / QK_K;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q6_K * GGML_RESTRICT x0 = x;
+        const block_q6_K * GGML_RESTRICT x1 = (const block_q6_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * GGML_RESTRICT y0 = y;
+        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * GGML_RESTRICT ql0 = x0->ql;
+            const uint8_t * GGML_RESTRICT ql1 = x1->ql;
+            const uint8_t * GGML_RESTRICT qh0 = x0->qh;
+            const uint8_t * GGML_RESTRICT qh1 = x1->qh;
+            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
+
+            const uint8x16_t mone = vdupq_n_u8(0x30);
+            const uint8x16_t  m4b = vdupq_n_u8(0x0f);
+
+            int32x4_t visum = vdupq_n_s32(0);
+
+            // process 8 blocks per iteration, totally 16 blocks
+            for (int j = 0; j < 2; ++j, qh0 += 32, ql0 += 64, qh1 += 32, ql1 += 64) {
+                int8x16_t vx0[8], vx1[8];
+
+                // de-quantize vx0[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh0);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql0);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx0[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx0[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx0[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx0[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx0[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx0[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx0[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // de-quantize vx1[8]
+                {
+                    const uint8x16x2_t qh_bits = vld1q_u8_x2(qh1);
+                    const uint8x16x4_t ql_bits = vld1q_u8_x4(ql1);
+
+                    uint8x16_t q6h_0 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 4));
+                    uint8x16_t q6h_1 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 4));
+                    uint8x16_t q6h_2 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[0], 2));
+                    uint8x16_t q6h_3 = vandq_u8(mone, vshlq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[0], m4b), q6h_0));
+                    vx1[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[1], m4b), q6h_1));
+                    vx1[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[2], m4b), q6h_2));
+                    vx1[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(ql_bits.val[3], m4b), q6h_3));
+
+                    q6h_0 = vandq_u8(mone, qh_bits.val[0]);
+                    q6h_1 = vandq_u8(mone, qh_bits.val[1]);
+                    q6h_2 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[0], 2));
+                    q6h_3 = vandq_u8(mone, vshrq_n_u8(qh_bits.val[1], 2));
+
+                    vx1[4] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[0], 4), q6h_0));
+                    vx1[5] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[1], 4), q6h_1));
+                    vx1[6] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[2], 4), q6h_2));
+                    vx1[7] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(ql_bits.val[3], 4), q6h_3));
+                }
+
+                // process 16 elements (one block with same scale) per iteration
+                // - vx = concat(ql, qh) - 32
+                // - r1,r2,r3,r4 = smmla(vx, vy)
+                for (int k = 0; k < 8; ++k) {
+                    const int blk = j * 8 + k;
+
+                    const int8x16_t vy0 = vld1q_s8(qy0);
+                    const int8x16_t vy1 = vld1q_s8(qy1);
+                    qy0 += 16;
+                    qy1 += 16;
+
+                    const int32x4_t block_scale = {
+                        x0->scales[blk],
+                        x0->scales[blk],
+                        x1->scales[blk],
+                        x1->scales[blk],
+                    };
+
+                    // calculate four results at once with outer product
+                    const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vx0[k]), vreinterpretq_s64_s8(vx1[k])));
+                    const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(vy0), vreinterpretq_s64_s8(vy1)));
+                    int32x4_t vr = vdupq_n_s32(0);
+                    vr = vmmlaq_s32(vr, vx_l, vy_l);
+                    vr = vmmlaq_s32(vr, vx_h, vy_h);
+
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int6*int8) <= 2^(8+8+6+8) = 30 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+#ifdef __ARM_FEATURE_SVE
+                const svbool_t pg16_8 = svptrue_pat_b16(SV_VL8);
+                const svbool_t pg8_8 = svptrue_pat_b8(SV_VL8);
+                const svint16_t y0_q8sums_0 = svld1_s16(pg16_8, y0->bsums);
+                const svint16_t y0_q8sums_1 = svld1_s16(pg16_8, y0->bsums + 8);
+                const svint16_t y1_q8sums_0 = svld1_s16(pg16_8, y1->bsums);
+                const svint16_t y1_q8sums_1 = svld1_s16(pg16_8, y1->bsums + 8);
+                const svint16_t x0_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x0->scales));
+                const svint16_t x0_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x0->scales + 8));
+                const svint16_t x1_q6scales_0 = svunpklo_s16(svld1_s8(pg8_8, x1->scales));
+                const svint16_t x1_q6scales_1 = svunpklo_s16(svld1_s8(pg8_8, x1->scales + 8));
+                const svint64_t zero = svdup_n_s64(0);
+                bias[0] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x0_q6scales_1)));
+                bias[1] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x0_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x0_q6scales_1)));
+                bias[2] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y0_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y0_q8sums_1, x1_q6scales_1)));
+                bias[3] = svaddv_s64(svptrue_b64(), svadd_s64_x(svptrue_b64(), svdot_s64(zero, y1_q8sums_0, x1_q6scales_0),
+                                                                               svdot_s64(zero, y1_q8sums_1, x1_q6scales_1)));
+#else
+                // NEON doesn't support int16 dot product, fallback to separated mul and add
+                const int16x8x2_t q8sums0 = vld1q_s16_x2(y0->bsums);
+                const int16x8x2_t q8sums1 = vld1q_s16_x2(y1->bsums);
+
+                int8x16_t scales_s8 = vld1q_s8(x0->scales);
+                const int16x8x2_t q6scales0 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+                scales_s8 = vld1q_s8(x1->scales);
+                const int16x8x2_t q6scales1 = {{vmovl_s8(vget_low_s8(scales_s8)), vmovl_s8(vget_high_s8(scales_s8))}};
+
+                int32x4_t prod;
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[0] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales0.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales0.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales0.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales0.val[1]))));
+                bias[1] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums0.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums0.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[2] = vaddvq_s32(prod);
+                prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[0]), vget_low_s16 (q6scales1.val[0])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[0]), vget_high_s16(q6scales1.val[0]))),
+                                 vaddq_s32(vmull_s16(vget_low_s16 (q8sums1.val[1]), vget_low_s16 (q6scales1.val[1])),
+                                           vmull_s16(vget_high_s16(q8sums1.val[1]), vget_high_s16(q6scales1.val[1]))));
+                bias[3] = vaddvq_s32(prod);
+
+#endif
+                const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
+
+                const float32x4_t superblock_scale = {
+                    GGML_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                };
+
+                visum = vsubq_s32(visum, vibias);
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
 #ifdef __ARM_FEATURE_SVE
     const int vector_length = ggml_cpu_get_sve_cnt()*8;
     float sum = 0;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -282,7 +282,11 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q6_K,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
         .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_IQ2_XXS] = {
         .from_float               = NULL,

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

@@ -4,16 +4,22 @@
 
 // KleidiAI micro-kernels
 #include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
-#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
-#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
-#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
 #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1vlx4_qsi4c32p4vlx4_1vlx4vl_sme2_mopa.h"
 #include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
+#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+
+#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
+
+#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
+#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
+#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
+
 #include "kai_common.h"
 
 #include "kernels.h"
@@ -61,6 +67,53 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon,
         },
         /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+    {
+        /* SME GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .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,
+        },
+        /* SME GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
+            /* .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,
+        },
+        /* .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,
+        },
+        /* .rhs_info = */ {
+            /* .packed_size = */ kai_get_rhs_packed_size_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+            /* .pack_func   = */ kai_run_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SME,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_F16,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__APPLE__)
@@ -105,6 +158,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_MATMUL_INT8)
@@ -148,6 +204,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #else
@@ -192,6 +251,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD | CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #if defined(__ARM_FEATURE_DOTPROD)
@@ -235,12 +297,33 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .pack_func   = */ kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
         },
         /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
     },
 #endif
 #endif
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature features) {
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor) {
+    ggml_kleidiai_kernels * kernel = nullptr;
+
+    if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
+        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 &&
+                gemm_gemv_kernels[i].rhs_type == tensor->src[0]->type &&
+                gemm_gemv_kernels[i].op_type  == tensor->type) {
+                kernel = &gemm_gemv_kernels[i];
+                break;
+            }
+        }
+    }
+
+    return kernel;
+}
+
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels); ++i) {

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

@@ -4,6 +4,10 @@
 
 #pragma once
 
+#include <functional>
+#include <variant>
+#include "ggml.h"
+
 enum cpu_feature {
     CPU_FEATURE_NONE    = 0,
     CPU_FEATURE_DOTPROD = 1,
@@ -26,26 +30,53 @@ struct kernel_info {
     size_t (*get_nr)(void);
     size_t (*get_kr)(void);
     size_t (*get_sr)(void);
-    size_t (*get_lhs_offset)(size_t m_idx, size_t k, size_t bl);
-    size_t (*get_rhs_packed_offset)(size_t n_idx, size_t k, size_t bl);
+    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);
-    void (*run_kernel)(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::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;
 };
 
 struct lhs_packing_info {
     size_t (*get_offset)(size_t m_idx, size_t lhs_stride);
-    size_t (*get_packed_offset)(size_t m_idx, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    size_t (*packed_size)(size_t m, size_t k, size_t bl, size_t mr, size_t kr, size_t sr);
-    void (*pack_func)(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::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;
 };
 
 struct rhs_packing_info {
-    size_t (*packed_size)(size_t n, size_t k, size_t nr, size_t kr, size_t bl);
-    void (*pack_func)(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::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;
+    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;
 };
 
 struct ggml_kleidiai_kernels {
@@ -55,6 +86,10 @@ struct ggml_kleidiai_kernels {
     rhs_packing_info rhs_info;
 
     cpu_feature required_cpu;
+    ggml_type lhs_type;
+    ggml_type rhs_type;
+    ggml_type op_type;
 };
 
-ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, const ggml_tensor * tensor);
+ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features);

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

@@ -3,7 +3,9 @@
 //
 #include <arm_neon.h>
 #include <assert.h>
+#include <atomic>
 #include <cfloat>
+#include <stdexcept>
 #include <stdint.h>
 #include <string.h>
 #if defined(__linux__)
@@ -34,8 +36,9 @@
 #include "ggml-common.h"
 
 struct ggml_kleidiai_context {
+    cpu_feature features;
     ggml_kleidiai_kernels * kernels;
-} static ctx = { NULL };
+} static ctx = { CPU_FEATURE_NONE, NULL };
 
 static void init_kleidiai_context(void) {
 
@@ -47,18 +50,18 @@ static void init_kleidiai_context(void) {
         const char *env_var = getenv("GGML_KLEIDIAI_SME");
         int sme_enabled = 0;
 
-        cpu_feature features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                                (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
+        ctx.features  = (ggml_cpu_has_dotprod()     ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_matmul_int8() ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
+                        (ggml_cpu_has_sve()         ? CPU_FEATURE_SVE     : CPU_FEATURE_NONE);
 
         if (env_var) {
             sme_enabled = atoi(env_var);
         }
 
         if (sme_enabled != 0) {
-            features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
+            ctx.features |= ggml_cpu_has_sme() ? CPU_FEATURE_SME : CPU_FEATURE_NONE;
         }
-        ctx.kernels = ggml_kleidiai_select_kernels(features);
+        ctx.kernels = ggml_kleidiai_select_kernels_q4_0(ctx.features);
     }
     ggml_critical_section_end();
 }
@@ -68,95 +71,275 @@ static inline int64_t ggml_ne(const ggml_tensor * tensor, int dim) {
     return tensor->ne[dim];
 }
 
+template<typename Ret, typename Variant, typename... Args>
+static Ret variant_call(const Variant & var, Args&&... args) {
+    return std::visit([&](auto&& func) -> Ret {
+        if constexpr (std::is_invocable_r_v<Ret, decltype(func), Args...>) {
+            return func(std::forward<Args>(args)...);
+        } else {
+            throw std::runtime_error("Invalid function type in variant_call");
+        }
+    }, var);
+}
+
 namespace ggml::cpu::kleidiai {
+
+static size_t round_down(size_t x, size_t y) {
+    return y == 0 ? x : x - (x % y);
+}
+
+static void transpose_f32kxn_f16nxk(size_t n, size_t k, float * dst, const uint16_t * src, size_t rhs_stride) {
+    size_t src_stride = rhs_stride / sizeof(uint16_t);
+    size_t dst_stride = n;
+
+    for (size_t k_idx = 0; k_idx < k; ++k_idx) {
+        for (size_t n_idx = 0; n_idx < n; ++n_idx) {
+            uint16_t v = *(src + k_idx + n_idx * src_stride);
+            *(dst + n_idx + k_idx * dst_stride) = kai_cast_f32_f16(v);
+        }
+    }
+}
+
 class tensor_traits : public ggml::cpu::tensor_traits {
     bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
-        GGML_ASSERT(ctx.kernels);
-        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, op);
+        GGML_ASSERT(kernels);
+        kernel_info * kernel = op->src[1]->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
 
         size_t k = op->src[0]->ne[0];
+        size_t n = op->src[0]->ne[1];
         size_t m = op->src[1]->ne[1];
 
         size_t mr = kernel->get_mr();
         size_t kr = kernel->get_kr();
         size_t sr = kernel->get_sr();
 
-        size = ctx.kernels->lhs_info.packed_size(m, k, QK4_0, mr, kr, sr);
+        if (kernels->rhs_type == GGML_TYPE_Q4_0) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, QK4_0, mr, kr, sr);
+        } else if (kernels->rhs_type == GGML_TYPE_F16) {
+            size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr) +
+                   variant_call<size_t>(kernels->rhs_info.packed_size, n, k) +
+                   k * n * sizeof(float) + n * sizeof(float);
+        } else {
+            GGML_ASSERT(false);
+        }
 
         return true;
     }
 
+
     bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * dst) override {
         if (dst->op == GGML_OP_MUL_MAT) {
-            const ggml_tensor * src0 = dst->src[0];
-            const ggml_tensor * src1 = dst->src[1];
+            if (dst->src[0]->type == GGML_TYPE_Q4_0) {
+                return compute_forward_q4_0(params, dst);
+            } else if (dst->src[0]->type == GGML_TYPE_F16) {
+                return compute_forward_kv_cache(params, dst);
+            }
+        }
+        return false;
+    }
 
-            GGML_TENSOR_BINARY_OP_LOCALS
+    bool compute_forward_kv_cache(ggml_compute_params * params, struct ggml_tensor * dst) {
+        static std::atomic_flag first_to_arrive = ATOMIC_FLAG_INIT;
 
-            GGML_ASSERT(ctx.kernels);
-            kernel_info * kernel = src1->ne[1] == 1 ? &ctx.kernels->gemv : &ctx.kernels->gemm;
-            lhs_packing_info * lhs_info = &ctx.kernels->lhs_info;
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
 
-            GGML_ASSERT(kernel);
+        GGML_TENSOR_BINARY_OP_LOCALS
 
-            const int ith = params->ith;
-            const int nth = params->nth;
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
 
-            const size_t k = ne00;
-            const size_t m = ne11;
-            const size_t n = ne01;
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        GGML_ASSERT(kernel);
 
-            const size_t n_step = kernel->get_n_step();
-            const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
-            const size_t n_start = ith * num_n_per_thread;
+        const int nth = params->nth;
+        const int ith = params->ith;
 
-            size_t n_to_process = num_n_per_thread;
-            if ((n_start + n_to_process) > n) {
-                n_to_process = n - n_start;
+        const int64_t lhs_batch_size0 = ne12;
+        const int64_t rhs_batch_size0 = ne02;
+        const int64_t batch_size      = rhs_batch_size0;
+
+        const int64_t r = lhs_batch_size0 / rhs_batch_size0;
+
+        const int64_t m = ne11 * r;
+        const int64_t n = ne01;
+        const int64_t k = ne00;
+
+        const size_t lhs_stride = src1->nb[1];
+        const size_t rhs_stride = src0->nb[1];
+        const size_t dst_stride = dst->nb[1];
+
+        const int64_t mr = static_cast<int64_t>(kernel->get_mr());
+        const int64_t nr = static_cast<int64_t>(kernel->get_nr());
+        const int64_t kr = static_cast<int64_t>(kernel->get_kr());
+        const int64_t sr = static_cast<int64_t>(kernel->get_sr());
+
+        const size_t lhs_packed_size = variant_call<size_t>(kernels->lhs_info.packed_size, m, k, mr, kr, sr);
+        const size_t rhs_packed_size = variant_call<size_t>(kernels->rhs_info.packed_size, n, k);
+        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);
+
+        uint8_t * lhs_packed = static_cast<uint8_t *>(params->wdata);
+        uint8_t * rhs_packed = lhs_packed + lhs_packed_size;
+        uint8_t * rhs_kxn    = rhs_packed + rhs_packed_size;
+        uint8_t * bias       = rhs_kxn + kxn_size;
+
+        for (int64_t batch_idx = 0; batch_idx < batch_size; ++batch_idx) {
+            const uint8_t * lhs_batch = static_cast<const uint8_t *>(src1->data) + batch_idx * m * lhs_stride;
+            const uint8_t * rhs_batch = static_cast<const uint8_t *>(src0->data) + batch_idx * n * rhs_stride;
+            uint8_t * dst_batch       = static_cast<uint8_t *>(dst->data) + batch_idx * m * dst_stride;
+
+            // LHS packing
+            {
+                const int64_t m_roundup_mr = kai_roundup(m, mr);
+                const int64_t num_threads  = KAI_MIN(m_roundup_mr / mr, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_m_per_thread0   = round_down(m_roundup_mr / num_threads, mr);
+                    const int64_t num_m_per_threadN_1 = m - (num_threads - 1) * num_m_per_thread0;
+
+                    const int64_t m_start          = ith * num_m_per_thread0;
+                    const int64_t num_m_per_thread = (ith == num_threads - 1) ? num_m_per_threadN_1 : num_m_per_thread0;
+
+                    const size_t lhs_offset        = variant_call<size_t>(kernels->gemm.get_lhs_offset, m_start, lhs_stride);
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernels->lhs_info.get_packed_offset, m_start, k, mr, kr, sr);
+
+                    const void * src_ptr = static_cast<const uint8_t *>(lhs_batch) + lhs_offset;
+                    void * dst_ptr       = static_cast<uint8_t *>(lhs_packed) + lhs_packed_offset;
+
+                    variant_call<void>(kernels->lhs_info.pack_func, num_m_per_thread, k, mr, kr, sr, 0, src_ptr, lhs_stride, dst_ptr);
+                }
             }
 
-            const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
-            uint8_t * lhs_packed       = (uint8_t*)params->wdata;
-            const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+            // RHS packing
+            if (first_to_arrive.test_and_set(std::memory_order_acquire) == false) {
+                // First thread to reach this point handles RHS packing
+                memset(bias, 0, n * sizeof(float));
+                transpose_f32kxn_f16nxk(n, k, reinterpret_cast<float *>(rhs_kxn),
+                                        reinterpret_cast<const uint16_t *>(rhs_batch), rhs_stride);
 
-            size_t mr = kernel->get_mr();
-            size_t kr = kernel->get_kr();
-            size_t sr = kernel->get_sr();
-
-            // Calculate number of columns to be processed per thread
-            const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
-            const size_t m_start = ith * num_m_per_thread;
-            size_t m_to_process = num_m_per_thread;
-            if ((m_start + m_to_process) > m) {
-                m_to_process = m - m_start;
-            }
-
-            if(m_start < m) {
-                // 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 = lhs_info->get_packed_offset(m_start, k, QK4_0, mr, kr, sr);
-                void * lhs_packed_ptr          = static_cast<void *>(lhs_packed + lhs_packed_offset);
-
-                lhs_info->pack_func(m_to_process, k, QK4_0, mr, kr, sr, 0, src_ptr, src_stride, lhs_packed_ptr);
+                variant_call<void>(kernels->rhs_info.pack_func, 1, n, k, nr, kr, sr, n * sizeof(float),
+                             rhs_kxn, bias, nullptr, rhs_packed, 0, nullptr);
             }
 
             ggml_barrier(params->threadpool);
 
-            // Perform the operation
-            const size_t dst_stride        = dst->nb[1];
-            const size_t lhs_packed_offset = lhs_info->get_packed_offset(0, k, QK4_0, mr, kr, sr);
-            const size_t rhs_packed_offset = kernel->get_rhs_packed_offset(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);
+            first_to_arrive.clear(std::memory_order_release);
 
-            kernel->run_kernel(m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr,
-                               dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
-            return true;
+            // Perform the matmul
+            {
+                const int64_t m_to_process = m;
+                const int64_t m_start      = 0;
+
+                const int64_t n_step      = static_cast<int64_t>(kernel->get_n_step());
+                const int64_t num_threads = KAI_MIN(n / n_step, nth);
+
+                if (ith < num_threads) {
+                    const int64_t num_n_per_thread0   = round_down(n / num_threads, n_step);
+                    const int64_t num_n_per_threadN_1 = n - (num_threads - 1) * num_n_per_thread0;
+
+                    const int64_t n_start      = ith * num_n_per_thread0;
+                    const int64_t n_to_process = (ith == num_threads - 1) ? num_n_per_threadN_1 : num_n_per_thread0;
+
+                    const size_t lhs_packed_offset = variant_call<size_t>(kernel->get_lhs_offset, m_start, k);
+                    const size_t rhs_packed_offset = variant_call<size_t>(kernel->get_rhs_packed_offset, n_start, k);
+                    const size_t dst_offset        = kernel->get_dst_offset(m_start, n_start, dst_stride);
+
+                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * rhs_ptr = rhs_packed + rhs_packed_offset;
+                    float * dst_ptr      = reinterpret_cast<float *>(dst_batch + dst_offset);
+
+                    variant_call<void>(kernel->run_kernel, m_to_process, n_to_process, k, lhs_ptr, rhs_ptr, dst_ptr, dst_stride, sizeof(float), -FLT_MAX, FLT_MAX);
+                }
+            }
+
+            if (batch_idx != batch_size - 1) {
+                // This barrier is necessary when the batch size is larger than 1. While processing a batch,
+                // the work data buffer (params->wdata) is used as temporary storage which means that only
+                // a single batch can be processed at any given time. No barrier is needed for the last
+                // batch since GGML inserts a barrier between the execution of every operator.
+                ggml_barrier(params->threadpool);
+            }
         }
-        return false;
+
+        return true;
+    }
+
+    bool compute_forward_q4_0(struct ggml_compute_params * params, struct ggml_tensor * dst) {
+        const ggml_tensor * src0 = dst->src[0];
+        const ggml_tensor * src1 = dst->src[1];
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        ggml_kleidiai_kernels *kernels = ggml_kleidiai_select_kernels(ctx.features, dst);
+        GGML_ASSERT(kernels);
+
+        kernel_info * kernel = src1->ne[1] == 1 ? &kernels->gemv : &kernels->gemm;
+        lhs_packing_info * lhs_info = &kernels->lhs_info;
+
+        GGML_ASSERT(kernel);
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+
+        size_t mr = kernel->get_mr();
+        size_t kr = kernel->get_kr();
+        size_t sr = kernel->get_sr();
+
+        const uint8_t * lhs        = static_cast<const uint8_t *>(src1->data);
+        uint8_t * lhs_packed       = (uint8_t*)params->wdata;
+        const uint8_t * rhs_packed = static_cast<const uint8_t *>(src0->data);
+
+        const size_t n_step = kernel->get_n_step();
+        const size_t num_n_per_thread = kai_roundup(kai_roundup(n, nth) / nth, n_step);
+        const size_t n_start = ith * num_n_per_thread;
+
+        size_t n_to_process = num_n_per_thread;
+        if ((n_start + n_to_process) > n) {
+            n_to_process = n - n_start;
+        }
+
+        // Calculate number of columns to be processed per thread
+        const size_t num_m_per_thread = kai_roundup(m, mr * nth) / nth;
+        const size_t m_start = ith * num_m_per_thread;
+        size_t m_to_process = num_m_per_thread;
+        if ((m_start + m_to_process) > m) {
+            m_to_process = m - m_start;
+        }
+
+        if (m_start < m) {
+            // 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);
+            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);
+        }
+
+        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 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);
+
+        variant_call<void>(kernel->run_kernel, m, n_to_process, k, QK4_0, lhs_ptr, rhs_ptr, dst_ptr, dst_stride,
+                           sizeof(float), -FLT_MAX, FLT_MAX);
+
+        return true;
     }
 
 public:
@@ -169,13 +352,13 @@ public:
         size_t sr      = ctx.kernels->gemm.get_sr();
 
 #ifndef NDEBUG
-        const size_t repacked_size = ctx.kernels->rhs_info.packed_size(n, k, nr, kr, QK4_0);
+        const size_t repacked_size = variant_call<size_t>(ctx.kernels->rhs_info.packed_size, n, k, nr, kr, QK4_0);
         GGML_ASSERT(repacked_size <= data_size && "repacked size larger than the packed size!");
 #endif
         struct kai_rhs_pack_qs4cxs1s0_param params;
         params.lhs_zero_point = 1;
         params.rhs_zero_point = 8;
-        ctx.kernels->rhs_info.pack_func(1, n, k, nr, kr, sr, QK4_0, (const uint8_t *)data, NULL, tensor->data, 0, &params);
+        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);
 
         return 0;
 
@@ -189,7 +372,7 @@ static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struc
 }
 }  // namespace ggml::cpu::kleidiai
 
-GGML_API enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+static enum ggml_status ggml_backend_cpu_kleidiai_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     tensor->extra = (void *) ggml::cpu::kleidiai::get_tensor_traits(buffer, tensor);
 
     GGML_UNUSED(buffer);
@@ -238,12 +421,11 @@ static size_t ggml_backend_cpu_kleidiai_buffer_type_get_alignment(ggml_backend_b
 namespace ggml::cpu::kleidiai {
 class extra_buffer_type : ggml::cpu::extra_buffer_type {
     bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
-        if (    op->op == GGML_OP_MUL_MAT &&
-                op->src[0]->type == GGML_TYPE_Q4_0 &&
-                op->src[0]->buffer &&
-                (ggml_n_dims(op->src[0]) == 2) &&
-                op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels
-                ) {
+        if (op->op == GGML_OP_MUL_MAT &&
+            op->src[0]->type == GGML_TYPE_Q4_0 &&
+            op->src[0]->buffer &&
+            (ggml_n_dims(op->src[0]) == 2) &&
+            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() && ctx.kernels) {
             if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
                 return false;
             }
@@ -260,6 +442,19 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
             if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
                 return (ggml::cpu::tensor_traits *) op->src[0]->extra;
             }
+            else if (ggml_kleidiai_select_kernels(ctx.features, op) &&
+                     op->src[0]->op == GGML_OP_VIEW &&
+                     (op->src[1]->op == GGML_OP_PERMUTE || op->src[1]->op ==  GGML_OP_SOFT_MAX) &&
+                     op->src[1]->ne[1] > 1) {
+                if ((op->src[0]->nb[0] != 2) ||
+                    (op->src[1]->nb[0] != 4) ||
+                    (op->src[0]->nb[1] * op->src[0]->ne[1] != op->src[0]->nb[2]) ||
+                    (op->src[1]->nb[1] * op->src[1]->ne[1] != op->src[1]->nb[2])) {
+                    return nullptr;
+                }
+
+                return ggml::cpu::kleidiai::get_tensor_traits(NULL, NULL);
+            }
         }
         return nullptr;
     }

ggml/src/ggml-cuda/acc.cu

@@ -1,47 +1,61 @@
 #include "acc.cuh"
 
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int64_t ne,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s11, const int64_t s12, const int64_t s13, const int64_t offset) {
+    const int64_t i = blockDim.x * blockIdx.x + threadIdx.x;
+
     if (i >= ne) {
         return;
     }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
+
+    int64_t src1_idx = i - offset;
+
+    int64_t tmp = src1_idx;
+    const int64_t i13 = tmp / s13;
+    tmp -= i13 * s13;
+    const int64_t i12 = tmp / s12;
+    tmp -= i12 * s12;
+    const int64_t i11 = tmp / s11;
+    tmp -= i11 * s11;
+    const int64_t i10 = tmp;
+
+    float val = x[i];
+    if (src1_idx >= 0 && i10 < ne10 && i11 < ne11 && i12 < ne12 && i13 < ne13) {
+        val += y[((i13*ne12 + i12) * ne11 + i11) * ne10 + i10];
     }
+    dst[i] = val;
 }
 
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int64_t n_elements,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
+        const int64_t s1, const int64_t s2, const int64_t s3, const int64_t offset, cudaStream_t stream) {
+    const int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, ne13, s1, s2, s3, offset);
 }
 
 void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
-    const float * src0_d = (const float *)src0->data;
-    const float * src1_d = (const float *)src1->data;
-    float * dst_d = (float *)dst->data;
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *)  dst->data;
+
     cudaStream_t stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
 
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(dst->nb[0] == ggml_element_size(dst));
+    GGML_ASSERT(ggml_is_contiguously_allocated(dst));
 
-    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+    const int64_t s1     = dst->op_params[0] / sizeof(float);
+    const int64_t s2     = dst->op_params[1] / sizeof(float);
+    const int64_t s3     = dst->op_params[2] / sizeof(float);
+    const int64_t offset = dst->op_params[3] / sizeof(float);
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3], s1, s2, s3, offset, stream);
 }

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

@@ -678,10 +678,14 @@ void launch_fattn(
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
+    const bool is_mla = DV == 512; // TODO better parameterization
+
     const ggml_tensor * Q = dst->src[0];
     const ggml_tensor * K = dst->src[1];
     const ggml_tensor * V = dst->src[2];
 
+    GGML_ASSERT(V || is_mla);
+
     const ggml_tensor * mask = dst->src[3];
 
     ggml_tensor * KQV = dst;
@@ -689,6 +693,10 @@ void launch_fattn(
     GGML_ASSERT(Q->type == GGML_TYPE_F32);
     GGML_ASSERT(KQV->type == GGML_TYPE_F32);
 
+    GGML_ASSERT(      Q->nb[0] == ggml_element_size(Q));
+    GGML_ASSERT(      K->nb[0] == ggml_element_size(K));
+    GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
+
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
     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");
@@ -713,10 +721,10 @@ void launch_fattn(
     size_t nb12 = K->nb[2];
     size_t nb13 = K->nb[3];
 
-    const char * V_data = (const char *) V->data;
-    size_t nb21 = V->nb[1];
-    size_t nb22 = V->nb[2];
-    size_t nb23 = V->nb[3];
+    const char * V_data = V ? (const char *) V->data : nullptr;
+    size_t nb21 = V ? V->nb[1] : nb11;
+    size_t nb22 = V ? V->nb[2] : nb12;
+    size_t nb23 = V ? V->nb[3] : nb13;
 
     if (need_f16_K && K->type != GGML_TYPE_F16) {
         GGML_ASSERT(ggml_is_contiguously_allocated(K));
@@ -733,7 +741,7 @@ void launch_fattn(
         nb13 = nb13*bs*sizeof(half)/ts;
     }
 
-    if (need_f16_V && V->type != GGML_TYPE_F16) {
+    if (V && need_f16_V && V->type != GGML_TYPE_F16) {
         GGML_ASSERT(ggml_is_contiguously_allocated(V));
         V_f16.alloc(ggml_nelements(V));
         to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);

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

@@ -33,9 +33,30 @@ struct fattn_mma_f16_config< 64,  64> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 32;
-    static constexpr int  nbatch_V2      = 32;
-    static constexpr int  nbatch_combine = 32;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 32;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 32;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 32;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 32;
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 32;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 32;
+    }
 };
 
 template <>
@@ -44,9 +65,30 @@ struct fattn_mma_f16_config< 80,  80> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 40;
-    static constexpr int  nbatch_V2      = 40;
-    static constexpr int  nbatch_combine = 40;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 40;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 40;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 40;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 40;
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 40;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 40;
+    }
 };
 
 template <>
@@ -55,9 +97,30 @@ struct fattn_mma_f16_config< 96,  96> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 48;
-    static constexpr int  nbatch_V2      = 48;
-    static constexpr int  nbatch_combine = 48;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 48;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 48;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 48;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 48;
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 48;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 48;
+    }
 };
 
 template <>
@@ -66,9 +129,30 @@ struct fattn_mma_f16_config<112, 112> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 56;
-    static constexpr int  nbatch_V2      = 56;
-    static constexpr int  nbatch_combine = 56;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 56;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 56;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 56;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 56;
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 56;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 56;
+    }
 };
 
 template <>
@@ -77,9 +161,30 @@ struct fattn_mma_f16_config<128, 128> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 64;
-    static constexpr int  nbatch_V2      = 64;
-    static constexpr int  nbatch_combine = 64;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 64;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 64;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 64;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 64;
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 64;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 64;
+    }
 };
 
 template <>
@@ -88,9 +193,38 @@ struct fattn_mma_f16_config<256, 256> {
     static constexpr int  nwarps_max     = 4;
     static constexpr bool Q_in_reg       = true;
     static constexpr int  nstages_target = 2;
-    static constexpr int  nbatch_K2      = 128;
-    static constexpr int  nbatch_V2      = 128;
-    static constexpr int  nbatch_combine = 128;
+
+    static int get_nbatch_K2_host(const int /*cc*/, const int /*ncols*/) {
+        return 128;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int /*ncols*/) {
+        return 128;
+    }
+
+    static int get_nbatch_V2_host(const int /*cc*/, const int /*ncols*/) {
+        return 128;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int /*ncols*/) {
+        return 128;
+    }
+
+    static int get_nbatch_combine_host(const int cc, const int ncols) {
+        if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING) {
+            return ncols <= 16 ? 128 : 64;
+        }
+        return 64;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int ncols) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+        return ncols <= 16 ? 128 : 64;
+#else
+        GGML_UNUSED(ncols);
+        return 128;
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+    }
 };
 
 template <>
@@ -99,9 +233,44 @@ struct fattn_mma_f16_config<576, 512> {
     static constexpr int  nwarps_max     = 8;
     static constexpr bool Q_in_reg       = false;
     static constexpr int  nstages_target = 1;
-    static constexpr int  nbatch_K2      = 160;
-    static constexpr int  nbatch_V2      = 128;
-    static constexpr int  nbatch_combine = 128;
+
+    static int get_nbatch_K2_host(const int cc, const int ncols) {
+        if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING) {
+            return ncols <= 16 ? 96 : 160;
+        }
+        return ncols <= 16 ? 288 : 160;
+    }
+
+    static constexpr __device__ int get_nbatch_K2_device(int ncols) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+        return ncols <= 16 ? 96 : 160;
+#else
+        return ncols <= 16 ? 288 : 160;
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+    }
+
+    static int get_nbatch_V2_host(const int cc, const int ncols) {
+        if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING) {
+            return ncols <= 16 ? 64 : 128;
+        }
+        return ncols <= 16 ? 256 : 128;
+    }
+
+    static constexpr __device__ int get_nbatch_V2_device(int ncols) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+        return ncols <= 16 ? 64 : 128;
+#else
+        return ncols <= 16 ? 256 : 128;
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+    }
+
+    static int get_nbatch_combine_host(const int /*cc*/, const int /*ncols*/) {
+        return 128;
+    }
+
+    static constexpr __device__ int get_nbatch_combine_device(int /*ncols*/) {
+        return 128;
+    }
 };
 
 // ------------------------------------------------------------------------------------------------------------------
@@ -120,7 +289,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
 
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
-        auto load = [&] __device__ (const int n) {
+        auto load = [&] __device__ (auto n) {
             const int stride_k = WARP_SIZE >> n;
             const int k0_start = stride_k == WARP_SIZE ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
             const int k0_stop  =                             chunks_per_row - chunks_per_row % (1*stride_k);
@@ -223,7 +392,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
     }
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool needs_fixup, bool is_fixup, bool last_iter>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup, bool last_iter>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -261,10 +430,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr int cols_per_warp   = ntiles * tile_B::I;
     constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles;
     constexpr int np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int ncols           = ncols1 * ncols2;
+    constexpr int nbatch_K2       = c::get_nbatch_K2_device(ncols);
+    constexpr int nbatch_V2       = c::get_nbatch_V2_device(ncols);
 
-    constexpr int stride_tile_Q = DKQ/2        + 4;
-    constexpr int stride_tile_K = c::nbatch_K2 + 4;
-    constexpr int stride_tile_V = c::nbatch_V2 + 4;
+    constexpr int stride_tile_Q = DKQ/2     + 4;
+    constexpr int stride_tile_K = nbatch_K2 + 4;
+
+    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
+    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
 
     const int k_VKQ_0 = kb0 * c::nbatch_fa;
     tile_C_KQ KQ_C[c::nbatch_fa/(np*tile_C_KQ::I) * ntiles];
@@ -275,12 +449,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     tile_C_KQ_16  * KQ_C_16  = (tile_C_KQ_16  *) KQ_C;
 
     if constexpr (nstages > 1) {
-        static_assert(c::nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading");
+        static_assert(!mla, "multi-stage loading not implemented for MLA");
+        static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading");
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
         __syncthreads();
         flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
-            (V_h2 + k_VKQ_0*stride_V, tile_V, c::nbatch_V2, stride_V);
+            (V_h2 + k_VKQ_0*stride_V, tile_V, nbatch_V2, stride_V);
     } else {
         constexpr bool use_cp_async = nstages == 1;
         if (ncols2 > 1 || mask_h2) {
@@ -289,8 +464,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 #pragma unroll
-    for (int k0_start = 0; k0_start < DKQ/2; k0_start += c::nbatch_K2) {
-        const int k0_stop = k0_start + c::nbatch_K2 < DKQ/2 ? k0_start + c::nbatch_K2 : DKQ/2;
+    for (int k0_start = 0; k0_start < DKQ/2; k0_start += nbatch_K2) {
+        const int k0_stop = k0_start + nbatch_K2 < DKQ/2 ? k0_start + nbatch_K2 : DKQ/2;
         const int k0_diff = k0_stop - k0_start;
 
         if (nstages <= 1) {
@@ -537,16 +712,21 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     (mask_h2 + (k_VKQ_0 + c::nbatch_fa)/2, tile_mask, stride_mask);
             }
             flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-                (K_h2 + (k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, c::nbatch_K2, stride_K);
+                (K_h2 + (k_VKQ_0 + c::nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K);
         }
     }
 
-#pragma unroll
-    for (int i0_start = 0; i0_start < DV; i0_start += 2*c::nbatch_V2) {
-        const int i0_stop = i0_start + 2*c::nbatch_V2 < DV ? i0_start + 2*c::nbatch_V2 : DV;
-        const int i0_diff = i0_stop - i0_start;
 
-        if (nstages <= 1) {
+    // For MLA K and V have the same data.
+    // Therefore, iterate over V in reverse and re-use the data if possible.
+    static_assert(!mla || nstages <= 1, "combination of MLA and multi-stage loading not implemented");
+    constexpr int reusable_cutoff = mla ? (DKQ - 1) - (DKQ - 1) % (2*nbatch_K2) - (DKQ - DV) : DV;
+#pragma unroll
+    for (int i0_stop = DV; i0_stop > 0; i0_stop -= 2*nbatch_V2) {
+        const int i0_start = i0_stop - 2*nbatch_V2 > 0 ? i0_stop - 2*nbatch_V2 : 0;
+        const int i0_diff  = i0_stop - i0_start;
+
+        if (nstages <= 1 && i0_start < reusable_cutoff) {
             constexpr bool use_cp_async = nstages == 1;
             flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, c::nbatch_fa, use_cp_async>
                 (V_h2 + k_VKQ_0*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V);
@@ -555,6 +735,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
             }
             __syncthreads();
         }
+        const half2 * tile_V_i = i0_start < reusable_cutoff ? tile_V : tile_V + (i0_start - reusable_cutoff)/2;
 
         // Calculate VKQ tile:
 #pragma unroll
@@ -565,7 +746,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*tile_A::J;
 
                 tile_A A;
-                load_ldmatrix_trans(A, tile_V + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
+                load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
                 if (ntiles == 1) {
                     mma(VKQ_C[i_VKQ_0/tile_C_VKQ::I], A, B[k00/(np*tile_A::J)]);
                 } else {
@@ -596,7 +777,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #endif // NEW_MMA_AVAILABLE
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool needs_fixup, bool is_fixup>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla, bool needs_fixup, bool is_fixup>
 static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const float2 * const __restrict__ Q_f2,
         const half2  * const __restrict__ K_h2,
@@ -632,13 +813,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     constexpr int cols_per_warp   = ntiles * tile_B::I;
     constexpr int cols_per_thread = ntiles == 1 ? 2 : ntiles;
     constexpr int np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
+    constexpr int nbatch_K2       = c::get_nbatch_K2_device(ncols);
+    constexpr int nbatch_V2       = c::get_nbatch_V2_device(ncols);
 
     static_assert(nwarps * (cols_per_warp/ncols2) % ncols1 == 0, "bad nwarps");
 
-    constexpr int stride_tile_Q = DKQ/2        + 4;
-    constexpr int stride_tile_K = c::nbatch_K2 + 4;
-    constexpr int stride_tile_V = c::nbatch_V2 + 4;
+    constexpr int stride_tile_Q = DKQ/2     + 4;
+    constexpr int stride_tile_K = nbatch_K2 + 4;
 
+    static_assert(!mla || nbatch_K2 >= nbatch_V2, "bad nbatch_K2, nbatch_V2 for MLA");
+    constexpr int stride_tile_V = mla ? stride_tile_K : nbatch_V2 + 4;
     constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V;
 
     extern __shared__ half2 tile_Q[];
@@ -726,26 +910,26 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
     // Preload mask and K data for first iteration when using cp_async with multiple stages:
     if constexpr (nstages > 1) {
-        static_assert(c::nbatch_K2 == DKQ/2, "batching not implemented for multi-stage pipeline");
+        static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi-stage pipeline");
         constexpr bool use_cp_async = true;
         if (ncols2 > 1 || mask_h2) {
             flash_attn_ext_f16_load_mask<ncols1, nwarps, c::nbatch_fa, use_cp_async>
                 (mask_h2 + kb0_start*c::nbatch_fa/2, tile_mask, stride_mask);
         }
         flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, c::nbatch_fa, use_cp_async>
-            (K_h2 + kb0_start*c::nbatch_fa*stride_K, tile_K, c::nbatch_K2, stride_K);
+            (K_h2 + kb0_start*c::nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K);
     }
 
     // Iterate over ne11 == previous tokens:
     for (int kb0 = kb0_start; kb0 < kb0_stop-1; ++kb0) {
         constexpr bool last_iter = false;
-        flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, needs_fixup, is_fixup, last_iter>
+        flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     }
     { // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
         constexpr bool last_iter = true;
-        flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, needs_fixup, is_fixup, last_iter>
+        flash_attn_ext_f16_iter<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup, last_iter>
             (Q_f2, K_h2, V_h2, mask_h2, dstk, dstk_fixup, scale, slope, logit_softcap,
              ne01, ne02, stride_K, stride_V, stride_mask, jt, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0_stop-1);
     }
@@ -774,7 +958,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     // It's also faster to do small writes to shared memory, then large write to VRAM than to do small writes to VRAM.
     // So also write VKQ accumulators to shared memory in column-major format if np == 1.
 
-    constexpr int nbatch_combine = c::Q_in_reg ? DV/2 : DV/4;
+    constexpr int nbatch_combine = c::get_nbatch_combine_device(ncols);
     constexpr int tile_stride    = nbatch_combine + 4;
     static_assert((DV/2) % nbatch_combine == 0, "bad nbatch_combine");
 
@@ -874,6 +1058,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             }
         }
 
+        __syncthreads();
+
         // Write back combined meta data:
 #pragma unroll
         for (int imeta = 0; imeta < nmeta; ++imeta) {
@@ -893,6 +1079,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             float2 * dstk_fixup_meta = dstk_fixup + (gridDim.x + blockIdx.x)*ncols;
             dstk_fixup_meta[(threadIdx.y/np)*cols_per_warp + threadIdx.x] = make_float2(KQ_cmn, KQ_crs);
         }
+    } else if (np > 1) {
+        // Warps with threadIdx.y % np == 0 execute a __syncthreads() in the if branch.
+        // Therefore, all other warps also need to execute a __syncthreads().
+        // Otherwise the points at which warps synchronize with each other would become misaligned.
+        __syncthreads();
     }
 
 #pragma unroll
@@ -1005,7 +1196,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 #endif // NEW_MMA_AVAILABLE
 }
 
-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, int ntiles, bool use_logit_softcap, bool mla>
 __launch_bounds__(nwarps*WARP_SIZE, 1)
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
@@ -1050,6 +1241,14 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
+#if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+    if (ncols1*ncols2 > 32) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+
+    static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 
     typedef fattn_mma_f16_config<DKQ, DV> c;
 
@@ -1060,9 +1259,10 @@ static __global__ void flash_attn_ext_f16(
     const int stride_Q1   = nb01 / sizeof(float2);
     const int stride_Q2   = nb02 / sizeof(float2);
     const int stride_K    = nb11 / sizeof(half2);
-    const int stride_V    = nb21 / sizeof(half2);
     const int stride_mask = nb31 / sizeof(half2);
 
+    const int stride_V = mla ? stride_K : nb21 / sizeof(half2);
+
     const int iter_k = ne11 / FATTN_KQ_STRIDE;
     const int iter_j = (ne01 + (ncols1 - 1)) / ncols1;
 
@@ -1085,10 +1285,11 @@ static __global__ void flash_attn_ext_f16(
 
         const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
         const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-        const half2  * V_h2    = (const half2  *) (V + nb22*(channel*ncols2 / gqa_ratio));
         const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
         float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
+        const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+
         const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
 
         const int kb0_start_kernel = kb0_start * kb_niter;
@@ -1097,12 +1298,12 @@ static __global__ void flash_attn_ext_f16(
         constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         if (kb0_start == 0) {
             constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         } else {
             constexpr bool needs_fixup = true; // CUDA block is working on the beginning of a tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, needs_fixup, is_fixup>
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
                 (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
                  ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
         }
@@ -1123,10 +1324,11 @@ static __global__ void flash_attn_ext_f16(
 
     const float2 * Q_f2    = (const float2 *) (Q + nb02* channel*ncols2);
     const half2  * K_h2    = (const half2  *) (K + nb12*(channel*ncols2 / gqa_ratio));
-    const half2  * V_h2    = (const half2  *) (V + nb22*(channel*ncols2 / gqa_ratio)); // K and V have same shape
     const half2  * mask_h2 = ncols2 > 1 || mask ? (const half2  *) mask + (nb31/sizeof(half2))*jt*ncols1 : nullptr;
     float2       * dstk    = ((float2 *) dst) + channel*(ncols2 * DV/2);
 
+    const half2 * V_h2 = mla ? K_h2 + (DKQ/2 - DV/2) : (const half2 *) (V + nb22*(channel*ncols2 / gqa_ratio));
+
     const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, channel, n_head_log2, m0, m1) : 1.0f;
 
     const int kb0_start_kernel = kb0_start * kb_niter;
@@ -1134,7 +1336,7 @@ static __global__ void flash_attn_ext_f16(
 
     constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     constexpr bool needs_fixup = false;
-    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, needs_fixup, is_fixup>
+    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla, needs_fixup, is_fixup>
         (Q_f2, K_h2, V_h2, mask_h2, dstk, dst_meta, scale, slope, logit_softcap,
          ne01, ne02, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, kb0_start_kernel, kb0_stop_kernel);
 #else
@@ -1160,10 +1362,6 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
 
     typedef fattn_mma_f16_config<DKQ, DV> c;
 
-    constexpr int nbatch_K2      = c::nbatch_K2      < 1 ? DKQ/2 : c::nbatch_K2;
-    constexpr int nbatch_V2      = c::nbatch_V2      < 1 ? DV /2 : c::nbatch_V2;
-    constexpr int nbatch_combine = c::nbatch_combine < 1 ? DV /2 : c::nbatch_combine;
-
     const int nstages = cp_async_available(cc) ? c::nstages_target : 0;
 
     constexpr int ncols         = ncols1 * ncols2;
@@ -1173,15 +1371,21 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     constexpr int nwarps_max_y  = c::nbatch_fa / tile_A::I;
     constexpr int nwarps        = nwarps_max_x*nwarps_max_y <= c::nwarps_max ? nwarps_max_x*nwarps_max_y : c::nwarps_max;
 
+    constexpr bool mla = DKQ == 576;
+
+    const int nbatch_K2      = c::get_nbatch_K2_host     (cc, ncols);
+    const int nbatch_V2      = c::get_nbatch_K2_host     (cc, ncols);
+    const int nbatch_combine = c::get_nbatch_combine_host(cc, ncols);
+
     static_assert(DKQ   % tile_B::J     == 0, "bad DKQ");
     static_assert(DV    % tile_A::J     == 0, "bad DV");
     static_assert(ncols % cols_per_warp == 0, "bad ncols");
 
-    const size_t nbytes_shared_KV_1stage = c::nbatch_fa         * std::max(c::nbatch_K2 + 4,  c::nbatch_V2 + 4) * sizeof(half2);
-    const size_t nbytes_shared_KV_2stage = c::nbatch_fa         *         (c::nbatch_K2 + 4 + c::nbatch_V2 + 4) * sizeof(half2);
-    const size_t nbytes_shared_Q         = ncols                * (DKQ/2 + 4)                                   * sizeof(half2);
-    const size_t nbytes_shared_mask      = ncols1               * (c::nbatch_fa/2 + 4)                          * sizeof(half2);
-    const size_t nbytes_shared_combine   = nwarps*cols_per_warp * (nbatch_combine + 4)                          * sizeof(half2);
+    const size_t nbytes_shared_KV_1stage = c::nbatch_fa         * std::max(nbatch_K2 + 4,  nbatch_V2 + 4) * sizeof(half2);
+    const size_t nbytes_shared_KV_2stage = c::nbatch_fa         *         (nbatch_K2 + 4 + nbatch_V2 + 4) * sizeof(half2);
+    const size_t nbytes_shared_Q         = ncols                * (DKQ/2 + 4)                             * sizeof(half2);
+    const size_t nbytes_shared_mask      = ncols1               * (c::nbatch_fa/2 + 4)                    * sizeof(half2);
+    const size_t nbytes_shared_combine   = nwarps*cols_per_warp * (nbatch_combine + 4)                    * sizeof(half2);
 
     const size_t nbytes_shared_KV = nstages <= 1 ? nbytes_shared_KV_1stage : nbytes_shared_KV_2stage;
 
@@ -1195,7 +1399,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1206,7 +1410,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap>;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, nwarps, ntiles, use_logit_softcap, mla>;
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};

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

@@ -168,6 +168,7 @@ static __global__ void flash_attn_vec_ext_f16(
     for (int j = 0; j < ncols; ++j) {
         KQ[j*D + tid] = -HALF_MAX_HALF;
     }
+    __syncthreads();
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};
 

ggml/src/ggml-cuda/fattn.cu

@@ -10,6 +10,7 @@
 
 template <int DKQ, int DV, int ncols2>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const ggml_tensor * Q = dst->src[0];
 
     if constexpr (ncols2 <= 8) {
@@ -24,7 +25,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
         return;
     }
 
-    if (Q->ne[1] <= 32/ncols2) {
+    if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || Q->ne[1] <= 32/ncols2) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
         return;
     }

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

@@ -1909,13 +1909,19 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
+    // If src0 is a temporary compute buffer it may have some padding that needs to be cleared for mul_mat_vec_q or mul_mat_q.
+    // But if src0 is also a view of another tensor then this cannot be done safely because it may overwrite valid tensor data.
+    // Therefore, in such cases use cuBLAS.
+    const bool bad_padding_clear = ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE
+        && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
+
     bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
-    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
     bool any_gpus_with_slow_fp16   = false;
@@ -3216,7 +3222,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 #endif // FLASH_ATTN_AVAILABLE
             if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
                 const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-                if (!new_mma_available(cc) || cc < GGML_CUDA_CC_AMPERE) {
+                if (!new_mma_available(cc)) {
                     return false;
                 }
                 const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];

ggml/src/ggml-cuda/mmq.cu

@@ -91,11 +91,11 @@ void ggml_cuda_mul_mat_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }
@@ -122,6 +122,7 @@ void ggml_cuda_mul_mat_q(
             const int64_t s13 = src1->nb[3] / ts_src1;
             quantize_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
                 ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+            CUDA_CHECK(cudaGetLastError());
         }
 
         const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
@@ -205,6 +206,7 @@ void ggml_cuda_mul_mat_q(
         const int64_t s13 = src1->nb[2] / ts_src1;
         quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
             ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+        CUDA_CHECK(cudaGetLastError());
     }
 
     const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));

ggml/src/ggml-cuda/mmvq.cu

@@ -515,11 +515,11 @@ void ggml_cuda_mul_mat_vec_q(
 
     // If src0 is a temporary compute buffer, clear any potential padding.
     if (ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
-        GGML_ASSERT(ggml_is_contiguously_allocated(src0));
-        GGML_ASSERT(!src0->view_src);
         const size_t size_data  = ggml_nbytes(src0);
         const size_t size_alloc = ggml_backend_buffer_get_alloc_size(src0->buffer, src0);
         if (size_alloc > size_data) {
+            GGML_ASSERT(ggml_is_contiguously_allocated(src0));
+            GGML_ASSERT(!src0->view_src);
             CUDA_CHECK(cudaMemsetAsync((char *) src0->data + size_data, 0, size_alloc - size_data, stream));
         }
     }

ggml/src/ggml-cuda/quantize.cu

@@ -56,13 +56,13 @@ static __global__ void quantize_mmq_q8_1(
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.y + threadIdx.x)*4;
 
     if (i0 >= ne0) {
         return;
     }
 
-    const int64_t i1 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
     const int64_t i2 = blockIdx.z % ne2;
     const int64_t i3 = blockIdx.z / ne2;
 
@@ -75,8 +75,8 @@ static __global__ void quantize_mmq_q8_1(
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
-    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.x*gridDim.y*blockDim.x/QK8_1); // first block of channel
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.x;                    // block index in channel
     const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
@@ -166,8 +166,9 @@ void quantize_mmq_q8_1_cuda(
     GGML_ASSERT(ne00 % 4 == 0);
     GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
-    const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
-    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
+    // ne1 tends to assume the highest values, therefore use it as the "x" dimension of the CUDA grid:
+    const int64_t block_num_y = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(ne1, block_num_y, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
     switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:

ggml/src/ggml-cuda/sum.cu

@@ -31,7 +31,7 @@ void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguously_allocated(src0));
 
     const float * src0_d = (const float *) src0->data;
     float * dst_d = (float *) dst->data;

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

@@ -207,6 +207,10 @@ typedef struct {
     float    attn_factor;
     float    beta_fast;
     float    beta_slow;
+    int32_t  sect_0;
+    int32_t  sect_1;
+    int32_t  sect_2;
+    int32_t  sect_3;
 } ggml_metal_kargs_rope;
 
 typedef struct {

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

@@ -332,6 +332,10 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16,
     GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,
+    GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32,
+    GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16,
+    GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32,
+    GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16,
     GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_F16,
@@ -411,6 +415,13 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,
@@ -1275,6 +1286,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F16,            mul_mm_id_iq4_xs_f16,            has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,                   rope_norm_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,                   rope_norm_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32,                  rope_multi_f32,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16,                  rope_multi_f16,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32,                 rope_vision_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16,                 rope_vision_f16,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,                   rope_neox_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,                   rope_neox_f16,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                      im2col_f16,                      true);
@@ -1354,6 +1369,13 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK192_HV128, flash_attn_ext_q8_0_hk192_hv128, has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256,        flash_attn_ext_q8_0_h256,        has_simdgroup_mm);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_HK576_HV512, flash_attn_ext_q8_0_hk576_hv512, has_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64,      flash_attn_ext_vec_f16_h64,      has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64,     flash_attn_ext_vec_bf16_h64,     has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64,     flash_attn_ext_vec_q4_0_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64,     flash_attn_ext_vec_q4_1_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64,     flash_attn_ext_vec_q5_0_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64,     flash_attn_ext_vec_q5_1_h64,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64,     flash_attn_ext_vec_q8_0_h64,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H96,      flash_attn_ext_vec_f16_h96,      has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H96,     flash_attn_ext_vec_bf16_h96,     has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H96,     flash_attn_ext_vec_q4_0_h96,     has_simdgroup_reduction);
@@ -1637,16 +1659,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_NORM:
             return has_simdgroup_reduction && (op->ne[0] % 4 == 0 && ggml_is_contiguous_1(op->src[0]));
         case GGML_OP_ROPE:
-            {
-                const int mode = ((const int32_t *) op->op_params)[2];
-                if (mode & GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                if (mode & GGML_ROPE_TYPE_VISION) {
-                    return false;
-                }
-                return true;
-            }
+            return true;
         case GGML_OP_IM2COL:
             return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
@@ -3826,6 +3839,7 @@ static bool ggml_metal_encode_node(
             } break;
         case GGML_OP_ROPE:
             {
+
                 // make sure we have one or more position id(ne10) per token(ne02)
                 GGML_ASSERT(ne10 % ne02 == 0);
                 GGML_ASSERT(ne10 >= ne02);
@@ -3852,20 +3866,42 @@ static bool ggml_metal_encode_node(
                 memcpy(&beta_fast,   (const int32_t *) dst->op_params +  9, sizeof(float));
                 memcpy(&beta_slow,   (const int32_t *) dst->op_params + 10, sizeof(float));
 
-                const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+                const bool is_neox   = mode & GGML_ROPE_TYPE_NEOX;
+                const bool is_mrope  = mode & GGML_ROPE_TYPE_MROPE;
+                const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+                // mrope
+                const int sect_0 = ((const int32_t *) dst->op_params)[11];
+                const int sect_1 = ((const int32_t *) dst->op_params)[12];
+                const int sect_2 = ((const int32_t *) dst->op_params)[13];
+                const int sect_3 = ((const int32_t *) dst->op_params)[14];
 
                 id<MTLComputePipelineState> pipeline = nil;
 
-                if (!is_neox) {
+                if (is_neox) {
                     switch (src0->type) {
-                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
-                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    };
+                } else if (is_mrope && !is_vision) {
+                    GGML_ASSERT(ne10*4 >= ne02); // need at least 4 pos per token
+                    switch (src0->type) {
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_MULTI_F16].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    };
+                } else if (is_vision) {
+                    GGML_ASSERT(ne10*4 >= ne02); // need at least 4 pos per token
+                    switch (src0->type) {
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_VISION_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_VISION_F16].pipeline; break;
                         default: GGML_ABORT("fatal error");
                     };
                 } else {
                     switch (src0->type) {
-                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
-                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
                         default: GGML_ABORT("fatal error");
                     };
                 }
@@ -3896,6 +3932,10 @@ static bool ggml_metal_encode_node(
                     /*.attn_factor =*/ attn_factor,
                     /*.beta_fast   =*/ beta_fast,
                     /*.beta_slow   =*/ beta_slow,
+                    /* sect_0      =*/ sect_0,
+                    /* sect_1      =*/ sect_1,
+                    /* sect_2      =*/ sect_2,
+                    /* sect_3      =*/ sect_3,
                 };
 
                 [encoder setComputePipelineState:pipeline];
@@ -4332,7 +4372,7 @@ static bool ggml_metal_encode_node(
                 // TODO: add vec kernels for (ne00%64 == 0) and maybe also for (ne00%32 == 0)
                 //       for now avoiding mainly to keep the number of templates/kernels a bit lower
                 //       these are now trivial to add after: https://github.com/ggml-org/llama.cpp/pull/12612
-                if (ne01 >= 4 || (ne00%128 != 0 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
+                if (ne01 >= 20 || (ne00%128 != 0 && ne00 != 64 && ne00 != 96 && ne00 != 192 && ne00 != 576)) {
                     switch (src1->type) {
                         case GGML_TYPE_F16:
                             {
@@ -4513,6 +4553,24 @@ static bool ggml_metal_encode_node(
                     use_vec_kernel = true;
 
                     switch (ne00) {
+                        case 64:
+                            {
+                                switch (src1->type) {
+                                    case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H64].pipeline; break;
+                                    case GGML_TYPE_BF16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H64].pipeline; break;
+                                    case GGML_TYPE_Q4_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H64].pipeline; break;
+                                    case GGML_TYPE_Q4_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H64].pipeline; break;
+                                    case GGML_TYPE_Q5_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H64].pipeline; break;
+                                    case GGML_TYPE_Q5_1: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H64].pipeline; break;
+                                    case GGML_TYPE_Q8_0: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H64].pipeline; break;
+                                    default:
+                                        {
+                                            GGML_LOG_ERROR("unsupported type: %d\n", src1->type);
+                                            GGML_LOG_ERROR("add template specialization for this type\n");
+                                            GGML_ABORT("add template specialization for this type");
+                                        }
+                                }
+                            } break;
                         case 96:
                             {
                                 switch (src1->type) {

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

@@ -2713,8 +2713,148 @@ kernel void kernel_rope_neox(
     }
 }
 
+template<typename T>
+kernel void kernel_rope_multi(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
+
+            // mrope theta calculations
+            // note: the rest is the same as kernel_rope_neox
+            const int sect_dims = args.sect_0 + args.sect_1 + args.sect_2 + args.sect_3;
+            const int sec_w01   = args.sect_0 + args.sect_1;               // end of section 1
+            const int sec_w012  = args.sect_0 + args.sect_1 + args.sect_2; // end of section 2
+            const int sector    = ic % sect_dims;
+
+            float theta_base;
+            if (sector < args.sect_0) {
+                theta_base = (float) pos[i2];
+            } else if (sector < sec_w01) {
+                theta_base = (float) pos[i2 + args.ne02];
+            } else if (sector < sec_w012) {
+                theta_base = (float) pos[i2 + args.ne02 * 2];
+            } else {
+                theta_base = (float) pos[i2 + args.ne02 * 3];
+            }
+            // end of mrope
+
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims/2];
+
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_vision(
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
+
+    device const int32_t * pos = (device const int32_t *) src1;
+
+    const float inv_ndims = -1.f/args.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < 2*args.n_dims) { // different from kernel_rope_multi
+            const int ic = i0/2;
+
+            // mrope theta calculations (only support 2 dimensions)
+            const int sect_dims = args.sect_0 + args.sect_1;
+            const int sector    = ic % sect_dims;
+
+            float p;
+            float theta_base;
+            if (sector < args.sect_1) {
+                p = (float) sector;
+                theta_base = (float) pos[i2];
+            } else {
+                p = (float) sector - args.sect_0;
+                theta_base = (float) pos[i2 + args.ne02];
+            }
+
+            const float theta = theta_base * pow(args.freq_base, 2.0f * inv_ndims * p);
+            // end of mrope
+
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[args.n_dims]; // different from kernel_rope_multi
+
+            dst_data[0]           = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims] = x0*sin_theta + x1*cos_theta; // different from kernel_rope_multi
+        } else {
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
 typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
 typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+typedef decltype(kernel_rope_multi<float>) kernel_rope_multi_t;
+typedef decltype(kernel_rope_vision<float>) kernel_rope_vision_t;
 
 template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
 template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
@@ -2722,6 +2862,12 @@ template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_
 template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
 template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
 
+template [[host_name("kernel_rope_multi_f32")]] kernel kernel_rope_multi_t kernel_rope_multi<float>;
+template [[host_name("kernel_rope_multi_f16")]] kernel kernel_rope_multi_t kernel_rope_multi<half>;
+
+template [[host_name("kernel_rope_vision_f32")]] kernel kernel_rope_vision_t kernel_rope_vision<float>;
+template [[host_name("kernel_rope_vision_f16")]] kernel kernel_rope_vision_t kernel_rope_vision<half>;
+
 typedef void (im2col_t)(
         device const float * x,
         device        char * dst,
@@ -3741,6 +3887,11 @@ kernel void kernel_flash_attn_ext_vec(
                 sm[tiisg] = pm[ic + tiisg];
             }
 
+            // skip -INF blocks
+            if (simd_max(sm[tiisg]) == -INFINITY) {
+                continue;
+            }
+
             // Q*K^T
             {
                 // each simdgroup processes 1 query and NE (NW/NL) head elements
@@ -3973,6 +4124,16 @@ kernel void kernel_flash_attn_ext_vec(
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 
+template [[host_name("kernel_flash_attn_ext_vec_f16_h64")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  64, 64, 8>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_flash_attn_ext_vec_bf16_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 64, 64, 8>;
+#endif
+template [[host_name("kernel_flash_attn_ext_vec_q4_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0,        8, dequantize_q4_0_t4, block_q4_0,  8, dequantize_q4_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q4_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_1,        8, dequantize_q4_1_t4, block_q4_1,  8, dequantize_q4_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_0,        8, dequantize_q5_0_t4, block_q5_0,  8, dequantize_q5_0_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q5_1_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q5_1,        8, dequantize_q5_1_t4, block_q5_1,  8, dequantize_q5_1_t4, 64, 64, 8>;
+template [[host_name("kernel_flash_attn_ext_vec_q8_0_h64")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0,        8, dequantize_q8_0_t4, block_q8_0,  8, dequantize_q8_0_t4, 64, 64, 8>;
+
 template [[host_name("kernel_flash_attn_ext_vec_f16_h96")]]  kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4,             1, dequantize_f16_t4,  half4,       1, dequantize_f16_t4,  96, 96, 4>;
 #if defined(GGML_METAL_USE_BF16)
 template [[host_name("kernel_flash_attn_ext_vec_bf16_h96")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4,           1, dequantize_bf16_t4, bfloat4,     1, dequantize_bf16_t4, 96, 96, 4>;

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -4855,8 +4855,6 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             if (!any_on_device) {
                 return false;
             }
-            GGML_ASSERT(ggml_is_contiguous(src0));
-            GGML_ASSERT(ggml_is_contiguous(src1));
             func = ggml_cl_add;
             break;
         case GGML_OP_MUL:

ggml/src/ggml-opt.cpp

@@ -28,16 +28,19 @@ struct ggml_opt_dataset {
 };
 
 struct ggml_opt_context {
-    ggml_backend_sched_t    backend_sched        = nullptr;
-    ggml_cgraph           * allocated_graph      = nullptr;
-    ggml_cgraph           * allocated_graph_copy = nullptr;
-    struct ggml_context   * ctx_static           = nullptr;
-    struct ggml_context   * ctx_static_cpu       = nullptr;
-    struct ggml_context   * ctx_compute          = nullptr;
-    struct ggml_context   * ctx_copy             = nullptr;
-    ggml_backend_buffer_t   buf_static           = nullptr;
-    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
-    std::mt19937            rng;
+    ggml_backend_sched_t       backend_sched        = nullptr;
+    ggml_cgraph              * allocated_graph      = nullptr;
+    ggml_cgraph              * allocated_graph_copy = nullptr;
+    struct ggml_context      * ctx_static           = nullptr;
+    struct ggml_context      * ctx_cpu              = nullptr;
+    struct ggml_context      * ctx_compute          = nullptr;
+    struct ggml_context      * ctx_copy             = nullptr;
+    ggml_backend_buffer_t      buf_static           = nullptr;
+    ggml_backend_buffer_t      buf_cpu              = nullptr;
+    std::mt19937               rng;
+    enum ggml_opt_loss_type    loss_type;
+    enum ggml_opt_build_type   build_type;
+    enum ggml_opt_build_type   build_type_alloc;
 
     struct ggml_tensor * inputs  = nullptr;
     struct ggml_tensor * outputs = nullptr;
@@ -50,6 +53,11 @@ struct ggml_opt_context {
     struct ggml_cgraph * gf      = nullptr;
     struct ggml_cgraph * gb_grad = nullptr;
     struct ggml_cgraph * gb_opt  = nullptr;
+    bool static_graphs           = false;
+    bool eval_ready              = false;
+    std::vector<struct ggml_tensor *> grad_accs;
+    std::vector<struct ggml_tensor *> grad_m;
+    std::vector<struct ggml_tensor *> grad_v;
 
     int64_t iter               = 1;
     int32_t opt_period         = 1;
@@ -73,7 +81,13 @@ struct ggml_opt_result {
 
 // ====== Dataset ======
 
-ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+ggml_opt_dataset_t ggml_opt_dataset_init(
+        enum ggml_type type_data,
+        enum ggml_type type_label,
+        int64_t        ne_datapoint,
+        int64_t        ne_label,
+        int64_t        ndata,
+        int64_t        ndata_shard) {
     GGML_ASSERT(ne_datapoint >  0);
     GGML_ASSERT(ne_label     >= 0);
     GGML_ASSERT(ndata        >  0);
@@ -92,11 +106,11 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label,
         result->ctx = ggml_init(params);
     }
 
-    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->data = ggml_new_tensor_2d(result->ctx, type_data, ne_datapoint, ndata);
     result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
 
     if (ne_label > 0) {
-        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->labels = ggml_new_tensor_2d(result->ctx, type_label, ne_label, ndata);
         result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
     } else {
         result->labels = nullptr;
@@ -119,6 +133,10 @@ void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
     delete dataset;
 }
 
+int64_t ggml_opt_dataset_ndata(ggml_opt_dataset_t dataset) {
+    return dataset->ndata;
+}
+
 struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
     return dataset->data;
 }
@@ -144,6 +162,8 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
     GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
     GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(                   data_batch->type == dataset->data->type);
+    GGML_ASSERT(!labels_batch || labels_batch->type == dataset->labels->type);
 
     const size_t nb_data_batch = ggml_nbytes(data_batch);
     GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
@@ -171,6 +191,31 @@ void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor *
     }
 }
 
+void ggml_opt_dataset_get_batch_host(ggml_opt_dataset_t dataset, void * data_batch, size_t nb_data_batch, void * labels_batch, int64_t ibatch) {
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data       = (const char *) dataset->data->data + ishard      *dataset->nbs_data;
+        char       * ptr_data_batch = (char       *) data_batch          + ishard_batch*dataset->nbs_data;
+        memcpy(ptr_data_batch, ptr_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels       = (const char *) dataset->labels->data + ishard      *dataset->nbs_labels;
+        char       * ptr_labels_batch = (char       *) labels_batch          + ishard_batch*dataset->nbs_labels;
+        memcpy(ptr_labels_batch, ptr_labels, dataset->nbs_labels);
+    }
+}
+
 // ====== Model / Context ======
 
 struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
@@ -187,17 +232,18 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     return result;
 }
 
+struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
+    return *((struct ggml_opt_optimizer_params *) userdata);
+}
+
 struct ggml_opt_params ggml_opt_default_params(
         ggml_backend_sched_t      backend_sched,
-        struct ggml_context     * ctx_compute,
-        struct ggml_tensor      * inputs,
-        struct ggml_tensor      * outputs,
         enum ggml_opt_loss_type   loss_type) {
     return {
         /*backend_sched   =*/ backend_sched,
-        /*ctx_compute     =*/ ctx_compute,
-        /*inputs          =*/ inputs,
-        /*logits          =*/ outputs,
+        /*ctx_compute     =*/ nullptr,
+        /*inputs          =*/ nullptr,
+        /*logits          =*/ nullptr,
         /*loss_type       =*/ loss_type,
         /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
         /*opt_period      =*/ 1,
@@ -266,195 +312,246 @@ static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
     return dst;
 }
 
-static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
-    GGML_ASSERT(graph);
-    if (opt_ctx->allocated_graph == graph) {
-        return;
-    }
+static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
+    GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
+    GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
-    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+    const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
+        !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
-    {
-        ggml_init_params params = {
-            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        ggml_free(opt_ctx->ctx_copy);
-        opt_ctx->ctx_copy = ggml_init(params);
-    }
-
-    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
-
-    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
-    opt_ctx->allocated_graph = graph;
-}
-
-ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
-    ggml_opt_context_t result = new struct ggml_opt_context;
-    result->backend_sched   = params.backend_sched;
-    result->ctx_compute     = params.ctx_compute;
-    result->inputs          = params.inputs;
-    result->outputs         = params.outputs;
-    result->opt_period      = params.opt_period;
-    result->get_opt_pars    = params.get_opt_pars;
-    result->get_opt_pars_ud = params.get_opt_pars_ud;
-
-    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
-    GGML_ASSERT(result->opt_period >= 1);
-
-    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
-        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
-
-    ggml_set_input(result->inputs);
-    ggml_set_output(result->outputs);
-
-    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
-    ggml_build_forward_expand(result->gf, result->outputs);
+    ggml_set_input(opt_ctx->inputs);
+    ggml_set_output(opt_ctx->outputs);
 
     int n_param = 0;
-    for (int i = 0; i < result->gf->n_nodes; ++i) {
-        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+    for (int i = 0; i < opt_ctx->gf->n_nodes; ++i) {
+        const struct ggml_tensor * node = opt_ctx->gf->nodes[i];
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             n_param++;
         }
+        GGML_ASSERT(!(node->flags & GGML_TENSOR_FLAG_LOSS) && "support for extra loss terms not implemented");
     }
 
-    {
+    if (!opt_ctx->ctx_static) {
         // The static context is used for:
-        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - gradients (1 per loss, 1 tensor per param if using gradient accumulation)
         //   - optimizer momenta (2 tensors per param)
-        //   - labels
-        //   - loss + its gradient (up to 5 tensors)
-        //   - pred
-        //   - ncorrect (2 tensors).
-        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
-        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        //   - labels (if using static graphs)
+        //   - loss (if using static graphs, up to 5 tensors)
+        //   - pred (if using static graphs)
+        //   - ncorrect (if using static graphs, 2 tensors).
+        constexpr size_t n_loss = 1;
+        const size_t tensors_per_param = (accumulate ? 1 : 0) +
+            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
+        const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static = ggml_init(params);
+        opt_ctx->ctx_static = ggml_init(params);
     }
+    GGML_ASSERT(opt_ctx->build_type <= opt_ctx->build_type_alloc);
+
     {
-        // The static cpu context is used for:
-        //   - optimizer parameters (1 for the entire context)
+        // The cpu context is allocated statically if using static graphs, dynamically otherwise.
+        // It is used for:
+        //   - optimizer parameters (1 shared for all optimizer invocations)
         const size_t size_meta = 1 * ggml_tensor_overhead();
         struct ggml_init_params params = {
             /*.mem_size   =*/ size_meta,
             /*.mem_buffer =*/ nullptr,
             /*.no_alloc   =*/ true,
         };
-        result->ctx_static_cpu = ggml_init(params);
+        ggml_free(opt_ctx->ctx_cpu);
+        opt_ctx->ctx_cpu = ggml_init(params);
+
+        ggml_backend_buffer_free(opt_ctx->buf_cpu);
+        opt_ctx->buf_cpu = nullptr;
     }
 
+    struct ggml_context * ctx_results = opt_ctx->static_graphs ? opt_ctx->ctx_static : opt_ctx->ctx_compute;
 
-    switch (params.loss_type) {
+    switch (opt_ctx->loss_type) {
         case GGML_OPT_LOSS_TYPE_MEAN: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean");
-            result->loss_per_datapoint = true;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_SUM: {
-            result->loss = ggml_sum(result->ctx_static, result->outputs);
-            ggml_set_name(result->loss, "loss_sum");
-            result->loss_per_datapoint = false;
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->outputs);
+            ggml_set_name(opt_ctx->loss, "loss_sum");
+            opt_ctx->loss_per_datapoint = false;
             break;
         }
         case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_cross_entropy");
-            if (result->opt_period > 1) {
-                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
-                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_cross_entropy_loss(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_cross_entropy");
+            if (opt_ctx->opt_period > 1) {
+                opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, 1.0f / opt_ctx->opt_period);
+                ggml_set_name(opt_ctx->loss, "loss_cross_entropy_scaled");
             }
-            result->loss_per_datapoint = true;
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
         case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
-            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
-            ggml_set_input(result->labels);
-            ggml_set_name(result->labels, "labels");
-            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
-            ggml_set_name(result->loss, "loss_error");
-            result->loss = ggml_sqr(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_squared_error");
-            result->loss = ggml_sum(result->ctx_static, result->loss);
-            ggml_set_name(result->loss, "loss_sum_squared_error");
-            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
-            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
-            ggml_set_name(result->loss, "loss_mean_squared_error");
-            result->loss_per_datapoint = true;
+            opt_ctx->labels = ggml_dup_tensor(ctx_results, opt_ctx->outputs);
+            ggml_set_input(opt_ctx->labels);
+            ggml_set_name(opt_ctx->labels, "labels");
+            opt_ctx->loss = ggml_sub(ctx_results, opt_ctx->outputs, opt_ctx->labels);
+            ggml_set_name(opt_ctx->loss, "loss_error");
+            opt_ctx->loss = ggml_sqr(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_squared_error");
+            opt_ctx->loss = ggml_sum(ctx_results, opt_ctx->loss);
+            ggml_set_name(opt_ctx->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (opt_ctx->opt_period * ggml_nelements(opt_ctx->outputs));
+            opt_ctx->loss = ggml_scale(ctx_results, opt_ctx->loss, scale);
+            ggml_set_name(opt_ctx->loss, "loss_mean_squared_error");
+            opt_ctx->loss_per_datapoint = true;
             break;
         }
     }
-    ggml_set_output(result->loss);
-    ggml_set_loss(result->loss);
-    ggml_build_forward_expand(result->gf, result->loss);
+    ggml_set_output(opt_ctx->loss);
+    ggml_set_loss(opt_ctx->loss);
+    ggml_build_forward_expand(opt_ctx->gf, opt_ctx->loss);
 
-    result->pred = ggml_argmax(result->ctx_static, result->outputs);
-    ggml_set_name(result->pred, "pred");
-    ggml_set_output(result->pred);
-    ggml_build_forward_expand(result->gf, result->pred);
+    if (opt_ctx->loss_type == GGML_OPT_LOSS_TYPE_CROSS_ENTROPY) {
+        opt_ctx->pred = ggml_argmax(ctx_results, opt_ctx->outputs);
+        ggml_set_name(opt_ctx->pred, "pred");
+        ggml_set_output(opt_ctx->pred);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->pred);
 
-    if (result->labels) {
-        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
-        ggml_set_name(result->ncorrect, "ncorrect");
-        ggml_set_output(result->ncorrect);
-        ggml_build_forward_expand(result->gf, result->ncorrect);
-    } else {
-        result->ncorrect = nullptr;
+        opt_ctx->ncorrect = ggml_count_equal(ctx_results, opt_ctx->pred, ggml_argmax(ctx_results, opt_ctx->labels));
+        ggml_set_name(opt_ctx->ncorrect, "ncorrect");
+        ggml_set_output(opt_ctx->ncorrect);
+        ggml_build_forward_expand(opt_ctx->gf, opt_ctx->ncorrect);
     }
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        return result;
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_FORWARD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        return;
     }
 
-    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
-    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
-    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+    if (opt_ctx->grad_accs.empty()) {
+        GGML_ASSERT(opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD);
 
-    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
-        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
-        ggml_graph_reset(result->gb_grad);
-        return result;
-    }
+        const int n_nodes = opt_ctx->gf->n_nodes;
+        opt_ctx->grad_accs.resize(n_nodes);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor * node = opt_ctx->gf->nodes[i];
+            if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
+                opt_ctx->grad_accs[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            } else {
+                opt_ctx->grad_accs[i] = nullptr;
+            }
+        }
 
-    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
-
-    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
-    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
-
-    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(result->adamw_params);
-    ggml_set_name(result->adamw_params, "adamw_params");
-
-    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
-        struct ggml_tensor * node = result->gb_opt->nodes[i];
-        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
-            ggml_build_forward_expand(result->gb_opt, opt_step);
+        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+            opt_ctx->grad_m.resize(n_nodes);
+            opt_ctx->grad_v.resize(n_nodes);
+            for (int i = 0; i < n_nodes; ++i) {
+                ggml_tensor * node = opt_ctx->gf->nodes[i];
+                if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+                    opt_ctx->grad_m[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                    opt_ctx->grad_v[i] = ggml_new_tensor(opt_ctx->ctx_static, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+                } else {
+                    opt_ctx->grad_m[i] = nullptr;
+                    opt_ctx->grad_v[i] = nullptr;
+                }
+            }
         }
     }
 
-    result->buf_static = ggml_backend_alloc_ctx_tensors(
-        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    opt_ctx->gb_grad = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gf, /*force_grads =*/ true);
+    ggml_build_backward_expand(opt_ctx->ctx_compute, opt_ctx->gb_grad, opt_ctx->grad_accs.data());
 
-    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+    if (opt_ctx->buf_static) {
+        if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+            return;
+        }
+    } else if (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_GRAD) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
 
-    ggml_graph_reset(result->gb_opt);
+    GGML_ASSERT(opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
+
+    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(opt_ctx->adamw_params);
+    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
+
+    for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
+
+        if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
+            struct ggml_tensor * m        = opt_ctx->grad_m[i];
+            struct ggml_tensor * v        = opt_ctx->grad_v[i];
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
+
+            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
+            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
+            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
+
+            ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
+        }
+    }
+
+    if (!opt_ctx->buf_static) {
+        opt_ctx->buf_static = ggml_backend_alloc_ctx_tensors(
+            opt_ctx->ctx_static, ggml_backend_sched_get_backend(opt_ctx->backend_sched, 0));
+        ggml_graph_reset(opt_ctx->gb_opt);
+    }
+
+    opt_ctx->buf_cpu = ggml_backend_alloc_ctx_tensors_from_buft(opt_ctx->ctx_cpu, ggml_backend_cpu_buffer_type());
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched    = params.backend_sched;
+    result->ctx_compute      = params.ctx_compute;
+    result->loss_type        = params.loss_type;
+    result->build_type       = params.build_type;
+    result->build_type_alloc = params.build_type;
+    result->inputs           = params.inputs;
+    result->outputs          = params.outputs;
+    result->opt_period       = params.opt_period;
+    result->get_opt_pars     = params.get_opt_pars;
+    result->get_opt_pars_ud  = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->opt_period >= 1);
+
+    result->static_graphs = result->ctx_compute;
+
+    if (!result->static_graphs) {
+        GGML_ASSERT(!result->inputs);
+        GGML_ASSERT(!result->outputs);
+        return result;
+    }
+
+    GGML_ASSERT(result->inputs);
+    GGML_ASSERT(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    ggml_opt_build(result);
 
     return result;
 }
@@ -464,9 +561,9 @@ void ggml_opt_free(ggml_opt_context_t opt_ctx) {
         return;
     }
     ggml_backend_buffer_free(opt_ctx->buf_static);
-    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_backend_buffer_free(opt_ctx->buf_cpu);
     ggml_free(opt_ctx->ctx_static);
-    ggml_free(opt_ctx->ctx_static_cpu);
+    ggml_free(opt_ctx->ctx_cpu);
     delete opt_ctx;
 }
 
@@ -582,8 +679,79 @@ void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, doubl
 
 // ====== Computation ======
 
-static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
-    if (graph != opt_ctx->gf) {
+void ggml_opt_prepare_alloc(
+        ggml_opt_context_t    opt_ctx,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * gf,
+        struct ggml_tensor  * inputs,
+        struct ggml_tensor  * outputs) {
+    GGML_ASSERT(!opt_ctx->static_graphs);
+    opt_ctx->ctx_compute = ctx_compute;
+    opt_ctx->gf          = gf;
+    opt_ctx->inputs      = inputs;
+    opt_ctx->outputs     = outputs;
+}
+
+void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
+    GGML_ASSERT(!opt_ctx->eval_ready);
+    if (opt_ctx->build_type == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period > 1 && opt_ctx->opt_i == 0) {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+    if (backward) {
+        const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+        opt_ctx->build_type = opt_i_next == 0 ? GGML_OPT_BUILD_TYPE_OPT : GGML_OPT_BUILD_TYPE_GRAD;
+    } else {
+        opt_ctx->build_type = GGML_OPT_BUILD_TYPE_FORWARD;
+    }
+
+    if (!opt_ctx->static_graphs) {
+        ggml_opt_build(opt_ctx);
+    }
+
+    struct ggml_cgraph * graph = nullptr;
+    switch (opt_ctx->build_type) {
+        case GGML_OPT_BUILD_TYPE_FORWARD: {
+            graph = opt_ctx->gf;
+        } break;
+        case GGML_OPT_BUILD_TYPE_GRAD: {
+            graph = opt_ctx->gb_grad;
+        } break;
+        case GGML_OPT_BUILD_TYPE_OPT: {
+            graph = opt_ctx->gb_opt;
+        } break;
+    }
+    GGML_ASSERT(graph);
+
+    if (opt_ctx->allocated_graph == graph) {
+        opt_ctx->eval_ready = true;
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    if (opt_ctx->static_graphs) {
+        ggml_init_params params = {
+            /*.mem_size   =*/ graph->size*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph->size, graph->grads),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+
+        opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+    } else {
+        opt_ctx->allocated_graph_copy = graph;
+    }
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+
+    opt_ctx->eval_ready = true;
+}
+
+void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
+    GGML_ASSERT(opt_ctx->eval_ready);
+    if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
         struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
 
         GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
@@ -609,9 +777,19 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
         adamw_par_data[6] = beta2h;
     }
 
-    ggml_opt_alloc_graph(opt_ctx, graph);
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
     opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+    opt_ctx->opt_i = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+
+    if (!opt_ctx->static_graphs) {
+        opt_ctx->gf                   = nullptr;
+        opt_ctx->gb_grad              = nullptr;
+        opt_ctx->gb_opt               = nullptr;
+        opt_ctx->allocated_graph      = nullptr;
+        opt_ctx->allocated_graph_copy = nullptr;
+    }
+
+    opt_ctx->eval_ready = false;
 
     if (!result) {
         return;
@@ -635,12 +813,14 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
     result->loss.push_back(loss);
 
-    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
-    std::vector<int32_t> pred(ndata);
-    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
-    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    if (opt_ctx->pred) {
+        GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+        std::vector<int32_t> pred(ndata);
+        ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+        result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+    }
 
-    if (!opt_ctx->labels || result->ncorrect < 0) {
+    if (!opt_ctx->ncorrect || result->ncorrect < 0) {
         result->ncorrect = -1;
         return;
     }
@@ -652,26 +832,6 @@ static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph,
     result->ncorrect += ncorrect;
 }
 
-void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
-}
-
-void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
-    if (opt_ctx->opt_period == 1) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        return;
-    }
-
-    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
-    if (opt_i_next == 0) {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
-        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
-    } else {
-        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
-    }
-    opt_ctx->opt_i = opt_i_next;
-}
-
 // ====== High-Level Functions ======
 
 void ggml_opt_epoch(
@@ -700,16 +860,18 @@ void ggml_opt_epoch(
     int64_t ibatch = 0;
     int64_t t_loop_start = ggml_time_us();
     for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ true);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward_backward(opt_ctx, result_train);
+        ggml_opt_eval(opt_ctx, result_train);
         if (callback_train) {
             callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
         }
     }
     t_loop_start = ggml_time_us();
     for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_alloc(opt_ctx, /*backward =*/ false);
         ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
-        ggml_opt_forward(opt_ctx, result_eval);
+        ggml_opt_eval(opt_ctx, result_eval);
         if (callback_eval) {
             callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
         }
@@ -726,13 +888,26 @@ void ggml_opt_epoch_callback_progress_bar(
         int64_t            t_start_us) {
     fprintf(stderr, "%s[", train ? "train: " : "val:   ");
 
-    constexpr int64_t bar_length = 25;
+    // The progress bar consists of partially filled blocks, unicode has 8 separate fill levels.
+    constexpr int64_t bar_length = 8;
+    const int64_t ibatch8 = 8 * ibatch;
     for (int64_t j = 0; j < bar_length; ++j) {
-        const int64_t ibatch_j = ibatch_max * j/bar_length;
-        if (ibatch_j < ibatch) {
-            fprintf(stderr, "=");
-        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
-            fprintf(stderr, ">");
+        if        (ibatch_max * (8*j + 8) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2588"); // full block
+        } else if (ibatch_max * (8*j + 7) / bar_length < ibatch8) {
+            fprintf(stderr, "\u2589"); // 7/8 filled
+        } else if (ibatch_max * (8*j + 6) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258A"); // 6/8 filled
+        } else if (ibatch_max * (8*j + 5) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258B"); // 5/8 filled
+        } else if (ibatch_max * (8*j + 4) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258C"); // 4/8 filled
+        } else if (ibatch_max * (8*j + 3) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258D"); // 3/8 filled
+        } else if (ibatch_max * (8*j + 2) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258E"); // 2/8 filled
+        } else if (ibatch_max * (8*j + 1) / bar_length < ibatch8) {
+            fprintf(stderr, "\u258F"); // 1/8 filled
         } else {
             fprintf(stderr, " ");
         }
@@ -764,8 +939,8 @@ void ggml_opt_epoch_callback_progress_bar(
     const int64_t t_eta_m = t_eta_s / 60;
     t_eta_s -= t_eta_m * 60;
 
-    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
-            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+    fprintf(stderr, "] data=%07" PRId64 "/%07" PRId64 " loss=%.5lf±%.5lf acc=%.2lf±%.2lf%% "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 " \r",
             idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
             t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
     if (ibatch == ibatch_max) {
@@ -806,7 +981,10 @@ void ggml_opt_fit(
 
     int64_t epoch = 1;
 
-    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, loss_type);
+    params.ctx_compute     = ctx_compute;
+    params.inputs          = inputs;
+    params.outputs         = outputs;
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;

ggml/src/ggml-sycl/CMakeLists.txt

@@ -49,35 +49,38 @@ endif()
 target_compile_options(ggml-sycl PRIVATE "-Wno-narrowing")
 
 # Link against oneDNN
-find_package(DNNL)
 set(GGML_SYCL_DNNL 0)
-if(DNNL_FOUND)
-    if (DEFINED ENV{ONEAPI_ROOT} AND NOT DEFINED DNNL_GPU_VENDOR)
-        # Assuming oneDNN packaged with oneapi release is used which
-        # supports only intel target
-        set(DNNL_GPU_VENDOR "INTEL")
-        if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
-            message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+if(GGML_SYCL_DNN)
+    find_package(DNNL)
+    if(DNNL_FOUND)
+        if (NOT DEFINED DNNL_GPU_VENDOR)
+            # default to intel target
+            set(DNNL_GPU_VENDOR "INTEL")
+            if(NOT "${GGML_SYCL_TARGET}" STREQUAL "INTEL")
+                message(WARNING "oneDNN builds bundled with oneapi release only support INTEL target")
+            endif()
         endif()
-    endif()
 
-    # Verify oneDNN was compiled for the same target as llama
-    if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
-        target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
-        set(GGML_SYCL_DNNL 1)
-        get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
-        foreach(CONFIG ${CONFIGS})
-            get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
-            message(STATUS "Found oneDNN: ${DNNL_LIB}")
-        endforeach()
+        # Verify oneDNN was compiled for the same target as llama
+        if("${GGML_SYCL_TARGET}" STREQUAL "${DNNL_GPU_VENDOR}")
+            target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+            set(GGML_SYCL_DNNL 1)
+            get_target_property(CONFIGS DNNL::dnnl IMPORTED_CONFIGURATIONS)
+            foreach(CONFIG ${CONFIGS})
+                get_target_property(DNNL_LIB DNNL::dnnl IMPORTED_LOCATION_${CONFIG})
+                message(STATUS "Found oneDNN: ${DNNL_LIB}")
+            endforeach()
+        else()
+            message(WARNING
+                "oneDNN must be compiled for the same target as llama.cpp.
+                 llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
+                 Disabling oneDNN support.")
+        endif()
     else()
-        message(WARNING
-            "oneDNN must be compiled for the same target as llama.cpp.
-             llama.cpp: ${GGML_SYCL_TARGET}, oneDNN: ${DNNL_GPU_VENDOR}.
-             Disabling oneDNN support.")
+        message(STATUS "oneDNN not found, disabling oneDNN support")
     endif()
 else()
-    message(STATUS "oneDNN not found, disabling oneDNN support")
+    message(STATUS "oneDNN support disabled by the user")
 endif()
 target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_DNNL=${GGML_SYCL_DNNL})
 
@@ -108,6 +111,9 @@ endif()
 if (GGML_SYCL_TARGET STREQUAL "INTEL")
     # Intel devices use Intel oneMKL directly instead of oneMath to avoid the limitation of linking Intel oneMKL statically
     # See https://github.com/uxlfoundation/oneMath/issues/654
+    if (CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+        set(SYCL_COMPILER ON)
+    endif()
     find_package(MKL REQUIRED)
     target_link_libraries(ggml-sycl PRIVATE MKL::MKL_SYCL::BLAS)
     target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_USE_INTEL_ONEMKL)

ggml/src/ggml-sycl/binbcast.cpp

@@ -1,93 +1,74 @@
 #include "binbcast.hpp"
 
+#include <array>
 #include <cstddef>
 #include <cstdint>
 #include <sycl/sycl.hpp>
 
+#include "dpct/helper.hpp"
 #include "ggml.h"
 
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
-    const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
-    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1));
-    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) /
-                   ne3;
-    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) %
-                   ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0;
-         i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __dpct_inline__ void k_bin_bcast_contiguous(const src0_t * __restrict__ src0, const src1_t * __restrict__ src1,
+                                                   dst_t * dst, std::size_t num_elements, const sycl::nd_item<1> & it) {
+    auto element_id   = it.get_global_id(0);
+    auto global_range = it.get_global_range(0);
+    for (; element_id < num_elements; element_id += global_range) {
+        auto  src0_float_val = sycl::vec(src0[element_id]).template convert<float, sycl::rounding_mode::rte>();
+        auto  src1_float_val = sycl::vec(src1[element_id]).template convert<float, sycl::rounding_mode::rte>();
+        float dst_val        = bin_op(src0_float_val[0], src1_float_val[0]);
+        auto  val_to_store   = sycl::vec(dst_val).template convert<dst_t, sycl::rounding_mode::rte>();
+        dst[element_id]      = val_to_store;
     }
 }
 
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
+template <float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __dpct_inline__ void k_bin_bcast(const src0_t * __restrict__ src0, const src1_t * __restrict__ src1, dst_t * dst,
+                                        int ne0, int ne1, int ne2, int ne3, int ne10, int ne11, int ne12, int ne13,
+                                        int s0, int s1, int s2, int s3, int s00, int s01, int s02, int s03, int s10,
+                                        int s11, int s12, int s13, std::size_t num_dst_elements,
+                                        const sycl::nd_item<1> & item_ct1) {
+    auto calculate_logical_index =
+        [](const std::array<int, 4> & dims, std::size_t element_id) __attribute__((always_inline))->std::array<int, 4> {
+        std::array<int, 4> logical_index;
+#pragma unroll(4)
+        for (int i = 3; i >= 0; i--) {
+            logical_index[i] = element_id % dims[i];
+            element_id /= dims[i];
+        }
+        return logical_index;
+    };
 
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
+    auto calculate_index = [](const std::array<int, 4> & dims, const std::array<int, 4> & strides,
+                              const std::array<int, 4> & indices) __attribute__((always_inline))
+                               ->std::size_t {
+        std::size_t index = 0;
+#pragma unroll(4)
+        for (int i = 0; i < 4; i++) {
+            auto index_i = indices[i];
+            if (indices[i] >= dims[i]) {
+                index_i = indices[i] % dims[i];
+            }
+            index += strides[i] * index_i;
+        }
+        return index;
+    };
 
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
+    auto element_id = item_ct1.get_global_id(0);
+    for (; element_id < num_dst_elements; element_id += item_ct1.get_global_range(0)) {
+        auto  logical_index  = calculate_logical_index({ ne3, ne2, ne1, ne0 }, element_id);
+        auto  src_0_index    = calculate_index({ ne3, ne2, ne1, ne0 }, { s03, s02, s01, s00 }, logical_index);
+        auto  src_1_index    = calculate_index({ ne13, ne12, ne11, ne10 }, { s13, s12, s11, s10 }, logical_index);
+        auto  dst_index      = calculate_index({ ne3, ne2, ne1, ne0 }, { s3, s2, s1, s0 }, logical_index);
+        auto  src0_float_val = sycl::vec(src0[src_0_index]).template convert<float, sycl::rounding_mode::rte>();
+        auto  src1_float_val = sycl::vec(src1[src_1_index]).template convert<float, sycl::rounding_mode::rte>();
+        float dst_val        = bin_op(src0_float_val[0], src1_float_val[0]);
+        auto  val_to_store   = sycl::vec(dst_val).template convert<dst_t, sycl::rounding_mode::rte>();
+        dst[dst_index]       = val_to_store;
     }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
 }
 
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_sycl {
+template <float (*bin_op)(const float, const float)> struct bin_bcast_sycl {
     template <typename src0_t, typename src1_t, typename dst_t>
     void operator()(const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd, const int64_t ne00,
                     const int64_t ne01, const int64_t ne02, const int64_t ne03, const int64_t ne10, const int64_t ne11,
@@ -96,165 +77,73 @@ struct bin_bcast_sycl {
                     const size_t nb10, const size_t nb11, const size_t nb12, const size_t nb13, const size_t nb0,
                     const size_t nb1, const size_t nb2, const size_t nb3, const bool src0_is_contiguous,
                     const bool src1_is_contiguous, const bool dst_is_contiguous, queue_ptr stream) {
-        int nr0 = ne10 / ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (src0_is_contiguous && src1_is_contiguous && dst_is_contiguous) {
+        auto check_bcast_required = [](const std::array<int64_t, 4> & src_dims,
+                                       const std::array<int64_t, 4> & dst_dims) -> bool {
             for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
+                if (dst_dims[i] > src_dims[i]) {
+                    return true;
                 }
             }
-        }
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
+            return false;
+        };
 
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
+        GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
+        GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
 
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
+        GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
+        GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
 
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
+        GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
+        GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
 
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
+        // dst strides in number of elements
+        size_t s0 = nb0 / sizeof(dst_t);
+        size_t s1 = nb1 / sizeof(dst_t);
+        size_t s2 = nb2 / sizeof(dst_t);
+        size_t s3 = nb3 / sizeof(dst_t);
 
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
+        // src1 strides in number of elements
+        size_t s10 = nb10 / sizeof(src0_t);
+        size_t s11 = nb11 / sizeof(src1_t);
+        size_t s12 = nb12 / sizeof(src1_t);
+        size_t s13 = nb13 / sizeof(src1_t);
 
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
+        // src0 strides in number of elements
+        size_t s00 = nb00 / sizeof(src0_t);
+        size_t s01 = nb01 / sizeof(src0_t);
+        size_t s02 = nb02 / sizeof(src0_t);
+        size_t s03 = nb03 / sizeof(src0_t);
 
-            GGML_UNUSED(s00);
+        std::size_t num_dst_elements = static_cast<std::size_t>(ne0) * static_cast<std::size_t>(ne1) *
+                                       static_cast<std::size_t>(ne2) * static_cast<std::size_t>(ne3);
+        std::size_t local_range  = 256;
+        std::size_t global_range = ceil_div(num_dst_elements, local_range) * local_range;
 
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
+        bool needs_broadcasting = check_bcast_required({ ne00, ne01, ne02, ne03 }, { ne0, ne1, ne2, ne3 }) ||
+                                  check_bcast_required({ ne10, ne11, ne12, ne13 }, { ne0, ne1, ne2, ne3 });
+        bool all_contiguous = src0_is_contiguous && src1_is_contiguous && dst_is_contiguous;
 
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            sycl::range<3> block_dims(1, 1, 1);
-            block_dims[2] = std::min<unsigned int>(hne0, block_size);
-            block_dims[1] = std::min<unsigned int>(
-                ne1, block_size / (unsigned int)block_dims[2]);
-            block_dims[0] = std::min(
-                std::min<unsigned int>(
-                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
-                                   (unsigned int)block_dims[1]),
-                64U);
-
-            sycl::range<3> block_nums(
-                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
-                (ne1 + block_dims[1] - 1) / block_dims[1],
-                (hne0 + block_dims[2] - 1) / block_dims[2]);
-
-            if (block_nums[0] > 65535) {
-                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                {
-                    dpct::has_capability_or_fail(stream->get_device(),
-                                                 {sycl::aspect::fp16});
-
-                    stream->parallel_for(
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
-                                          sycl::range<3>(1, 1, block_size)),
-                        [=](sycl::nd_item<3> item_ct1) {
-                            k_bin_bcast_unravel<bin_op>(
-                                src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
-                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
-                                s03, s11, s12, s13, item_ct1);
-                        });
-                }
-            } else {
-                /*
-                DPCT1049:16: The work-group size passed to the SYCL kernel may
-                exceed the limit. To get the device limit, query
-                info::device::max_work_group_size. Adjust the work-group size if
-                needed.
-                */
-                dpct::has_capability_or_fail(stream->get_device(),
-                                             {sycl::aspect::fp16});
-
-                stream->parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
-                        k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
-                                            ne2, ne3, ne10, ne11, ne12, ne13,
-                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
-                                            item_ct1);
-                    });
-            }
+        if (! needs_broadcasting && all_contiguous) {
+            stream->submit([&](sycl::handler & cgh) {
+                cgh.parallel_for(sycl::nd_range<1>({ global_range }, { local_range }), [=](sycl::nd_item<1> it) {
+                    k_bin_bcast_contiguous<bin_op>(src0_dd, src1_dd, dst_dd, num_dst_elements, it);
+                });
+            });
+        } else {
+            stream->submit([&](sycl::handler & cgh) {
+                cgh.parallel_for(sycl::nd_range<1>({ global_range }, { local_range }), [=](sycl::nd_item<1> it) {
+                    k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3, ne10, ne11, ne12, ne13, s0, s1,
+                                        s2, s3, s00, s01, s02, s03, s10, s11, s12, s13, num_dst_elements, it);
+                });
+            });
         }
     }
 };

ggml/src/ggml-sycl/convert.cpp

@@ -183,6 +183,24 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k,
     }
 }
 
+template <typename dst_t>
+static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    const size_t  local_size  = 32;
+    const size_t  global_size = nb * local_size;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    stream->submit([&](sycl::handler & cgh) {
+        sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
+
+        cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
+                         [=](sycl::nd_item<1> item_ct1) {
+                             dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
+                         });
+    });
+}
+
 template <typename dst_t>
 static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -504,7 +522,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
         case GGML_TYPE_Q3_K:
             return dequantize_row_q3_K_sycl;
         case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q4_K_sycl_reorder;
+            } else {
+                return dequantize_row_q4_K_sycl;
+            }
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:
@@ -556,7 +578,12 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
         case GGML_TYPE_Q3_K:
             return dequantize_row_q3_K_sycl;
         case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_sycl;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q4_K_sycl_reorder;
+            } else {
+                return dequantize_row_q4_K_sycl;
+            }
         case GGML_TYPE_Q5_K:
             return dequantize_row_q5_K_sycl;
         case GGML_TYPE_Q6_K:

ggml/src/ggml-sycl/dequantize.hpp

@@ -357,6 +357,28 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8
 }
 #endif
 
+template <typename dst_t>
+inline void dequantize_q4_K_common(dst_t * __restrict__ y, const uint8_t * __restrict__ qs_ptr, const float dall,
+                                   const float dmin, uint8_t * __restrict__ scales_local, int il, int ir) {
+    const int is = 2 * il;
+    constexpr int n  = 4;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, scales_local, sc, m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+
+    get_scale_min_k4(is + 1, scales_local, sc, m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    sycl::vec<uint8_t, n> q_vec = vec_aligned_load<uint8_t, n>(qs_ptr + 32 * il + n * ir);
+    for (int l = 0; l < n; ++l) {
+        y[l + 0]  = d1 * (q_vec[l] & 0xF) - m1;
+        y[l + 32] = d2 * (q_vec[l] >> 4) - m2;
+    }
+}
+
 template<typename dst_t>
 static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                   uint8_t* scales_local, const sycl::nd_item<3> &item_ct1) {
@@ -365,36 +387,22 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
     const int64_t i = item_ct1.get_group(2);
 
 #if QK_K == 256
-    // assume 32 threads
     const int64_t tid = item_ct1.get_local_id(2);
-    const int64_t il  = tid/8;
-    const int64_t ir  = tid%8;
-    const int64_t is  = 2*il;
-    const int64_t n   = 4;
+    const int64_t il  = tid / 8;
+    const int64_t ir  = tid % 8;
 
-    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+    dst_t * y = yy + i * QK_K + 64 * il + 4 * ir;
 
     const sycl::half2 dm = x[i].dm;
     const float dall = dm[0];
     const float dmin = dm[1];
 
-    if (tid < 12)
+    if (tid < 12) {
         scales_local[tid] = x[i].scales[tid];
-    item_ct1.barrier(sycl::access::fence_space::local_space);
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, scales_local, sc, m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, scales_local, sc, m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
-
-    sycl::vec<uint8_t, n> q_vec = vec_aligned_load<uint8_t, n>(x[i].qs + 32*il + n*ir);
-    for (int l = 0; l < n; ++l) {
-        y[l + 0] = d1 * (q_vec[l] & 0xF) - m1;
-        y[l +32] = d2 * (q_vec[l] >>  4) - m2;
     }
+
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+    dequantize_q4_K_common(y, x[i].qs, dall, dmin, scales_local, il, ir);
 #else
     const int64_t tid = item_ct1.get_local_id(2);
     const uint8_t * q = x[i].qs;
@@ -406,6 +414,36 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_block_q4_K_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy, uint8_t * scales_local,
+                                          const sycl::nd_item<1> & item_ct1, int64_t nb) {
+    const int64_t i   = item_ct1.get_group(0);     // block index
+    const int64_t tid = item_ct1.get_local_id(0);  // thread index within block
+    const int64_t il  = tid / 8;
+    const int64_t ir  = tid % 8;
+
+    dst_t * y = yy + i * QK_K + 64 * il + 4 * ir;
+
+    const uint8_t * base          = static_cast<const uint8_t *>(vx);
+    const size_t    qs_offset     = i * (QK_K / 2);
+    const size_t    scales_offset = nb * (QK_K / 2) + i * K_SCALE_SIZE;
+    const size_t    dm_offset     = nb * (QK_K / 2) + nb * K_SCALE_SIZE + i * sizeof(ggml_half2);
+
+    const uint8_t *    qs_ptr     = base + qs_offset;
+    const uint8_t *    scales_ptr = base + scales_offset;
+    ggml_half2         dm_values  = *reinterpret_cast<const ggml_half2 *>(base + dm_offset);
+
+    const float dall = dm_values.x();
+    const float dmin = dm_values.y();
+
+    if (tid < 12) {
+        scales_local[tid] = scales_ptr[tid];
+    }
+
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+    dequantize_q4_K_common(y, qs_ptr, dall, dmin, scales_local, il, ir);
+}
+
 template<typename dst_t>
 static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                   const sycl::nd_item<3> &item_ct1) {

ggml/src/ggml-sycl/dmmv.cpp

@@ -1129,7 +1129,13 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             break;
         case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                // reorder is currently not supported for dmmv
+                GGML_ABORT("Unimplemented dequantize case case for q4_k reorder");
+            } else {
+                dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            }
             break;
         case GGML_TYPE_Q5_K:
             dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);

ggml/src/ggml-sycl/element_wise.cpp

@@ -655,7 +655,6 @@ inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -688,7 +687,6 @@ inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -722,7 +720,6 @@ inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -754,7 +751,6 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -786,7 +782,6 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -818,7 +813,6 @@ inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml_tensor
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -850,7 +844,6 @@ inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -883,7 +876,6 @@ inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -917,7 +909,6 @@ inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggml_tenso
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -949,7 +940,6 @@ inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_tensor
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -981,7 +971,6 @@ inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1013,7 +1002,6 @@ inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1045,7 +1033,6 @@ inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_tensor *
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1078,7 +1065,6 @@ inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1110,7 +1096,6 @@ inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1142,7 +1127,6 @@ inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1174,7 +1158,6 @@ inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tensor * dst
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1206,7 +1189,6 @@ inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1241,7 +1223,6 @@ inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml_tensor
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1273,7 +1254,6 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1315,7 +1295,6 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, ggml_tensor *
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1350,7 +1329,6 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, ggml_tensor * dst)
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 
@@ -1388,7 +1366,6 @@ inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * ds
             }
         default:
             GGML_ABORT("GGML tensor type not supported!\n");
-            break;
     }
 }
 

ggml/src/ggml-sycl/gemm.hpp

@@ -32,16 +32,36 @@ public:
         else static_assert(0);
     }
 
-    static inline void row_gemm(ggml_backend_sycl_context & ctx, bool a_trans, bool b_trans, int m, int n, int k,
-                                const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+    // matrix A has m rows, k columns
+    // matrix B has k rows, n columns
+    // nra - number of elements to skip when moving into next row in A
+    // nrb - number of elements to skip when moving into next row in B
+    // nca - number of elements to skip when moving into next column in A
+    // ncb - number of elements to skip when moving into next column in B
+    // stride_a - number of elements to skip when moving to next A matrix
+    // stride_b - number of elements to skip when moving to next B matrix
+    // batches_a - number of A matrices
+    // batches_b - number of B matrices
+    static void gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
+        const void * a, dt at, dnnl_dim_t nra, dnnl_dim_t nca, dnnl_dim_t stride_a,
+        const void * b, dt bt, dnnl_dim_t nrb, dnnl_dim_t ncb, dnnl_dim_t stride_b,
+        void * c, dt ct, const queue_ptr & q, dnnl_dim_t batches_a, dnnl_dim_t batches_b) {
+
         auto stream = ctx.stream_dnnl(q);
         auto eng = ctx.engine_dnnl(q);
-        dnnl::memory::dims a_dims = { m, k };
-        dnnl::memory::dims b_dims = { k, n };
-        dnnl::memory::dims c_dims = { m, n };
-        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
-        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
-        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::ab);
+
+        // { # strides, # rows, # columns }
+        dnnl::memory::dims a_dims = { batches_a, m, k };
+        dnnl::memory::dims b_dims = { batches_b, k, n };
+        dnnl::memory::dims c_dims = { std::max(batches_a, batches_b), m, n };
+
+        // { # elements to skip to next stride, # elements to skip to next row, # elements to skip to next column }
+        dnnl::memory::dims a_strides = { stride_a, nra, nca };
+        dnnl::memory::dims b_strides = { stride_b, nrb, ncb };
+
+        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_strides);
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_strides);
+        const auto c_md    = dnnl::memory::desc(c_dims, ct, tag::abc);
 
         dnnl::primitive_attr primitive_attr;
         primitive_attr.set_scratchpad_mode(dnnl::scratchpad_mode::user);
@@ -63,6 +83,15 @@ public:
 
         matmul_prim.execute(stream, matmul_args);
     }
+
+    // matrices A and B are column major, both having k rows
+    // matrix A has m column, matrix B has n columns
+    // output: column major matrix C = A transposed * B
+    static void row_gemm(ggml_backend_sycl_context & ctx, int m, int n, int k,
+        const void * a, dt at, const void * b, dt bt, void * c, dt ct, const queue_ptr & q) {
+
+        gemm(ctx, m, n, k, a, at, k, 1, k * m, b, bt, 1, k, n * k, c, ct, q, 1, 1);
+    }
 };
 
 #endif

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

@@ -49,6 +49,7 @@ static bool g_sycl_loaded = false;
 int g_ggml_sycl_debug = 0;
 int g_ggml_sycl_disable_optimize = 0;
 int g_ggml_sycl_disable_graph = 0;
+int g_ggml_sycl_disable_dnn = 0;
 int g_ggml_sycl_prioritize_dmmv = 0;
 
 static ggml_sycl_device_info ggml_sycl_init() {
@@ -196,12 +197,22 @@ static void ggml_check_sycl() try {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
         g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
+        g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
         g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         GGML_LOG_INFO("Running with Environment Variables:\n");
         GGML_LOG_INFO("  GGML_SYCL_DEBUG: %d\n", g_ggml_sycl_debug);
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_OPT: %d\n", g_ggml_sycl_disable_optimize);
+#ifdef GGML_SYCL_GRAPH
         GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: %d\n", g_ggml_sycl_disable_graph);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_GRAPH: graph disabled by compile flag\n");
+#endif
+#if GGML_SYCL_DNNL
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: %d\n", g_ggml_sycl_disable_dnn);
+#else
+        GGML_LOG_INFO("  GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n");
+#endif
         GGML_LOG_INFO("  GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
         GGML_LOG_INFO("Build with Macros:\n");
 #if defined(GGML_SYCL_FORCE_MMQ)
@@ -341,7 +352,7 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         assert(tensor->view_src->buffer->buft == buffer->buft);
         return GGML_STATUS_SUCCESS;
     }
-    if (tensor->type == GGML_TYPE_Q4_0 && !g_ggml_sycl_disable_optimize) {
+    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K) && !g_ggml_sycl_disable_optimize) {
         ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
         tensor->extra                 = extra;
         ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
@@ -1985,19 +1996,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne10 = src1->ne[0];
-
+    GGML_ASSERT(ne00 == ne10);
 
     const int64_t row_diff = row_high - row_low;
 
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
-#if !GGML_SYCL_DNNL
-    const int64_t ne0 = dst->ne[0];
+
+    const int64_t ne0 = dst->ne[0]; // used by MKL only
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = id == ctx.device ? ne0 : row_diff;
-#endif
+    int ldc = id == ctx.device ? ne0 : row_diff; // used by MKL only
 
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -2033,25 +2043,29 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
         ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
-#if !GGML_SYCL_DNNL
-        const sycl::half alpha_f16 = 1.0f;
-        const sycl::half beta_f16  = 0.0f;
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
-            *stream, oneapi::math::transpose::trans,
-            oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
-            &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
-            src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
-            dst_f16.get(), dpct::library_data_t::real_half, ldc,
-            dpct::library_data_t::real_half)));
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-#else
-        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ptr,
-                                  DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
-                                  dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
-        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
-        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+#if GGML_SYCL_DNNL
+        if (!g_ggml_sycl_disable_dnn) {
+            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ptr,
+                                      DnnlGemmWrapper::to_dt<sycl::half>(), src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+                                      dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>(), stream);
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
+            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+        }
+        else
 #endif
+        {
+            const sycl::half alpha_f16 = 1.0f;
+            const sycl::half beta_f16  = 0.0f;
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
+                *stream, oneapi::math::transpose::trans,
+                oneapi::math::transpose::nontrans, row_diff, src1_ncols, ne10,
+                &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
+                src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
+                dst_f16.get(), dpct::library_data_t::real_half, ldc,
+                dpct::library_data_t::real_half)));
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16, dst);
+            to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
+        }
     }
     else {
         // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp32 path\n");
@@ -2072,18 +2086,22 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
-#if !GGML_SYCL_DNNL
-        const float alpha = 1.0f;
-        const float beta  = 0.0f;
-        SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
-            get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
-            src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
-            dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
-#else
-        DnnlGemmWrapper::row_gemm(ctx, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i,
-                                  DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
-                                  dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
+#if GGML_SYCL_DNNL
+        if (!g_ggml_sycl_disable_dnn) {
+            DnnlGemmWrapper::row_gemm(ctx, src1_ncols, row_diff, ne10, src1_ddf1_i,
+                                      DnnlGemmWrapper::to_dt<float>(), src0_ddf_i, DnnlGemmWrapper::to_dt<float>(),
+                                      dst_dd_i, DnnlGemmWrapper::to_dt<float>(), stream);
+        }
+        else
 #endif
+        {
+            const float alpha = 1.0f;
+            const float beta  = 0.0f;
+            SYCL_CHECK(CHECK_TRY_ERROR(oneapi::math::blas::column_major::gemm(
+                get_onemath_backend(*stream), oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, row_diff,
+                src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10,
+                dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
+        }
     }
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddq_i);
@@ -2697,7 +2715,7 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-static void k_compute_batched_ptrs(const sycl::half * src0_as_f16, const sycl::half * src1_as_f16, char * dst,
+static void k_compute_batched_ptrs(const sycl::half * src0_as_f16, const sycl::half * src1_as_f16, void * dst,
                                    const void ** ptrs_src, void ** ptrs_dst, int64_t ne12, int64_t ne13, int64_t ne23,
                                    size_t nb02, size_t nb03, size_t nb12, size_t nb13, size_t nbd2, size_t nbd3,
                                    int64_t r2, int64_t r3, const sycl::nd_item<3> & item_ct1) {
@@ -2713,7 +2731,7 @@ static void k_compute_batched_ptrs(const sycl::half * src0_as_f16, const sycl::h
 
     const uint8_t * src0_bytes = reinterpret_cast<const uint8_t *>(src0_as_f16);
     const uint8_t * src1_bytes = reinterpret_cast<const uint8_t *>(src1_as_f16);
-    uint8_t *       dst_bytes  = reinterpret_cast<uint8_t *>(dst);
+    uint8_t *       dst_bytes  = static_cast<uint8_t *>(dst);
 
     ptrs_src[0 * ne23 + i12 + i13 * ne12] = src0_bytes + i02 * nb02 + i03 * nb03;
     ptrs_src[1 * ne23 + i12 + i13 * ne12] = src1_bytes + i12 * nb12 + i13 * nb13;
@@ -2726,6 +2744,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -2766,7 +2785,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     }
 
     ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool());
-    char *                           dst_t = reinterpret_cast<char *>(dst_ddf);
 
     dpct::library_data_t mkl_compute_type = dpct::library_data_t::real_float;
     dpct::library_data_t mkl_data_type    = dpct::library_data_t::real_float;
@@ -2783,42 +2801,83 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
 
     GGML_ASSERT(ne12 % ne02 == 0);
     GGML_ASSERT(ne13 % ne03 == 0);
+    GGML_ASSERT(ne01 == static_cast<int64_t>(nb1/nb0));
+    GGML_ASSERT(ne10 == ne00);
 
     // broadcast factors
     const int64_t r2 = ne12 / ne02;
     const int64_t r3 = ne13 / ne03;
 
-    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
-        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
-                                                    oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-                                                    src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
-                                                    src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_t,
-                                                    mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
-    } else {
-        const int ne23 = ne12 * ne13;
+#if GGML_SYCL_DNNL
+    if (!g_ggml_sycl_disable_dnn) {
+        auto dnn_gemm = [&ctx, queue, ne11, ne01, ne10, nb00, nb01, nb02, s11, s12]
+            (const sycl::half* src1, const sycl::half* src0, float* dst, const dnnl_dim_t batches_a, const dnnl_dim_t batches_b) {
 
-        ggml_sycl_pool_alloc<const void *>         ptrs_src(ctx.pool(), 2 * ne23);
-        ggml_sycl_pool_alloc<void *>               ptrs_dst(ctx.pool(), 1 * ne23);
-        ggml_sycl_pool_alloc<matrix_info_t<float>> matrix_info(ctx.host_pool(), 1);
+            DnnlGemmWrapper::gemm(ctx, ne11,ne01, ne10,
+                            src1, DnnlGemmWrapper::to_dt<sycl::half>(), s11, 1, s12,
+                            src0, DnnlGemmWrapper::to_dt<sycl::half>(), 1, nb01/nb00, nb02/nb00,
+                            dst, DnnlGemmWrapper::to_dt<float>(), queue, batches_a, batches_b);
+        };
 
-        sycl::range<3> block_dims(1, ne12, ne13);
-        queue->submit([&](sycl::handler & cgh) {
-            const void ** ptrs_src_get = ptrs_src.get();
-            void **       ptrs_dst_get = ptrs_dst.get();
-            size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
-            size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
-            cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-                k_compute_batched_ptrs(src0_f16, src1_f16, dst_t, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
-                                       nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
+        if (r2 == 1 && r3 == 1) {
+            if (ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+                dnn_gemm(src1_f16, src0_f16, dst_ddf, ne12*ne13, ne02 * ne03);
+            }
+            else {
+                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie03*nb03)/sizeof(sycl::half)); // nb is in bytes
+                    const sycl::half* src1_f16_shifted = src1_f16 + ie03*s13;
+                    float* dst_shifted = dst_ddf + ((ie03*nb3)/sizeof(float));
+                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, ne12, ne02);
+                }
+            }
+        } else {
+            // iterate over batches from smaller set of matrices (matrix 0)
+            for (int64_t ie02 = 0; ie02 < ne02; ++ie02) {
+                for (int64_t ie03 = 0; ie03 < ne03; ++ie03) {
+                    const sycl::half* src0_f16_shifted = src0_f16 + ((ie02*nb02 + ie03*nb03)/sizeof(sycl::half));
+                    const sycl::half* src1_f16_shifted = src1_f16 + ie02*s12*r2 + ie03*s13*r3;
+                    float* dst_shifted = dst_ddf + ((ie02*nb2*r2 + ie03*nb3*r3)/sizeof(float));
+                    dnn_gemm(src1_f16_shifted, src0_f16_shifted, dst_shifted, r2*r3, 1);
+                }
+            }
+        }
+    }
+    else
+#endif
+    {
+        if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+            // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(*queue, oneapi::math::transpose::trans,
+                                                        oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                                                        src0_f16, dpct::library_data_t::real_half, nb01 / nb00, nb02 / nb00,
+                                                        src1_f16, dpct::library_data_t::real_half, s11, s12, beta, dst_ddf,
+                                                        mkl_data_type, ne0, ne1 * ne0, ne12 * ne13, mkl_compute_type)));
+        } else {
+            const int ne23 = ne12 * ne13;
+
+            ggml_sycl_pool_alloc<const void *>         ptrs_src(ctx.pool(), 2 * ne23);
+            ggml_sycl_pool_alloc<void *>               ptrs_dst(ctx.pool(), 1 * ne23);
+            ggml_sycl_pool_alloc<matrix_info_t<float>> matrix_info(ctx.host_pool(), 1);
+
+            sycl::range<3> block_dims(1, ne12, ne13);
+            queue->submit([&](sycl::handler & cgh) {
+                const void ** ptrs_src_get = ptrs_src.get();
+                void **       ptrs_dst_get = ptrs_dst.get();
+                size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
+                size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
+                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                    k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
+                                           nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
+                });
             });
-        });
 
-        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
-            *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
-            (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
-            (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, s11, beta,
-            (void **) (ptrs_dst.get() + 0 * ne23), mkl_data_type, ne0, ne23, mkl_compute_type, matrix_info.get())));
+            SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
+                *queue, oneapi::math::transpose::trans, oneapi::math::transpose::nontrans, ne01, ne11, ne10, alpha,
+                (const void **) (ptrs_src.get() + 0 * ne23), dpct::library_data_t::real_half, nb01 / nb00,
+                (const void **) (ptrs_src.get() + 1 * ne23), dpct::library_data_t::real_half, s11, beta,
+                (void **) (ptrs_dst.get() + 0 * ne23), mkl_data_type, ne0, ne23, mkl_compute_type, matrix_info.get())));
+        }
     }
 } catch (const sycl::exception & exc) {
     std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
@@ -2841,6 +2900,8 @@ inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
             return true;
+        case GGML_TYPE_Q4_K:
+            return !g_ggml_sycl_prioritize_dmmv;
         default:
             return false;
     }
@@ -2858,6 +2919,7 @@ inline bool ggml_sycl_supports_reorder_dmmv(enum ggml_type type) {
 inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_K:
             return true;
         default:
             return false;
@@ -2883,16 +2945,16 @@ static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     }
 }
 
-static void reorder_qw(char *data_device, const int ncols, const int nrows,
-                size_t size, size_t offset, dpct::queue_ptr stream) {
-    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
+static void reorder_qw_q4_0(uint8_t * data_device, const int ncols, const int nrows, size_t size, size_t offset,
+                            dpct::queue_ptr stream) {
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
     SYCL_CHECK(
         CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
             .wait()));
     GGML_ASSERT((size % sizeof(block_q4_0) == 0));
     GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
     int offset_blks = offset / sizeof(block_q4_0);
-    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;
+    auto qs_ptr      = data_device + offset_blks * QK4_0 / 2;
     auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
 
     stream->parallel_for(
@@ -2906,18 +2968,59 @@ static void reorder_qw(char *data_device, const int ncols, const int nrows,
                 *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
             }
             *(d_ptr + ib) = x[ib].d;
-        });
+        }).wait_and_throw();
+
+    sycl::free(tmp_buf, *stream);
+}
+
+static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
+    GGML_ASSERT(size % sizeof(block_q4_K) == 0);
+    GGML_ASSERT(offset % sizeof(block_q4_K) == 0);
+
+    const int nblocks = size / sizeof(block_q4_K);
+
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+
+    auto * qs_ptr     = data_device;
+    auto * scales_ptr = qs_ptr + QK_K / 2 * nblocks;
+    auto * dm_ptr     = (sycl::half2 *) (scales_ptr + K_SCALE_SIZE * nblocks);
+
+    stream->parallel_for(nblocks, [=](auto i) {
+        const block_q4_K * x  = (const block_q4_K *) tmp_buf;
+        const int          ib = i;
+
+        for (int j = 0; j < QK_K / 2; ++j) {
+            qs_ptr[ib * (QK_K / 2) + j] = x[ib].qs[j];
+        }
+
+        for (int j = 0; j < K_SCALE_SIZE; ++j) {
+            scales_ptr[ib * K_SCALE_SIZE + j] = x[ib].scales[j];
+        }
+
+        dm_ptr[ib] = x[ib].dm;
+    }).wait_and_throw();
 
     sycl::free(tmp_buf, *stream);
 }
 
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
-    char*data_device = (char*)src0->data;
+    uint8_t * data_device = (uint8_t *) src0->data;
     size_t ncols = src0->ne[0];
     size_t nrows = src0->ne[1];
     size_t size = ggml_nbytes(src0);
 
-    reorder_qw(data_device, ncols, nrows, size, 0, stream);
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            reorder_qw_q4_0(data_device, ncols, nrows, size, 0, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            reorder_qw_q4_k(data_device, size, 0, stream);
+            break;
+        default:
+            GGML_ABORT("reorder_qw() called with unsupported type");
+            break;
+    }
 }
 
 static bool should_reorder_tensor(ggml_backend_sycl_context& ctx, const ggml_tensor * dst) {
@@ -2960,8 +3063,18 @@ static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor *
     extra->optimized_feature.reorder = true;  // Used to decode/dequan in next steps and avoid re-reordering
 }
 
-static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 
+static bool can_use_dequantize_mul_mat_vec(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    return ggml_sycl_supports_dmmv(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
+           src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
+}
+
+static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    return ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 &&
+           src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+}
+
+static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
     int64_t min_compute_capability = INT_MAX;
 
@@ -2984,13 +3097,9 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     }
 
     // check data types and tensor shapes for custom matrix multiplication kernels:
-    bool use_dequantize_mul_mat_vec = ggml_sycl_supports_dmmv(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
+    bool use_dequantize_mul_mat_vec = can_use_dequantize_mul_mat_vec(src0, src1, dst);
 
-    bool use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
-        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+    bool use_mul_mat_vec_q = can_use_mul_mat_vec_q(src0, src1, dst);
 
     bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
@@ -3713,7 +3822,8 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
             return GGML_STATUS_SUCCESS;
         }
 
-        sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream()));
+        sycl_ex::command_graph model_sycl_graph(*(sycl_ctx->stream()), {sycl_ex::property::graph::assume_buffer_outlives_graph{}});
+
         model_sycl_graph.begin_recording(*(sycl_ctx->stream()));
         ggml_backend_sycl_graph_compute_impl(sycl_ctx, cgraph);
         model_sycl_graph.end_recording();

ggml/src/ggml-sycl/mmvq.cpp

@@ -24,6 +24,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
     const int     blocks_per_row              = ncols / block_traits::qk;
     constexpr int blocks_per_subgroup         = ceil_div(block_traits::vdr_mmvq * WARP_SIZE, block_traits::qi);
     constexpr int block_elements_per_subgroup = block_traits::qi / block_traits::vdr_mmvq;
+    const int     nblocks                     = nrows * (ncols / block_traits::qk);
 
     static_assert(blocks_per_subgroup > 0);
     static_assert(block_elements_per_subgroup > 0);
@@ -45,7 +46,7 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
             // x block quant index when casting the quants to int
             const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);
 
-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
         }
     }
 
@@ -739,6 +740,27 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     }
 }
 
+static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+    const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+
+    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                            [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
+                                                                                            nrows, nd_item);
+                            });
+    });
+}
+
+
 static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -1035,7 +1057,12 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 break;
             case GGML_TYPE_Q4_K:
-                mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    reorder_mul_mat_vec_q4_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
                 break;
             case GGML_TYPE_Q5_K:
                 mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);

ggml/src/ggml-sycl/quants.hpp

@@ -56,6 +56,28 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q4_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI4_K;
+        static constexpr uint32_t qr       = QR4_K;
+        static constexpr uint32_t vdr_mmvq = 2;
+    };
+
+    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+
+    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks = (nrows * (ncols / traits::qk));
+        return (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2));
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+
+    constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
+
+    constexpr size_t get_dm_offset(int nblocks) { return get_total_qs_bytes(nblocks) + nblocks * K_SCALE_SIZE; }
+};
+
 }  // namespace ggml_sycl_reordered
 
 #endif  // GGML_SYCL_QUANTS_HPP

ggml/src/ggml-sycl/vecdotq.hpp

@@ -285,7 +285,7 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     }
 
     __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int /* nblocks */) {
         const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
         const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
         int             v[q4_0_traits::vdr_mmvq];
@@ -303,6 +303,67 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
     };
 };
 
+static inline float vec_dot_q4_K_q8_1_common(const int * __restrict__ q4, const uint16_t * __restrict__ scales,
+                                             const ggml_half2 & dm, const block_q8_1 * __restrict__ bq8_1,
+                                             const int &        iqs) {
+    int   v[2];
+    int   u[2 * QR4_K];
+    float d8[QR4_K];
+
+    v[0] = q4[0];
+    v[1] = q4[4];
+
+    uint16_t  aux[2];
+    const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+    if (j < 2) {
+        aux[0] = scales[j + 0] & 0x3f3f;
+        aux[1] = scales[j + 2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+    }
+
+    const uint8_t * sc = (const uint8_t *) aux;
+    const uint8_t * m  = sc + 2;
+
+    const int bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+
+    for (int i = 0; i < QR4_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i]                   = bq8i->ds[0];
+
+        const int * q8 = (const int *) bq8i->qs + ((iqs / 2) % 4);
+        u[2 * i + 0]   = q8[0];
+        u[2 * i + 1]   = q8[4];
+    }
+
+    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, dm, d8);
+}
+
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q4_K;
+
+    using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
+    using q4_k_traits = typename q4_k_block::traits;
+
+    float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
+                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int nblocks) {
+        const int ib = ibx_offset / (QK_K / 2);
+
+        const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
+        const uint8_t *    qs             = base + ibx_offset;
+        const int          total_qs_bytes = nblocks * (QK_K / 2);
+        const uint8_t *    scs            = base + total_qs_bytes + ib * K_SCALE_SIZE;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset);
+
+        const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+        const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
+        const uint16_t * scales     = (const uint16_t *) scs;
+
+        return vec_dot_q4_K_q8_1_common(q4, scales, *dms, bq8_1, iqs);
+    }
+};
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -649,52 +710,17 @@ vec_dot_q3_K_q8_1(const void *__restrict__ vbq,
     return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
 }
 
-static __dpct_inline__ float
-vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
-                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
-
+static __dpct_inline__ float vec_dot_q4_K_q8_1(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1,
+                                               const int & iqs) {
 #ifndef GGML_QKK_64
+
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
-    int    v[2];
-    int    u[2*QR4_K];
-    float d8[QR4_K];
+    const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
+    const int *      q4         = (const int *) (bq4_K->qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
+    const uint16_t * scales     = (const uint16_t *) bq4_K->scales;
 
-    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
-    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
-
-    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
-    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
-    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
-    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
-
-    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
-    v[0] = q4[0];
-    v[1] = q4[4];
-
-    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
-    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
-    const uint8_t * sc = (const uint8_t *)aux;
-    const uint8_t * m  = sc + 2;
-
-    for (int i = 0; i < QR4_K; ++i) {
-        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = bq8i->ds[0];
-
-        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
-        u[2*i+0] = q8[0];
-        u[2*i+1] = q8[4];
-    }
-
-    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
+    return vec_dot_q4_K_q8_1_common(q4, scales, bq4_K->dm, bq8_1, iqs);
 
 #else
 

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -15,6 +15,32 @@ function(detect_host_compiler)
     set(HOST_CXX_COMPILER "${HOST_CXX_COMPILER}" PARENT_SCOPE)
 endfunction()
 
+# Function to test shader extension support
+# Parameters:
+#  EXTENSION_NAME - Name of the extension to test (e.g., "GL_EXT_integer_dot_product")
+#  TEST_SHADER_FILE - Path to the test shader file
+#  RESULT_VARIABLE - Name of the variable to set (ON/OFF) based on test result
+function(test_shader_extension_support EXTENSION_NAME TEST_SHADER_FILE RESULT_VARIABLE)
+    execute_process(
+        COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${TEST_SHADER_FILE}"
+        OUTPUT_VARIABLE glslc_output
+        ERROR_VARIABLE glslc_error
+    )
+
+    if (${glslc_error} MATCHES ".*extension not supported: ${EXTENSION_NAME}.*")
+        message(STATUS "${EXTENSION_NAME} not supported by glslc")
+        set(${RESULT_VARIABLE} OFF PARENT_SCOPE)
+    else()
+        message(STATUS "${EXTENSION_NAME} supported by glslc")
+        set(${RESULT_VARIABLE} ON PARENT_SCOPE)
+        add_compile_definitions(${RESULT_VARIABLE})
+
+        # Ensure the extension support is forwarded to vulkan-shaders-gen
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -D${RESULT_VARIABLE}=ON)
+        set(VULKAN_SHADER_GEN_CMAKE_ARGS "${VULKAN_SHADER_GEN_CMAKE_ARGS}" PARENT_SCOPE)
+    endif()
+endfunction()
+
 if (Vulkan_FOUND)
     message(STATUS "Vulkan found")
 
@@ -23,69 +49,35 @@ if (Vulkan_FOUND)
                              ../../include/ggml-vulkan.h
                             )
 
-    # Compile a test shader to determine whether GL_KHR_cooperative_matrix is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
+    set(VULKAN_SHADER_GEN_CMAKE_ARGS
+        -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
+        -DCMAKE_RUNTIME_OUTPUT_DIRECTORY=${CMAKE_RUNTIME_OUTPUT_DIRECTORY}
+    )
 
-    if (${glslc_error} MATCHES ".*extension not supported: GL_KHR_cooperative_matrix.*")
-        message(STATUS "GL_KHR_cooperative_matrix not supported by glslc")
-        set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT OFF)
-    else()
-        message(STATUS "GL_KHR_cooperative_matrix supported by glslc")
-        set(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT ON)
-        add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
-    endif()
+    # Test all shader extensions
+    test_shader_extension_support(
+        "GL_KHR_cooperative_matrix"
+        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
+        "GGML_VULKAN_COOPMAT_GLSLC_SUPPORT"
+    )
 
-    # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
+    test_shader_extension_support(
+        "GL_NV_cooperative_matrix2"
+        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
+        "GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT"
+    )
 
-    if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
-        message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
-        set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT OFF)
-    else()
-        message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
-        set(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT ON)
-        add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-    endif()
+    test_shader_extension_support(
+        "GL_EXT_integer_dot_product"
+        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp"
+        "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT"
+    )
 
-    # Compile a test shader to determine whether GL_EXT_integer_dot_product is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
-
-    if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_integer_dot_product.*")
-        message(STATUS "GL_EXT_integer_dot_product not supported by glslc")
-        set(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT OFF)
-    else()
-        message(STATUS "GL_EXT_integer_dot_product supported by glslc")
-        set(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT ON)
-        add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
-    endif()
-
-    # Compile a test shader to determine whether GL_EXT_bfloat16 is supported.
-    # If it's not, there will be an error to stderr.
-    # If it's supported, set a define to indicate that we should compile those shaders
-    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
-                    OUTPUT_VARIABLE glslc_output
-                    ERROR_VARIABLE glslc_error)
-
-    if (${glslc_error} MATCHES ".*extension not supported: GL_EXT_bfloat16.*")
-        message(STATUS "GL_EXT_bfloat16 not supported by glslc")
-        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT OFF)
-    else()
-        message(STATUS "GL_EXT_bfloat16 supported by glslc")
-        set(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT ON)
-        add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
-    endif()
+    test_shader_extension_support(
+        "GL_EXT_bfloat16"
+        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+        "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT"
+    )
 
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
@@ -124,16 +116,8 @@ if (Vulkan_FOUND)
         add_compile_definitions(GGML_VULKAN_RUN_TESTS)
     endif()
 
-    if (NOT CMAKE_CROSSCOMPILING)
-        add_subdirectory(vulkan-shaders)
-        if (MSVC)
-            foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
-                string(TOUPPER ${CONFIG} CONFIG)
-                set_target_properties(vulkan-shaders-gen PROPERTIES
-                    RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
-            endforeach()
-        endif()
-    else()
+    # Set up toolchain for host compilation whether cross-compiling or not
+    if (CMAKE_CROSSCOMPILING)
         if (GGML_VULKAN_SHADERS_GEN_TOOLCHAIN)
             set(HOST_CMAKE_TOOLCHAIN_FILE ${GGML_VULKAN_SHADERS_GEN_TOOLCHAIN})
         else()
@@ -146,25 +130,31 @@ if (Vulkan_FOUND)
             configure_file(${CMAKE_CURRENT_SOURCE_DIR}/cmake/host-toolchain.cmake.in ${CMAKE_BINARY_DIR}/host-toolchain.cmake @ONLY)
             set(HOST_CMAKE_TOOLCHAIN_FILE ${CMAKE_BINARY_DIR}/host-toolchain.cmake)
         endif()
-        message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
-
-        include(ExternalProject)
-        # Native build through ExternalProject_Add
-        ExternalProject_Add(
-            vulkan-shaders-gen
-            SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
-            CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE}
-                    -DCMAKE_INSTALL_PREFIX=${CMAKE_BINARY_DIR}
-                    -DGGML_VULKAN_COOPMAT_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT_GLSLC_SUPPORT}
-                    -DGGML_VULKAN_COOPMAT2_GLSLC_SUPPORT=${GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT}
-                    -DGGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT=${GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT}
-                    -DGGML_VULKAN_BFLOAT16_GLSLC_SUPPORT=${GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT}
-            BUILD_COMMAND ${CMAKE_COMMAND} --build .
-            INSTALL_COMMAND ${CMAKE_COMMAND} --install .
-            INSTALL_DIR ${CMAKE_BINARY_DIR}
-        )
-        ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
+    else()
+        # For non-cross-compiling, use empty toolchain (use host compiler)
+        set(HOST_CMAKE_TOOLCHAIN_FILE "")
     endif()
+
+    # Always use ExternalProject_Add approach
+    include(ExternalProject)
+
+    # Add toolchain file if cross-compiling
+    if (CMAKE_CROSSCOMPILING)
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DCMAKE_TOOLCHAIN_FILE=${HOST_CMAKE_TOOLCHAIN_FILE})
+        message(STATUS "vulkan-shaders-gen toolchain file: ${HOST_CMAKE_TOOLCHAIN_FILE}")
+    endif()
+
+    # Native build through ExternalProject_Add
+    ExternalProject_Add(
+        vulkan-shaders-gen
+        SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders
+        CMAKE_ARGS ${VULKAN_SHADER_GEN_CMAKE_ARGS}
+        BUILD_COMMAND ${CMAKE_COMMAND} --build .
+        INSTALL_COMMAND ${CMAKE_COMMAND} --install .
+        INSTALL_DIR ${CMAKE_BINARY_DIR}
+    )
+    ExternalProject_Add_StepTargets(vulkan-shaders-gen build install)
+
     set (_ggml_vk_host_suffix $<IF:$<STREQUAL:${CMAKE_HOST_SYSTEM_NAME},Windows>,.exe,>)
     set (_ggml_vk_genshaders_cmd ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/vulkan-shaders-gen${_ggml_vk_host_suffix})
     set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
@@ -175,9 +165,8 @@ if (Vulkan_FOUND)
     file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
     set (_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen)
 
-    if (CMAKE_CROSSCOMPILING)
-        set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
-    endif()
+    # Add build and install dependencies for all builds
+    set(_ggml_vk_shader_deps ${_ggml_vk_shader_deps} vulkan-shaders-gen-build vulkan-shaders-gen-install)
 
     add_custom_command(
         OUTPUT ${_ggml_vk_header}

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -288,6 +288,9 @@ struct vk_device_struct {
     bool coopmat_acc_f32_support {};
     bool coopmat_acc_f16_support {};
     bool coopmat_bf16_support {};
+    bool coopmat_support_16x16x16_f16acc {};
+    bool coopmat_support_16x16x16_f32acc {};
+    bool coopmat1_fa_support {};
     uint32_t coopmat_m;
     uint32_t coopmat_n;
     uint32_t coopmat_k;
@@ -410,6 +413,13 @@ struct vk_device_struct {
     vk_pipeline pipeline_flash_attn_f32_f16_D128_cm2[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D256_cm2[GGML_TYPE_COUNT][2][2][2];
 
+    vk_pipeline pipeline_flash_attn_f32_f16_D64_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D80_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D96_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D112_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D128_cm1[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D256_cm1[GGML_TYPE_COUNT][2][2][2];
+
     vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
     vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
@@ -1588,19 +1598,36 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
     );
 }
 
+enum FaCodePath {
+    FA_SCALAR,
+    FA_COOPMAT1,
+    FA_COOPMAT2,
+};
+
 // number of rows/cols for flash attention shader
 static constexpr uint32_t flash_attention_num_small_rows = 32;
 static constexpr uint32_t scalar_flash_attention_num_small_rows = 1;
 static constexpr uint32_t scalar_flash_attention_num_large_rows = 8;
 
-static uint32_t get_fa_num_small_rows(bool scalar) {
-    return scalar ? scalar_flash_attention_num_small_rows : flash_attention_num_small_rows;
+// The FA coopmat1 shader assumes 16x16x16 matrix multiply support.
+// 128 threads split into four subgroups, each subgroup does 1/4
+// of the Bc dimension.
+static constexpr uint32_t coopmat1_flash_attention_num_large_rows = 16;
+static constexpr uint32_t scalar_flash_attention_Bc = 64;
+static constexpr uint32_t scalar_flash_attention_workgroup_size = 128;
+
+static uint32_t get_fa_num_small_rows(FaCodePath path) {
+    if (path == FA_COOPMAT2) {
+        return flash_attention_num_small_rows;
+    } else {
+        return scalar_flash_attention_num_small_rows;
+    }
 }
 
-static std::array<uint32_t, 2> fa_rows_cols(bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
     GGML_UNUSED(clamp);
 
-    if (scalar) {
+    if (path == FA_SCALAR) {
         if (small_rows) {
             return {scalar_flash_attention_num_small_rows, 64};
         } else {
@@ -1608,9 +1635,17 @@ static std::array<uint32_t, 2> fa_rows_cols(bool scalar, uint32_t D, uint32_t cl
         }
     }
 
+    if (path == FA_COOPMAT1) {
+        if (small_rows) {
+            return {scalar_flash_attention_num_small_rows, scalar_flash_attention_Bc};
+        } else {
+            return {coopmat1_flash_attention_num_large_rows, scalar_flash_attention_Bc};
+        }
+    }
+
     // small rows, large cols
     if (small_rows) {
-        return {get_fa_num_small_rows(scalar), 32};
+        return {get_fa_num_small_rows(FA_COOPMAT2), 32};
     }
 
     // small cols to reduce register count
@@ -1907,17 +1942,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
-    auto const &fa_wg_denoms = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
-        return {fa_rows_cols(scalar, D, clamp, type, small_rows)[0], 1, 1};
+    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+        return {fa_rows_cols(path, D, clamp, type, small_rows)[0], 1, 1};
     };
 
-    auto const &fa_spec_constants = [&](bool scalar, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
         // For large number of rows, 128 invocations seems to work best.
         // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
         // can't use 256 for D==80.
         // For scalar, use 128 (arbitrary)
-        uint32_t wg_size = scalar ? 128 : ((small_rows && (D % 32) == 0) ? 256 : 128);
-        auto rows_cols = fa_rows_cols(scalar, D, clamp, type, small_rows);
+        uint32_t wg_size = (path == FA_SCALAR || path == FA_COOPMAT1)
+                            ? scalar_flash_attention_workgroup_size
+                            : ((small_rows && (D % 32) == 0) ? 256 : 128);
+        auto rows_cols = fa_rows_cols(path, D, clamp, type, small_rows);
 
         // D_split can't be larger than a subgroup because we use subgroupShuffle to reduce it.
         // D_split can't be larger than the LSB of D divided by 4 due to vectorization in the shader.
@@ -1929,36 +1966,43 @@ static void ggml_vk_load_shaders(vk_device& device) {
         return {wg_size, rows_cols[0], rows_cols[1], (D), clamp, D_split};
     };
 
-#define CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, D) \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,false), fa_spec_constants(SCALAR, D,1,TYPE,false), 1, true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,false), fa_spec_constants(SCALAR, D,0,TYPE,false), fa_rows_cols(SCALAR,D,0,TYPE,false)[1], true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,1,TYPE,true), fa_spec_constants(SCALAR, D,1,TYPE,true), 1, true);     \
-        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(SCALAR, D,0,TYPE,true), fa_spec_constants(SCALAR, D,0,TYPE,true), fa_rows_cols(SCALAR,D,0,TYPE,true)[1], true);     \
+#define CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, D) \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,false), fa_spec_constants(FAPATH, D,1,TYPE,false), 1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC #SUFFIX,           flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,false), fa_spec_constants(FAPATH, D,0,TYPE,false), fa_rows_cols(FAPATH,D,0,TYPE,false)[1], true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,1,TYPE,true), fa_spec_constants(FAPATH, D,1,TYPE,true),   1,                                      true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D ## SUFFIX[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC #SUFFIX, flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _len,         flash_attn_f32_f16_ ## NAMELC ##     SUFFIX ## _data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, D,0,TYPE,true), fa_spec_constants(FAPATH, D,0,TYPE,true),   fa_rows_cols(FAPATH,D,0,TYPE,true)[1],  true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? 32 : 0));     \
 
-#define CREATE_FA(TYPE, NAMELC, SCALAR, SUFFIX) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 64) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 80) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 96) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 112) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 128) \
-        CREATE_FA2(TYPE, NAMELC, SCALAR, SUFFIX, 256)
+#define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 64) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 80) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 96) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 112) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 128) \
+        CREATE_FA2(TYPE, NAMELC, FAPATH, SUFFIX, 256)
 
-    CREATE_FA(GGML_TYPE_F16, f16, true, )
-    CREATE_FA(GGML_TYPE_Q4_0, q4_0, true, )
-    CREATE_FA(GGML_TYPE_Q8_0, q8_0, true, )
+    CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, )
+    CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, )
+    CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_SCALAR, )
+#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+    if (device->coopmat1_fa_support) {
+        CREATE_FA(GGML_TYPE_F16, f16, FA_COOPMAT1, _cm1)
+        CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_COOPMAT1, _cm1)
+        CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_COOPMAT1, _cm1)
+    }
+#endif
 #if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     if (device->coopmat2) {
-        CREATE_FA(GGML_TYPE_F16, f16, false, _cm2)
-        CREATE_FA(GGML_TYPE_Q4_0, q4_0, false, _cm2)
-        CREATE_FA(GGML_TYPE_Q4_1, q4_1, false, _cm2)
-        CREATE_FA(GGML_TYPE_Q5_0, q5_0, false, _cm2)
-        CREATE_FA(GGML_TYPE_Q5_1, q5_1, false, _cm2)
-        CREATE_FA(GGML_TYPE_Q8_0, q8_0, false, _cm2)
-        CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, false, _cm2)
+        CREATE_FA(GGML_TYPE_F16, f16, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_Q4_1, q4_1, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_Q5_0, q5_0, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_Q5_1, q5_1, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_COOPMAT2, _cm2)
+        CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl, FA_COOPMAT2, _cm2)
     }
 #endif
 #undef CREATE_FA2
@@ -2041,17 +2085,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // Create 6 variants, {s,m,l}x{unaligned,aligned}
 #define CREATE_MM(TYPE, PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm1_len, NAMELC ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
         if (device->mul_mat ## ID ## _l[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
         if (device->mul_mat ## ID ## _m[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
         if (device->mul_mat ## ID ## _s[TYPE]) \
-            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm1_len, NAMELC ## _aligned ## F16ACC ## _cm1_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
 
         // Create 2 variants, {f16,f32} accumulator
 #define CREATE_MM2(TYPE, PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
@@ -3009,6 +3053,11 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
 #if defined(VK_KHR_cooperative_matrix)
         device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix;
+
+        // coopmat1 fa shader currently assumes 32 invocations per subgroup
+        device->coopmat1_fa_support = device->coopmat_support && device->subgroup_require_full_support &&
+                                      device->subgroup_size_control && device->subgroup_min_size <= 32 &&
+                                      device->subgroup_max_size >= 32;
 #endif
 
         if (coopmat2_support) {
@@ -3143,6 +3192,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                             // Only enable if shape is identical
                             device->coopmat_acc_f32_support = true;
                         }
+                        if (prop.MSize == 16 && prop.NSize == 16 && prop.KSize == 16) {
+                            device->coopmat_support_16x16x16_f32acc = true;
+                        }
                     } else if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat16 &&
                                (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat16) {
                         // coopmat sizes not set yet
@@ -3155,6 +3207,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
                             // Only enable if shape is identical
                             device->coopmat_acc_f16_support = true;
                         }
+                        if (prop.MSize == 16 && prop.NSize == 16 && prop.KSize == 16) {
+                            device->coopmat_support_16x16x16_f16acc = true;
+                        }
                     }
                 } else if ((vk::ComponentTypeKHR)prop.AType      == vk::ComponentTypeKHR::eSint8 &&
                            (vk::ComponentTypeKHR)prop.BType      == vk::ComponentTypeKHR::eSint8 &&
@@ -5688,6 +5743,36 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t D, bool f32acc) {
+    // Needs to be kept up to date on shader changes
+    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
+    const uint32_t Br = scalar_flash_attention_num_large_rows;
+    const uint32_t Bc = scalar_flash_attention_Bc;
+
+    const uint32_t acctype = f32acc ? 4 : 2;
+    const uint32_t f16vec4 = 8;
+
+    const uint32_t tmpsh = wg_size * sizeof(float);
+    const uint32_t tmpshv4 = wg_size * 4 * acctype;
+
+    const uint32_t Qf = Br * (D / 4 + 2) * f16vec4;
+
+    const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+    const uint32_t sfsh = Bc * sfshstride * acctype;
+
+    const uint32_t kshstride = D / 4 + 2;
+    const uint32_t ksh = Bc * kshstride * f16vec4;
+
+    const uint32_t slope = Br * sizeof(float);
+
+    const uint32_t total_size = tmpsh + tmpshv4 + Qf + sfsh + ksh + slope;
+    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;
+
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(D=" << D << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);
+
+    return supported;
+}
+
 static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, ggml_tensor * dst, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_flash_attn((" << q << ", name=" << q->name << ", type=" << q->type << ", ne0=" << q->ne[0] << ", ne1=" << q->ne[1] << ", ne2=" << q->ne[2] << ", ne3=" << q->ne[3] << ", nb0=" << q->nb[0] << ", nb1=" << q->nb[1] << ", nb2=" << q->nb[2] << ", nb3=" << q->nb[3];
     std::cerr << "), (" << k << ", name=" << k->name << ", type=" << k->type << ", ne0=" << k->ne[0] << ", ne1=" << k->ne[1] << ", ne2=" << k->ne[2] << ", ne3=" << k->ne[3] << ", nb0=" << k->nb[0] << ", nb1=" << k->nb[1] << ", nb2=" << k->nb[2] << ", nb3=" << k->nb[3];
@@ -5738,7 +5823,19 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     assert(q->type == GGML_TYPE_F32);
     assert(k->type == v->type);
 
-    bool scalar = !ctx->device->coopmat2;
+    FaCodePath path = ctx->device->coopmat2 ? FA_COOPMAT2 :
+                      ctx->device->coopmat1_fa_support ? FA_COOPMAT1 : FA_SCALAR;
+
+    if (path == FA_COOPMAT1) {
+        const bool coopmat_shape_supported = (dst->op_params[3] == GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f32acc) ||
+                                             (dst->op_params[3] != GGML_PREC_F32 && ctx->device->coopmat_support_16x16x16_f16acc);
+
+        const bool coopmat_shmem_supported = ggml_vk_flash_attn_coopmat_shmem_support(ctx->device, D, dst->op_params[3] == GGML_PREC_F32);
+
+        if (!coopmat_shape_supported || !coopmat_shmem_supported) {
+            path = FA_SCALAR;
+        }
+    }
 
     uint32_t gqa_ratio = 1;
     uint32_t qk_ratio = neq2 / nek2;
@@ -5746,9 +5843,21 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     uint32_t workgroups_y = (uint32_t)neq2;
     uint32_t workgroups_z = (uint32_t)neq3;
 
-    // For scalar FA, we can use the "large" size to accommodate qga.
-    // For coopmat FA, we always use the small size (which is still pretty large for gqa).
-    const uint32_t max_gqa = scalar ? scalar_flash_attention_num_large_rows : get_fa_num_small_rows(false);
+    // For scalar/coopmat1 FA, we can use the "large" size to accommodate qga.
+    // For coopmat2 FA, we always use the small size (which is still pretty large for gqa).
+    uint32_t max_gqa;
+    switch (path) {
+    case FA_SCALAR:
+    case FA_COOPMAT1:
+        // We may switch from coopmat1 to scalar, so use the scalar limit for both
+        max_gqa = scalar_flash_attention_num_large_rows;
+        break;
+    case FA_COOPMAT2:
+        max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
+        break;
+    default:
+        GGML_ASSERT(0);
+    }
 
     if (N == 1 && qk_ratio > 1 && qk_ratio <= max_gqa &&
         qk_ratio * nek2 == neq2 && nek2 == nev2 && neq3 == 1 && nek3 == 1 && nev3 == 1) {
@@ -5761,11 +5870,23 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 
     vk_pipeline *pipelines;
-    // XXX TODO other backends may be changing accumulator precision to default to f32 soon
-    bool f32acc = scalar || dst->op_params[3] == GGML_PREC_F32;
-    bool small_rows = N <= get_fa_num_small_rows(scalar);
+    bool small_rows = N <= get_fa_num_small_rows(path);
 
-    if (scalar) {
+    // coopmat1 does not actually support "small rows" (it needs 16 rows).
+    // So use scalar instead.
+    if (small_rows && path == FA_COOPMAT1) {
+        path = FA_SCALAR;
+    }
+
+    // scalar is faster than coopmat2 when N==1
+    if (N == 1 && path == FA_COOPMAT2) {
+        path = FA_SCALAR;
+    }
+
+    bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
+
+    switch (path) {
+    case FA_SCALAR:
         switch (D) {
         case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
         case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
@@ -5777,7 +5898,21 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
             GGML_ASSERT(!"unsupported D value");
             return;
         }
-    } else {
+        break;
+    case FA_COOPMAT1:
+        switch (D) {
+        case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm1[k->type][f32acc][small_rows][0]; break;
+        case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm1[k->type][f32acc][small_rows][0]; break;
+        case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96_cm1[k->type][f32acc][small_rows][0]; break;
+        case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112_cm1[k->type][f32acc][small_rows][0]; break;
+        case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128_cm1[k->type][f32acc][small_rows][0]; break;
+        case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256_cm1[k->type][f32acc][small_rows][0]; break;
+        default:
+            GGML_ASSERT(!"unsupported D value");
+            return;
+        }
+        break;
+    case FA_COOPMAT2:
         switch (D) {
         case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64_cm2[k->type][f32acc][small_rows][0]; break;
         case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80_cm2[k->type][f32acc][small_rows][0]; break;
@@ -5789,6 +5924,9 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
             GGML_ASSERT(!"unsupported D value");
             return;
         }
+        break;
+    default:
+        GGML_ASSERT(0);
     }
     assert(pipelines);
 

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -5,18 +5,35 @@ find_package (Threads REQUIRED)
 
 if (GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+    message(STATUS "Enabling coopmat glslc support")
 endif()
 if (GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    message(STATUS "Enabling coopmat2 glslc support")
 endif()
 if (GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT)
+    message(STATUS "Enabling dot glslc support")
 endif()
 if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
+    message(STATUS "Enabling bfloat16 glslc support")
 endif()
+
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
+
+# Configure output directories for MSVC builds
+if(MSVC)
+    # Get the main project's runtime output directory if possible
+    if(DEFINED CMAKE_RUNTIME_OUTPUT_DIRECTORY)
+        foreach(CONFIG ${CMAKE_CONFIGURATION_TYPES})
+            string(TOUPPER ${CONFIG} CONFIG)
+            set_target_properties(${TARGET} PROPERTIES
+                RUNTIME_OUTPUT_DIRECTORY_${CONFIG} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+        endforeach()
+    endif()
+endif()

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -9,59 +9,13 @@
 #extension GL_KHR_shader_subgroup_shuffle : enable
 
 #include "types.comp"
+#include "flash_attn_base.comp"
 
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 1) const uint32_t Br = 1;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-
-layout (constant_id = 5) const uint32_t D_split = 16;
 const uint32_t D_per_thread = D / D_split;
 
-const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split;
+const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
 
-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;
 
 layout (binding = 0) readonly buffer Q {float data_q[];};
 layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
@@ -70,39 +24,6 @@ layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
 layout (binding = 2) readonly buffer V {float16_t data_v[];};
 layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
 layout (binding = 3) readonly buffer M {float16_t data_m[];};
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#if defined(A_TYPE_PACKED16)
-#define BINDING_IDX_K 0
-#define BINDING_IDX_V 1
-layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
-#endif
-
-#if defined(DATA_A_Q4_0)
-#define BLOCK_BYTE_SIZE 18
-
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
-    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
-    uint shift = (iqs & 0x10) >> 2;
-    vui_lo >>= shift;
-    vui_hi >>= shift;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-#define BLOCK_BYTE_SIZE 34
-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
-    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
-    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
-
-    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
-}
-#endif
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
 
 // Store the output when doing grouped query attention.
 // Rows index by Q's dimension 2, and the first N rows are valid.
@@ -113,29 +34,8 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
-shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
-shared vec4 tmpshv4[gl_WorkGroupSize.x];
+shared FLOAT_TYPE tmpsh[WorkGroupSize];
+shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
 shared vec4 Qf[Br][D / 4];
@@ -145,58 +45,12 @@ void main() {
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint32_t tid = gl_LocalInvocationIndex;
-    const uint32_t N = p.N;
-    const uint32_t KV = p.KV;
+    init_indices();
 
+    const uint32_t tid = gl_LocalInvocationIndex;
     const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
     const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
 
-    uint32_t i = gl_WorkGroupID.x;
-    uint32_t split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    const uint32_t Tr = CEIL_DIV(N, Br);
-
-    const uint32_t start_j = split_k_index * p.split_kv / Bc;
-    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    const uint32_t iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    const uint32_t rk2 = p.neq2/p.nek2;
-    const uint32_t rk3 = p.neq3/p.nek3;
-
-    const uint32_t rv2 = p.neq2/p.nev2;
-    const uint32_t rv3 = p.neq3/p.nev3;
-
-    // k indices
-    const uint32_t ik3 = iq3 / rk3;
-    const uint32_t ik2 = iq2 / rk2;
-
-    // v indices
-    const uint32_t iv3 = iq3 / rv3;
-    const uint32_t iv2 = iq2 / rv2;
-
-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    uint32_t k_stride = p.nb11;
-    uint32_t v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
-
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
     [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -0,0 +1,162 @@
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
+layout (constant_id = 1) const uint32_t Br = 1;
+layout (constant_id = 2) const uint32_t Bc = 32;
+layout (constant_id = 3) const uint32_t D = 32;
+layout (constant_id = 4) const uint32_t Clamp = 0;
+layout (constant_id = 5) const uint32_t D_split = 16;
+
+
+layout (push_constant) uniform parameter {
+    uint32_t N;
+    uint32_t KV;
+
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+
+    uint32_t neq2;
+    uint32_t neq3;
+    uint32_t nek2;
+    uint32_t nek3;
+    uint32_t nev2;
+    uint32_t nev3;
+    uint32_t nem1;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t nb21;
+    uint32_t nb22;
+    uint32_t nb23;
+    uint32_t nb31;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t mask;
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+
+    uint32_t gqa_ratio;
+    uint32_t split_kv;
+    uint32_t k_num;
+} p;
+
+layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+
+#if defined(A_TYPE_PACKED16)
+#define BINDING_IDX_K 0
+#define BINDING_IDX_V 1
+layout (binding = 1) readonly buffer KV_PACKED16 {A_TYPE_PACKED16 data_packed16[];} kv_packed[2];
+#endif
+
+#if defined(DATA_A_Q4_0)
+#define BLOCK_BYTE_SIZE 18
+
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    uint vui_lo = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
+    uint vui_hi = uint(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
+    uint shift = (iqs & 0x10) >> 2;
+    vui_lo >>= shift;
+    vui_hi >>= shift;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+#define BLOCK_BYTE_SIZE 34
+vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+    const i8vec2 v0 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
+    const i8vec2 v1 = unpack8(int32_t(kv_packed[binding_idx].data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
+
+    return float(kv_packed[binding_idx].data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+}
+#endif
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+
+// Store column zero. This is used to save per-row m and L values for split_k.
+ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    if (r < N && c == 0) {
+        uint32_t offset = iq2 + r;
+        data_o[o_offset + offset] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Load the slope matrix, indexed by Q's dimension 2.
+ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
+{
+    const uint32_t h = iq2 + (r % p.gqa_ratio);
+
+    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+
+    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
+}
+
+uint32_t i, N, KV, split_k_index, Tr, start_j, end_j,
+         iq2, iq3, rk2, rk3, rv2, rv3, ik2, ik3, iv2, iv3,
+         q_stride, k_stride, v_stride, m_stride;
+
+void init_indices()
+{
+    N = p.N;
+    KV = p.KV;
+
+    i = gl_WorkGroupID.x;
+    split_k_index = 0;
+
+    if (p.k_num > 1) {
+        i = 0;
+        split_k_index = gl_WorkGroupID.x;
+    }
+
+    Tr = CEIL_DIV(N, Br);
+
+    start_j = split_k_index * p.split_kv / Bc;
+    end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
+
+    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
+    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
+    iq2 = gl_WorkGroupID.y * p.gqa_ratio;
+    iq3 = gl_WorkGroupID.z;
+
+    // broadcast factors
+    rk2 = p.neq2/p.nek2;
+    rk3 = p.neq3/p.nek3;
+
+    rv2 = p.neq2/p.nev2;
+    rv3 = p.neq3/p.nev3;
+
+    // k indices
+    ik3 = iq3 / rk3;
+    ik2 = iq2 / rk2;
+
+    // v indices
+    iv3 = iq3 / rv3;
+    iv2 = iq2 / rv2;
+
+    // nb?1 are already divided by the type size and are in units of elements.
+    // When using grouped query attention, Q is indexed by iq2, so the stride
+    // should be nb02 (which is in bytes).
+    q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
+    k_stride = p.nb11;
+    v_stride = p.nb21;
+    // When using grouped query attention, all rows use the same mask (stride 0).
+    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
+    // that prevents the compiler from folding the "&" through the select
+    // and breaking the alignment detection.
+    m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
+}

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -0,0 +1,360 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+
+#extension GL_KHR_shader_subgroup_basic : enable
+#extension GL_KHR_memory_scope_semantics : enable
+#extension GL_KHR_cooperative_matrix : enable
+
+#include "types.comp"
+#include "flash_attn_base.comp"
+
+const uint32_t D_per_thread = D / D_split;
+const uint32_t row_split = 4;
+const uint32_t rows_per_thread = Br / row_split;
+const uint32_t cols_per_iter = gl_WorkGroupSize.x / D_split / row_split;
+const uint32_t cols_per_thread = Bc / cols_per_iter;
+
+
+layout (binding = 0) readonly buffer Q {float data_q[];};
+layout (binding = 0) readonly buffer QV4 {vec4 data_qv4[];};
+layout (binding = 1) readonly buffer K {float16_t data_k[];};
+layout (binding = 1) readonly buffer KV4 {f16vec4 data_kv4[];};
+layout (binding = 2) readonly buffer V {float16_t data_v[];};
+layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
+layout (binding = 3) readonly buffer M {float16_t data_m[];};
+
+// Store the output when doing grouped query attention.
+// Rows index by Q's dimension 2, and the first N rows are valid.
+D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    uint32_t offset = (iq2 + r) * D + c;
+    data_o[o_offset + offset] = D_TYPE(elem);
+    return elem;
+}
+
+// These need to be supported N,M values for a MatBc x MatBr x 16 coopmatmuladd
+const uint32_t MatBr = 16;
+const uint32_t MatBc = 16;
+
+shared FLOAT_TYPE tmpsh[gl_WorkGroupSize.x];
+shared ACC_TYPEV4 tmpshv4[gl_WorkGroupSize.x];
+
+const uint32_t qstride = D / 4 + 2; // in units of f16vec4
+shared f16vec4 Qf[Br * qstride];
+
+// Avoid padding for D==256 to make it fit in 48KB shmem.
+const uint32_t sfshstride = (D <= 128) ? (Br + 8) : Br;
+shared ACC_TYPE sfsh[Bc * sfshstride];
+
+const uint32_t kshstride = D / 4 + 2; // in units of f16vec4
+shared f16vec4 ksh[Bc * kshstride];
+
+shared float slope[Br];
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
+#endif
+
+    init_indices();
+
+    const uint32_t tid = gl_LocalInvocationIndex;
+
+    const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
+    const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
+    const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
+    const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
+
+#define tile_row(r) (row_tid * rows_per_thread + (r))
+
+    uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
+
+    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (D / 4);
+        uint32_t r = (idx + tid) / (D / 4);
+        if (r < Br && d < D / 4 &&
+            i * Br + r < N) {
+            Qf[r * qstride + d] = f16vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
+        }
+    }
+    barrier();
+
+    ACC_TYPEV4 Of[rows_per_thread][D_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] = ACC_TYPEV4(0.0);
+        }
+    }
+
+    float Lf[rows_per_thread], Mf[rows_per_thread];
+
+    // Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
+    const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
+
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        Lf[r] = 0;
+        Mf[r] = NEG_FLT_MAX_OVER_2;
+    }
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        if (tid < Br) {
+            uint r = tid;
+            slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
+        }
+        barrier();
+    } else {
+        if (tid < Br) {
+            uint r = tid;
+            slope[r] = 1.0;
+        }
+        barrier();
+    }
+
+#if BLOCK_SIZE > 1
+    uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / BLOCK_BYTE_SIZE;
+    uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / BLOCK_BYTE_SIZE;
+#else
+    uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
+    uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
+#endif
+
+    [[dont_unroll]]
+    for (uint32_t j = start_j; j < end_j; ++j) {
+
+        [[unroll]] for (uint32_t idx = 0; idx < Bc * D / 4; idx += gl_WorkGroupSize.x) {
+            uint32_t d = (idx + tid) % (D / 4);
+            uint32_t c = (idx + tid) / (D / 4);
+            if (c < Bc && d < D / 4) {
+#if BLOCK_SIZE > 1
+                uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
+                uint ib = coord / BLOCK_SIZE;
+                uint iqs = (coord % BLOCK_SIZE);
+                f16vec4 K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+#else
+                f16vec4 K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+#endif
+
+                ksh[c * kshstride + d] = K_Tf;
+            }
+        }
+        barrier();
+
+        // K * Q^T -> S^T: Bc x D * D x Br -> Bc x Br
+        // Bc split across workgroup (four subgroups), loop over D in chunks of 16: 16 x 16 * 16 x 16 -> 16 x 16
+        // This is written transposed in order to allow for N being 8 if implementations need it
+        coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
+        coopmat<float16_t, gl_ScopeSubgroup, MatBc, 16, gl_MatrixUseA> KMat;
+        coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
+
+        for (uint32_t d = 0; d < D / 16; ++d) {
+            coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
+
+            uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
+            coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+
+            SfMat = coopMatMulAdd(KMat, QMat, SfMat);
+        }
+
+        uint coord = gl_SubgroupID * MatBc * sfshstride;
+        coopMatStore(SfMat, sfsh, coord, sfshstride, gl_CooperativeMatrixLayoutRowMajor);
+        barrier();
+
+        if (p.logit_softcap != 0.0f) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) / Br;
+                uint32_t r = (idx + tid) % Br;
+                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
+                    sfsh[c * sfshstride + r] = ACC_TYPE(p.logit_softcap * tanh(sfsh[c * sfshstride + r]));
+                }
+            }
+            barrier();
+        }
+
+        if (p.mask != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
+                uint32_t c = (idx + tid) % Bc;
+                uint32_t r = (idx + tid) / Bc;
+                if (idx + tid < Bc * Br || idx + gl_WorkGroupSize.x <= Bc * Br) {
+                    sfsh[c * sfshstride + r] += ACC_TYPE(slope[r] * float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]));
+                }
+            }
+            barrier();
+        }
+
+        float eMf[rows_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float rowmaxf = sfsh[tile_row(r) + (0 * cols_per_iter + col_tid) * sfshstride];
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                rowmaxf = max(rowmaxf, float(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride]));
+            }
+            float Moldf = Mf[r];
+
+            // M = max(rowmax, Mold)
+            // P = e^(S - M)
+            // eM = e^(Mold - M)
+            Mf[r] = max(rowmaxf, Moldf);
+            eMf[r] = exp(Moldf - Mf[r]);
+        }
+
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            }
+        }
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Lf[r] = eMf[r]*Lf[r];
+        }
+
+        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+            float Pf[rows_per_thread];
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Pf[r] = exp(sfsh[tile_row(r) + (c * cols_per_iter + col_tid) * sfshstride] - Mf[r]);
+                Lf[r] += Pf[r];
+            }
+            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+#if BLOCK_SIZE > 1
+                uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                uint ib = coord / BLOCK_SIZE;
+                uint iqs = (coord % BLOCK_SIZE);
+                vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
+#endif
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                }
+            }
+        }
+
+        barrier();
+    }
+
+    // reduce across threads
+
+    float rowmaxf[rows_per_thread], eMf[rows_per_thread], Moldf[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        FLOAT_TYPE M = Mf[r];
+        tmpsh[tid] = M;
+        // Compute max across the row
+        barrier();
+        [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
+            M = max(M, tmpsh[tid ^ s]);
+            barrier();
+            tmpsh[tid] = M;
+            barrier();
+        }
+        rowmaxf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
+        barrier();
+    }
+
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        Moldf[r] = Mf[r];
+
+        // M = max(rowmax, Mold)
+        // eM = e^(Mold - M)
+        Mf[r] = max(rowmaxf[r], Moldf[r]);
+        eMf[r] = exp(Moldf[r] - Mf[r]);
+
+        Lf[r] = eMf[r]*Lf[r];
+    }
+
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        FLOAT_TYPE L = Lf[r];
+        tmpsh[tid] = L;
+        // Compute sum across the row
+        barrier();
+        [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
+            L += tmpsh[tid ^ s];
+            barrier();
+            tmpsh[tid] = L;
+            barrier();
+        }
+        Lf[r] = tmpsh[d_tid + row_tid * threads_per_rowgroup];
+        barrier();
+    }
+
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+
+            Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            tmpshv4[tid] = Of[r][d];
+
+            barrier();
+            [[unroll]] for (int s = int(gl_WorkGroupSize.x / row_split) / 2; s >= D_split; s >>= 1) {
+                Of[r][d] += tmpshv4[tid ^ s];
+                barrier();
+                tmpshv4[tid] = Of[r][d];
+                barrier();
+            }
+            Of[r][d] = tmpshv4[d_tid + row_tid * threads_per_rowgroup];
+            barrier();
+        }
+    }
+
+    // If there is split_k, then the split_k resolve shader does the final
+    // division by L. Store the intermediate O value and per-row m and L values.
+    if (p.k_num > 1) {
+        uint32_t o_offset = D * p.ne1 * split_k_index;
+
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            if (tile_row(r) < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
+                    }
+                }
+            }
+        }
+
+        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            if (tile_row(r) < N) {
+                perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
+                perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+            }
+        }
+
+        return;
+    }
+
+    float Lfrcp[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        Lfrcp[r] = 1.0 / Lf[r];
+    }
+
+    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] *= float16_t(Lfrcp[r]);
+        }
+    }
+
+    uint32_t o_offset = iq3*p.ne2*p.ne1;
+
+    if (p.gqa_ratio > 1) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            if (tile_row(r) < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        perElemOpGqaStore(tile_row(r), 4*(d * D_split + d_tid) + comp, float(Of[r][d][comp]), o_offset, iq2, N);
+                    }
+                }
+            }
+        }
+    } else {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            if (i * Br + tile_row(r) < N) {
+                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
+                        data_o[o_offset + iq2 * D + (i * Br + tile_row(r)) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                    }
+                }
+            }
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp

@@ -18,62 +18,12 @@
 
 #include "types.comp"
 #include "dequant_funcs_cm2.comp"
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 1) const uint32_t Br = 32;
-layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = gl_CooperativeMatrixClampModeConstantNV;
-
-layout (push_constant) uniform parameter {
-    uint32_t N;
-    uint32_t KV;
-
-    uint32_t ne1;
-    uint32_t ne2;
-    uint32_t ne3;
-
-    uint32_t neq2;
-    uint32_t neq3;
-    uint32_t nek2;
-    uint32_t nek3;
-    uint32_t nev2;
-    uint32_t nev3;
-    uint32_t nem1;
-
-    uint32_t nb01;
-    uint32_t nb02;
-    uint32_t nb03;
-    uint32_t nb11;
-    uint32_t nb12;
-    uint32_t nb13;
-    uint32_t nb21;
-    uint32_t nb22;
-    uint32_t nb23;
-    uint32_t nb31;
-
-    float scale;
-    float max_bias;
-    float logit_softcap;
-
-    uint32_t mask;
-    uint32_t n_head_log2;
-    float m0;
-    float m1;
-
-    uint32_t gqa_ratio;
-    uint32_t split_kv;
-    uint32_t k_num;
-} p;
+#include "flash_attn_base.comp"
 
 layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
 layout (binding = 1) readonly buffer K {uint8_t data_k[];};
 layout (binding = 2) readonly buffer V {uint8_t data_v[];};
 layout (binding = 3) readonly buffer M {uint8_t data_m[];};
-layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
-
-#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
 
 ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
     return max(x, y);
@@ -118,67 +68,12 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
     return elem;
 }
 
-// Store column zero. This is used to save per-row m and L values for split_k.
-ACC_TYPE perElemOpStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
-{
-    if (r < N && c == 0) {
-        uint32_t offset = iq2 + r;
-        data_o[o_offset + offset] = D_TYPE(elem);
-    }
-    return elem;
-}
-
-// Load the slope matrix, indexed by Q's dimension 2.
-ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t iq2)
-{
-    const uint32_t h = iq2 + (r % p.gqa_ratio);
-
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
-
-    return ACC_TYPE(pow(base, ACC_TYPE(exph)));
-}
-
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint32_t N = p.N;
-    const uint32_t KV = p.KV;
-
-    uint32_t i = gl_WorkGroupID.x;
-    uint32_t split_k_index = 0;
-
-    if (p.k_num > 1) {
-        i = 0;
-        split_k_index = gl_WorkGroupID.x;
-    }
-
-    const uint32_t Tr = CEIL_DIV(N, Br);
-
-    const uint32_t start_j = split_k_index * p.split_kv / Bc;
-    const uint32_t end_j = CEIL_DIV(min(KV, (split_k_index + 1) * p.split_kv), Bc);
-
-    // When not using grouped query attention, all rows share the same iq2, equal to gl_WorkGroupID.y.
-    // When using grouped query attention, each workgroup does gqa_ratio consecutive values of iq2.
-    const uint32_t iq2 = gl_WorkGroupID.y * p.gqa_ratio;
-    const uint32_t iq3 = gl_WorkGroupID.z;
-
-    // broadcast factors
-    const uint32_t rk2 = p.neq2/p.nek2;
-    const uint32_t rk3 = p.neq3/p.nek3;
-
-    const uint32_t rv2 = p.neq2/p.nev2;
-    const uint32_t rv3 = p.neq3/p.nev3;
-
-    // k indices
-    const uint32_t ik3 = iq3 / rk3;
-    const uint32_t ik2 = iq2 / rk2;
-
-    // v indices
-    const uint32_t iv3 = iq3 / rv3;
-    const uint32_t iv2 = iq2 / rv2;
+    init_indices();
 
     tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
     tensorLayoutNV<2, Clamp> tensorLayoutK = createTensorLayoutNV(2, Clamp);
@@ -195,17 +90,6 @@ void main() {
     tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
     tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
 
-    // nb?1 are already divided by the type size and are in units of elements.
-    // When using grouped query attention, Q is indexed by iq2, so the stride
-    // should be nb02 (which is in bytes).
-    uint32_t q_stride = p.gqa_ratio > 1 ? (p.nb02 / 4) : p.nb01;
-    uint32_t k_stride = p.nb11;
-    uint32_t v_stride = p.nb21;
-    // When using grouped query attention, all rows use the same mask (stride 0).
-    // "p.gqa_ratio >> 16" is just a roundabout way of writing zero
-    // that prevents the compiler from folding the "&" through the select
-    // and breaking the alignment detection.
-    uint32_t m_stride = (p.gqa_ratio > 1) ? (p.gqa_ratio >> 16) : KV;
     // hint to the compiler that strides are aligned for the aligned variant of the shader
     if (Clamp != gl_CooperativeMatrixClampModeConstantNV)
     {

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -215,7 +215,7 @@ static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
 void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
-    std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_coopmat" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
+    std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_cm1" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
     std::string out_fname = join_paths(output_dir, name + ".spv");
     std::string in_path = join_paths(input_dir, in_fname);
 
@@ -424,6 +424,7 @@ void process_shaders() {
     // flash attention
     for (const auto& f16acc : {false, true}) {
         std::string acctype = f16acc ? "float16_t" : "float";
+        std::string acctypev4 = f16acc ? "f16vec4" : "vec4";
 
         for (const auto& tname : type_names) {
             if (tname == "f32") {
@@ -440,6 +441,16 @@ void process_shaders() {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
                     merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
             }
+#endif
+#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
+            if (tname == "f16") {
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                    merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"ACC_TYPEV4", acctypev4}, {"COOPMAT", "1"}}), true, true, false, f16acc);
+            } else if (tname == "q4_0" || tname == "q8_0") {
+                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                    merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"ACC_TYPEV4", acctypev4}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
+            }
 #endif
             if (tname == "f16") {
                 string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",

ggml/src/ggml.c

@@ -5499,7 +5499,7 @@ static void ggml_compute_backward(
             // tensor = src0 * 1 + src1 * 0
             if (src0_needs_grads) {
                 // dsrc0 = dtensor * 1
-                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad, src0));
             }
             if (src1_needs_grads) {
                 // dsrc1 = dtensor * 0 -> noop
@@ -5780,10 +5780,9 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
 }
 
 void ggml_build_backward_expand(
-        struct ggml_context * ctx_static,
-        struct ggml_context * ctx_compute,
-        struct ggml_cgraph  * cgraph,
-        bool                  accumulate) {
+        struct ggml_context *  ctx,
+        struct ggml_cgraph  *  cgraph,
+        struct ggml_tensor  ** grad_accs) {
     GGML_ASSERT(cgraph->n_nodes > 0);
     GGML_ASSERT(cgraph->grads);
     GGML_ASSERT(cgraph->grad_accs);
@@ -5856,21 +5855,24 @@ void ggml_build_backward_expand(
         GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
-        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
-        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
-        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
-        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
-            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
-            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
-            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
+        const size_t ihash = ggml_hash_find(&cgraph->visited_hash_set, node);
+        GGML_ASSERT(ihash != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, ihash));
+        if (grad_accs && grad_accs[i]) {
+            cgraph->grad_accs[ihash] = grad_accs[i];
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
+        } else if (node->flags & GGML_TENSOR_FLAG_LOSS) {
+            // loss tensors always need a gradient accumulator
+            cgraph->grad_accs[ihash] = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, node->ne);
+            cgraph->grads[ihash]     = cgraph->grad_accs[ihash];
         }
-        grads_needed[igrad] = true;
+        grads_needed[ihash] = true;
     }
 
     for (int i = n_nodes_f - 1; i >= 0; --i) {
         // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
-        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
+        ggml_compute_backward(ctx, cgraph, i, grads_needed);
     }
 
     free(grads_needed);
@@ -6016,8 +6018,8 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     }
 }
 
-struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
-    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL);
+struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads) {
+    struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads || force_grads);
     ggml_graph_cpy(cgraph, result);
     return result;
 }
@@ -6036,6 +6038,9 @@ struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
 }
 
 void ggml_graph_reset(struct ggml_cgraph * cgraph) {
+    if (!cgraph) {
+        return;
+    }
     GGML_ASSERT(cgraph->grads != NULL);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -6345,8 +6350,8 @@ void ggml_set_output(struct ggml_tensor * tensor) {
     tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
 }
 
-void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) {
-    GGML_UNUSED(ctx); // TODO: remove this parameter
+void ggml_set_param(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->op == GGML_OP_NONE);
     tensor->flags |= GGML_TENSOR_FLAG_PARAM;
 }
 

ggml/src/gguf.cpp

@@ -299,10 +299,10 @@ bool gguf_read_emplace_helper(const struct gguf_reader & gr, std::vector<struct
                 return false;
             }
         } catch (std::length_error &) {
-            fprintf(stderr, "%s: encountered length_error while reading value for key '%s'\n", __func__, key.c_str());
+            GGML_LOG_ERROR("%s: encountered length_error while reading value for key '%s'\n", __func__, key.c_str());
             return false;
         } catch (std::bad_alloc &) {
-            fprintf(stderr, "%s: encountered bad_alloc error while reading value for key '%s'\n", __func__, key.c_str());
+            GGML_LOG_ERROR("%s: encountered bad_alloc error while reading value for key '%s'\n", __func__, key.c_str());
             return false;
         }
         kv.emplace_back(key, value);
@@ -328,14 +328,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         ok = ok && gr.read(magic, 4);
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to read magic\n", __func__);
+            GGML_LOG_ERROR("%s: failed to read magic\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
 
         for (uint32_t i = 0; i < magic.size(); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                fprintf(stderr, "%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
+                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 gguf_free(ctx);
                 return nullptr;
             }
@@ -348,11 +348,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
     if (ok && gr.read(ctx->version)) {
         if (ctx->version == 1) {
-            fprintf(stderr, "%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
+            GGML_LOG_ERROR("%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
             ok = false;
         }
         if (ctx->version > GGUF_VERSION) {
-            fprintf(stderr, "%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
+            GGML_LOG_ERROR("%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
                 __func__, ctx->version, GGUF_VERSION);
             ok = false;
         }
@@ -363,7 +363,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     if (ok && gr.read(n_tensors)) {
         static_assert(sizeof(size_t) <= 8 && sizeof(gguf_tensor_info) >= 2, "int64_t insufficient for indexing");
         if (n_tensors < 0 || n_tensors > int64_t(SIZE_MAX/sizeof(gguf_tensor_info))) {
-            fprintf(stderr, "%s: number of tensors is %" PRIi64 " but must be in [0, %zu]\n",
+            GGML_LOG_ERROR("%s: number of tensors is %" PRIi64 " but must be in [0, %zu]\n",
                 __func__, n_tensors, SIZE_MAX/sizeof(gguf_tensor_info));
             ok = false;
         }
@@ -374,7 +374,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     if (ok && gr.read(n_kv)) {
         static_assert(sizeof(size_t) <= 8 && sizeof(gguf_tensor_info) >= 2, "int64_t insufficient for indexing");
         if (n_kv < 0 || n_kv > int64_t(SIZE_MAX/sizeof(gguf_kv))) {
-            fprintf(stderr, "%s: number of key value pairs is %" PRIi64 " but must be in [0, %zu]\n",
+            GGML_LOG_ERROR("%s: number of key value pairs is %" PRIi64 " but must be in [0, %zu]\n",
                     __func__, n_kv, SIZE_MAX/sizeof(gguf_kv));
             ok = false;
         }
@@ -383,7 +383,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     }
 
     if (!ok) {
-        fprintf(stderr, "%s: failed to read header\n", __func__);
+        GGML_LOG_ERROR("%s: failed to read header\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -399,15 +399,15 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             try {
                 ok = ok && gr.read(key);
             } catch (std::length_error &) {
-                fprintf(stderr, "%s: encountered length_error while reading key %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered length_error while reading key %" PRIi64 "\n", __func__, i);
                 ok = false;
             } catch (std::bad_alloc &) {
-                fprintf(stderr, "%s: encountered bad_alloc error while reading key %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered bad_alloc error while reading key %" PRIi64 "\n", __func__, i);
                 ok = false;
             }
             for (size_t j = 0; ok && j < ctx->kv.size(); ++j) {
                 if (key == ctx->kv[j].key) {
-                    fprintf(stderr, "%s: duplicate key '%s' for tensors %zu and %" PRIi64 " \n", __func__, key.c_str(), j, i);
+                    GGML_LOG_ERROR("%s: duplicate key '%s' for tensors %zu and %" PRIi64 " \n", __func__, key.c_str(), j, i);
                     ok = false;
                 }
             }
@@ -441,14 +441,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                 case GGUF_TYPE_ARRAY:
                 default:
                     {
-                        fprintf(stderr, "%s: key '%s' has invalid GGUF type %d\n", __func__, key.c_str(), type);
+                        GGML_LOG_ERROR("%s: key '%s' has invalid GGUF type %d\n", __func__, key.c_str(), type);
                         ok = false;
                     } break;
             }
         }
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to read key-value pairs\n", __func__);
+            GGML_LOG_ERROR("%s: failed to read key-value pairs\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
@@ -458,7 +458,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         ctx->alignment = alignment_idx == -1 ? GGUF_DEFAULT_ALIGNMENT : gguf_get_val_u32(ctx, alignment_idx);
 
         if (ctx->alignment == 0 || (ctx->alignment & (ctx->alignment - 1)) != 0) {
-            fprintf(stderr, "%s: alignment %zu is not a power of 2\n", __func__, ctx->alignment);
+            GGML_LOG_ERROR("%s: alignment %zu is not a power of 2\n", __func__, ctx->alignment);
             gguf_free(ctx);
             return nullptr;
         }
@@ -474,14 +474,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             try {
                 ok = ok && gr.read(name);
             } catch (std::length_error &) {
-                fprintf(stderr, "%s: encountered length_error while reading tensor name %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered length_error while reading tensor name %" PRIi64 "\n", __func__, i);
                 ok = false;
             } catch (std::bad_alloc &) {
-                fprintf(stderr, "%s: encountered bad_alloc error while reading tensor name %" PRIi64 "\n", __func__, i);
+                GGML_LOG_ERROR("%s: encountered bad_alloc error while reading tensor name %" PRIi64 "\n", __func__, i);
                 ok = false;
             }
             if (name.length() >= GGML_MAX_NAME) {
-                fprintf(stderr, "%s: tensor name %" PRIi64 " is too long: %zu >= %d\n", __func__, i, name.length(), GGML_MAX_NAME);
+                GGML_LOG_ERROR("%s: tensor name %" PRIi64 " is too long: %zu >= %d\n", __func__, i, name.length(), GGML_MAX_NAME);
                 ok = false;
                 break;
             }
@@ -490,7 +490,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             // make sure there are no duplicate tensor names
             for (int64_t j = 0; ok && j < i; ++j) {
                 if (strcmp(info.t.name, ctx->info[j].t.name) == 0) {
-                    fprintf(stderr, "%s: duplicate tensor name '%s' for tensors %" PRIi64 " and %" PRIi64 "\n", __func__, info.t.name, j, i);
+                    GGML_LOG_ERROR("%s: duplicate tensor name '%s' for tensors %" PRIi64 " and %" PRIi64 "\n", __func__, info.t.name, j, i);
                     ok = false;
                     break;
                 }
@@ -505,7 +505,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             uint32_t n_dims = -1;
             ok = ok && gr.read(n_dims);
             if (n_dims > GGML_MAX_DIMS) {
-                fprintf(stderr, "%s: tensor '%s' has invalid number of dimensions: %" PRIu32 " > %" PRIu32 "\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid number of dimensions: %" PRIu32 " > %" PRIu32 "\n",
                     __func__, info.t.name, n_dims, GGML_MAX_DIMS);
                 ok = false;
                 break;
@@ -518,7 +518,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
                 // check that all ne are non-negative
                 if (info.t.ne[j] < 0) {
-                    fprintf(stderr, "%s: tensor '%s' dimension %" PRIu32 " has invalid number of elements: %" PRIi64 " < 0\n",
+                    GGML_LOG_ERROR("%s: tensor '%s' dimension %" PRIu32 " has invalid number of elements: %" PRIi64 " < 0\n",
                         __func__, info.t.name, j, info.t.ne[j]);
                     ok = false;
                     break;
@@ -530,7 +530,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
                        (INT64_MAX/info.t.ne[2] <= info.t.ne[0]*info.t.ne[1]) ||
                        (INT64_MAX/info.t.ne[3] <= info.t.ne[0]*info.t.ne[1]*info.t.ne[2]))) {
 
-                fprintf(stderr, "%s: total number of elements in tensor '%s' with shape "
+                GGML_LOG_ERROR("%s: total number of elements in tensor '%s' with shape "
                     "(%" PRIi64 ", %" PRIi64 ", %" PRIi64 ", %" PRIi64 ") is >= %" PRIi64 "\n",
                     __func__, info.t.name, info.t.ne[0], info.t.ne[1], info.t.ne[2], info.t.ne[3], INT64_MAX);
                 ok = false;
@@ -547,7 +547,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that tensor type is within defined range
             if (info.t.type < 0 || info.t.type >= GGML_TYPE_COUNT) {
-                fprintf(stderr, "%s: tensor '%s' has invalid ggml type %d (%s)\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d (%s)\n",
                     __func__, info.t.name, info.t.type, ggml_type_name(info.t.type));
                 ok = false;
                 break;
@@ -557,7 +557,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
             // check that row size is divisible by block size
             if (blck_size == 0 || info.t.ne[0] % blck_size != 0) {
-                fprintf(stderr, "%s: tensor '%s' of type %d (%s) has %" PRId64 " elements per row, "
+                GGML_LOG_ERROR("%s: tensor '%s' of type %d (%s) has %" PRId64 " elements per row, "
                     "not a multiple of block size (%" PRId64 ")\n",
                     __func__, info.t.name, (int) info.t.type, ggml_type_name(info.t.type), info.t.ne[0], blck_size);
                 ok = false;
@@ -582,7 +582,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
     }
 
     if (!ok) {
-        fprintf(stderr, "%s: failed to read tensor info\n", __func__);
+        GGML_LOG_ERROR("%s: failed to read tensor info\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -590,7 +590,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
     // we require the data section to be aligned, so take into account any padding
     if (fseek(file, GGML_PAD(ftell(file), ctx->alignment), SEEK_SET) != 0) {
-        fprintf(stderr, "%s: failed to seek to beginning of data section\n", __func__);
+        GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
         gguf_free(ctx);
         return nullptr;
     }
@@ -604,9 +604,9 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         for (size_t i = 0; i < ctx->info.size(); ++i) {
             const gguf_tensor_info & ti = ctx->info[i];
             if (ti.offset != ctx->size) {
-                fprintf(stderr, "%s: tensor '%s' has offset %" PRIu64 ", expected %zu\n",
+                GGML_LOG_ERROR("%s: tensor '%s' has offset %" PRIu64 ", expected %zu\n",
                     __func__, ti.t.name, ti.offset, ctx->size);
-                fprintf(stderr, "%s: failed to read tensor data\n", __func__);
+                GGML_LOG_ERROR("%s: failed to read tensor data\n", __func__);
                 gguf_free(ctx);
                 return nullptr;
             }
@@ -634,7 +634,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
         *params.ctx = ggml_init(pdata);
         if (*params.ctx == nullptr) {
-            fprintf(stderr, "%s: failed to initialize ggml context for storing tensors\n", __func__);
+            GGML_LOG_ERROR("%s: failed to initialize ggml context for storing tensors\n", __func__);
             gguf_free(ctx);
             return nullptr;
         }
@@ -656,7 +656,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
             ok = ok && gr.read(data->data, ctx->size);
 
             if (!ok) {
-                fprintf(stderr, "%s: failed to read tensor data binary blob\n", __func__);
+                GGML_LOG_ERROR("%s: failed to read tensor data binary blob\n", __func__);
                 ggml_free(ctx_data);
                 *params.ctx = nullptr;
                 gguf_free(ctx);
@@ -689,7 +689,7 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
         }
 
         if (!ok) {
-            fprintf(stderr, "%s: failed to create tensors\n", __func__);
+            GGML_LOG_ERROR("%s: failed to create tensors\n", __func__);
             ggml_free(ctx_data);
             *params.ctx = nullptr;
             gguf_free(ctx);
@@ -706,7 +706,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     FILE * file = ggml_fopen(fname, "rb");
 
     if (!file) {
-        fprintf(stderr, "%s: failed to open GGUF file '%s'\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open GGUF file '%s'\n", __func__, fname);
         return nullptr;
     }
 
@@ -1305,7 +1305,7 @@ bool gguf_write_to_file(const struct gguf_context * ctx, const char * fname, boo
     FILE * file = ggml_fopen(fname, "wb");
 
     if (!file) {
-        fprintf(stderr, "%s: failed to open file '%s' for writing GGUF data\n", __func__, fname);
+        GGML_LOG_ERROR("%s: failed to open file '%s' for writing GGUF data\n", __func__, fname);
         return false;
     }
 

gguf-py/gguf/constants.py

@@ -483,7 +483,9 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_EMBD_PATCH     = auto()
     V_ENC_EMBD_POS       = auto()
     V_ENC_ATTN_Q         = auto()
+    V_ENC_ATTN_Q_NORM    = auto()
     V_ENC_ATTN_K         = auto()
+    V_ENC_ATTN_K_NORM    = auto()
     V_ENC_ATTN_V         = auto()
     V_ENC_INPUT_NORM     = auto()
     V_ENC_OUTPUT         = auto()
@@ -742,7 +744,9 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
+    MODEL_TENSOR.V_ENC_ATTN_Q_NORM:         "v.blk.{bid}.attn_q_norm",
     MODEL_TENSOR.V_ENC_ATTN_K:              "v.blk.{bid}.attn_k",
+    MODEL_TENSOR.V_ENC_ATTN_K_NORM:         "v.blk.{bid}.attn_k_norm",
     MODEL_TENSOR.V_ENC_ATTN_V:              "v.blk.{bid}.attn_v",
     MODEL_TENSOR.V_ENC_INPUT_NORM:          "v.blk.{bid}.ln1",
     MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
@@ -782,7 +786,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_ATTN_Q,
+        MODEL_TENSOR.V_ENC_ATTN_Q_NORM,
         MODEL_TENSOR.V_ENC_ATTN_K,
+        MODEL_TENSOR.V_ENC_ATTN_K_NORM,
         MODEL_TENSOR.V_ENC_ATTN_V,
         MODEL_TENSOR.V_ENC_INPUT_NORM,
         MODEL_TENSOR.V_ENC_OUTPUT,
@@ -1899,6 +1905,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,
         MODEL_TENSOR.FFN_UP_EXP,
+        MODEL_TENSOR.FFN_GATE_SHEXP,
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
     ],
     MODEL_ARCH.CHAMELEON: [
         MODEL_TENSOR.TOKEN_EMBD,

gguf-py/gguf/scripts/gguf_editor_gui.py

@@ -823,6 +823,7 @@ class GGUFEditorWindow(QMainWindow):
         self.modified = False
         self.metadata_changes = {}  # Store changes to apply when saving
         self.metadata_to_remove = set()  # Store keys to remove when saving
+        self.on_metadata_changed_is_connected = False
 
         self.setup_ui()
 
@@ -941,9 +942,11 @@ class GGUFEditorWindow(QMainWindow):
             return
 
         # Disconnect to prevent triggering during loading
-        with warnings.catch_warnings():
-            warnings.filterwarnings('ignore')
-            self.metadata_table.itemChanged.disconnect(self.on_metadata_changed)
+        if self.on_metadata_changed_is_connected:
+            with warnings.catch_warnings():
+                warnings.filterwarnings('ignore')
+                self.metadata_table.itemChanged.disconnect(self.on_metadata_changed)
+            self.on_metadata_changed_is_connected = False
 
         for i, (key, field) in enumerate(self.reader.fields.items()):
             self.metadata_table.insertRow(i)
@@ -1021,6 +1024,7 @@ class GGUFEditorWindow(QMainWindow):
 
         # Reconnect after loading
         self.metadata_table.itemChanged.connect(self.on_metadata_changed)
+        self.on_metadata_changed_is_connected = True
 
     def extract_array_values(self, field: ReaderField) -> list:
         """Extract all values from an array field."""

gguf-py/gguf/tensor_mapping.py

@@ -68,7 +68,7 @@ class TensorNameMap:
             "output_layer",              # chatglm
             "head",                      # rwkv
             "head.out",                  # wavtokenizer
-            "language_model.lm_head",    # llama4
+            "lm_head",                   # llama4
         ),
 
         # Output norm
@@ -91,7 +91,7 @@ class TensorNameMap:
             "rwkv.ln_out",                             # rwkv6
             "model.ln_out",                            # rwkv7
             "backbone.final_layer_norm",               # wavtokenizer
-            "language_model.model.norm",               # llama4
+            "model.norm",                              # llama4
         ),
 
         # Rope frequencies
@@ -133,7 +133,7 @@ class TensorNameMap:
             "transformer.layers.{bid}.attn_norm",                   # openelm
             "rwkv.blocks.{bid}.ln1",                                # rwkv6
             "model.layers.{bid}.ln1",                               # rwkv7
-            "language_model.model.layers.{bid}.input_layernorm",    # llama4
+            "model.layers.{bid}.input_layernorm",                   # llama4
         ),
 
         # Attention norm 2
@@ -173,7 +173,7 @@ class TensorNameMap:
             "model.layers.{bid}.attention.wq",                           # internlm2
             "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
             "transformer.h.{bid}.attn.attention.q_proj",                 # exaone
-            "language_model.model.layers.{bid}.self_attn.q_proj",        # llama4
+            "model.layers.{bid}.self_attn.q_proj",                       # llama4
         ),
 
         # Attention key
@@ -188,7 +188,7 @@ class TensorNameMap:
             "model.layers.{bid}.attention.wk",                         # internlm2
             "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
             "transformer.h.{bid}.attn.attention.k_proj",               # exaone
-            "language_model.model.layers.{bid}.self_attn.k_proj",      # llama4
+            "model.layers.{bid}.self_attn.k_proj",                     # llama4
         ),
 
         # Attention value
@@ -202,7 +202,7 @@ class TensorNameMap:
             "model.layers.{bid}.attention.wv",                           # internlm2
             "transformer.decoder_layer.{bid}.multi_head_attention.value",# Grok
             "transformer.h.{bid}.attn.attention.v_proj",                 # exaone
-            "language_model.model.layers.{bid}.self_attn.v_proj",        # llama4
+            "model.layers.{bid}.self_attn.v_proj",                       # llama4
         ),
 
         # Attention output
@@ -229,7 +229,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.self_attention.dense",                    # chatglm
             "transformer.layers.{bid}.attn.out_proj",                       # openelm
             "transformer.h.{bid}.attn.attention.out_proj",                  # exaone
-            "language_model.model.layers.{bid}.self_attn.o_proj",           # llama4
+            "model.layers.{bid}.self_attn.o_proj",                          # llama4
         ),
 
         # Attention output norm
@@ -268,7 +268,7 @@ class TensorNameMap:
             "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
             "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
             "transformer.layers.{bid}.ffn_norm",                             # openelm
-            "language_model.model.layers.{bid}.post_attention_layernorm",    # llama4
+            "model.layers.{bid}.post_attention_layernorm",                   # llama4
         ),
 
         # Post feed-forward norm
@@ -289,7 +289,7 @@ class TensorNameMap:
             "transformer.decoder_layer.{bid}.router",           # Grok
             "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
             "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
-            "language_model.model.layers.{bid}.feed_forward.router", # llama4
+            "model.layers.{bid}.feed_forward.router",           # llama4
             "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
         ),
 
@@ -329,7 +329,7 @@ class TensorNameMap:
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
             "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
             "transformer.h.{bid}.mlp.c_fc_1",                         # exaone
-            "language_model.model.layers.{bid}.feed_forward.up_proj", # llama4
+            "model.layers.{bid}.feed_forward.up_proj",                # llama4
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -338,14 +338,14 @@ class TensorNameMap:
             "transformer.blocks.{bid}.ffn.experts.mlp.v1",    # dbrx
             "model.layers.{bid}.mlp.experts.up_proj",         # qwen2moe olmoe (merged)
             "model.layers.{bid}.block_sparse_moe.experts.w3", # phimoe (merged)
-            "language_model.model.layers.{bid}.feed_forward.experts.up_proj", # llama4
+            "model.layers.{bid}.feed_forward.experts.up_proj", # llama4
             "encoder.layers.{bid}.mlp.experts.mlp.w1",        # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_UP_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.up_proj",  # qwen2moe
-            "model.layers.{bid}.mlp.shared_experts.up_proj", # deepseek deepseek2
-            "language_model.model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4
+            "model.layers.{bid}.mlp.shared_expert.up_proj",          # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.up_proj",         # deepseek deepseek2
+            "model.layers.{bid}.feed_forward.shared_expert.up_proj", # llama4
         ),
 
         # AWQ-activation gate
@@ -366,22 +366,22 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.linear_1",           # refact
             "model.layers.{bid}.residual_mlp.w1",         # arctic
             "transformer.h.{bid}.mlp.c_fc_0",             # exaone
-            "language_model.model.layers.{bid}.feed_forward.gate_proj", # llama4
+            "model.layers.{bid}.feed_forward.gate_proj",  # llama4
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
-            "layers.{bid}.feed_forward.experts.w1",           # mixtral (merged)
-            "transformer.decoder_layer.{bid}.moe.linear",     # Grok (merged)
-            "transformer.blocks.{bid}.ffn.experts.mlp.w1",    # dbrx
-            "model.layers.{bid}.mlp.experts.gate_proj",       # qwen2moe olmoe (merged)
-            "model.layers.{bid}.block_sparse_moe.experts.w1", # phimoe (merged)
-            "language_model.model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
+            "layers.{bid}.feed_forward.experts.w1",              # mixtral (merged)
+            "transformer.decoder_layer.{bid}.moe.linear",        # Grok (merged)
+            "transformer.blocks.{bid}.ffn.experts.mlp.w1",       # dbrx
+            "model.layers.{bid}.mlp.experts.gate_proj",          # qwen2moe olmoe (merged)
+            "model.layers.{bid}.block_sparse_moe.experts.w1",    # phimoe (merged)
+            "model.layers.{bid}.feed_forward.experts.gate_proj", # llama4
         ),
 
         MODEL_TENSOR.FFN_GATE_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.gate_proj",  # qwen2moe
-            "model.layers.{bid}.mlp.shared_experts.gate_proj", # deepseek deepseek2
-            "language_model.model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4
+            "model.layers.{bid}.mlp.shared_expert.gate_proj",          # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.gate_proj",         # deepseek deepseek2
+            "model.layers.{bid}.feed_forward.shared_expert.gate_proj", # llama4
         ),
 
         # Feed-forward down
@@ -410,7 +410,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.mlp.down_layer",                     # jina-bert-v2
             "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
             "model.layers.h.{bid}.mlp.c_proj",                        # exaone
-            "language_model.model.layers.{bid}.feed_forward.down_proj", # llama4
+            "model.layers.{bid}.feed_forward.down_proj",              # llama4
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -420,14 +420,15 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.experts.down_proj",          # qwen2moe olmoe (merged)
             "model.layers.{bid}.block_sparse_moe.output_linear", # granitemoe
             "model.layers.{bid}.block_sparse_moe.experts.w2",    # phimoe (merged)
-            "language_model.model.layers.{bid}.feed_forward.experts.down_proj", # llama4
+            "model.layers.{bid}.feed_forward.experts.down_proj", # llama4
             "encoder.layers.{bid}.mlp.experts.mlp.w2",           # nomic-bert-moe
         ),
 
         MODEL_TENSOR.FFN_DOWN_SHEXP: (
-            "model.layers.{bid}.mlp.shared_expert.down_proj",  # qwen2moe
-            "model.layers.{bid}.mlp.shared_experts.down_proj", # deepseek deepseek2
-            "language_model.model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
+            "model.layers.{bid}.mlp.shared_expert.down_proj",          # qwen2moe
+            "model.layers.{bid}.mlp.shared_experts.down_proj",         # deepseek deepseek2
+            "model.layers.{bid}.feed_forward.shared_expert.down_proj", # llama4
+            "model.layers.{bid}.shared_mlp.output_linear",             # granitemoe
         ),
 
         MODEL_TENSOR.ATTN_Q_NORM: (
@@ -938,6 +939,10 @@ class TensorNameMap:
             "visual.blocks.{bid}.attn.q", # qwen2vl, generated
         ),
 
+        MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
+            "vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL
+        ),
+
         MODEL_TENSOR.V_ENC_ATTN_K: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
@@ -946,6 +951,10 @@ class TensorNameMap:
             "visual.blocks.{bid}.attn.k", # qwen2vl, generated
         ),
 
+        MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
+            "vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL
+        ),
+
         MODEL_TENSOR.V_ENC_ATTN_V: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
             "vpm.encoder.layers.{bid}.self_attn.v_proj",

include/llama.h

@@ -4,6 +4,7 @@
 #include "ggml.h"
 #include "ggml-cpu.h"
 #include "ggml-backend.h"
+#include "ggml-opt.h"
 
 #include <stddef.h>
 #include <stdint.h>
@@ -112,6 +113,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
         LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
+        LLAMA_VOCAB_PRE_TYPE_SEED_CODER     = 35,
     };
 
     enum llama_rope_type {
@@ -343,7 +345,7 @@ extern "C" {
         float    yarn_beta_fast;   // YaRN low correction dim
         float    yarn_beta_slow;   // YaRN high correction dim
         uint32_t yarn_orig_ctx;    // YaRN original context size
-        float    defrag_thold;     // defragment the KV cache if holes/size > thold, < 0 disabled (default)
+        float    defrag_thold;     // defragment the KV cache if holes/size > thold, <= 0 disabled (default)
 
         ggml_backend_sched_eval_callback cb_eval;
         void * cb_eval_user_data;
@@ -362,6 +364,7 @@ extern "C" {
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
         bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]
         bool no_perf;     // whether to measure performance timings
+        bool op_offload;  // whether to offload host tensor operations to device
     };
 
     // model quantization parameters
@@ -443,6 +446,10 @@ extern "C" {
                                  size_t    n_paths,
               struct llama_model_params    params);
 
+    LLAMA_API void llama_model_save_to_file(
+            const struct llama_model * model,
+                        const char * path_model);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -1431,6 +1438,37 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
+    //
+    // training
+    //
+
+    // function that returns whether or not a given tensor contains trainable parameters
+    typedef bool (*llama_opt_param_filter)(const struct ggml_tensor * tensor, void * userdata);
+
+    // always returns true
+    LLAMA_API bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata);
+
+    struct llama_opt_params {
+        uint32_t n_ctx_train; // assumed context size post training, use context size specified in llama_context if 0
+
+        llama_opt_param_filter param_filter; // callback for determining which tensors contain trainable parameters
+        void * param_filter_ud;              // userdata for determining which tensors contain trainable parameters
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);
+
+    LLAMA_API void llama_opt_epoch(
+            struct llama_context    * lctx,
+            ggml_opt_dataset_t        dataset,
+            ggml_opt_result_t         result_train,
+            ggml_opt_result_t         result_eval,
+            int64_t                   idata_split,
+            ggml_opt_epoch_callback   callback_train,
+            ggml_opt_epoch_callback   callback_eval);
+
 #ifdef __cplusplus
 }
 #endif

scripts/compare-llama-bench.py

@@ -7,6 +7,10 @@ import sys
 import os
 from glob import glob
 import sqlite3
+import json
+import csv
+from typing import Optional, Union
+from collections.abc import Iterator, Sequence
 
 try:
     import git
@@ -17,6 +21,28 @@ except ImportError as e:
 
 logger = logging.getLogger("compare-llama-bench")
 
+# All llama-bench SQL fields
+DB_FIELDS = [
+    "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
+    "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
+    "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
+    "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
+    "defrag_thold",
+    "use_mmap",     "embeddings",   "no_op_offload",  "n_prompt",   "n_gen",        "n_depth",
+    "test_time",    "avg_ns",       "stddev_ns",      "avg_ts",     "stddev_ts",
+]
+
+DB_TYPES = [
+    "TEXT",    "INTEGER", "TEXT",    "TEXT",    "TEXT",    "TEXT",
+    "TEXT",    "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
+    "TEXT",    "INTEGER", "INTEGER", "TEXT",    "TEXT",    "INTEGER",
+    "TEXT",    "INTEGER", "INTEGER", "INTEGER", "TEXT",    "TEXT",
+    "REAL",
+    "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER", "INTEGER",
+    "TEXT",    "INTEGER", "INTEGER", "REAL",    "REAL",
+]
+assert len(DB_FIELDS) == len(DB_TYPES)
+
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
     "cpu_info", "gpu_info", "backends", "n_gpu_layers", "tensor_buft_overrides", "model_filename", "model_type",
@@ -42,7 +68,7 @@ DEFAULT_HIDE = ["model_filename"]  # Always hide these properties by default.
 GPU_NAME_STRIP = ["NVIDIA GeForce ", "Tesla ", "AMD Radeon "]  # Strip prefixes for smaller tables.
 MODEL_SUFFIX_REPLACE = {" - Small": "_S", " - Medium": "_M", " - Large": "_L"}
 
-DESCRIPTION = """Creates tables from llama-bench data written to an SQLite database. Example usage (Linux):
+DESCRIPTION = """Creates tables from llama-bench data written to multiple JSON/CSV files, a single JSONL file or SQLite database. Example usage (Linux):
 
 $ git checkout master
 $ make clean && make llama-bench
@@ -70,12 +96,13 @@ help_c = (
 )
 parser.add_argument("-c", "--compare", help=help_c)
 help_i = (
-    "Input SQLite file for comparing commits. "
+    "JSON/JSONL/SQLite/CSV files for comparing commits. "
+    "Specify multiple times to use multiple input files (JSON/CSV only). "
     "Defaults to 'llama-bench.sqlite' in the current working directory. "
     "If no such file is found and there is exactly one .sqlite file in the current directory, "
     "that file is instead used as input."
 )
-parser.add_argument("-i", "--input", help=help_i)
+parser.add_argument("-i", "--input", action="append", help=help_i)
 help_o = (
     "Output format for the table. "
     "Defaults to 'pipe' (GitHub compatible). "
@@ -86,7 +113,7 @@ parser.add_argument("-o", "--output", help=help_o, default="pipe")
 help_s = (
     "Columns to add to the table. "
     "Accepts a comma-separated list of values. "
-    f"Legal values: {', '.join(KEY_PROPERTIES[:-2])}. "
+    f"Legal values: {', '.join(KEY_PROPERTIES[:-3])}. "
     "Defaults to model name (model_type) and CPU and/or GPU name (cpu_info, gpu_info) "
     "plus any column where not all data points are the same. "
     "If the columns are manually specified, then the results for each unique combination of the "
@@ -110,119 +137,321 @@ if unknown_args:
     sys.exit(1)
 
 input_file = known_args.input
-if input_file is None and os.path.exists("./llama-bench.sqlite"):
-    input_file = "llama-bench.sqlite"
-if input_file is None:
+if not input_file and os.path.exists("./llama-bench.sqlite"):
+    input_file = ["llama-bench.sqlite"]
+if not input_file:
     sqlite_files = glob("*.sqlite")
     if len(sqlite_files) == 1:
-        input_file = sqlite_files[0]
+        input_file = sqlite_files
 
-if input_file is None:
+if not input_file:
     logger.error("Cannot find a suitable input file, please provide one.\n")
     parser.print_help()
     sys.exit(1)
 
-connection = sqlite3.connect(input_file)
-cursor = connection.cursor()
 
-build_len_min: int = cursor.execute("SELECT MIN(LENGTH(build_commit)) from test;").fetchone()[0]
-build_len_max: int = cursor.execute("SELECT MAX(LENGTH(build_commit)) from test;").fetchone()[0]
+class LlamaBenchData:
+    repo: Optional[git.Repo]
+    build_len_min: int
+    build_len_max: int
+    build_len: int = 8
+    builds: list[str] = []
+    check_keys = set(KEY_PROPERTIES + ["build_commit", "test_time", "avg_ts"])
 
-if build_len_min != build_len_max:
-    logger.warning(f"{input_file} contains commit hashes of differing lengths. It's possible that the wrong commits will be compared. "
-                   "Try purging the the database of old commits.")
-    cursor.execute(f"UPDATE test SET build_commit = SUBSTRING(build_commit, 1, {build_len_min});")
+    def __init__(self):
+        try:
+            self.repo = git.Repo(".", search_parent_directories=True)
+        except git.InvalidGitRepositoryError:
+            self.repo = None
 
-build_len: int = build_len_min
+    def _builds_init(self):
+        self.build_len = self.build_len_min
 
-builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
-builds = list(map(lambda b: b[0], builds))  # list[tuple[str]] -> list[str]
-
-if not builds:
-    raise RuntimeError(f"{input_file} does not contain any builds.")
-
-try:
-    repo = git.Repo(".", search_parent_directories=True)
-except git.InvalidGitRepositoryError:
-    repo = None
-
-
-def find_parent_in_data(commit: git.Commit):
-    """Helper function to find the most recent parent measured in number of commits for which there is data."""
-    heap: list[tuple[int, git.Commit]] = [(0, commit)]
-    seen_hexsha8 = set()
-    while heap:
-        depth, current_commit = heapq.heappop(heap)
-        current_hexsha8 = commit.hexsha[:build_len]
-        if current_hexsha8 in builds:
-            return current_hexsha8
-        for parent in commit.parents:
-            parent_hexsha8 = parent.hexsha[:build_len]
-            if parent_hexsha8 not in seen_hexsha8:
-                seen_hexsha8.add(parent_hexsha8)
-                heapq.heappush(heap, (depth + 1, parent))
-    return None
-
-
-def get_all_parent_hexsha8s(commit: git.Commit):
-    """Helper function to recursively get hexsha8 values for all parents of a commit."""
-    unvisited = [commit]
-    visited   = []
-
-    while unvisited:
-        current_commit = unvisited.pop(0)
-        visited.append(current_commit.hexsha[:build_len])
-        for parent in current_commit.parents:
-            if parent.hexsha[:build_len] not in visited:
-                unvisited.append(parent)
-
-    return visited
-
-
-def get_commit_name(hexsha8: str):
-    """Helper function to find a human-readable name for a commit if possible."""
-    if repo is None:
-        return hexsha8
-    for h in repo.heads:
-        if h.commit.hexsha[:build_len] == hexsha8:
-            return h.name
-    for t in repo.tags:
-        if t.commit.hexsha[:build_len] == hexsha8:
-            return t.name
-    return hexsha8
-
-
-def get_commit_hexsha8(name: str):
-    """Helper function to search for a commit given a human-readable name."""
-    if repo is None:
+    def _check_keys(self, keys: set) -> Optional[set]:
+        """Private helper method that checks against required data keys and returns missing ones."""
+        if not keys >= self.check_keys:
+            return self.check_keys - keys
         return None
-    for h in repo.heads:
-        if h.name == name:
-            return h.commit.hexsha[:build_len]
-    for t in repo.tags:
-        if t.name == name:
-            return t.commit.hexsha[:build_len]
-    for c in repo.iter_commits("--all"):
-        if c.hexsha[:build_len] == name[:build_len]:
-            return c.hexsha[:build_len]
-    return None
+
+    def find_parent_in_data(self, commit: git.Commit) -> Optional[str]:
+        """Helper method to find the most recent parent measured in number of commits for which there is data."""
+        heap: list[tuple[int, git.Commit]] = [(0, commit)]
+        seen_hexsha8 = set()
+        while heap:
+            depth, current_commit = heapq.heappop(heap)
+            current_hexsha8 = commit.hexsha[:self.build_len]
+            if current_hexsha8 in self.builds:
+                return current_hexsha8
+            for parent in commit.parents:
+                parent_hexsha8 = parent.hexsha[:self.build_len]
+                if parent_hexsha8 not in seen_hexsha8:
+                    seen_hexsha8.add(parent_hexsha8)
+                    heapq.heappush(heap, (depth + 1, parent))
+        return None
+
+    def get_all_parent_hexsha8s(self, commit: git.Commit) -> Sequence[str]:
+        """Helper method to recursively get hexsha8 values for all parents of a commit."""
+        unvisited = [commit]
+        visited   = []
+
+        while unvisited:
+            current_commit = unvisited.pop(0)
+            visited.append(current_commit.hexsha[:self.build_len])
+            for parent in current_commit.parents:
+                if parent.hexsha[:self.build_len] not in visited:
+                    unvisited.append(parent)
+
+        return visited
+
+    def get_commit_name(self, hexsha8: str) -> str:
+        """Helper method to find a human-readable name for a commit if possible."""
+        if self.repo is None:
+            return hexsha8
+        for h in self.repo.heads:
+            if h.commit.hexsha[:self.build_len] == hexsha8:
+                return h.name
+        for t in self.repo.tags:
+            if t.commit.hexsha[:self.build_len] == hexsha8:
+                return t.name
+        return hexsha8
+
+    def get_commit_hexsha8(self, name: str) -> Optional[str]:
+        """Helper method to search for a commit given a human-readable name."""
+        if self.repo is None:
+            return None
+        for h in self.repo.heads:
+            if h.name == name:
+                return h.commit.hexsha[:self.build_len]
+        for t in self.repo.tags:
+            if t.name == name:
+                return t.commit.hexsha[:self.build_len]
+        for c in self.repo.iter_commits("--all"):
+            if c.hexsha[:self.build_len] == name[:self.build_len]:
+                return c.hexsha[:self.build_len]
+        return None
+
+    def builds_timestamp(self, reverse: bool = False) -> Union[Iterator[tuple], Sequence[tuple]]:
+        """Helper method that gets rows of (build_commit, test_time) sorted by the latter."""
+        return []
+
+    def get_rows(self, properties: list[str], hexsha8_baseline: str, hexsha8_compare: str) -> Sequence[tuple]:
+        """
+        Helper method that gets table rows for some list of properties.
+        Rows are created by combining those where all provided properties are equal.
+        The resulting rows are then grouped by the provided properties and the t/s values are averaged.
+        The returned rows are unique in terms of property combinations.
+        """
+        return []
+
+
+class LlamaBenchDataSQLite3(LlamaBenchData):
+    connection: sqlite3.Connection
+    cursor: sqlite3.Cursor
+
+    def __init__(self):
+        super().__init__()
+        self.connection = sqlite3.connect(":memory:")
+        self.cursor = self.connection.cursor()
+        self.cursor.execute(f"CREATE TABLE test({', '.join(' '.join(x) for x in zip(DB_FIELDS, DB_TYPES))});")
+
+    def _builds_init(self):
+        if self.connection:
+            self.build_len_min = self.cursor.execute("SELECT MIN(LENGTH(build_commit)) from test;").fetchone()[0]
+            self.build_len_max = self.cursor.execute("SELECT MAX(LENGTH(build_commit)) from test;").fetchone()[0]
+
+            if self.build_len_min != self.build_len_max:
+                logger.warning("Data contains commit hashes of differing lengths. It's possible that the wrong commits will be compared. "
+                               "Try purging the the database of old commits.")
+                self.cursor.execute(f"UPDATE test SET build_commit = SUBSTRING(build_commit, 1, {self.build_len_min});")
+
+            builds = self.cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
+            self.builds = list(map(lambda b: b[0], builds))  # list[tuple[str]] -> list[str]
+        super()._builds_init()
+
+    def builds_timestamp(self, reverse: bool = False) -> Union[Iterator[tuple], Sequence[tuple]]:
+        data = self.cursor.execute(
+            "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall()
+        return reversed(data) if reverse else data
+
+    def get_rows(self, properties: list[str], hexsha8_baseline: str, hexsha8_compare: str) -> Sequence[tuple]:
+        select_string = ", ".join(
+            [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
+        equal_string = " AND ".join(
+            [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
+                f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
+        )
+        group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
+        query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
+                 f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
+        return self.cursor.execute(query).fetchall()
+
+
+class LlamaBenchDataSQLite3File(LlamaBenchDataSQLite3):
+    def __init__(self, data_file: str):
+        super().__init__()
+
+        self.connection.close()
+        self.connection = sqlite3.connect(data_file)
+        self.cursor = self.connection.cursor()
+        self._builds_init()
+
+    @staticmethod
+    def valid_format(data_file: str) -> bool:
+        connection = sqlite3.connect(data_file)
+        cursor = connection.cursor()
+
+        try:
+            if cursor.execute("PRAGMA schema_version;").fetchone()[0] == 0:
+                raise sqlite3.DatabaseError("The provided input file does not exist or is empty.")
+        except sqlite3.DatabaseError as e:
+            logger.debug(f'"{data_file}" is not a valid SQLite3 file.', exc_info=e)
+            cursor = None
+
+        connection.close()
+        return True if cursor else False
+
+
+class LlamaBenchDataJSONL(LlamaBenchDataSQLite3):
+    def __init__(self, data_file: str):
+        super().__init__()
+
+        with open(data_file, "r", encoding="utf-8") as fp:
+            for i, line in enumerate(fp):
+                parsed = json.loads(line)
+
+                for k in parsed.keys() - set(DB_FIELDS):
+                    del parsed[k]
+
+                if (missing_keys := self._check_keys(parsed.keys())):
+                    raise RuntimeError(f"Missing required data key(s) at line {i + 1}: {', '.join(missing_keys)}")
+
+                self.cursor.execute(f"INSERT INTO test({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
+
+        self._builds_init()
+
+    @staticmethod
+    def valid_format(data_file: str) -> bool:
+        try:
+            with open(data_file, "r", encoding="utf-8") as fp:
+                for line in fp:
+                    json.loads(line)
+                    break
+        except Exception as e:
+            logger.debug(f'"{data_file}" is not a valid JSONL file.', exc_info=e)
+            return False
+
+        return True
+
+
+class LlamaBenchDataJSON(LlamaBenchDataSQLite3):
+    def __init__(self, data_files: list[str]):
+        super().__init__()
+
+        for data_file in data_files:
+            with open(data_file, "r", encoding="utf-8") as fp:
+                parsed = json.load(fp)
+
+                for i, entry in enumerate(parsed):
+                    for k in entry.keys() - set(DB_FIELDS):
+                        del entry[k]
+
+                    if (missing_keys := self._check_keys(entry.keys())):
+                        raise RuntimeError(f"Missing required data key(s) at entry {i + 1}: {', '.join(missing_keys)}")
+
+                    self.cursor.execute(f"INSERT INTO test({', '.join(entry.keys())}) VALUES({', '.join('?' * len(entry))});", tuple(entry.values()))
+
+        self._builds_init()
+
+    @staticmethod
+    def valid_format(data_files: list[str]) -> bool:
+        if not data_files:
+            return False
+
+        for data_file in data_files:
+            try:
+                with open(data_file, "r", encoding="utf-8") as fp:
+                    json.load(fp)
+            except Exception as e:
+                logger.debug(f'"{data_file}" is not a valid JSON file.', exc_info=e)
+                return False
+
+        return True
+
+
+class LlamaBenchDataCSV(LlamaBenchDataSQLite3):
+    def __init__(self, data_files: list[str]):
+        super().__init__()
+
+        for data_file in data_files:
+            with open(data_file, "r", encoding="utf-8") as fp:
+                for i, parsed in enumerate(csv.DictReader(fp)):
+                    keys = set(parsed.keys())
+
+                    for k in keys - set(DB_FIELDS):
+                        del parsed[k]
+
+                    if (missing_keys := self._check_keys(keys)):
+                        raise RuntimeError(f"Missing required data key(s) at line {i + 1}: {', '.join(missing_keys)}")
+
+                    self.cursor.execute(f"INSERT INTO test({', '.join(parsed.keys())}) VALUES({', '.join('?' * len(parsed))});", tuple(parsed.values()))
+
+        self._builds_init()
+
+    @staticmethod
+    def valid_format(data_files: list[str]) -> bool:
+        if not data_files:
+            return False
+
+        for data_file in data_files:
+            try:
+                with open(data_file, "r", encoding="utf-8") as fp:
+                    for parsed in csv.DictReader(fp):
+                        break
+            except Exception as e:
+                logger.debug(f'"{data_file}" is not a valid CSV file.', exc_info=e)
+                return False
+
+        return True
+
+
+bench_data = None
+if len(input_file) == 1:
+    if LlamaBenchDataSQLite3File.valid_format(input_file[0]):
+        bench_data = LlamaBenchDataSQLite3File(input_file[0])
+    elif LlamaBenchDataJSON.valid_format(input_file):
+        bench_data = LlamaBenchDataJSON(input_file)
+    elif LlamaBenchDataJSONL.valid_format(input_file[0]):
+        bench_data = LlamaBenchDataJSONL(input_file[0])
+    elif LlamaBenchDataCSV.valid_format(input_file):
+        bench_data = LlamaBenchDataCSV(input_file)
+else:
+    if LlamaBenchDataJSON.valid_format(input_file):
+        bench_data = LlamaBenchDataJSON(input_file)
+    elif LlamaBenchDataCSV.valid_format(input_file):
+        bench_data = LlamaBenchDataCSV(input_file)
+
+if not bench_data:
+    raise RuntimeError("No valid (or some invalid) input files found.")
+
+if not bench_data.builds:
+    raise RuntimeError(f"{input_file} does not contain any builds.")
 
 
 hexsha8_baseline = name_baseline = None
 
 # If the user specified a baseline, try to find a commit for it:
 if known_args.baseline is not None:
-    if known_args.baseline in builds:
+    if known_args.baseline in bench_data.builds:
         hexsha8_baseline = known_args.baseline
     if hexsha8_baseline is None:
-        hexsha8_baseline = get_commit_hexsha8(known_args.baseline)
+        hexsha8_baseline = bench_data.get_commit_hexsha8(known_args.baseline)
         name_baseline = known_args.baseline
     if hexsha8_baseline is None:
         logger.error(f"cannot find data for baseline={known_args.baseline}.")
         sys.exit(1)
 # Otherwise, search for the most recent parent of master for which there is data:
-elif repo is not None:
-    hexsha8_baseline = find_parent_in_data(repo.heads.master.commit)
+elif bench_data.repo is not None:
+    hexsha8_baseline = bench_data.find_parent_in_data(bench_data.repo.heads.master.commit)
 
     if hexsha8_baseline is None:
         logger.error("No baseline was provided and did not find data for any master branch commits.\n")
@@ -235,27 +464,25 @@ else:
     sys.exit(1)
 
 
-name_baseline = get_commit_name(hexsha8_baseline)
+name_baseline = bench_data.get_commit_name(hexsha8_baseline)
 
 hexsha8_compare = name_compare = None
 
 # If the user has specified a compare value, try to find a corresponding commit:
 if known_args.compare is not None:
-    if known_args.compare in builds:
+    if known_args.compare in bench_data.builds:
         hexsha8_compare = known_args.compare
     if hexsha8_compare is None:
-        hexsha8_compare = get_commit_hexsha8(known_args.compare)
+        hexsha8_compare = bench_data.get_commit_hexsha8(known_args.compare)
         name_compare = known_args.compare
     if hexsha8_compare is None:
         logger.error(f"cannot find data for compare={known_args.compare}.")
         sys.exit(1)
 # Otherwise, search for the commit for llama-bench was most recently run
 # and that is not a parent of master:
-elif repo is not None:
-    hexsha8s_master = get_all_parent_hexsha8s(repo.heads.master.commit)
-    builds_timestamp = cursor.execute(
-        "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall()
-    for (hexsha8, _) in reversed(builds_timestamp):
+elif bench_data.repo is not None:
+    hexsha8s_master = bench_data.get_all_parent_hexsha8s(bench_data.repo.heads.master.commit)
+    for (hexsha8, _) in bench_data.builds_timestamp(reverse=True):
         if hexsha8 not in hexsha8s_master:
             hexsha8_compare = hexsha8
             break
@@ -270,26 +497,7 @@ else:
     parser.print_help()
     sys.exit(1)
 
-name_compare = get_commit_name(hexsha8_compare)
-
-
-def get_rows(properties):
-    """
-    Helper function that gets table rows for some list of properties.
-    Rows are created by combining those where all provided properties are equal.
-    The resulting rows are then grouped by the provided properties and the t/s values are averaged.
-    The returned rows are unique in terms of property combinations.
-    """
-    select_string = ", ".join(
-        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "tb.n_depth", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
-    equal_string = " AND ".join(
-        [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
-            f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
-    )
-    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt", "tb.n_depth"])
-    query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
-             f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
-    return cursor.execute(query).fetchall()
+name_compare = bench_data.get_commit_name(hexsha8_compare)
 
 
 # If the user provided columns to group the results by, use them:
@@ -297,16 +505,16 @@ if known_args.show is not None:
     show = known_args.show.split(",")
     unknown_cols = []
     for prop in show:
-        if prop not in KEY_PROPERTIES[:-2]:  # Last two values are n_prompt, n_gen.
+        if prop not in KEY_PROPERTIES[:-3]:  # Last three values are n_prompt, n_gen, n_depth.
             unknown_cols.append(prop)
     if unknown_cols:
         logger.error(f"Unknown values for --show: {', '.join(unknown_cols)}")
         parser.print_usage()
         sys.exit(1)
-    rows_show = get_rows(show)
+    rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
 # Otherwise, select those columns where the values are not all the same:
 else:
-    rows_full = get_rows(KEY_PROPERTIES)
+    rows_full = bench_data.get_rows(KEY_PROPERTIES, hexsha8_baseline, hexsha8_compare)
     properties_different = []
     for i, kp_i in enumerate(KEY_PROPERTIES):
         if kp_i in DEFAULT_SHOW or kp_i in ["n_prompt", "n_gen", "n_depth"]:
@@ -318,7 +526,7 @@ else:
 
     show = []
     # Show CPU and/or GPU by default even if the hardware for all results is the same:
-    if "n_gpu_layers" not in properties_different:
+    if rows_full and "n_gpu_layers" not in properties_different:
         ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
 
         if ngl != 99 and "cpu_info" not in properties_different:
@@ -336,7 +544,11 @@ else:
             show.remove(prop)
         except ValueError:
             pass
-    rows_show = get_rows(show)
+    rows_show = bench_data.get_rows(show, hexsha8_baseline, hexsha8_compare)
+
+if not rows_show:
+    logger.error(f"No comparable data was found between {name_baseline} and {name_compare}.\n")
+    sys.exit(1)
 
 table = []
 for row in rows_show:

scripts/sync-ggml.last

@@ -1 +1 @@
-b59bddafe278877dfa22a80e53a637513862babb
+9b048bb72b811f50b0c30d9e5c84d6ff9f4bf005

src/CMakeLists.txt

@@ -23,6 +23,7 @@ add_library(llama
             llama-memory.cpp
             llama-mmap.cpp
             llama-model-loader.cpp
+            llama-model-saver.cpp
             llama-model.cpp
             llama-quant.cpp
             llama-sampling.cpp

src/llama-arch.cpp

@@ -1481,6 +1481,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { 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_GATE_SHEXP,  "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,  "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,    "blk.%d.ffn_up_shexp" },
         },
     },
     {

src/llama-context.cpp

@@ -93,6 +93,7 @@ llama_context::llama_context(
     }
 
     cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
+    cparams.op_offload = params.op_offload;
 
     const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max;
 
@@ -243,7 +244,7 @@ llama_context::llama_context(
             }
         }
 
-        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel));
+        sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, pipeline_parallel, cparams.op_offload));
 
         if (pipeline_parallel) {
             LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get()));
@@ -358,7 +359,9 @@ llama_context::llama_context(
     }
 }
 
-llama_context::~llama_context() = default;
+llama_context::~llama_context() {
+    ggml_opt_free(opt_ctx);
+}
 
 void llama_context::synchronize() {
     ggml_backend_sched_synchronize(sched.get());
@@ -1701,10 +1704,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
         }
     }
 
-    LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
     llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-    kv_self->state_write(io);
+    if (kv_self != nullptr) {
+        LLAMA_LOG_DEBUG("%s: - writing KV self\n", __func__);
+        kv_self->state_write(io);
+    }
 
     return io.n_bytes();
 }
@@ -1787,10 +1792,13 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
         }
     }
 
-    LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        LLAMA_LOG_DEBUG("%s: - reading KV self\n", __func__);
 
-    kv_self->state_read(io);
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
+        kv_self->state_read(io);
+    }
 
     return io.n_bytes();
 }
@@ -1798,9 +1806,11 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
 size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-    kv_self->state_write(io, seq_id);
+        kv_self->state_write(io, seq_id);
+    }
 
     return io.n_bytes();
 }
@@ -1808,9 +1818,11 @@ size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id s
 size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
     GGML_UNUSED(seq_id);
 
-    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+    if (memory) {
+        llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
 
-    kv_self->state_read(io, seq_id);
+        kv_self->state_read(io, seq_id);
+    }
 
     return io.n_bytes();
 }
@@ -1838,6 +1850,215 @@ void llama_context::perf_reset() {
     t_p_eval_us = n_p_eval = 0;
 }
 
+//
+// training
+//
+
+static void llama_set_param(struct ggml_tensor * tensor, llama_opt_param_filter param_filter, void * userdata) {
+    if (!tensor || tensor->type != GGML_TYPE_F32) {
+        return;
+    }
+    if (!param_filter(tensor, userdata)) {
+        return;
+    }
+    if (strcmp(tensor->name, "token_embd.weight") == 0) {
+        return; // FIXME
+    }
+    if (strcmp(tensor->name, "rope_freqs.weight") == 0) {
+        return; // FIXME
+    }
+    ggml_set_param(tensor);
+}
+
+void llama_context::opt_init(struct llama_model * model, struct llama_opt_params lopt_params) {
+    GGML_ASSERT(!opt_ctx);
+    model->hparams.n_ctx_train = lopt_params.n_ctx_train > 0 ? lopt_params.n_ctx_train : n_ctx();
+    const uint32_t n_batch     = std::min(this->n_batch(),  model->hparams.n_ctx_train);
+    const uint32_t n_ubatch    = std::min(this->n_ubatch(), n_batch);
+    GGML_ASSERT(model->hparams.n_ctx_train % n_batch  == 0);
+    GGML_ASSERT(n_batch                    % n_ubatch == 0);
+
+    ggml_opt_params opt_params = ggml_opt_default_params(sched.get(), GGML_OPT_LOSS_TYPE_CROSS_ENTROPY);
+    opt_params.opt_period      = n_batch / n_ubatch;
+    opt_params.get_opt_pars    = lopt_params.get_opt_pars;
+    opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
+
+    opt_ctx = ggml_opt_init(opt_params);
+
+    llama_opt_param_filter param_filter = lopt_params.param_filter;
+    void * param_filter_ud              = lopt_params.param_filter_ud;
+
+  //llama_set_param(model->tok_embd,        param_filter, param_filter_ud); // FIXME
+    llama_set_param(model->type_embd,       param_filter, param_filter_ud);
+    llama_set_param(model->pos_embd,        param_filter, param_filter_ud);
+    llama_set_param(model->tok_norm,        param_filter, param_filter_ud);
+    llama_set_param(model->tok_norm_b,      param_filter, param_filter_ud);
+    llama_set_param(model->output_norm,     param_filter, param_filter_ud);
+    llama_set_param(model->output_norm_b,   param_filter, param_filter_ud);
+    llama_set_param(model->output,          param_filter, param_filter_ud);
+    llama_set_param(model->output_b,        param_filter, param_filter_ud);
+    llama_set_param(model->output_norm_enc, param_filter, param_filter_ud);
+    llama_set_param(model->cls,             param_filter, param_filter_ud);
+    llama_set_param(model->cls_b,           param_filter, param_filter_ud);
+    llama_set_param(model->cls_out,         param_filter, param_filter_ud);
+    llama_set_param(model->cls_out_b,       param_filter, param_filter_ud);
+
+    for (struct llama_layer & layer : model->layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            llama_set_param(reinterpret_cast<struct ggml_tensor **>(&layer)[i], param_filter, param_filter_ud);
+        }
+    }
+}
+
+void llama_context::opt_epoch_iter(
+        ggml_opt_dataset_t               dataset,
+        ggml_opt_result_t                result,
+        const std::vector<llama_token> & tokens,
+        const std::vector<llama_token> & labels_sparse,
+        llama_batch                    & batch,
+        ggml_opt_epoch_callback          callback,
+        bool                             train,
+        int64_t                          idata_in_loop,
+        int64_t                          ndata_in_loop,
+        int64_t                          t_loop_start) {
+    GGML_ASSERT(opt_ctx);
+    const uint32_t n_ctx    = llama_model_n_ctx_train(&model);
+    const uint32_t n_batch  = std::min(this->n_batch(),  n_ctx);
+    const uint32_t n_ubatch = std::min(this->n_ubatch(), n_batch);
+
+    llama_kv_cache * kv_self = static_cast<llama_kv_cache *>(memory.get());
+
+    kv_self->clear();
+    llama_kv_cache_guard kv_guard(kv_self);
+
+    for (uint32_t pos_ctx = 0; pos_ctx < n_ctx; pos_ctx += n_batch) {
+        batch.n_tokens = n_batch;
+        for (uint32_t pos_batch = 0; pos_batch < n_batch; ++pos_batch) {
+            batch.token   [pos_batch]    = tokens[pos_ctx + pos_batch];
+            batch.pos     [pos_batch]    = pos_ctx + pos_batch;
+            batch.n_seq_id[pos_batch]    = 1;
+            batch.seq_id  [pos_batch][0] = 0;
+            batch.logits  [pos_batch]    = true;
+        }
+
+        const auto n_tokens_all = batch.n_tokens;
+
+        n_queued_tokens += n_tokens_all;
+
+        // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
+        const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
+
+        embd_seq.clear();
+
+        int64_t n_outputs_all = n_tokens_all;
+
+        llama_sbatch sbatch = kv_self->sbatch_init(batch, /*logits_all =*/ true);
+
+        // reserve output buffer
+        if (output_reserve(n_outputs_all) < n_outputs_all) {
+            LLAMA_LOG_ERROR("%s: could not reserve space for batch with %" PRId64 " outputs\n", __func__, n_outputs_all);
+            GGML_ABORT("TODO: handle this error");
+        };
+
+        for (uint32_t pos_batch = 0; pos_batch < n_batch; pos_batch += n_ubatch) {
+            llama_ubatch ubatch = kv_self->ubatch_next(sbatch, cparams.n_ubatch, embd_pooled);
+
+            n_outputs = ubatch.n_tokens;
+
+            // TODO: not sure if this is needed
+            if (!kv_self->find_slot(ubatch)) {
+                LLAMA_LOG_WARN("%s: failed to find KV cache slot for ubatch of size %d\n", __func__, ubatch.n_tokens);
+
+                GGML_ABORT("TODO: handle this error");
+            }
+
+            auto * gf = graph_init();
+            auto res = graph_build(ctx_compute.get(), gf, ubatch, LLM_GRAPH_TYPE_DEFAULT);
+
+            struct ggml_context * ctx_compute_opt;
+            {
+                const size_t size_gf = ggml_graph_size(gf);
+                const size_t size_meta = 4*size_gf*ggml_tensor_overhead() + 2*ggml_graph_overhead_custom(size_gf, /*grads = */ true);
+                struct ggml_init_params params = {
+                    /*.mem_size   =*/ size_meta,
+                    /*.mem_buffer =*/ nullptr,
+                    /*.no_alloc   =*/ true,
+                };
+                ctx_compute_opt = ggml_init(params);
+            }
+            ggml_opt_prepare_alloc(opt_ctx, ctx_compute_opt, gf, res->get_tokens(), res->get_logits());
+            ggml_opt_alloc(opt_ctx, train);
+            res->set_inputs(&ubatch);
+            {
+                struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
+                GGML_ASSERT(labels->ne[1] == n_ubatch);
+                ggml_set_zero(labels);
+                const float onef = 1.0f;
+                for (uint32_t pos_ubatch = 0; pos_ubatch < n_ubatch; ++pos_ubatch) {
+                    const uint32_t ilabel = pos_ctx + pos_batch + pos_ubatch;
+                    GGML_ASSERT(labels_sparse[ilabel] < labels->ne[0]);
+                    ggml_backend_tensor_set(labels, &onef, (pos_ubatch*labels->ne[0] + labels_sparse[ilabel])*sizeof(float), sizeof(float));
+                }
+            }
+            ggml_opt_eval(opt_ctx, result);
+            if (callback) {
+                callback(train, opt_ctx, dataset, result, idata_in_loop + (pos_ctx + pos_batch)/n_ubatch + 1, ndata_in_loop, t_loop_start);
+            }
+            ggml_free(ctx_compute_opt);
+        }
+    }
+
+    kv_guard.commit();
+}
+
+void llama_context::opt_epoch(
+        ggml_opt_dataset_t        dataset,
+        ggml_opt_result_t         result_train,
+        ggml_opt_result_t         result_eval,
+        int64_t                   idata_split,
+        ggml_opt_epoch_callback   callback_train,
+        ggml_opt_epoch_callback   callback_eval) {
+    const uint32_t n_ctx    = this->n_ctx();
+    const uint32_t n_batch  = std::min(cparams.n_batch,  n_ctx);
+    const uint32_t n_ubatch = std::min(cparams.n_ubatch, n_batch);
+    const  int64_t ndata    = ggml_opt_dataset_ndata(dataset);
+
+    GGML_ASSERT(idata_split >= 0);
+    GGML_ASSERT(idata_split <= ndata);
+
+    const uint32_t ubatch_per_ctx = n_ctx / n_ubatch;
+
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    std::vector<llama_token>        tokens(n_ctx);
+    std::vector<llama_token> labels_sparse(n_ctx);
+
+    int64_t idata = 0;
+
+    int64_t t_loop_start = ggml_time_us();
+    int64_t ndata_in_loop = idata_split*ubatch_per_ctx;
+    for (; idata < idata_split; ++idata) {
+        constexpr bool train = true;
+        const int64_t idata_in_loop = idata*ubatch_per_ctx;
+
+        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
+        opt_epoch_iter(dataset, result_train, tokens, labels_sparse, batch,
+            callback_train, train, idata_in_loop, ndata_in_loop, t_loop_start);
+    }
+
+    t_loop_start = ggml_time_us();
+    ndata_in_loop = (ndata - idata_split)*ubatch_per_ctx;
+    for (; idata < ndata; ++idata) {
+        constexpr bool train = false;
+        const int64_t idata_in_loop = (idata - idata_split)*ubatch_per_ctx;
+
+        ggml_opt_dataset_get_batch_host(dataset, tokens.data(), n_ctx*sizeof(llama_token), labels_sparse.data(), idata);
+        opt_epoch_iter(dataset, result_eval, tokens, labels_sparse, batch,
+            callback_eval, train, idata_in_loop, ndata_in_loop, t_loop_start);
+    }
+
+    llama_batch_free(batch);
+}
+
 //
 // interface implementation
 //
@@ -1871,6 +2092,7 @@ llama_context_params llama_context_default_params() {
         /*.offload_kqv                 =*/ true,
         /*.flash_attn                  =*/ false,
         /*.no_perf                     =*/ true,
+        /*.op_offload                  =*/ true,
     };
 
     return result;
@@ -2455,3 +2677,34 @@ void llama_perf_context_print(const llama_context * ctx) {
 void llama_perf_context_reset(llama_context * ctx) {
     ctx->perf_reset();
 }
+
+//
+// training
+//
+
+bool llama_opt_param_filter_all(const struct ggml_tensor * tensor, void * userdata) {
+    GGML_UNUSED(tensor);
+    GGML_UNUSED(userdata);
+    return true;
+}
+
+void llama_opt_init(struct llama_context * ctx, struct llama_model * model, struct llama_opt_params lopt_params) {
+    ctx->opt_init(model, lopt_params);
+}
+
+void llama_opt_epoch(
+        struct llama_context    * ctx,
+        ggml_opt_dataset_t        dataset,
+        ggml_opt_result_t         result_train,
+        ggml_opt_result_t         result_eval,
+        int64_t                   idata_split,
+        ggml_opt_epoch_callback   callback_train,
+        ggml_opt_epoch_callback   callback_eval) {
+    ctx->opt_epoch(
+        dataset,
+        result_train,
+        result_eval,
+        idata_split,
+        callback_train,
+        callback_eval);
+}

src/llama-context.h

@@ -7,6 +7,7 @@
 #include "llama-adapter.h"
 
 #include "ggml-cpp.h"
+#include "ggml-opt.h"
 
 #include <map>
 #include <vector>
@@ -133,6 +134,32 @@ struct llama_context {
     llama_perf_context_data perf_get_data() const;
     void perf_reset();
 
+    //
+    // training
+    //
+
+    void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
+
+    void opt_epoch(
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,
+            ggml_opt_result_t       result_eval,
+            int64_t                 idata_split,
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    void opt_epoch_iter(
+            ggml_opt_dataset_t               dataset,
+            ggml_opt_result_t                result,
+            const std::vector<llama_token> & tokens,
+            const std::vector<llama_token> & labels_sparse,
+            llama_batch                    & batch,
+            ggml_opt_epoch_callback          callback,
+            bool                             train,
+            int64_t                          idata_in_loop,
+            int64_t                          ndata_in_loop,
+            int64_t                          t_loop_start);
+
 private:
     //
     // output
@@ -212,6 +239,9 @@ private:
 
     ggml_context_ptr ctx_compute;
 
+    // training
+    ggml_opt_context_t opt_ctx = nullptr;
+
     ggml_threadpool_t threadpool       = nullptr;
     ggml_threadpool_t threadpool_batch = nullptr;
 

src/llama-cparams.h

@@ -30,6 +30,7 @@ struct llama_cparams {
     bool flash_attn;
     bool no_perf;
     bool warmup;
+    bool op_offload;
 
     enum llama_pooling_type pooling_type;
 

src/llama-graph.cpp

@@ -971,6 +971,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
         inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
         //cb(inp->tokens, "inp_tokens", -1);
         ggml_set_input(inp->tokens);
+        res->t_tokens = inp->tokens;
 
         cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
 

src/llama-graph.h

@@ -298,6 +298,7 @@ class llm_graph_result_i {
 public:
     virtual ~llm_graph_result_i() = default;
 
+    virtual ggml_tensor * get_tokens()      = 0;
     virtual ggml_tensor * get_logits()      = 0;
     virtual ggml_tensor * get_embd()        = 0;
     virtual ggml_tensor * get_embd_pooled() = 0;
@@ -312,6 +313,7 @@ class llm_graph_result : public llm_graph_result_i {
 public:
     virtual ~llm_graph_result() = default;
 
+    ggml_tensor * get_tokens()      override { return t_tokens; }
     ggml_tensor * get_logits()      override { return t_logits; }
     ggml_tensor * get_embd()        override { return t_embd; }
     ggml_tensor * get_embd_pooled() override { return t_embd_pooled; }
@@ -328,6 +330,7 @@ public:
     }
 
     // important graph nodes
+    ggml_tensor * t_tokens      = nullptr;
     ggml_tensor * t_logits      = nullptr;
     ggml_tensor * t_embd        = nullptr;
     ggml_tensor * t_embd_pooled = nullptr;

src/llama-kv-cache.cpp

@@ -441,6 +441,13 @@ void llama_kv_cache_unified::defrag_sched(float thold) {
 
 void llama_kv_cache_unified::set_full() {
     n = size;
+
+    // when simulating a full KV cache, the specific value of the "head" pointer is not important because it does not
+    //   affect the shapes of the tensors in the compute graph - it only affects the offsets of the K/V views.
+    //   we should only guarantee that the head position won't cause out-of-bounds view of the K, V tensors, so
+    //   setting it to 0 is the simplest way to achieve that
+    // ref: https://github.com/ggml-org/llama.cpp/issues/13359
+    head = 0;
 }
 
 llama_sbatch llama_kv_cache_unified::sbatch_init(
@@ -1712,6 +1719,7 @@ void llama_kv_cache_recurrent::defrag_sched(float thold) {
 
 void llama_kv_cache_recurrent::set_full() {
     n = size;
+    head = 0;
 }
 
 llama_sbatch llama_kv_cache_recurrent::sbatch_init(

src/llama-kv-cache.h

@@ -171,11 +171,8 @@ public:
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
     void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
-    // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_impl also uses it, so it
-    // cannot be freely changed after a slot has been allocated.
-    uint32_t head = 0;
-    uint32_t size = 0;
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
     uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build
@@ -343,11 +340,8 @@ public:
     void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
     void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
 
-    // Note: The value of head isn't only used to optimize searching
-    // for a free KV slot. llama_decode_impl also uses it, so it
-    // cannot be freely changed after a slot has been allocated.
-    uint32_t head = 0;
-    uint32_t size = 0;
+    uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
+    uint32_t size = 0; // total number of cells, shared across all sequences
     uint32_t used = 0; // used cells (i.e. at least one seq_id)
 
     // computed before each graph build

src/llama-model-loader.cpp

@@ -301,12 +301,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         result.resize(arr_info.length);
@@ -330,12 +330,12 @@ namespace GGUFMeta {
             GGUFMeta::GKV<GGUFMeta::ArrayInfo>::get_kv(meta.get(), kid);
 
         switch (arr_info.gt) {
-            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T, float>::value)); break;
-            case GGUF_TYPE_INT32:   GGML_ASSERT(
-                                            (std::is_same<T,  int32_t>::value) ||
-                                            (std::is_same<T, uint32_t>::value));  break;
+            case GGUF_TYPE_UINT32:
+            case GGUF_TYPE_INT32:   GGML_ASSERT((std::is_same<T,  int32_t>::value) ||
+                                                (std::is_same<T, uint32_t>::value)); break;
+            case GGUF_TYPE_FLOAT32: GGML_ASSERT((std::is_same<T,    float>::value)); break;
             default:
-                throw std::runtime_error(format("%s is not a float32, int32 array", key.c_str()));
+                throw std::runtime_error(format("%s is not a float32/uint32/int32 array", key.c_str()));
         }
 
         if (arr_info.length > N_MAX) {
@@ -469,7 +469,7 @@ llama_model_loader::llama_model_loader(
 
     meta.reset(gguf_init_from_file(fname.c_str(), params));
     if (!meta) {
-        throw std::runtime_error(format("%s: failed to load model from %s\n", __func__, fname.c_str()));
+        throw std::runtime_error(format("%s: failed to load model from %s", __func__, fname.c_str()));
     }
 
     get_key(llm_kv(LLM_KV_GENERAL_ARCHITECTURE), arch_name, false);
@@ -528,7 +528,7 @@ llama_model_loader::llama_model_loader(
             };
             gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
             if (!ctx_gguf) {
-                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, fname_split));
+                throw std::runtime_error(format("%s: failed to load GGUF split from %s", __func__, fname_split));
             }
 
             // check idx
@@ -822,13 +822,18 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         mappings.reserve(files.size());
         mmaps_used.reserve(files.size());
         for (const auto & file : files) {
-            auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
-            if (!reg) {
-                throw std::runtime_error(format("%s: no CPU backend found", __func__));
+            bool is_numa = false;
+
+            auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+            if (dev) {
+                auto * reg = ggml_backend_dev_backend_reg(dev);
+                auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
+                if (is_numa_fn) {
+                    is_numa = is_numa_fn();
+                }
             }
 
-            auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
-            std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
+            std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa);
             mmaps_used.emplace_back(mapping->size(), 0);
             if (mlock_mmaps) {
                 std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());

src/llama-model-saver.cpp

@@ -0,0 +1,281 @@
+#include "llama-model-saver.h"
+
+#include "gguf.h"
+
+#include "llama.h"
+#include "llama-hparams.h"
+#include "llama-model.h"
+#include "llama-vocab.h"
+
+#include <string>
+
+llama_model_saver::llama_model_saver(const struct llama_model & model) : model(model), llm_kv(model.arch) {
+    gguf_ctx = gguf_init_empty();
+}
+
+llama_model_saver::~llama_model_saver() {
+    gguf_free(gguf_ctx);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const uint32_t value) {
+    gguf_set_val_u32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const int32_t value) {
+    gguf_set_val_i32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const float value) {
+    gguf_set_val_f32(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const bool value) {
+    gguf_set_val_bool(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const char * value) {
+    gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), value);
+}
+
+[[noreturn]]
+void llama_model_saver::add_kv(const enum llm_kv key, const char value) {
+    GGML_UNUSED(key);
+    GGML_UNUSED(value);
+    GGML_ABORT("fatal error"); // this should never be called, only needed to make the template below compile
+}
+
+template <typename Container>
+void llama_model_saver::add_kv(const enum llm_kv key, const Container & value, const bool per_layer) {
+    const size_t n_values = per_layer ? size_t(model.hparams.n_layer) : value.size();
+    GGML_ASSERT(n_values <= value.size());
+
+    if (n_values == 0) {
+        return;
+    }
+
+    if (per_layer) {
+        bool all_values_the_same = true;
+        for (size_t i = 1; i < n_values; ++i) {
+            if (value[i] != value[0]) {
+                all_values_the_same = false;
+                break;
+            }
+        }
+        if (all_values_the_same) {
+            add_kv(key, value[0]);
+            return;
+        }
+    }
+
+    if (std::is_same<typename Container::value_type, uint8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int8_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT8, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, uint32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_UINT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, int32_t>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_INT32, value.data(), n_values);
+    } else if (std::is_same<typename Container::value_type, float>::value) {
+        gguf_set_arr_data(gguf_ctx, llm_kv(key).c_str(), GGUF_TYPE_FLOAT32, value.data(), n_values);
+    } else if (std::is_same<Container, std::string>::value) {
+        gguf_set_val_str(gguf_ctx, llm_kv(key).c_str(), reinterpret_cast<const char *>(value.data()));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void llama_model_saver::add_kv(const enum llm_kv key, const std::vector<std::string> & value) {
+    std::vector<const char *> tmp(value.size());
+    for (size_t i = 0; i < value.size(); ++i) {
+        tmp[i] = value[i].c_str();
+    }
+    gguf_set_arr_str(gguf_ctx, llm_kv(key).c_str(), tmp.data(), tmp.size());
+}
+
+void llama_model_saver::add_tensor(const struct ggml_tensor * tensor) {
+    if (!tensor) {
+        return;
+    }
+    if (gguf_find_tensor(gguf_ctx, tensor->name) >= 0) {
+        GGML_ASSERT(std::string(tensor->name) == "rope_freqs.weight"); // FIXME
+        return;
+    }
+    gguf_add_tensor(gguf_ctx, tensor);
+}
+
+void llama_model_saver::add_kv_from_model() {
+    const llama_hparams & hparams = model.hparams;
+    const llama_vocab   & vocab   = model.vocab;
+
+    const int32_t n_vocab = vocab.n_tokens();
+    std::vector<std::string> tokens(n_vocab);
+    std::vector<float>       scores(n_vocab);
+    std::vector<int32_t>     token_types(n_vocab);
+
+    for (int32_t id = 0; id < n_vocab; ++id) {
+        const llama_vocab::token_data & token_data = vocab.get_token_data(id);
+
+        tokens[id] = token_data.text;
+        scores[id] = token_data.score;
+
+        switch(token_data.attr) {
+            case LLAMA_TOKEN_ATTR_UNKNOWN:      token_types[id] = LLAMA_TOKEN_TYPE_UNKNOWN;      break;
+            case LLAMA_TOKEN_ATTR_UNUSED:       token_types[id] = LLAMA_TOKEN_TYPE_UNUSED;       break;
+            case LLAMA_TOKEN_ATTR_NORMAL:       token_types[id] = LLAMA_TOKEN_TYPE_NORMAL;       break;
+            case LLAMA_TOKEN_ATTR_CONTROL:      token_types[id] = LLAMA_TOKEN_TYPE_CONTROL;      break;
+            case LLAMA_TOKEN_ATTR_USER_DEFINED: token_types[id] = LLAMA_TOKEN_TYPE_USER_DEFINED; break;
+            case LLAMA_TOKEN_ATTR_BYTE:         token_types[id] = LLAMA_TOKEN_TYPE_BYTE;         break;
+            case LLAMA_TOKEN_ATTR_UNDEFINED:
+            default:                            token_types[id] = LLAMA_TOKEN_TYPE_UNDEFINED;    break;
+        }
+    }
+
+    // add_kv(LLM_KV_GENERAL_TYPE,                      ???);
+    add_kv(LLM_KV_GENERAL_ARCHITECTURE,              model.arch_name());
+    // add_kv(LLM_KV_GENERAL_QUANTIZATION_VERSION,      ???);
+    // add_kv(LLM_KV_GENERAL_ALIGNMENT,                 ???);
+    add_kv(LLM_KV_GENERAL_NAME,                      model.name);
+    // add_kv(LLM_KV_GENERAL_AUTHOR,                    ???);
+    // add_kv(LLM_KV_GENERAL_VERSION,                   ???);
+    // add_kv(LLM_KV_GENERAL_URL,                       ???);
+    // add_kv(LLM_KV_GENERAL_DESCRIPTION,               ???);
+    // add_kv(LLM_KV_GENERAL_LICENSE,                   ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_URL,                ???);
+    // add_kv(LLM_KV_GENERAL_SOURCE_HF_REPO,            ???);
+
+    add_kv(LLM_KV_VOCAB_SIZE,                        vocab.n_tokens());
+    add_kv(LLM_KV_CONTEXT_LENGTH,                    hparams.n_ctx_train);
+    add_kv(LLM_KV_EMBEDDING_LENGTH,                  hparams.n_embd);
+    add_kv(LLM_KV_BLOCK_COUNT,                       hparams.n_layer);
+    add_kv(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead);
+    add_kv(LLM_KV_FEED_FORWARD_LENGTH,               hparams.n_ff_arr, true);
+    add_kv(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    add_kv(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    add_kv(LLM_KV_USE_PARALLEL_RESIDUAL,             hparams.use_par_res);
+    // add_kv(LLM_KV_TENSOR_DATA_LAYOUT,                ???);
+    add_kv(LLM_KV_EXPERT_COUNT,                      hparams.n_expert);
+    add_kv(LLM_KV_EXPERT_USED_COUNT,                 hparams.n_expert_used);
+    add_kv(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    add_kv(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale);
+    add_kv(LLM_KV_POOLING_TYPE,                      uint32_t(hparams.pooling_type));
+    add_kv(LLM_KV_LOGIT_SCALE,                       hparams.f_logit_scale);
+    add_kv(LLM_KV_DECODER_START_TOKEN_ID,            hparams.dec_start_token_id);
+    add_kv(LLM_KV_ATTN_LOGIT_SOFTCAPPING,            hparams.f_attn_logit_softcapping);
+    add_kv(LLM_KV_FINAL_LOGIT_SOFTCAPPING,           hparams.f_final_logit_softcapping);
+    add_kv(LLM_KV_SWIN_NORM,                         hparams.swin_norm);
+    add_kv(LLM_KV_RESCALE_EVERY_N_LAYERS,            hparams.rescale_every_n_layers);
+    add_kv(LLM_KV_TIME_MIX_EXTRA_DIM,                hparams.time_mix_extra_dim);
+    add_kv(LLM_KV_TIME_DECAY_EXTRA_DIM,              hparams.time_decay_extra_dim);
+    add_kv(LLM_KV_RESIDUAL_SCALE,                    hparams.f_residual_scale);
+    add_kv(LLM_KV_EMBEDDING_SCALE,                   hparams.f_embedding_scale);
+
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT,              hparams.n_head_arr, true);
+    add_kv(LLM_KV_ATTENTION_HEAD_COUNT_KV,           hparams.n_head_kv_arr, true);
+    add_kv(LLM_KV_ATTENTION_MAX_ALIBI_BIAS,          hparams.f_max_alibi_bias);
+    add_kv(LLM_KV_ATTENTION_CLAMP_KQV,               hparams.f_clamp_kqv);
+    add_kv(LLM_KV_ATTENTION_KEY_LENGTH,              hparams.n_embd_head_k);
+    add_kv(LLM_KV_ATTENTION_VALUE_LENGTH,            hparams.n_embd_head_v);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,           hparams.f_norm_eps);
+    add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    add_kv(LLM_KV_ATTENTION_CAUSAL,                  hparams.causal_attn);
+    add_kv(LLM_KV_ATTENTION_Q_LORA_RANK,             hparams.n_lora_q);
+    add_kv(LLM_KV_ATTENTION_KV_LORA_RANK,            hparams.n_lora_kv);
+    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);
+    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
+    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
+
+    const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
+
+    add_kv(LLM_KV_ROPE_DIMENSION_COUNT,              hparams.n_rot);
+    add_kv(LLM_KV_ROPE_FREQ_BASE,                    hparams.rope_freq_base_train);
+    // add_kv(LLM_KV_ROPE_SCALE_LINEAR,                 rope_scaling_factor); // old name
+    add_kv(LLM_KV_ROPE_SCALING_TYPE,                 llama_rope_scaling_type_name(hparams.rope_scaling_type_train));
+    add_kv(LLM_KV_ROPE_SCALING_FACTOR,               rope_scaling_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ATTN_FACTOR,          hparams.rope_attn_factor);
+    add_kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN,         hparams.n_ctx_orig_yarn);
+    add_kv(LLM_KV_ROPE_SCALING_FINETUNED,            hparams.rope_finetuned);
+    add_kv(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,         hparams.rope_yarn_log_mul);
+
+    // TODO: implement split file support
+    // add_kv(LLM_KV_SPLIT_NO,                          ???);
+    // add_kv(LLM_KV_SPLIT_COUNT,                       ???);
+    // add_kv(LLM_KV_SPLIT_TENSORS_COUNT,               ???);
+
+    add_kv(LLM_KV_SSM_INNER_SIZE,                    hparams.ssm_d_inner);
+    add_kv(LLM_KV_SSM_CONV_KERNEL,                   hparams.ssm_d_conv);
+    add_kv(LLM_KV_SSM_STATE_SIZE,                    hparams.ssm_d_state);
+    add_kv(LLM_KV_SSM_TIME_STEP_RANK,                hparams.ssm_dt_rank);
+    add_kv(LLM_KV_SSM_DT_B_C_RMS,                    hparams.ssm_dt_b_c_rms);
+
+    add_kv(LLM_KV_WKV_HEAD_SIZE,                     hparams.wkv_head_size);
+
+    add_kv(LLM_KV_TOKENIZER_MODEL,                   vocab.get_tokenizer_model());
+    add_kv(LLM_KV_TOKENIZER_PRE,                     vocab.get_tokenizer_pre());
+    add_kv(LLM_KV_TOKENIZER_LIST,                    tokens);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE,              token_types);
+    add_kv(LLM_KV_TOKENIZER_TOKEN_TYPE_COUNT,        vocab.n_token_types());
+    add_kv(LLM_KV_TOKENIZER_SCORES,                  scores);
+    add_kv(LLM_KV_TOKENIZER_MERGES,                  vocab.get_bpe_merges());
+    // FIXME llama_token is type i32 but when reading in a GGUF file u32 is expected, not an issue for writing though
+    add_kv(LLM_KV_TOKENIZER_BOS_ID,                  uint32_t(vocab.token_bos()));
+    add_kv(LLM_KV_TOKENIZER_EOS_ID,                  uint32_t(vocab.token_eos()));
+    add_kv(LLM_KV_TOKENIZER_EOT_ID,                  uint32_t(vocab.token_eot()));
+    add_kv(LLM_KV_TOKENIZER_EOM_ID,                  uint32_t(vocab.token_eom()));
+    add_kv(LLM_KV_TOKENIZER_UNK_ID,                  uint32_t(vocab.token_unk()));
+    add_kv(LLM_KV_TOKENIZER_SEP_ID,                  uint32_t(vocab.token_sep()));
+    add_kv(LLM_KV_TOKENIZER_PAD_ID,                  uint32_t(vocab.token_pad()));
+    // add_kv(LLM_KV_TOKENIZER_CLS_ID,                  uint32_t(vocab.token_bos())); // deprecated
+    // add_kv(LLM_KV_TOKENIZER_MASK_ID,                 ???);
+    add_kv(LLM_KV_TOKENIZER_ADD_BOS,                 vocab.get_add_bos());
+    add_kv(LLM_KV_TOKENIZER_ADD_EOS,                 vocab.get_add_eos());
+    add_kv(LLM_KV_TOKENIZER_ADD_PREFIX,              vocab.get_add_space_prefix());
+    add_kv(LLM_KV_TOKENIZER_REMOVE_EXTRA_WS,         vocab.get_remove_extra_whitespaces());
+    add_kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP,    vocab.get_precompiled_charsmap());
+    // add_kv(LLM_KV_TOKENIZER_HF_JSON,                 ???);
+    // add_kv(LLM_KV_TOKENIZER_RWKV,                    ???);
+    add_kv(LLM_KV_TOKENIZER_FIM_PRE_ID,              uint32_t(vocab.token_fim_pre()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SUF_ID,              uint32_t(vocab.token_fim_suf()));
+    add_kv(LLM_KV_TOKENIZER_FIM_MID_ID,              uint32_t(vocab.token_fim_mid()));
+    add_kv(LLM_KV_TOKENIZER_FIM_PAD_ID,              uint32_t(vocab.token_fim_pad()));
+    add_kv(LLM_KV_TOKENIZER_FIM_REP_ID,              uint32_t(vocab.token_fim_rep()));
+    add_kv(LLM_KV_TOKENIZER_FIM_SEP_ID,              uint32_t(vocab.token_fim_sep()));
+
+    // TODO: implement LoRA support
+    // add_kv(LLM_KV_ADAPTER_TYPE,                      ???);
+    // add_kv(LLM_KV_ADAPTER_LORA_ALPHA,                ???);
+
+    // deprecated
+    // add_kv(LLM_KV_TOKENIZER_PREFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_SUFFIX_ID,               ???);
+    // add_kv(LLM_KV_TOKENIZER_MIDDLE_ID,               ???);
+}
+
+void llama_model_saver::add_tensors_from_model() {
+    if (std::string(model.output->name) != std::string(model.tok_embd->name)) {
+        add_tensor(model.tok_embd); // some models use the same tensor for tok_embd and output
+    }
+    add_tensor(model.type_embd);
+    add_tensor(model.pos_embd);
+    add_tensor(model.tok_norm);
+    add_tensor(model.tok_norm_b);
+    add_tensor(model.output_norm);
+    add_tensor(model.output_norm_b);
+    add_tensor(model.output);
+    add_tensor(model.output_b);
+    add_tensor(model.output_norm_enc);
+    add_tensor(model.cls);
+    add_tensor(model.cls_b);
+    add_tensor(model.cls_out);
+    add_tensor(model.cls_out_b);
+
+    for (const struct llama_layer & layer : model.layers) {
+        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
+            add_tensor(reinterpret_cast<const struct ggml_tensor * const *>(&layer)[i]);
+        }
+    }
+}
+
+void llama_model_saver::save(const std::string & path_model) {
+    gguf_write_to_file(gguf_ctx, path_model.c_str(), false);
+}
+

src/llama-model-saver.h

@@ -0,0 +1,37 @@
+#pragma once
+
+#include "llama.h"
+#include "llama-arch.h"
+
+#include <vector>
+
+struct llama_model_saver {
+    struct gguf_context * gguf_ctx = nullptr;
+    const struct llama_model & model;
+    const struct LLM_KV llm_kv;
+
+    llama_model_saver(const struct llama_model & model);
+    ~llama_model_saver();
+
+    void add_kv(enum llm_kv key, uint32_t     value);
+    void add_kv(enum llm_kv key, int32_t      value);
+    void add_kv(enum llm_kv key, float        value);
+    void add_kv(enum llm_kv key, bool         value);
+    void add_kv(enum llm_kv key, const char * value);
+
+    [[noreturn]]
+    void add_kv(enum llm_kv key, char value); // needed to make the template below compile
+
+    template <typename Container>
+    void add_kv(enum llm_kv key, const Container & value, bool per_layer = false);
+
+    void add_kv(enum llm_kv key, const std::vector<std::string> & value);
+
+    void add_tensor(const struct ggml_tensor * tensor);
+
+    void add_kv_from_model();
+
+    void add_tensors_from_model();
+
+    void save(const std::string & path_model);
+};

src/llama-model.cpp

@@ -117,6 +117,10 @@ static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_
     { LLAMA_ROPE_SCALING_TYPE_LONGROPE,   "longrope"   },
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type) {
+    return LLAMA_ROPE_SCALING_TYPES.at(rope_scaling_type);
+}
+
 static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
     for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) {
         if (kv.second == name) {
@@ -1385,6 +1389,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     // Add additional layer/vocab/etc checks here for other model sizes
                     default: type = LLM_TYPE_UNKNOWN;
                 }
+
+                // For Granite MoE Shared
+                ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
             } break;
         case LLM_ARCH_CHAMELEON:
             {
@@ -1768,6 +1775,13 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
                             layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
                             layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+                            // For Granite MoE Shared
+                            if (hparams.n_ff_shexp > 0) {
+                                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+                            }
                         }
                     }
                 } break;
@@ -4264,7 +4278,7 @@ uint64_t llama_model::n_elements() const {
 }
 
 void llama_model::print_info() const {
-    const char * rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train);
+    const std::string rope_scaling_type = llama_rope_scaling_type_name(hparams.rope_scaling_type_train);
 
     auto print_f = [](const std::function<uint32_t(uint32_t)> & f, uint32_t n) {
         bool is_var = false;
@@ -4325,7 +4339,7 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: causal attn      = %d\n",     __func__, hparams.causal_attn);
         LLAMA_LOG_INFO("%s: pooling type     = %d\n",     __func__, hparams.pooling_type);
         LLAMA_LOG_INFO("%s: rope type        = %d\n",     __func__, hparams.rope_type);
-        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type);
+        LLAMA_LOG_INFO("%s: rope scaling     = %s\n",     __func__, rope_scaling_type.c_str());
         LLAMA_LOG_INFO("%s: freq_base_train  = %.1f\n",   __func__, hparams.rope_freq_base_train);
         LLAMA_LOG_INFO("%s: freq_scale_train = %g\n",     __func__, hparams.rope_freq_scale_train);
         LLAMA_LOG_INFO("%s: n_ctx_orig_yarn  = %u\n",     __func__, hparams.n_ctx_orig_yarn);
@@ -4381,10 +4395,13 @@ void llama_model::print_info() const {
         LLAMA_LOG_INFO("%s: n_ff_exp         = %d\n",     __func__, hparams.n_ff_exp);
     }
 
-    if (arch == LLM_ARCH_MINICPM || arch == LLM_ARCH_GRANITE || arch == LLM_ARCH_GRANITE_MOE) {
+    if (arch == LLM_ARCH_MINICPM ||
+        arch == LLM_ARCH_GRANITE ||
+        arch == LLM_ARCH_GRANITE_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
+        LLAMA_LOG_INFO("%s: n_ff_shexp        = %d\n", __func__, hparams.n_ff_shexp);
     }
 
     if (arch == LLM_ARCH_BAILINGMOE) {
@@ -4594,11 +4611,6 @@ struct llm_build_llama : public llm_graph_context {
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
@@ -4670,11 +4682,6 @@ struct llm_build_llama : public llm_graph_context {
                 cb(cur, "ffn_moe_out", il);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "ffn_out", il);
 
@@ -4697,11 +4704,6 @@ struct llm_build_llama : public llm_graph_context {
         // lm_head
         cur = build_lora_mm(model.output, cur);
 
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
         cb(cur, "result_output", -1);
         res->t_logits = cur;
 
@@ -4812,11 +4814,6 @@ struct llm_build_deci : public llm_graph_context {
                 continue;
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             // modified to support attention-free layer of Llama-3_1-Nemotron-51B
             ggml_tensor * ffn_inp = cur;
             if (n_head > 0) {
@@ -4840,11 +4837,6 @@ struct llm_build_deci : public llm_graph_context {
                 cb(cur, "ffn_out", il);
             }
 
-            // For Granite architecture
-            if (hparams.f_residual_scale) {
-                cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
-            }
-
             cur = ggml_add(ctx0, cur, ffn_inp);
             cb(cur, "ffn_out", il);
 
@@ -4867,11 +4859,6 @@ struct llm_build_deci : public llm_graph_context {
         // lm_head
         cur = build_lora_mm(model.output, cur);
 
-        // For Granite architecture
-        if (hparams.f_logit_scale) {
-            cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
-        }
-
         cb(cur, "result_output", -1);
         res->t_logits = cur;
 
@@ -12210,6 +12197,194 @@ struct llm_build_arwkv7 : public llm_build_rwkv7_base {
     }
 };
 
+
+struct llm_build_granite : public llm_graph_context {
+    llm_build_granite(
+        const llama_model & model,
+        const llm_graph_params & params,
+        ggml_cgraph * gf,
+        const bool use_rope = true)
+        : llm_graph_context(params) {
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        ggml_tensor * cur;
+        ggml_tensor * inpL;
+
+        inpL = build_inp_embd(model.tok_embd);
+
+        // inp_pos - built only if rope enabled
+        ggml_tensor * inp_pos = nullptr;
+        if (use_rope) {
+            inp_pos = build_inp_pos();
+        }
+
+        auto * inp_attn = build_attn_inp_kv_unified();
+
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        for (int il = 0; il < n_layer; ++il) {
+            ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = build_norm(inpL,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and (optionally) RoPE them
+                ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+                if (use_rope) {
+                    ggml_tensor * rope_factors = model.get_rope_factors(n_ctx_per_seq, il);
+                    Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+
+                    Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+                }
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+
+                cur = build_attn(inp_attn, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il);
+                cb(cur, "attn_out", il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network (non-MoE)
+            if (model.layers[il].ffn_gate_inp == nullptr) {
+
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = build_ffn(cur,
+                        model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                        model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                        model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                cb(cur, "ffn_out", il);
+
+            } else {
+                // MoE branch
+                cur = build_norm(ffn_inp,
+                        model.layers[il].ffn_norm, NULL,
+                        LLM_NORM_RMS, il);
+                cb(cur, "ffn_norm", il);
+
+                ggml_tensor * moe_out = 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,
+                        nullptr,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, true,
+                        false, 0.0,
+                        LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                        il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // For Granite MoE Shared
+                if (hparams.n_ff_shexp > 0) {
+                    ggml_tensor * ffn_shexp = build_ffn(cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                } else {
+                    cur = moe_out;
+                }
+            }
+
+            // For Granite architectures - scale residual
+            cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = build_cvec(cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = build_norm(cur,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, -1);
+
+        cb(cur, "result_norm", -1);
+        res->t_embd = cur;
+
+        // lm_head
+        cur = build_lora_mm(model.output, cur);
+
+        // For Granite architectures - scale logits
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
+        cb(cur, "result_output", -1);
+        res->t_logits = cur;
+
+        ggml_build_forward_expand(gf, cur);
+    }
+};
+
 // ref: https://github.com/facebookresearch/chameleon
 // based on the original build_llama() function, changes:
 //   * qk-norm
@@ -12917,8 +13092,6 @@ llm_graph_result_ptr llama_model::build_graph(
         case LLM_ARCH_LLAMA:
         case LLM_ARCH_LLAMA4:
         case LLM_ARCH_MINICPM:
-        case LLM_ARCH_GRANITE:
-        case LLM_ARCH_GRANITE_MOE:
             {
                 llm = std::make_unique<llm_build_llama>(*this, params, gf);
             } break;
@@ -13149,6 +13322,11 @@ llm_graph_result_ptr llama_model::build_graph(
             {
                 llm = std::make_unique<llm_build_arwkv7>(*this, params, gf);
             } break;
+        case LLM_ARCH_GRANITE:
+        case LLM_ARCH_GRANITE_MOE:
+            {
+                llm = std::make_unique<llm_build_granite>(*this, params, gf);
+            } break;
         case LLM_ARCH_CHAMELEON:
             {
                 llm = std::make_unique<llm_build_chameleon>(*this, params, gf);

src/llama-model.h

@@ -96,6 +96,8 @@ enum llm_type {
     LLM_TYPE_235B_A22B,
 };
 
+std::string llama_rope_scaling_type_name(llama_rope_scaling_type rope_scaling_type);
+
 struct llama_layer_posnet {
     // resnet
     struct ggml_tensor * norm1   = nullptr;

src/llama-quant.cpp

@@ -14,6 +14,12 @@
 #include <thread>
 #include <unordered_map>
 
+// Quantization types. Changes to this struct must be replicated in quantize.cpp
+struct tensor_quantization {
+    std::string name;
+    ggml_type quant = GGML_TYPE_COUNT;
+};
+
 static void zeros(std::ofstream & file, size_t n) {
     char zero = 0;
     for (size_t i = 0; i < n; ++i) {
@@ -48,12 +54,6 @@ struct quantize_state_impl {
         {}
 };
 
-// changes to this struct must be replicated in quantize.cpp
-struct tensor_quantization {
-    std::string name;
-    ggml_type quant = GGML_TYPE_COUNT;
-};
-
 static void llama_tensor_dequantize_impl(
     ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
     const size_t nelements, const int nthread
@@ -519,7 +519,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         nthread = std::thread::hardware_concurrency();
     }
 
-    // mmap consistently increases speed Linux, and also increases speed on Windows with
+    // mmap consistently increases speed on Linux, and also increases speed on Windows with
     // hot cache. It may cause a slowdown on macOS, possibly related to free memory.
 #if defined(__linux__) || defined(_WIN32)
     constexpr bool use_mmap = true;
@@ -529,7 +529,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     llama_model_kv_override * kv_overrides = nullptr;
     if (params->kv_overrides) {
-        auto v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
+        auto * v = (std::vector<llama_model_kv_override>*)params->kv_overrides;
         kv_overrides = v->data();
     }
 
@@ -796,17 +796,19 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 // unless the user specifies a type
                 if (params->tensor_types) {
                     const std::vector<tensor_quantization> & tensor_types = *static_cast<const std::vector<tensor_quantization> *>(params->tensor_types);
+                    const std::string tensor_name(tensor->name);
                     for (const auto & [tname, qtype] : tensor_types) {
-                        if (std::regex pattern(tname); std::regex_search(tensor->name, pattern)) {
-                            if (qtype != new_type) {
-                                LLAMA_LOG_DEBUG("(overriding %s -> %s), ", ggml_type_name(new_type), ggml_type_name(qtype));
+                        if (std::regex pattern(tname); std::regex_search(tensor_name, pattern)) {
+                            if  (qtype != new_type) {
+                                LLAMA_LOG_DEBUG("(overriding %s) ", ggml_type_name(new_type));
+                                new_type = qtype;
+                                break; // if two or more types are specified for the tensor, first match wins
                             }
-                            new_type = qtype;
-                            break;
                         }
                     }
                 }
             }
+
             if (params->token_embedding_type < GGML_TYPE_COUNT && strcmp(tensor->name, "token_embd.weight") == 0) {
                 new_type = params->token_embedding_type;
             }

src/llama-vocab.cpp

@@ -1,5 +1,7 @@
 #include "llama-vocab.h"
 
+#include "ggml.h"
+#include "gguf.h"
 #include "llama-impl.h"
 #include "llama-model-loader.h"
 
@@ -415,6 +417,13 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                     "'(?:[sSdDmMtT]|[lL][lL]|[vV][eE]|[rR][eE])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]|\\s+(?!\\S)|\\s+",
                 };
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_SEED_CODER:
+                regex_exprs = {
+                    // original regex from tokenizer.json
+                    // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\r\n]+|\\s*[\r\n]+|\\s+(?!\\S)|\\s+"
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1}| ?[^\\s\\p{L}\\p{N}\\r\\n]+|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
+                };
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -1227,6 +1236,9 @@ struct fragment_buffer_variant {
 struct llama_vocab::impl {
     uint32_t n_token_types = 0; // for BERT-style token types
 
+    std::string tokenizer_model;
+    std::string tokenizer_pre;
+
     enum llama_vocab_type     type     = LLAMA_VOCAB_TYPE_SPM;
     enum llama_vocab_pre_type pre_type = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
 
@@ -1362,9 +1374,6 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
     // determine vocab type
     {
-        std::string tokenizer_model;
-        std::string tokenizer_pre;
-
         ml.get_key(LLM_KV_TOKENIZER_MODEL, tokenizer_model);
         ml.get_key(LLM_KV_TOKENIZER_PRE,   tokenizer_pre, false);
 
@@ -1459,7 +1468,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             const int precompiled_charsmap_keyidx = gguf_find_key(ctx, kv(LLM_KV_TOKENIZER_PRECOMPILED_CHARSMAP).c_str());
             if (precompiled_charsmap_keyidx != -1) {
-                size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
+                const gguf_type pc_type = gguf_get_arr_type(ctx, precompiled_charsmap_keyidx);
+                GGML_ASSERT(pc_type == GGUF_TYPE_INT8 || pc_type == GGUF_TYPE_UINT8);
+
+                const size_t n_precompiled_charsmap = gguf_get_arr_n(ctx, precompiled_charsmap_keyidx);
                 const char * pc = (const char *) gguf_get_arr_data(ctx, precompiled_charsmap_keyidx);
                 precompiled_charsmap.assign(pc, pc + n_precompiled_charsmap);
 #ifdef IS_BIG_ENDIAN
@@ -1634,6 +1646,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 tokenizer_pre == "bailingmoe") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_BAILINGMOE;
                 clean_spaces = false;
+            } else if (
+                tokenizer_pre == "seed-coder") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SEED_CODER;
+                clean_spaces = false;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -2778,6 +2794,14 @@ void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);
 }
 
+std::string llama_vocab::get_tokenizer_model() const {
+    return pimpl->tokenizer_model;
+}
+
+std::string llama_vocab::get_tokenizer_pre() const {
+    return pimpl->tokenizer_pre;
+}
+
 enum llama_vocab_type llama_vocab::get_type() const {
     return pimpl->type;
 }
@@ -3000,6 +3024,20 @@ int llama_vocab::find_bpe_rank(const std::string & token_left, const std::string
     return it->second;
 }
 
+std::vector<std::string> llama_vocab::get_bpe_merges() const {
+    std::vector<std::string> result(pimpl->bpe_ranks.size());
+
+    for (const auto & pair : pimpl->bpe_ranks) {
+        result[pair.second] = pair.first.first + " " + pair.first.second;
+    }
+
+    return result;
+}
+
+std::vector<char> llama_vocab::get_precompiled_charsmap() const {
+    return pimpl->precompiled_charsmap;
+}
+
 int32_t llama_vocab::tokenize(
                   const char * text,
                      int32_t   text_len,

src/llama-vocab.h

@@ -21,6 +21,9 @@ struct llama_vocab {
 
     void load(llama_model_loader & ml, const LLM_KV & kv);
 
+    std::string get_tokenizer_model() const;
+    std::string get_tokenizer_pre() const;
+
     enum llama_vocab_type     get_type()     const;
     enum llama_vocab_pre_type get_pre_type() const;
 
@@ -80,6 +83,9 @@ struct llama_vocab {
     int max_token_len() const;
 
     int find_bpe_rank(const std::string & token_left, const std::string & token_right) const;
+    std::vector<std::string> get_bpe_merges() const;
+
+    std::vector<char> get_precompiled_charsmap() const;
 
     int32_t tokenize(
                    const char * text,

src/llama.cpp

@@ -4,6 +4,7 @@
 #include "llama-mmap.h"
 #include "llama-vocab.h"
 #include "llama-model-loader.h"
+#include "llama-model-saver.h"
 #include "llama-model.h"
 
 #include "ggml.h"
@@ -139,6 +140,11 @@ static struct llama_model * llama_model_load_from_file_impl(
         struct llama_model_params params) {
     ggml_time_init();
 
+    if (!params.vocab_only && ggml_backend_reg_count() == 0) {
+        LLAMA_LOG_ERROR("%s: no backends are loaded. hint: use ggml_backend_load() or ggml_backend_load_all() to load a backend before calling this function\n", __func__);
+        return nullptr;
+    }
+
     unsigned cur_percentage = 0;
     if (params.progress_callback == NULL) {
         params.progress_callback_user_data = &cur_percentage;
@@ -253,6 +259,13 @@ struct llama_model * llama_model_load_from_splits(
     return llama_model_load_from_file_impl(splits.front(), splits, params);
 }
 
+void llama_model_save_to_file(const struct llama_model * model, const char * path_model) {
+    llama_model_saver ms(*model);
+    ms.add_kv_from_model();
+    ms.add_tensors_from_model();
+    ms.save(path_model);
+}
+
 //
 // chat templates
 //
@@ -338,3 +351,4 @@ const char * llama_print_system_info(void) {
 
     return s.c_str();
 }
+

tests/CMakeLists.txt

@@ -144,6 +144,7 @@ endif()
 
 llama_build_and_test(test-log.cpp)
 llama_build_and_test(test-chat-template.cpp)
+llama_build_and_test(test-regex-partial.cpp)
 
 # this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
 if (NOT WIN32)

tests/test-backend-ops.cpp

@@ -823,7 +823,7 @@ struct test_case {
 
         ggml_build_forward_expand(gf, out);
         ggml_graph_cpy(gf, gb);
-        ggml_build_backward_expand(ctx.get(), ctx.get(), gb, false);
+        ggml_build_backward_expand(ctx.get(), gb, nullptr);
         if (expect.size() != 1 || expect[0] != 0.0f) {
             GGML_ASSERT(ggml_graph_n_nodes(gb) > ggml_graph_n_nodes(gf));
             for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != NULL; t = ggml_get_next_tensor(ctx.get(), t)) {
@@ -1026,7 +1026,7 @@ struct test_example : public test_case {
         // Step 3: return the output tensor.
         return out;
     }
-    // In order to also check the gradients for your op, add calls like ggml_set_param(ctx, a)
+    // In order to also check the gradients for your op, add calls like ggml_set_param(a)
     // immediately after you create the tensors.
     // This is optional and only makes sense if a backward pass has actually been implemented for the new op.
 };
@@ -1058,7 +1058,7 @@ struct test_unary : public test_case {
             auto ne = ne_a; ne[0] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
 
@@ -1067,7 +1067,7 @@ struct test_unary : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (grad_supported) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -1133,7 +1133,7 @@ struct test_get_rows : public test_case {
 
         const bool grad_supported = ggml_is_matrix(in) && ggml_is_vector(rows);
         if (grad_supported) {
-            ggml_set_param(ctx, in);
+            ggml_set_param(in);
             // rows is a constant input -> no gradients
         }
 
@@ -1322,7 +1322,7 @@ struct test_repeat : public test_case {
         ggml_set_name(target, "target");
 
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         ggml_tensor * out = ggml_repeat(ctx, src, target);
@@ -1406,7 +1406,7 @@ struct test_dup : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         if (_use_permute) {
@@ -1442,7 +1442,7 @@ struct test_set : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         auto ne_dst = ne;
@@ -1450,7 +1450,7 @@ struct test_set : public test_case {
             ne_dst[i] *= 2;
         }
         ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, ne_dst.data());
-        ggml_set_param(ctx, dst);
+        ggml_set_param(dst);
         ggml_set_name(dst, "dst");
 
         size_t offset = 0;
@@ -1498,7 +1498,7 @@ struct test_cpy : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         if (_src_use_permute) {
@@ -1536,7 +1536,7 @@ struct test_cont : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * src = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, src);
+        ggml_set_param(src);
         ggml_set_name(src, "src");
 
         src = ggml_transpose(ctx, src);
@@ -1583,8 +1583,8 @@ struct test_bin_bcast : public test_case {
         // The backward pass supports broadcasting only for GGML_ADD:
         const bool grad_supported = op == ggml_add || ggml_are_same_shape(a, b);
         if (grad_supported) {
-            ggml_set_param(ctx, a);
-            ggml_set_param(ctx, b);
+            ggml_set_param(a);
+            ggml_set_param(b);
         }
 
         ggml_tensor * out = op(ctx, a, b);
@@ -1632,11 +1632,11 @@ struct test_add1 : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * b = ggml_new_tensor_1d(ctx, type, 1);
-        // ggml_set_param(ctx, b); // TODO: implement
+        // ggml_set_param(b); // TODO: implement
         ggml_set_name(b, "b");
 
         ggml_tensor * out = ggml_add1(ctx, a, b);
@@ -1667,7 +1667,7 @@ struct test_scale : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_scale(ctx, a, scale);
@@ -1762,7 +1762,7 @@ struct test_rms_norm : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         if (v) {
@@ -2028,9 +2028,9 @@ struct test_mul_mat : public test_case {
             b = ggml_new_tensor_4d(ctx, type_b, ne_b[per[0]], ne_b[per[1]], ne_b[per[2]], ne_b[per[3]]);
             if (!ggml_is_quantized(type_a)) {
                 if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+                    ggml_set_param(a);
                 }
-                ggml_set_param(ctx, b);
+                ggml_set_param(b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2040,22 +2040,29 @@ struct test_mul_mat : public test_case {
             ggml_set_name(a, "a_permuted");
             ggml_set_name(b, "b_permuted");
         } else {
-
             if (v) {
                 a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0],       bs[1]);
                 b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);
 
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
+                }
+
                 a = ggml_view_4d(ctx, a, k, m, bs[0],       bs[1],       a->nb[1], a->nb[2], a->nb[3], 0);
                 b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
             } else {
                 a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0],       bs[1]);
                 b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
-            }
-            if (!ggml_is_quantized(type_a)) {
-                if (bs[1] == 1 && nr[1] == 1) {
-                    ggml_set_param(ctx, a);
+
+                if (!ggml_is_quantized(type_a)) {
+                    if (bs[1] == 1 && nr[1] == 1) {
+                        ggml_set_param(a);
+                    }
+                    ggml_set_param(b);
                 }
-                ggml_set_param(ctx, b);
             }
             ggml_set_name(a, "a");
             ggml_set_name(b, "b");
@@ -2204,7 +2211,7 @@ struct test_sqr : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sqr(ctx, a);
@@ -2233,7 +2240,7 @@ struct test_sqrt : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sqrt(ctx, a);
@@ -2273,7 +2280,7 @@ struct test_log : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_log(ctx, a);
@@ -2309,7 +2316,7 @@ struct test_sin : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sin(ctx, a);
@@ -2352,7 +2359,7 @@ struct test_cos : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_cos(ctx, a);
@@ -2432,7 +2439,7 @@ struct test_diag_mask_inf : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_diag_mask_inf(ctx, a, n_past);
@@ -2471,7 +2478,7 @@ struct test_soft_max : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * mask = nullptr;
@@ -2553,7 +2560,7 @@ struct test_rope : public test_case {
             auto ne = ne_a; ne[0] *= 2; ne[1] *= 4; ne[2] *= 3;
             a = ggml_new_tensor(ctx, type, 4, ne.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
 
@@ -2562,7 +2569,7 @@ struct test_rope : public test_case {
         } else {
             a = ggml_new_tensor(ctx, type, 4, ne_a.data());
             if (forward) {
-                ggml_set_param(ctx, a);
+                ggml_set_param(a);
             }
             ggml_set_name(a, "a");
         }
@@ -2676,7 +2683,7 @@ struct test_pool2d : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
 
         ggml_tensor * out = ggml_pool_2d(ctx, input, pool_type, k0, k1, s0, s1, p0, p1);
@@ -2752,7 +2759,7 @@ struct test_im2col : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * input = ggml_new_tensor(ctx, type_input, 4, ne_input.data());
-        ggml_set_param(ctx, input);
+        ggml_set_param(input);
         ggml_set_name(input, "input");
 
         ggml_tensor * kernel = ggml_new_tensor(ctx, type_kernel, 4, ne_kernel.data());
@@ -2929,7 +2936,7 @@ struct test_sum : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sum(ctx, a);
@@ -2958,7 +2965,7 @@ struct test_sum_rows : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_sum_rows(ctx, a);
@@ -2983,7 +2990,7 @@ struct test_mean : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * out = ggml_mean(ctx, a);
@@ -3129,11 +3136,11 @@ struct test_acc : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
-        ggml_set_param(ctx, a);
+        ggml_set_param(a);
         ggml_set_name(a, "a");
 
         ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
-        ggml_set_param(ctx, b);
+        ggml_set_param(b);
         ggml_set_name(b, "b");
 
         ggml_tensor * out = ggml_acc(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], b->nb[1]);
@@ -3370,7 +3377,7 @@ struct test_cross_entropy_loss : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * logits = ggml_new_tensor(ctx, type, 4, ne.data());
-        ggml_set_param(ctx, logits);
+        ggml_set_param(logits);
         ggml_set_name(logits, "logits");
 
         ggml_tensor * labels = ggml_new_tensor(ctx, type, 4, ne.data());
@@ -3452,7 +3459,7 @@ struct test_opt_step_adamw : public test_case {
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);
-        ggml_set_param(ctx, a); // Despite tensor a having gradients the output tensor will not.
+        ggml_set_param(a); // Despite tensor a having gradients the output tensor will not.
         ggml_set_name(a, "a");
 
         ggml_tensor * grad = ggml_new_tensor_4d(ctx, type, ne[0], ne[1], ne[2], ne[3]);

tests/test-chat.cpp

@@ -832,7 +832,9 @@ static void test_template_output_parsers() {
         assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY,
                       common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
         assert_equals(COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
-                      common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+            common_chat_templates_apply(tmpls.get(), inputs_tools).format);
+        assert_equals(COMMON_CHAT_FORMAT_CONTENT_ONLY,
+                        common_chat_templates_apply(tmpls.get(), inputs_no_tools).format);
 
         test_templates(tmpls.get(), end_tokens, message_assist, tools, "Hello, world!\nWhat's up?", /* expect_grammar_triggered= */ false);
         test_templates(tmpls.get(), end_tokens, message_assist_call, tools,

tests/test-opt.cpp

@@ -57,7 +57,8 @@ static helper_ctx_data helper_get_ctx_data(
         enum ggml_opt_loss_type loss_type          = GGML_OPT_LOSS_TYPE_SUM) {
     std::vector<ggml_opt_dataset_t> datasets(ndata);
     for (int64_t ndata_shard = 1; ndata_shard <= ndata; ++ndata_shard) {
-        ggml_opt_dataset_t dataset = ggml_opt_dataset_init(ne_datapoint, ne_label, ndata, ndata_shard);
+        ggml_opt_dataset_t dataset = ggml_opt_dataset_init(
+            GGML_TYPE_F32, GGML_TYPE_F32, ne_datapoint, ne_label, ndata, ndata_shard);
 
         float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
         float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
@@ -74,7 +75,8 @@ static helper_ctx_data helper_get_ctx_data(
         datasets[ndata_shard-1] = dataset;
     }
 
-    ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(1, 0, ndata, /*ndata_shard =*/ 1);
+    ggml_opt_dataset_t dataset_unsupervised = ggml_opt_dataset_init(
+        GGML_TYPE_F32, GGML_TYPE_F32, 1, 0, ndata, /*ndata_shard =*/ 1);
 
     float * data = ggml_get_data_f32(ggml_opt_dataset_data(dataset_unsupervised));
 
@@ -113,7 +115,7 @@ static helper_ctx_data helper_get_ctx_data(
 
     struct ggml_tensor * weights = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(weights, "weights");
-    ggml_set_param(ctx_static, weights);
+    ggml_set_param(weights);
 
     struct ggml_tensor * intermediary = ggml_add(ctx_compute, inputs, weights);
 
@@ -127,8 +129,11 @@ static helper_ctx_data helper_get_ctx_data(
     GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
     const int32_t opt_period = nbatch_logical / nbatch_physical;
 
-    struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
-    opt_params.opt_period = opt_period;
+    struct ggml_opt_params opt_params = ggml_opt_default_params(backend_sched, loss_type);
+    opt_params.ctx_compute = ctx_compute;
+    opt_params.inputs      = inputs;
+    opt_params.outputs     = outputs;
+    opt_params.opt_period  = opt_period;
     if (!optimizer_defaults) {
         opt_params.get_opt_pars = helper_get_test_opt_pars;
     }
@@ -264,8 +269,9 @@ static std::pair<int, int> test_grad(ggml_backend_sched_t backend_sched, ggml_ba
 
     for (int idata = 0; idata < ndata; ++idata) {
         const float idataf = idata;
+        ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
         ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-        ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+        ggml_opt_eval(cd.opt_ctx, cd.result);
         ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, sizeof(float));
     }
 
@@ -334,8 +340,9 @@ static std::pair<int, int> test_forward_backward(
     } else {
         for (int idata = 0; idata < ndata; ++idata) {
             const float idataf = idata;
+            ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false);
             ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-            ggml_opt_forward(cd.opt_ctx, cd.result);
+            ggml_opt_eval(cd.opt_ctx, cd.result);
             ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
         }
     }
@@ -367,7 +374,8 @@ static std::pair<int, int> test_forward_backward(
     float w0;
     ggml_backend_tensor_get(cd.weights, &w0, 0, sizeof(float));
     for (int i = 0; i < 10; ++i) {
-        ggml_opt_forward_backward(cd.opt_ctx, nullptr);
+        ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
+        ggml_opt_eval(cd.opt_ctx, cd.result);
     }
     ggml_backend_tensor_set(cd.weights, &w0, 0, sizeof(float));
 
@@ -387,8 +395,9 @@ static std::pair<int, int> test_forward_backward(
     } else {
         for (int idata = 0; idata < ndata; ++idata) {
             const float idataf = idata;
+            ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
             ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-            ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+            ggml_opt_eval(cd.opt_ctx, cd.result);
             ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
         }
     }
@@ -492,14 +501,16 @@ static std::pair<int, int> test_idata_split(ggml_backend_sched_t backend_sched,
             int idata = 0;
             for (; idata < idata_split; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
                 ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
             }
             for (; idata < ndata; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ false);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, ggml_nbytes(cd.inputs));
-                ggml_opt_forward(cd.opt_ctx, cd.result2);
+                ggml_opt_eval(cd.opt_ctx, cd.result2);
                 ggml_backend_tensor_get(loss, loss_history.data() + idata, 0, sizeof(float));
             }
         }
@@ -573,7 +584,6 @@ static std::pair<int, int> test_gradient_accumulation(
 
     struct helper_ctx_data cd = helper_get_ctx_data(
         backend_sched, backend, /*init_opt_ctx =*/ true, /*optimizer_defaults =*/ false, /*nbatch_logical =*/ 6, nbatch_physical, loss_type);
-    struct ggml_tensor * loss = ggml_opt_loss(cd.opt_ctx);
 
     std::vector<float> grad_history(ndata);
     for (int64_t idata = 0; idata < ndata; ++idata) {
@@ -584,15 +594,17 @@ static std::pair<int, int> test_gradient_accumulation(
         if (nbatch_physical == 1) {
             for (int idata = 0; idata < ndata; ++idata) {
                 const float idataf = idata;
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, &idataf, 0, 1*sizeof(float));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
                 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata, 0, 1*sizeof(float));
             }
         } else if (nbatch_physical == 2) {
             for (int idata = 0; idata < ndata; idata += 2) {
                 const float idataf[2] = {float(idata + 0), float(idata + 1)};
+                ggml_opt_alloc(cd.opt_ctx, /*backward =*/ true);
                 ggml_backend_tensor_set(cd.inputs, idataf, 0, 2*sizeof(float));
-                ggml_opt_forward_backward(cd.opt_ctx, cd.result);
+                ggml_opt_eval(cd.opt_ctx, cd.result);
 
                 grad_history[idata + 0] = 0.0f;
                 ggml_backend_tensor_get(ggml_opt_grad_acc(cd.opt_ctx, cd.weights), grad_history.data() + idata + 1, 0, 1*sizeof(float));
@@ -617,7 +629,7 @@ static std::pair<int, int> test_gradient_accumulation(
                 }
                 subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0, atol);
                 subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0, atol);
-                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0, atol);
             } else if (loss_type == GGML_OPT_LOSS_TYPE_MEAN) {
                 if (nbatch_physical == 1) {
                     subtest_ok = subtest_ok && almost_equal(grad_history[0], 1.0/ndata, atol);
@@ -630,7 +642,7 @@ static std::pair<int, int> test_gradient_accumulation(
                 }
                 subtest_ok = subtest_ok && almost_equal(grad_history[1], 2.0/ndata, atol);
                 subtest_ok = subtest_ok && almost_equal(grad_history[3], 4.0/ndata, atol);
-                subtest_ok = subtest_ok && almost_equal(grad_history[5], 0.0/ndata, atol);
+                subtest_ok = subtest_ok && almost_equal(grad_history[5], 6.0/ndata, atol);
             } else {
                 GGML_ASSERT(false);
             }
@@ -692,7 +704,8 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
     std::mt19937 gen(12345);
     std::normal_distribution<float> nd{0.0f, 0.1f};
 
-    ggml_opt_dataset_t dataset = ggml_opt_dataset_init(1, 1, ndata_regression, ndata_regression);
+    ggml_opt_dataset_t dataset = ggml_opt_dataset_init(
+        GGML_TYPE_F32, GGML_TYPE_F32, 1, 1, ndata_regression, ndata_regression);
 
     float * data   = ggml_get_data_f32(ggml_opt_dataset_data(  dataset));
     float * labels = ggml_get_data_f32(ggml_opt_dataset_labels(dataset));
@@ -733,15 +746,14 @@ static std::pair<int, int> test_regression(ggml_backend_sched_t backend_sched, g
 
     struct ggml_tensor * a = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(a, "a");
-    ggml_set_param(ctx_static, a);
+    ggml_set_param(a);
 
     struct ggml_tensor * b = ggml_new_tensor_1d(ctx_static, GGML_TYPE_F32, 1);
     ggml_set_name(b, "b");
-    ggml_set_param(ctx_static, b);
+    ggml_set_param(b);
 
     struct ggml_tensor * f = ggml_add(ctx_compute, ggml_mul(ctx_compute, x, a), b);
     ggml_set_name(f, "f");
-    ggml_set_param(ctx_static, f);
 
     ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors(ctx_static, backend);
     const float a0 = 1.0f;
@@ -853,7 +865,7 @@ int main(void) {
         backends_modded.insert(backends_modded.end(), backends.begin(), backends.end());
 
         ggml_backend_sched_t backend_sched = ggml_backend_sched_new(
-            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false);
+            backends_modded.data(), nullptr, backends_modded.size(), GGML_DEFAULT_GRAPH_SIZE, false, true);
 
         printf("Backend %zu/%zu: %s\n", i + 1, dev_count, ggml_backend_dev_name(devs[i]));
         printf("  Device description: %s\n", ggml_backend_dev_description(devs[i]));