common : add more aliases for sampler CLI params (#19797 )

* common : add more aliases for sampler CLI params
ci : update the ROCm/HIP toolchain versions [no ci] (#19891 )
2026-06-28 00:27:39 +02:00 · 2026-02-25 16:34:25 +01:00 · 2026-02-25 15:54:49 +01:00 · 2026-02-25 15:15:42 +02:00 · 2026-02-25 15:14:27 +02:00 · 2026-02-25 12:57:34 +01:00
192 changed files with 17202 additions and 11801 deletions
@@ -1,8 +1,8 @@
 ARG UBUNTU_VERSION=24.04

 # This needs to generally match the container host's environment.
-ARG ROCM_VERSION=7.0
-ARG AMDGPU_VERSION=7.0
+ARG ROCM_VERSION=7.2
+ARG AMDGPU_VERSION=7.2

 # Target the ROCm build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
@@ -11,13 +11,12 @@ ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-co
 FROM ${BASE_ROCM_DEV_CONTAINER} AS build

 # Unless otherwise specified, we make a fat build.
-# List from https://github.com/ggml-org/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-# gfx803, gfx900, gfx906, gfx1032, gfx1101, gfx1102,not officialy supported
-# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-6.4.1/reference/system-requirements.html
+# check https://rocm.docs.amd.com/projects/install-on-linux/en/docs-7.2.0/reference/system-requirements.html
+# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityrad/native_linux/native_linux_compatibility.html
+# check https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/compatibility/compatibilityryz/native_linux/native_linux_compatibility.html

-ARG ROCM_DOCKER_ARCH='gfx803;gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1010;gfx1030;gfx1032;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201;gfx1151'
-#ARG ROCM_DOCKER_ARCH='gfx1151'
+ARG ROCM_DOCKER_ARCH='gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201'

 # Set ROCm architectures
 ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
@@ -11,5 +11,5 @@ runs:
    - name: Setup ROCm
      uses: ./.github/actions/install-exe
      with:
-        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-WinSvr2022-For-HIP.exe
+        url: https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ inputs.version }}-Win11-For-HIP.exe
        args: -install
@@ -68,7 +68,7 @@ jobs:

    env:
      # Make sure this is in sync with build.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"

    steps:
      - name: Clone
@@ -1175,10 +1175,8 @@ jobs:
    runs-on: windows-2022

    env:
-      # The ROCm version must correspond to the version used in the HIP SDK.
-      ROCM_VERSION: "6.4.2"
      # Make sure this is in sync with build-cache.yml
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"

    steps:
      - name: Clone
@@ -1188,7 +1186,7 @@ jobs:
      - name: Grab rocWMMA package
        id: grab_rocwmma
        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/${{ env.ROCM_VERSION }}/pool/main/r/rocwmma-dev/rocwmma-dev_1.7.0.60402-120~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
          7z x rocwmma.deb
          7z x data.tar

@@ -1231,7 +1229,7 @@ jobs:
          cmake -G "Unix Makefiles" -B build -S . `
            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-${{ env.ROCM_VERSION }}/include/" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/" `
            -DCMAKE_BUILD_TYPE=Release `
            -DLLAMA_BUILD_BORINGSSL=ON `
            -DROCM_DIR="${env:HIP_PATH}" `
@@ -516,17 +516,113 @@ jobs:
          path: llama-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

+  ubuntu-22-rocm:
+    runs-on: ubuntu-22.04
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.2"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1151;gfx1150;gfx1200;gfx1201"
+            build: 'x64'
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential git cmake wget
+
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} jammy main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk
+
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2'
+        id: therock_env
+        run: |
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
+            -DGGML_HIP=ON \
+            -DHIP_PLATFORM=amd \
+            -DGGML_HIP_ROCWMMA_FATTN=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
+          name: llama-bin-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}.tar.gz
+
  windows-hip:
    runs-on: windows-2022

    env:
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"

    strategy:
      matrix:
        include:
          - name: "radeon"
-            gpu_targets: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+            gpu_targets: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"

    steps:
      - name: Clone
@@ -536,7 +632,7 @@ jobs:
      - name: Grab rocWMMA package
        id: grab_rocwmma
        run: |
-          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.0.1/pool/main/r/rocwmma-dev/rocwmma-dev_2.0.0.70001-42~24.04_amd64.deb"
+          curl -o rocwmma.deb "https://repo.radeon.com/rocm/apt/7.2/pool/main/r/rocwmma-dev/rocwmma-dev_2.2.0.70200-43~24.04_amd64.deb"
          7z x rocwmma.deb
          7z x data.tar

@@ -559,7 +655,7 @@ jobs:
        run: |
          $ErrorActionPreference = "Stop"
          write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
          write-host "Installing AMD HIP SDK"
          $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
          $completed = $proc.WaitForExit(600000)
@@ -593,20 +689,20 @@ jobs:
          cmake -G "Unix Makefiles" -B build -S . `
            -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" `
            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" `
-            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.0.1/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
+            -DCMAKE_CXX_FLAGS="-I$($PWD.Path.Replace('\', '/'))/opt/rocm-7.2.0/include/ -Wno-ignored-attributes -Wno-nested-anon-types" `
            -DCMAKE_BUILD_TYPE=Release `
            -DGGML_BACKEND_DL=ON `
            -DGGML_NATIVE=OFF `
            -DGGML_CPU=OFF `
-            -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}" `
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" `
            -DGGML_HIP_ROCWMMA_FATTN=ON `
            -DGGML_HIP=ON `
            -DLLAMA_BUILD_BORINGSSL=ON
          cmake --build build --target ggml-hip -j ${env:NUMBER_OF_PROCESSORS}
          md "build\bin\rocblas\library\"
          md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
          cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
@@ -784,6 +880,7 @@ jobs:
      - windows-cuda
      - windows-sycl
      - windows-hip
+      - ubuntu-22-rocm
      - ubuntu-22-cpu
      - ubuntu-22-vulkan
      - macOS-arm64
@@ -868,6 +965,7 @@ jobs:
            **Linux:**
            - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)
            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)

            **Windows:**
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)

@@ -1578,7 +1578,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    ).set_sparam());
    add_opt(common_arg(
-        {"--temp"}, "N",
+        {"--temp", "--temperature"}, "N",
        string_format("temperature (default: %.2f)", (double)params.sampling.temp),
        [](common_params & params, const std::string & value) {
            params.sampling.temp = std::stof(value);
@@ -1611,7 +1611,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    ).set_sparam());
    add_opt(common_arg(
-        {"--top-nsigma"}, "N",
+        {"--top-nsigma", "--top-n-sigma"}, "N",
        string_format("top-n-sigma sampling (default: %.2f, -1.0 = disabled)", params.sampling.top_n_sigma),
        [](common_params & params, const std::string & value) {
            params.sampling.top_n_sigma = std::stof(value);
@@ -1634,7 +1634,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        }
    ).set_sparam());
    add_opt(common_arg(
-        {"--typical"}, "N",
+        {"--typical", "--typical-p"}, "N",
        string_format("locally typical sampling, parameter p (default: %.2f, 1.0 = disabled)", (double)params.sampling.typ_p),
        [](common_params & params, const std::string & value) {
            params.sampling.typ_p = std::stof(value);
@@ -803,7 +803,7 @@ inline void parse_msg_with_xml_tool_calls(common_chat_msg_parser & builder, cons
        }

        // remove potential partial suffix
-        if (builder.pos() == builder.input().size()) {
+        if (builder.pos() == builder.input().size() && builder.is_partial()) {
            if (unclosed_reasoning_content.empty()) {
                rstrip(content);
                trim_potential_partial_word(content);
@@ -893,23 +893,6 @@ static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) {
    builder.consume_reasoning_with_xml_tool_calls(form, "<think>", "</think>");
 }

-static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) {
-    static const xml_tool_call_format form = ([]() {
-        xml_tool_call_format form {};
-        form.scope_start = "<tool_call>";
-        form.tool_start  = "<function=";
-        form.tool_sep    = ">";
-        form.key_start   = "<parameter=";
-        form.key_val_sep = ">";
-        form.val_end     = "</parameter>";
-        form.tool_end    = "</function>";
-        form.scope_end   = "</tool_call>";
-        form.trim_raw_argval = true;
-        return form;
-    })();
-    builder.consume_reasoning_with_xml_tool_calls(form);
-}
-
 static void common_chat_parse_kimi_k2(common_chat_msg_parser & builder) {
    static const xml_tool_call_format form = ([]() {
        xml_tool_call_format form {};
@@ -1590,9 +1573,6 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
        case COMMON_CHAT_FORMAT_KIMI_K2:
            common_chat_parse_kimi_k2(builder);
            break;
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML:
-            common_chat_parse_qwen3_coder_xml(builder);
-            break;
        case COMMON_CHAT_FORMAT_APRIEL_1_5:
            common_chat_parse_apriel_1_5(builder);
            break;
@@ -65,14 +65,25 @@ json common_chat_msg::to_json_oaicompat(bool concat_typed_text) const {
    } else if (!content_parts.empty()) {
        if (concat_typed_text) {
            std::string text;
+            bool last_was_media_marker = false;
+            // join parts with newline, do not add newline before or after media markers
            for (const auto & part : content_parts) {
-                if (part.type != "text") {
+                bool add_new_line = true;
+                if (part.type == "text") {
+                    add_new_line = !last_was_media_marker && !text.empty();
+                    last_was_media_marker = false;
+                } else if (part.type == "media_marker") {
+                    add_new_line = false;
+                    last_was_media_marker = true;
+                } else {
                    LOG_WRN("Ignoring content part type: %s\n", part.type.c_str());
                    continue;
                }
-                if (!text.empty()) {
+
+                if (add_new_line) {
                    text += '\n';
                }
+
                text += part.text;
            }
            jmsg["content"] = text;
@@ -319,7 +330,7 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
                            throw std::invalid_argument("Missing content part type: " + part.dump());
                        }
                        const auto & type = part.at("type");
-                        if (type != "text") {
+                        if (type != "text" && type != "media_marker") {
                            throw std::invalid_argument("Unsupported content part type: " + type.dump());
                        }
                        common_chat_msg_content_part msg_part;
@@ -725,7 +736,6 @@ const char * common_chat_format_name(common_chat_format format) {
        case COMMON_CHAT_FORMAT_MINIMAX_M2: return "MiniMax-M2";
        case COMMON_CHAT_FORMAT_GLM_4_5: return "GLM 4.5";
        case COMMON_CHAT_FORMAT_KIMI_K2: return "Kimi K2";
-        case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: return "Qwen3 Coder";
        case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
        case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
        case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
@@ -1511,14 +1521,17 @@ static common_chat_params common_chat_params_init_nemotron_v2(const common_chat_
    return data;
 }

-static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_template & tmpl, const struct templates_params & inputs) {
+static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_template & tmpl, const struct templates_params & inputs) {
    common_chat_params data;

    data.prompt = apply(tmpl, inputs);
    data.format = COMMON_CHAT_FORMAT_PEG_CONSTRUCTED;

+    // Nemotron Nano 3 and Step-3.5-Flash use the Qwen3 Coder tool calling with thinking
+    bool supports_reasoning = (tmpl.source().find("<think>") != std::string::npos);
+
    // Handle thinking tags appropriately based on inputs.enable_thinking
-    if (string_ends_with(data.prompt, "<think>\n")) {
+    if (supports_reasoning && string_ends_with(data.prompt, "<think>\n")) {
        if (!inputs.enable_thinking) {
            data.prompt += "</think>";
        } else {
@@ -1527,19 +1540,21 @@ static common_chat_params common_chat_params_init_nemotron_v3(const common_chat_
    }

    data.preserved_tokens = {
-        "<think>",
-        "</think>",
        "<tool_call>",
        "</tool_call>",
    };

+    if (supports_reasoning) {
+        data.preserved_tokens.insert(data.preserved_tokens.end(), {"<think>", "</think>"});
+    }
+
    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
    auto include_grammar = true;

    auto parser = build_chat_peg_constructed_parser([&](auto & p) {
        auto reasoning = p.eps();
-        if (inputs.enable_thinking && extract_reasoning) {
+        if (supports_reasoning && inputs.enable_thinking && extract_reasoning) {
            auto reasoning_content = p.reasoning(p.until("</think>")) + ("</think>" | p.end());
            if (data.thinking_forced_open) {
                reasoning = reasoning_content;
@@ -1877,38 +1892,6 @@ static common_chat_params common_chat_params_init_minimax_m2(const common_chat_t
    return data;
 }

-static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_chat_template & tmpl, const struct templates_params & params) {
-    common_chat_params data;
-    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
-
-    data.prompt = apply(tmpl, params);
-    data.format = COMMON_CHAT_FORMAT_QWEN3_CODER_XML;
-
-    data.preserved_tokens = {
-        "<tool_call>",
-        "</tool_call>",
-        "<function=",
-        "</function>",
-        "<parameter=",
-        "</parameter>",
-    };
-
-    // build grammar for tool call
-    static const xml_tool_call_format form {
-        /* form.scope_start = */ "<tool_call>\n",
-        /* form.tool_start  = */ "<function=",
-        /* form.tool_sep    = */ ">\n",
-        /* form.key_start   = */ "<parameter=",
-        /* form.key_val_sep = */ ">\n",
-        /* form.val_end     = */ "\n</parameter>\n",
-        /* form.tool_end    = */ "</function>\n",
-        /* form.scope_end   = */ "</tool_call>",
-    };
-    build_grammar_xml_tool_call(data, params.tools, form);
-
-    return data;
-}
-
 static common_chat_params common_chat_params_init_kimi_k2(const common_chat_template & tmpl, const struct templates_params & params) {
    common_chat_params data;
    data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -2032,6 +2015,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
        if (has_reasoning_content && has_tool_calls) {
            auto adjusted_message = msg;
            adjusted_message["thinking"] = msg.at("reasoning_content");
+            adjusted_message.erase("content");
            adjusted_messages.push_back(adjusted_message);
        } else {
            adjusted_messages.push_back(msg);
@@ -3129,19 +3113,13 @@ static common_chat_params common_chat_templates_apply_jinja(
    }

    // Qwen3-Coder XML format detection (must come before Hermes 2 Pro)
-    // Detect via explicit XML markers unique to Qwen3-Coder to avoid false positives in other templates.
-    // Require presence of <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Detect via XML markers: <tool_call>, <function=...>, and <parameter=...> blocks.
+    // Also matches Step-3.5-Flash and Nemotron 3 Nano which use the same output format.
    if (src.find("<tool_call>") != std::string::npos &&
-        src.find("<function>") != std::string::npos &&
        src.find("<function=") != std::string::npos &&
-        src.find("<parameters>") != std::string::npos &&
        src.find("<parameter=") != std::string::npos) {
        workaround::func_args_not_string(params.messages);
-        // Nemotron 3 Nano 30B A3B
-        if (src.find("<think>") != std::string::npos) {
-            return common_chat_params_init_nemotron_v3(tmpl, params);
-        }
-        return common_chat_params_init_qwen3_coder_xml(tmpl, params);
+        return common_chat_params_init_qwen3_coder(tmpl, params);
    }

    // Xiaomi MiMo format detection (must come before Hermes 2 Pro)
@@ -3307,7 +3285,7 @@ static common_chat_params common_chat_templates_apply_legacy(
    for (const auto & msg : inputs.messages) {
        auto content = msg.content;
        for (const auto & part : msg.content_parts) {
-            if (part.type != "text") {
+            if (part.type != "text" && part.type != "media_marker") {
                LOG_WRN("Ignoring non-text content part: %s\n", part.type.c_str());
                continue;
            }
@@ -128,7 +128,6 @@ enum common_chat_format {
    COMMON_CHAT_FORMAT_GLM_4_5,
    COMMON_CHAT_FORMAT_MINIMAX_M2,
    COMMON_CHAT_FORMAT_KIMI_K2,
-    COMMON_CHAT_FORMAT_QWEN3_CODER_XML,
    COMMON_CHAT_FORMAT_APRIEL_1_5,
    COMMON_CHAT_FORMAT_XIAOMI_MIMO,
    COMMON_CHAT_FORMAT_SOLAR_OPEN,
@@ -1760,3 +1760,65 @@ float lr_opt::get_lr(float epoch) const {
    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
    return r;
 }
+
+bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos) {
+    llama_batch batch = llama_batch_get_one(&last_token, 1);
+    batch.pos = &pos;
+    if (llama_decode(ctx, batch)) {
+        LOG_ERR("%s: failed to replay last token\n", __func__);
+        return false;
+    }
+    return true;
+}
+
+bool common_prompt_batch_decode(
+              struct llama_context * ctx,
+    const std::vector<llama_token> & tokens,
+                               int & n_past,
+                               int   n_batch,
+                  std::string_view   state_path,
+                              bool   save_state) {
+    const int n_eval = tokens.size();
+    if (n_eval == 0) {
+        return true;
+    }
+
+    if (save_state && n_eval > 1) {
+        const int n_tokens_before_last = n_eval - 1;
+
+        GGML_ASSERT(n_eval <= n_batch);
+
+        // Decode all but the last token so we can save the memory state before decoding the last token.
+        // This is done so we can restore the session state later and replay the last token.
+        // Memory implementations in recurrent/hybrid models don't support removing tokens from their
+        // memory, so we can't just remove the last token from the memory and replay the last token which
+        // is the reason for this logic.
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_tokens_before_last))) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        n_past += n_tokens_before_last;
+
+        llama_state_save_file(ctx, state_path.data(), tokens.data(), n_tokens_before_last);
+        LOG_INF("saved session before last token to %s, n_tokens = %d\n", state_path.data(), n_tokens_before_last);
+
+        llama_token last_token = tokens.back();
+        llama_batch batch = llama_batch_get_one(&last_token, 1);
+        int32_t pos = n_past;
+        batch.pos = &pos;
+
+        if (llama_decode(ctx, batch)) {
+            LOG_ERR("%s : failed to eval last token\n", __func__);
+            return false;
+        }
+        n_past++;
+    } else {
+        if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(tokens.data()), n_eval))) {
+            LOG_ERR("%s : failed to eval\n", __func__);
+            return false;
+        }
+        n_past += n_eval;
+    }
+
+    return true;
+}
@@ -804,6 +804,23 @@ void common_batch_add(
    const std::vector<llama_seq_id> & seq_ids,
                               bool   logits);

+// decodes a single batch of tokens for a prompt and manages session tokens
+//
+// Note: We save state before the last token so that we can replay it to ensure
+// compatibility with all memory types. Recurrent/hybrid models cannot remove
+// tokens from memory, so this approach works across all model architectures.
+bool common_prompt_batch_decode(
+              struct llama_context * ctx,
+    const std::vector<llama_token> & embd,
+                               int & n_past,
+                               int   n_batch,
+                  std::string_view   state_path,
+                              bool   save_state);
+
+// replays the last token after loading state to regenerate logits
+// used after loading session state to ensure the sampling context has valid logits
+bool common_replay_last_token(struct llama_context * ctx, llama_token last_token, int32_t pos);
+
 //
 // Vocab utils
 //
@@ -85,7 +85,7 @@ value identifier::execute_impl(context & ctx) {
    auto builtins = global_builtins();
    if (!it->is_undefined()) {
        if (ctx.is_get_stats) {
-            it->stats.used = true;
+            value_t::stats_t::mark_used(it);
        }
        JJ_DEBUG("Identifier '%s' found, type = %s", val.c_str(), it->type().c_str());
        return it;
@@ -277,7 +277,7 @@ value binary_expression::execute_impl(context & ctx) {
 static value try_builtin_func(context & ctx, const std::string & name, value & input, bool undef_on_missing = false) {
    JJ_DEBUG("Trying built-in function '%s' for type %s", name.c_str(), input->type().c_str());
    if (ctx.is_get_stats) {
-        input->stats.used = true;
+        value_t::stats_t::mark_used(input);
        input->stats.ops.insert(name);
    }
    auto builtins = input->get_builtins();
@@ -448,7 +448,7 @@ value for_statement::execute_impl(context & ctx) {

    // mark the variable being iterated as used for stats
    if (ctx.is_get_stats) {
-        iterable_val->stats.used = true;
+        value_t::stats_t::mark_used(iterable_val);
        iterable_val->stats.ops.insert("array_access");
    }

@@ -470,7 +470,7 @@ value for_statement::execute_impl(context & ctx) {
            items.push_back(std::move(tuple));
        }
        if (ctx.is_get_stats) {
-            iterable_val->stats.used = true;
+            value_t::stats_t::mark_used(iterable_val);
            iterable_val->stats.ops.insert("object_access");
        }
    } else {
@@ -480,7 +480,7 @@ value for_statement::execute_impl(context & ctx) {
            items.push_back(item);
        }
        if (ctx.is_get_stats) {
-            iterable_val->stats.used = true;
+            value_t::stats_t::mark_used(iterable_val);
            iterable_val->stats.ops.insert("array_access");
        }
    }
@@ -817,8 +817,9 @@ value member_expression::execute_impl(context & ctx) {
    }

    if (ctx.is_get_stats && val && object && property) {
-        val->stats.used = true;
-        object->stats.used = true;
+        value_t::stats_t::mark_used(val);
+        value_t::stats_t::mark_used(object);
+        value_t::stats_t::mark_used(property);
        if (is_val<value_int>(property)) {
            object->stats.ops.insert("array_access");
        } else if (is_val<value_string>(property)) {
@@ -4,6 +4,7 @@
 // for converting from JSON to jinja values
 #include <nlohmann/json.hpp>

+#include <sstream>
 #include <string>
 #include <cctype>
 #include <vector>
@@ -160,6 +161,11 @@ static value tojson(const func_args & args) {
    value val_separators = args.get_kwarg_or_pos("separators",   3);
    value val_sort       = args.get_kwarg_or_pos("sort_keys",    4);
    int indent = -1;
+    if (args.ctx.is_get_stats) {
+        // mark as used (recursively) for stats
+        auto val_input = args.get_pos(0);
+        value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+    }
    if (is_val<value_int>(val_indent)) {
        indent = static_cast<int>(val_indent->as_int());
    }
@@ -715,8 +721,46 @@ const func_builtins & value_string_t::get_builtins() const {
            return args.get_pos(0);
        }},
        {"tojson", tojson},
-        {"indent", [](const func_args &) -> value {
-            throw not_implemented_exception("String indent builtin not implemented");
+        {"indent", [](const func_args &args) -> value {
+            args.ensure_count(1, 4);
+            value val_input  = args.get_pos(0);
+            value val_width  = args.get_kwarg_or_pos("width", 1);
+            const bool first = args.get_kwarg_or_pos("first", 2)->as_bool(); // undefined == false
+            const bool blank = args.get_kwarg_or_pos("blank", 3)->as_bool(); // undefined == false
+            if (!is_val<value_string>(val_input)) {
+                throw raised_exception("indent() first argument must be a string");
+            }
+            std::string indent;
+            if (is_val<value_int>(val_width)) {
+                indent.assign(val_width->as_int(), ' ');
+            } else if (is_val<value_string>(val_width)) {
+                indent = val_width->as_string().str();
+            } else {
+                indent = "    ";
+            }
+            std::string indented;
+            std::string input = val_input->as_string().str();
+            std::istringstream iss = std::istringstream(input);
+            std::string line;
+            while (std::getline(iss, line)) {
+                if (!indented.empty()) {
+                    indented.push_back('\n');
+                }
+                if ((indented.empty() ? first : (!line.empty() || blank))) {
+                    indented += indent;
+                }
+                indented += line;
+            }
+            if (!input.empty() && input.back() == '\n') {
+                indented.push_back('\n');
+                if (blank) {
+                    indented += indent;
+                }
+            }
+
+            auto res = mk_val<value_string>(indented);
+            res->val_str.mark_input_based_on(val_input->as_string());
+            return res;
        }},
        {"join", [](const func_args &) -> value {
            throw not_implemented_exception("String join builtin not implemented");
@@ -852,6 +896,11 @@ const func_builtins & value_array_t::get_builtins() const {
        }},
        {"string", [](const func_args & args) -> value {
            args.ensure_vals<value_array>();
+            if (args.ctx.is_get_stats) {
+                // mark as used (recursively) for stats
+                auto val_input = args.get_pos(0);
+                value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+            }
            return mk_val<value_string>(args.get_pos(0)->as_string());
        }},
        {"tojson", tojson},
@@ -1007,6 +1056,11 @@ const func_builtins & value_object_t::get_builtins() const {
        {"tojson", tojson},
        {"string", [](const func_args & args) -> value {
            args.ensure_vals<value_object>();
+            if (args.ctx.is_get_stats) {
+                // mark as used (recursively) for stats
+                auto val_input = args.get_pos(0);
+                value_t::stats_t::mark_used(const_cast<value&>(val_input), true);
+            }
            return mk_val<value_string>(args.get_pos(0)->as_string());
        }},
        {"length", [](const func_args & args) -> value {
@@ -1319,4 +1373,21 @@ std::string value_to_string_repr(const value & val) {
    }
 }

+// stats utility
+void value_t::stats_t::mark_used(value & val, bool deep) {
+    val->stats.used = true;
+    if (deep) {
+        if (is_val<value_array>(val)) {
+            for (auto & item : val->val_arr) {
+                mark_used(item, deep);
+            }
+        } else if (is_val<value_object>(val)) {
+            for (auto & pair : val->val_obj) {
+                mark_used(pair.first, deep);
+                mark_used(pair.second, deep);
+            }
+        }
+    }
+}
+
 } // namespace jinja
@@ -118,6 +118,8 @@ struct value_t {
        bool used = false;
        // ops can be builtin calls or operators: "array_access", "object_access"
        std::set<std::string> ops;
+        // utility to recursively mark value and its children as used
+        static void mark_used(value & val, bool deep = false);
    } stats;

    value_t() = default;
@@ -1163,6 +1163,9 @@ class TextModel(ModelBase):
        if chkhsh == "b53802fb28e26d645c3a310b34bfe07da813026ec7c7716883404d5e0f8b1901":
            # ref: https://huggingface.co/core42/jais-13b
            res = "jais"
+        if chkhsh == "bc5108ee1eb6a3d600cadd065f63190fbd0554dbc9e4bbd6a0d977970afc8d2a":
+            # ref: https://huggingface.co/inceptionai/Jais-2-8B-Chat
+            res = "jais-2"
        if chkhsh == "7b3e7548e4308f52a76e8229e4e6cc831195d0d1df43aed21ac6c93da05fec5f":
            # ref: https://huggingface.co/WisdomShell/CodeShell-7B
            res = "codeshell"
@@ -1271,6 +1274,9 @@ class TextModel(ModelBase):
        if chkhsh == "b4b8ca1f9769494fbd956ebc4c249de6131fb277a4a3345a7a92c7dd7a55808d":
            # ref: https://huggingface.co/jdopensource/JoyAI-LLM-Flash
            res = "joyai-llm"
+        if chkhsh == "e4d54df1ebc1f2b91acd986c5b51aa50837d5faf7c7398e73c1f9e9ee5d19869":
+            # ref: https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601
+            res = "kanana2"

        if res is None:
            logger.warning("\n")
@@ -3730,6 +3736,13 @@ class Ernie4_5Model(TextModel):
    def set_vocab(self):
        self._set_vocab_sentencepiece()

+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()

@@ -3739,6 +3752,10 @@ class Ernie4_5Model(TextModel):
        if (head_dim := self.hparams.get("head_dim")) is None:
            head_dim = self.hparams["hidden_size"] // num_heads

+        if "mlp_AR" in name or "vision_model" in name:
+            # skip vision model and projector tensors
+            return
+
        if "ernie." in name:
            name = name.replace("ernie.", "model.")
        # split the qkv weights
@@ -3848,6 +3865,48 @@ class Ernie4_5MoeModel(Ernie4_5Model):
                raise ValueError(f"Unprocessed experts: {experts}")


+@ModelBase.register("PaddleOCRVLForConditionalGeneration")
+class PaddleOCRModel(Ernie4_5Model):
+    model_arch = gguf.MODEL_ARCH.PADDLEOCR
+
+
+@ModelBase.register("PaddleOCRVisionModel")
+class PaddleOCRVisionModel(MmprojModel):
+    # PaddleOCR-VL uses a modified version of Siglip
+    min_pixels: int = 0
+    max_pixels: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.min_pixels = self.preprocessor_config["min_pixels"]
+        self.max_pixels = self.preprocessor_config["max_pixels"]
+        self.hparams_vision["image_size"] = int(math.sqrt(self.max_pixels))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        hparams = self.hparams_vision
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PADDLEOCR)
+        self.gguf_writer.add_vision_max_pixels(self.max_pixels)
+        self.gguf_writer.add_vision_min_pixels(self.min_pixels)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("rms_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        name = name.replace("visual.", "model.")
+
+        if "vision_model" in name or "mlp_AR" in name:
+            if "packing_position_embedding" in name:
+                return # unused
+            elif "vision_model.head" in name:
+                # we don't yet support image embeddings for this model
+                return
+            else:
+                yield from super().modify_tensors(data_torch, name, bid)
+        return # skip other tensors
+
+
@ModelBase.register(
    "Qwen2VLModel",
    "Qwen2VLForConditionalGeneration",
@@ -4584,7 +4643,7 @@ class Qwen3VLVisionModel(MmprojModel):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration")
+@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")
 class Glm4VVisionModel(Qwen3VLVisionModel):
    def set_gguf_parameters(self):
        MmprojModel.set_gguf_parameters(self) # skip Qwen3VLVisionModel parameters
@@ -8633,6 +8692,17 @@ class T5EncoderModel(TextModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("Jais2ForCausalLM")
+class Jais2Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.JAIS2
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        head_dim = hparams.get("head_dim", hparams["hidden_size"] // hparams["num_attention_heads"])
+        self.gguf_writer.add_rope_dimension_count(head_dim)
+
+
@ModelBase.register("JAISLMHeadModel")
 class JaisModel(TextModel):
    model_arch = gguf.MODEL_ARCH.JAIS
@@ -8776,7 +8846,7 @@ class Glm4Model(TextModel):
            n_head = self.hparams["num_attention_heads"]
            n_kv_head = self.hparams["num_key_value_heads"]
            n_embd = self.hparams["hidden_size"]
-            head_dim = n_embd // n_head
+            head_dim = self.hparams.get("head_dim", n_embd // n_head)
            # because llama.cpp M-RoPE kernel only supports Neox ordering, we have to permute the weights here
            if name.endswith(("q_proj.weight", "q_proj.bias")):
                data_torch = Glm4Model.normal_to_neox(data_torch, n_head, n_head, head_dim, self.partial_rotary_factor)
@@ -8785,6 +8855,27 @@ class Glm4Model(TextModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("GlmOcrForConditionalGeneration")
+class GlmOCRModel(Glm4Model):
+    model_arch = gguf.MODEL_ARCH.GLM4
+    use_mrope = False
+    partial_rotary_factor = 0.5
+
+    # Note: GLM-OCR is the same as GLM4, but with an extra NextN/MTP prediction layer
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # GLM-OCR has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+            self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
+
+
@ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")
 class Glm4MoeModel(TextModel):
    model_arch = gguf.MODEL_ARCH.GLM4_MOE
@@ -10705,7 +10796,7 @@ class LFM2Model(TextModel):
    def set_gguf_parameters(self):
        # set num_key_value_heads only for attention layers
        self.hparams["num_key_value_heads"] = [
-            self.hparams["num_key_value_heads"] if layer_type == "full_attention" else 0
+            self.hparams["num_key_value_heads"] if layer_type != "conv" else 0
            for layer_type in self.hparams["layer_types"]
        ]

@@ -10891,6 +10982,28 @@ class LFM2AudioModel(ConformerAudioModel):
        yield from super().modify_tensors(data_torch, name, bid)


+@ModelBase.register("Lfm25AudioTokenizer")
+class LFM25AudioTokenizer(LFM2Model):
+    model_arch = gguf.MODEL_ARCH.LFM2
+
+    def set_vocab(self):
+        self._set_vocab_none()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])
+        self.gguf_writer.add_embedding_length_out(self.hparams["output_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "istft.window" or name.startswith("emb.emb"):
+            return
+
+        if name.startswith("lin"):
+            name = name.replace("lin", "dense_2_out")
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("SmallThinkerForCausalLM")
 class SmallThinkerModel(TextModel):
    model_arch = gguf.MODEL_ARCH.SMALLTHINKER
@@ -10982,13 +11095,17 @@ class ModernBertModel(BertModel):
        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # these layers act as MLM head, so we don't need them
-        if name.startswith("decoder."):
-            return
-
        if name.startswith("model."):
            name = name[6:]

+        if self.cls_out_labels:
+            # For BertForSequenceClassification (direct projection layer)
+            if name == "classifier.weight":
+                name = "classifier.out_proj.weight"
+
+            if name == "classifier.bias":
+                name = "classifier.out_proj.bias"
+
        yield from super().modify_tensors(data_torch, name, bid)


@@ -114,6 +114,7 @@ models = [
    {"name": "gemma",            "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2b", },
    {"name": "gemma-2",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
    {"name": "jais",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
+    {"name": "jais-2",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inceptionai/Jais-2-8B-Chat", },
    {"name": "t5",               "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
    {"name": "codeshell",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
    {"name": "tekken",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
@@ -151,6 +152,7 @@ models = [
    {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
    {"name": "qwen35",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3.5-9B-Instruct", },
    {"name": "joyai-llm",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jdopensource/JoyAI-LLM-Flash", },
+    {"name": "kanana2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct-2601", },
 ]

 # some models are known to be broken upstream, so we will skip them as exceptions
@@ -246,7 +246,7 @@ cmake --build build --config release

 1. **Retrieve and prepare model**

-    You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration.
+    You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model prepration.

    **Notes**:

@@ -281,7 +281,7 @@ as `-cl-fp32-correctly-rounded-divide-sqrt`

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q4_0.gguf?download=true) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).

 ##### Check device

@@ -569,7 +569,7 @@ Once it is completed, final results will be in **build/Release/bin**

 #### Retrieve and prepare model

-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).
+You can refer to the general [*Obtaining and quantizing models*](../../README.md#obtaining-and-quantizing-models) guide for model preparation, or download an already quantized model like [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) or [Meta-Llama-3-8B-Instruct-Q4_0.gguf](https://huggingface.co/aptha/Meta-Llama-3-8B-Instruct-Q4_0-GGUF/resolve/main/Meta-Llama-3-8B-Instruct-Q4_0.gguf).

 ##### Check device

@@ -31,7 +31,7 @@ Legend:
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
@@ -96,13 +96,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -8760,22 +8760,14 @@
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=32","support","0","no","WebGPU"
 "WebGPU: WebGPU","ADD_ID","type_a=f32,type_b=f32,n_embd=129,n_experts=8,n_experts_used=4,n_token=129","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f16,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f16,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -8786,22 +8778,14 @@
 "WebGPU: WebGPU","ROUND","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[10,5,4,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[10,5,4,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","CEIL","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROUND","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","TRUNC","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","LOG","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
-"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","CLAMP","type=f32,ne=[7,1,5,3],min=-0.500000,max=0.500000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","LEAKY_RELU","type=f32,ne_a=[7,1,5,3],negative_slope=0.100000","support","0","no","WebGPU"
 "WebGPU: WebGPU","FLOOR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
@@ -18901,3 +18885,27 @@
 "WebGPU: WebGPU","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
 "WebGPU: WebGPU","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","0","no","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f16,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[10,5,4,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[10,3,3,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[10,2,2,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQR","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SQRT","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","SIN","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[7,1,5,3]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","COS","type=f32,ne=[1024,1024,1,1]","support","1","yes","WebGPU"
@@ -77,7 +77,10 @@ causal-verify-embeddings: causal-run-original-embeddings causal-run-converted-em
 	@./scripts/causal/compare-embeddings-logits.sh

 causal-inspect-original-model:
-	@./scripts/utils/inspect-org-model.py
+	@./scripts/utils/inspect-org-model.py --list-all -s
+
+causal-list-original-model-tensors:
+	@./scripts/utils/inspect-org-model.py --list-all-short -s

 causal-inspect-converted-model:
 	@./scripts/utils/inspect-converted-model.sh
@@ -153,7 +156,7 @@ embedding-verify-logits-st: embedding-run-original-model-st embedding-run-conver

 embedding-inspect-original-model:
 	$(call validate_embedding_model_path,embedding-inspect-original-model)
-	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH}
+	@EMBEDDING_MODEL_PATH="$(EMBEDDING_MODEL_PATH)" ./scripts/utils/inspect-org-model.py -m ${EMBEDDING_MODEL_PATH} --list-all -s

 embedding-inspect-converted-model:
 	@CONVERTED_EMBEDDING_MODEL="$(CONVERTED_EMBEDDING_MODEL)" ./scripts/utils/inspect-converted-model.sh ${CONVERTED_EMBEDDING_MODEL}
@@ -1,67 +1,290 @@
 #!/usr/bin/env python3

 import argparse
-import os
 import json
+import os
+import re
+import struct
+import sys
+from pathlib import Path
+from typing import Optional
 from safetensors import safe_open
-from collections import defaultdict

-parser = argparse.ArgumentParser(description='Process model with specified path')
-parser.add_argument('--model-path', '-m', help='Path to the model')
-args = parser.parse_args()

-model_path = os.environ.get('MODEL_PATH', args.model_path)
-if model_path is None:
-    parser.error("Model path must be specified either via --model-path argument or MODEL_PATH environment variable")
+MODEL_SAFETENSORS_FILE = "model.safetensors"
+MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"

-# Check if there's an index file (multi-file model)
-index_path = os.path.join(model_path, "model.safetensors.index.json")
-single_file_path = os.path.join(model_path, "model.safetensors")
+DTYPE_SIZES = {
+    "F64": 8, "I64": 8, "U64": 8,
+    "F32": 4, "I32": 4, "U32": 4,
+    "F16": 2, "BF16": 2, "I16": 2, "U16": 2,
+    "I8": 1, "U8": 1, "BOOL": 1,
+    "F8_E4M3": 1, "F8_E5M2": 1,
+}

-if os.path.exists(index_path):
-    # Multi-file model
-    print("Multi-file model detected")
+SIZE_UNITS = ['B', 'KB', 'MB', 'GB', 'TB']

-    with open(index_path, 'r') as f:
-        index_data = json.load(f)

-    # Get the weight map (tensor_name -> file_name)
-    weight_map = index_data.get("weight_map", {})
+def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
+    index_file = model_path / MODEL_SAFETENSORS_INDEX

-    # Group tensors by file for efficient processing
-    file_tensors = defaultdict(list)
-    for tensor_name, file_name in weight_map.items():
-        file_tensors[file_name].append(tensor_name)
+    if index_file.exists():
+        with open(index_file, 'r') as f:
+            index = json.load(f)
+            return index.get("weight_map", {})

-    print("Tensors in model:")
+    return None

-    # Process each shard file
-    for file_name, tensor_names in file_tensors.items():
-        file_path = os.path.join(model_path, file_name)
-        print(f"\n--- From {file_name} ---")

-        with safe_open(file_path, framework="pt") as f:
-            for tensor_name in sorted(tensor_names):
-                tensor = f.get_tensor(tensor_name)
-                print(f"- {tensor_name} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+def get_all_tensor_names(model_path: Path) -> list[str]:
+    weight_map = get_weight_map(model_path)

-elif os.path.exists(single_file_path):
-    # Single file model (original behavior)
-    print("Single-file model detected")
+    if weight_map is not None:
+        return list(weight_map.keys())

-    with safe_open(single_file_path, framework="pt") as f:
-        keys = f.keys()
-        print("Tensors in model:")
-        for key in sorted(keys):
-            tensor = f.get_tensor(key)
-            print(f"- {key} : shape = {tensor.shape}, dtype = {tensor.dtype}")
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            with safe_open(single_file, framework="pt", device="cpu") as f:
+                return list(f.keys())
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)

-else:
-    print(f"Error: Neither 'model.safetensors.index.json' nor 'model.safetensors' found in {model_path}")
-    print("Available files:")
-    if os.path.exists(model_path):
-        for item in sorted(os.listdir(model_path)):
-            print(f"  {item}")
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        return weight_map.get(tensor_name)
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        return single_file.name
+
+    return None
+
+
+def read_safetensors_header(file_path: Path) -> dict:
+    with open(file_path, 'rb') as f:
+        header_size = struct.unpack('<Q', f.read(8))[0]
+        return json.loads(f.read(header_size))
+
+
+def get_tensor_size_bytes(tensor_meta: dict) -> int:
+    offsets = tensor_meta.get("data_offsets")
+    if offsets and len(offsets) == 2:
+        return offsets[1] - offsets[0]
+    n_elements = 1
+    for d in tensor_meta.get("shape", []):
+        n_elements *= d
+    return n_elements * DTYPE_SIZES.get(tensor_meta.get("dtype", "F32"), 4)
+
+
+def format_size(size_bytes: int) -> str:
+    val = float(size_bytes)
+    for unit in SIZE_UNITS[:-1]:
+        if val < 1024.0:
+            return f"{val:.2f} {unit}"
+        val /= 1024.0
+    return f"{val:.2f} {SIZE_UNITS[-1]}"
+
+
+def get_all_tensor_metadata(model_path: Path) -> dict[str, dict]:
+    weight_map = get_weight_map(model_path)
+
+    if weight_map is not None:
+        file_to_tensors: dict[str, list[str]] = {}
+        for tensor_name, file_name in weight_map.items():
+            file_to_tensors.setdefault(file_name, []).append(tensor_name)
+
+        all_metadata: dict[str, dict] = {}
+        for file_name, tensor_names in file_to_tensors.items():
+            try:
+                header = read_safetensors_header(model_path / file_name)
+                for tensor_name in tensor_names:
+                    if tensor_name in header:
+                        all_metadata[tensor_name] = header[tensor_name]
+            except Exception as e:
+                print(f"Warning: Could not read header from {file_name}: {e}", file=sys.stderr)
+        return all_metadata
+
+    single_file = model_path / MODEL_SAFETENSORS_FILE
+    if single_file.exists():
+        try:
+            header = read_safetensors_header(single_file)
+            return {k: v for k, v in header.items() if k != "__metadata__"}
+        except Exception as e:
+            print(f"Error reading {single_file}: {e}")
+            sys.exit(1)
+
+    print(f"Error: No safetensors files found in {model_path}")
+    sys.exit(1)
+
+
+def normalize_tensor_name(tensor_name: str) -> str:
+    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
+    normalized = re.sub(r'\.\d+$', '.#', normalized)
+    return normalized
+
+
+def list_all_tensors(
+    model_path: Path,
+    short: bool = False,
+    show_sizes: bool = False,
+):
+    tensor_names = get_all_tensor_names(model_path)
+
+    metadata: Optional[dict[str, dict]] = None
+    if show_sizes:
+        metadata = get_all_tensor_metadata(model_path)
+
+    total_bytes = 0
+
+    if short:
+        seen: dict[str, str] = {}
+        for tensor_name in sorted(tensor_names):
+            normalized = normalize_tensor_name(tensor_name)
+            if normalized not in seen:
+                seen[normalized] = tensor_name
+        display_pairs = list(sorted(seen.items()))
+        name_width = max((len(n) for n, _ in display_pairs), default=0)
+        for normalized, first_name in display_pairs:
+            if metadata and first_name in metadata:
+                m = metadata[first_name]
+                size = get_tensor_size_bytes(m)
+                total_bytes += size
+                print(f"{normalized:{name_width}}  {m.get('dtype', '?'):6s}  {str(m.get('shape', '')):30s}  {format_size(size)}")
+            else:
+                print(normalized)
    else:
-        print(f"  Directory {model_path} does not exist")
-    exit(1)
+        name_width = max((len(n) for n in tensor_names), default=0)
+        for tensor_name in sorted(tensor_names):
+            if metadata and tensor_name in metadata:
+                m = metadata[tensor_name]
+                size = get_tensor_size_bytes(m)
+                total_bytes += size
+                print(f"{tensor_name:{name_width}}  {m.get('dtype', '?'):6s}  {str(m.get('shape', '')):30s}  {format_size(size)}")
+            else:
+                print(tensor_name)
+
+    if show_sizes:
+        print(f"\nTotal: {format_size(total_bytes)}")
+
+
+def print_tensor_info(model_path: Path, tensor_name: str, num_values: Optional[int] = None):
+    tensor_file = find_tensor_file(model_path, tensor_name)
+
+    if tensor_file is None:
+        print(f"Error: Could not find tensor '{tensor_name}' in model index")
+        print(f"Model path: {model_path}")
+        sys.exit(1)
+
+    file_path = model_path / tensor_file
+
+    try:
+        header = read_safetensors_header(file_path)
+        tensor_meta = header.get(tensor_name, {})
+        dtype_str = tensor_meta.get("dtype")
+
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            if tensor_name in f.keys():
+                tensor_slice = f.get_slice(tensor_name)
+                shape = tensor_slice.get_shape()
+                print(f"Tensor: {tensor_name}")
+                print(f"File:   {tensor_file}")
+                print(f"Shape:  {shape}")
+                if dtype_str:
+                    print(f"Dtype:  {dtype_str}")
+                if tensor_meta:
+                    print(f"Size:   {format_size(get_tensor_size_bytes(tensor_meta))}")
+                if num_values is not None:
+                    tensor = f.get_tensor(tensor_name)
+                    if not dtype_str:
+                        print(f"Dtype:  {tensor.dtype}")
+                    flat = tensor.flatten()
+                    n = min(num_values, flat.numel())
+                    print(f"Values: {flat[:n].tolist()}")
+            else:
+                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
+                sys.exit(1)
+
+    except FileNotFoundError:
+        print(f"Error: The file '{file_path}' was not found.")
+        sys.exit(1)
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        sys.exit(1)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Print tensor information from a safetensors model"
+    )
+    parser.add_argument(
+        "tensor_name",
+        nargs="?",
+        help="Name of the tensor to inspect"
+    )
+    parser.add_argument(
+        "-m", "--model-path",
+        type=Path,
+        help="Path to the model directory (default: MODEL_PATH environment variable)"
+    )
+    parser.add_argument(
+        "-l", "--list-all-short",
+        action="store_true",
+        help="List unique tensor patterns (layer numbers replaced with #)"
+    )
+    parser.add_argument(
+        "-la", "--list-all",
+        action="store_true",
+        help="List all tensor names with actual layer numbers"
+    )
+    parser.add_argument(
+        "-n", "--num-values",
+        nargs="?",
+        const=10,
+        default=None,
+        type=int,
+        metavar="N",
+        help="Print the first N values of the tensor flattened (default: 10 if flag is given without a number)"
+    )
+    parser.add_argument(
+        "-s", "--sizes",
+        action="store_true",
+        help="Show dtype, shape, and size for each tensor when listing"
+    )
+
+    args = parser.parse_args()
+
+    model_path = args.model_path
+    if model_path is None:
+        model_path_str = os.environ.get("MODEL_PATH")
+        if model_path_str is None:
+            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
+            sys.exit(1)
+        model_path = Path(model_path_str)
+
+    if not model_path.exists():
+        print(f"Error: Model path does not exist: {model_path}")
+        sys.exit(1)
+
+    if not model_path.is_dir():
+        print(f"Error: Model path is not a directory: {model_path}")
+        sys.exit(1)
+
+    if args.list_all_short or args.list_all:
+        list_all_tensors(model_path, short=args.list_all_short, show_sizes=args.sizes)
+    else:
+        if args.tensor_name is None:
+            print("Error: tensor_name is required when not using --list-all-short or --list-all")
+            sys.exit(1)
+        print_tensor_info(model_path, args.tensor_name, args.num_values)
+
+
+if __name__ == "__main__":
+    main()
@@ -1,174 +0,0 @@
-#!/usr/bin/env python3
-
-import argparse
-import json
-import os
-import re
-import sys
-from pathlib import Path
-from typing import Optional
-from safetensors import safe_open
-
-
-MODEL_SAFETENSORS_FILE = "model.safetensors"
-MODEL_SAFETENSORS_INDEX = "model.safetensors.index.json"
-
-
-def get_weight_map(model_path: Path) -> Optional[dict[str, str]]:
-    index_file = model_path / MODEL_SAFETENSORS_INDEX
-
-    if index_file.exists():
-        with open(index_file, 'r') as f:
-            index = json.load(f)
-            return index.get("weight_map", {})
-
-    return None
-
-
-def get_all_tensor_names(model_path: Path) -> list[str]:
-    weight_map = get_weight_map(model_path)
-
-    if weight_map is not None:
-        return list(weight_map.keys())
-
-    single_file = model_path / MODEL_SAFETENSORS_FILE
-    if single_file.exists():
-        try:
-            with safe_open(single_file, framework="pt", device="cpu") as f:
-                return list(f.keys())
-        except Exception as e:
-            print(f"Error reading {single_file}: {e}")
-            sys.exit(1)
-
-    print(f"Error: No safetensors files found in {model_path}")
-    sys.exit(1)
-
-
-def find_tensor_file(model_path: Path, tensor_name: str) -> Optional[str]:
-    weight_map = get_weight_map(model_path)
-
-    if weight_map is not None:
-        return weight_map.get(tensor_name)
-
-    single_file = model_path / MODEL_SAFETENSORS_FILE
-    if single_file.exists():
-        return single_file.name
-
-    return None
-
-
-def normalize_tensor_name(tensor_name: str) -> str:
-    normalized = re.sub(r'\.\d+\.', '.#.', tensor_name)
-    normalized = re.sub(r'\.\d+$', '.#', normalized)
-    return normalized
-
-
-def list_all_tensors(model_path: Path, unique: bool = False):
-    tensor_names = get_all_tensor_names(model_path)
-
-    if unique:
-        seen = set()
-        for tensor_name in sorted(tensor_names):
-            normalized = normalize_tensor_name(tensor_name)
-            if normalized not in seen:
-                seen.add(normalized)
-                print(normalized)
-    else:
-        for tensor_name in sorted(tensor_names):
-            print(tensor_name)
-
-
-def print_tensor_info(model_path: Path, tensor_name: str, num_values: Optional[int] = None):
-    tensor_file = find_tensor_file(model_path, tensor_name)
-
-    if tensor_file is None:
-        print(f"Error: Could not find tensor '{tensor_name}' in model index")
-        print(f"Model path: {model_path}")
-        sys.exit(1)
-
-    file_path = model_path / tensor_file
-
-    try:
-        with safe_open(file_path, framework="pt", device="cpu") as f:
-            if tensor_name in f.keys():
-                tensor_slice = f.get_slice(tensor_name)
-                shape = tensor_slice.get_shape()
-                print(f"Tensor: {tensor_name}")
-                print(f"File:   {tensor_file}")
-                print(f"Shape:  {shape}")
-                if num_values is not None:
-                    tensor = f.get_tensor(tensor_name)
-                    print(f"Dtype:  {tensor.dtype}")
-                    flat = tensor.flatten()
-                    n = min(num_values, flat.numel())
-                    print(f"Values: {flat[:n].tolist()}")
-            else:
-                print(f"Error: Tensor '{tensor_name}' not found in {tensor_file}")
-                sys.exit(1)
-
-    except FileNotFoundError:
-        print(f"Error: The file '{file_path}' was not found.")
-        sys.exit(1)
-    except Exception as e:
-        print(f"An error occurred: {e}")
-        sys.exit(1)
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="Print tensor information from a safetensors model"
-    )
-    parser.add_argument(
-        "tensor_name",
-        nargs="?",  # optional (if --list is used for example)
-        help="Name of the tensor to inspect"
-    )
-    parser.add_argument(
-        "-m", "--model-path",
-        type=Path,
-        help="Path to the model directory (default: MODEL_PATH environment variable)"
-    )
-    parser.add_argument(
-        "-l", "--list",
-        action="store_true",
-        help="List unique tensor patterns in the model (layer numbers replaced with #)"
-    )
-    parser.add_argument(
-        "-n", "--num-values",
-        nargs="?",
-        const=10,
-        default=None,
-        type=int,
-        metavar="N",
-        help="Print the first N values of the tensor flattened (default: 10 if flag is given without a number)"
-    )
-
-    args = parser.parse_args()
-
-    model_path = args.model_path
-    if model_path is None:
-        model_path_str = os.environ.get("MODEL_PATH")
-        if model_path_str is None:
-            print("Error: --model-path not provided and MODEL_PATH environment variable not set")
-            sys.exit(1)
-        model_path = Path(model_path_str)
-
-    if not model_path.exists():
-        print(f"Error: Model path does not exist: {model_path}")
-        sys.exit(1)
-
-    if not model_path.is_dir():
-        print(f"Error: Model path is not a directory: {model_path}")
-        sys.exit(1)
-
-    if args.list:
-        list_all_tensors(model_path, unique=True)
-    else:
-        if args.tensor_name is None:
-            print("Error: tensor_name is required when not using --list")
-            sys.exit(1)
-        print_tensor_info(model_path, args.tensor_name, args.num_values)
-
-
-if __name__ == "__main__":
-    main()
@@ -5,12 +5,15 @@
 #include <vector>
 #include <cstdio>

+
 int main(int argc, char ** argv) {
    common_params params;

    params.prompt = "The quick brown fox";
    params.sampling.seed = 1234;

+    const std::string_view state_file = "dump_state.bin";
+
    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
        return 1;
    }
@@ -53,35 +56,16 @@ int main(int argc, char ** argv) {
    // tokenize prompt
    auto tokens = common_tokenize(ctx, params.prompt, true);

-    // prepare the batch
-    llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
-    for (size_t i = 0; i < tokens.size(); i++) {
-        common_batch_add(batch, tokens[i], i, {0}, false);
+    const bool save_state = true;
+    if (!common_prompt_batch_decode(ctx, tokens, n_past, params.n_batch, state_file, save_state)) {
+        return 1;
    }
-    batch.logits[batch.n_tokens - 1] = true; // generate next token
-
-    // evaluate prompt
-    llama_decode(ctx, batch);
-    n_past += batch.n_tokens;
-
-    // save state (rng, logits, embedding and kv_cache) to file
-    {
-        std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
-        const size_t written = llama_state_get_data(ctx, state_mem.data(), state_mem.size());
-
-        FILE *fp_write = fopen("dump_state.bin", "wb");
-        fwrite(state_mem.data(), 1, written, fp_write);
-        fclose(fp_write);
-
-        fprintf(stderr, "%s : serialized state into %zd out of a maximum of %zd bytes\n", __func__, written, state_mem.size());
-    }
-
-    // save state (last tokens)
-    const auto n_past_saved = n_past;

    // first run
    printf("\nfirst run: %s", params.prompt.c_str());

+    llama_batch batch = llama_batch_init(1, 0, 1);
+
    for (auto i = 0; i < params.n_predict; i++) {
        auto next_token     = llama_sampler_sample(smpl, ctx, -1);
        auto next_token_str = common_token_to_piece(ctx, next_token);
@@ -111,27 +95,23 @@ int main(int argc, char ** argv) {

    printf("\nsecond run: %s", params.prompt.c_str());

-    // load state (rng, logits, embedding and kv_cache) from file
-    {
-        std::vector<uint8_t> state_mem;
+    // load state from file
+    std::vector<llama_token> unused_sts(tokens.size()); // unused session tokens.
+    size_t n_token_count_out = 0;

-        FILE * fp_read = fopen("dump_state.bin", "rb");
-        fseek(fp_read, 0, SEEK_END);
-        state_mem.resize(ftell(fp_read));
-        fseek(fp_read, 0, SEEK_SET);
-        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        fclose(fp_read);
-
-        if (read != llama_state_set_data(ctx2, state_mem.data(), state_mem.size())) {
-            fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            return 1;
-        }
-
-        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
+    if (!llama_state_load_file(ctx2, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
+        fprintf(stderr, "\n%s : failed to load state\n", __func__);
+        return 1;
    }

+    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
+
    // restore state (last tokens)
-    n_past = n_past_saved;
+    n_past = n_token_count_out;
+    if (!common_replay_last_token(ctx2, tokens.back(), n_past)) {
+        return 1;
+    }
+    ++n_past;

    // second run
    for (auto i = 0; i < params.n_predict; i++) {
@@ -160,7 +140,9 @@ int main(int argc, char ** argv) {
    }

    // make new context
-    llama_context * ctx3 = llama_init_from_model(model, common_context_params_to_llama(params));
+    auto params_ctx3 = common_context_params_to_llama(params);
+    params_ctx3.n_seq_max = 2;
+    llama_context * ctx3 = llama_init_from_model(model, params_ctx3);

    llama_sampler * smpl3 = llama_sampler_chain_init(sparams);

@@ -169,26 +151,21 @@ int main(int argc, char ** argv) {
    printf("\nsingle seq run: %s", params.prompt.c_str());

    // load state (rng, logits, embedding and kv_cache) from file
-    {
-        std::vector<uint8_t> state_mem;
+    n_token_count_out = 0;

-        FILE * fp_read = fopen("dump_state.bin", "rb");
-        fseek(fp_read, 0, SEEK_END);
-        state_mem.resize(ftell(fp_read));
-        fseek(fp_read, 0, SEEK_SET);
-        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        fclose(fp_read);
-
-        if (read != llama_state_set_data(ctx3, state_mem.data(), state_mem.size())) {
-            fprintf(stderr, "\n%s : failed to read state\n", __func__);
-            return 1;
-        }
-
-        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
+    if (!llama_state_load_file(ctx3, state_file.data(), unused_sts.data(), unused_sts.size(), &n_token_count_out)) {
+        fprintf(stderr, "\n%s : failed to load state\n", __func__);
+        return 1;
    }

+    fprintf(stderr, "%s : loaded state with %zu tokens\n", __func__, n_token_count_out);
+
    // restore state (last tokens)
-    n_past = n_past_saved;
+    n_past = n_token_count_out;
+    if (!common_replay_last_token(ctx3, tokens.back(), n_past)) {
+        return 1;
+    }
+    ++n_past;

    // save seq 0 and load into seq 1
    {
@@ -730,10 +730,6 @@ extern "C" {
    GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
    GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row

-    GGML_DEPRECATED(
-    GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
-    "use ggml_row_size() instead");
-
    GGML_API const char * ggml_type_name(enum ggml_type type);
    GGML_API const char * ggml_op_name  (enum ggml_op   op);
    GGML_API const char * ggml_op_symbol(enum ggml_op   op);
@@ -42,6 +42,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -55,9 +56,10 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
-#define ggml_gemm_q6_K_8x8_q8_K_generic   ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
 #define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
@@ -77,6 +79,7 @@
 #define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
 #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -86,6 +89,7 @@
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -110,6 +114,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -123,6 +128,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -148,6 +154,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -161,6 +168,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -171,15 +179,9 @@
 #elif defined(__riscv)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
-#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
-#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
 #define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
-#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
 #define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
-#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
-#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
-#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
@@ -193,6 +195,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -205,6 +208,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -236,6 +240,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -249,6 +254,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -282,6 +288,7 @@
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
 #define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
 #define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
 #define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
@@ -295,6 +302,7 @@
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
 #define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
 #define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
 #define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
@@ -785,6 +785,165 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }

+void ggml_gemv_q5_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    col_groups = ncols_interleaved / 4;  // 0123 and 4567
+    const uint8x16_t m4b        = vdupq_n_u8(0x0f);
+    const uint8x16_t mone       = vdupq_n_u8(1);
+    const uint8x16_t mtwo       = vdupq_n_u8(2);
+
+    // 1x8 tile = 2 x 4
+    float32x4_t acc_f32[col_groups];
+
+    const block_q8_K * GGML_RESTRICT q8_ptr = (const block_q8_K *) vy;
+
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int i = 0; i < col_groups; i++) {
+            acc_f32[i] = vdupq_n_f32(0);
+        }
+
+        for (int b = 0; b < nb; b++) {
+            float32x4_t q5_d_0        = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d));      // d0 d1 d2 d3
+            float32x4_t q5_d_1        = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d + 4));  // d4 d5 d6 d7
+            float32x4_t q8_d          = vdupq_n_f32(q8_ptr[b].d);
+            float32x4_t sb_scale_0123 = vmulq_f32(q5_d_0, q8_d);
+            float32x4_t sb_scale_4567 = vmulq_f32(q5_d_1, q8_d);
+            float32x4_t q5_dmin_0     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin));      // dmin 0..3
+            float32x4_t q5_dmin_1     = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin + 4));  // dmin 4..7
+            float32x4_t sb_min_0123   = vmulq_f32(q5_dmin_0, q8_d);
+            float32x4_t sb_min_4567   = vmulq_f32(q5_dmin_1, q8_d);
+
+            // interleaved bias_acc: [0]->r0 0123, [1]->r0 4567
+            int32x4_t bias_acc[2] = { vdupq_n_s32(0), vdupq_n_s32(0) };
+            int32x4_t acc_lo[col_groups];
+            int32x4_t acc_hi[col_groups];
+
+            // Each bsum is 16 elements, pairwise add leaves us with the 8 bsums of the entire block
+            const int16x8_t bsums = vpaddq_s16(vld1q_s16(q8_ptr[b].bsums), vld1q_s16(q8_ptr[b].bsums + 8));
+            int16_t         bsums_arr[8];
+            vst1q_s16(bsums_arr, bsums);
+
+            uint8x16_t qh[col_groups][8];
+            for (int c = 0; c < col_groups; c++) {
+                for (int i = 0; i < 8; i++) {
+                    qh[c][i] = vld1q_u8(q5_ptr[b].qh + i * 32 + 16 * c);
+                }
+            }
+
+            for (int sb = 0; sb < QK_K / 64; sb++) {
+                for (int i = 0; i < col_groups; i++) {
+                    acc_lo[i] = vdupq_n_s32(0);
+                    acc_hi[i] = vdupq_n_s32(0);
+                }
+                // Need scales for the low and high nibbles
+                // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                int16x8_t q5sb_mins[2];
+                int16x8_t q5sb_scales[2];
+                for (int i = 0; i < 2; i++) {
+                    int8_t    aux_q5sb[8];
+                    const int offset = sb * 24 + i * 12;
+                    decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], aux_q5sb);
+                    q5sb_scales[i] = vmovl_s8(vld1_s8(aux_q5sb));
+                }
+
+                int8x16_t q8_qs[4];
+                for (int i = 0; i < 4; i++) {
+                    q8_qs[i] = vld1q_s8(q8_ptr[b].qs + sb * 64 + i * 16);
+                }
+
+                for (int c = 0; c < col_groups; c++) {
+                    uint8x16_t q5_cols[8];
+                    uint8x16_t hbit_lo[8];
+                    uint8x16_t hbit_hi[8];
+                    int8x16_t  q5_lo[8];
+                    int8x16_t  q5_hi[8];
+
+                    for (int i = 0; i < 8; i++) {
+                        q5_cols[i] = vld1q_u8(q5_ptr[b].qs + sb * QK_K + i * 32 + 16 * c);
+                        hbit_lo[i] = vandq_u8(qh[c][i], mone);
+                        hbit_hi[i] = vshlq_n_u8(vandq_u8(qh[c][i], mtwo), 3);
+                        qh[c][i]   = vshrq_n_u8(qh[c][i], 2);
+                        q5_lo[i]   = vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q5_cols[i], m4b), hbit_lo[i], 4));
+                        q5_hi[i]   = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5_cols[i], 4), hbit_hi[i]));
+                    }
+
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[0], q8_qs[0], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[1], q8_qs[0], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[2], q8_qs[0], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[3], q8_qs[0], 3);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[4], q8_qs[1], 0);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[5], q8_qs[1], 1);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[6], q8_qs[1], 2);
+                    acc_lo[c] = vdotq_laneq_s32(acc_lo[c], q5_lo[7], q8_qs[1], 3);
+
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[0], q8_qs[2], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[1], q8_qs[2], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[2], q8_qs[2], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[3], q8_qs[2], 3);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[4], q8_qs[3], 0);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[5], q8_qs[3], 1);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[6], q8_qs[3], 2);
+                    acc_hi[c] = vdotq_laneq_s32(acc_hi[c], q5_hi[7], q8_qs[3], 3);
+                }
+
+                // Scales
+                // row c0123 blk0 and blk1
+                const int16x4_t   sc_0123_lo = vget_low_s16(q5sb_scales[0]);
+                const int16x4_t   sc_0123_hi = vget_low_s16(q5sb_scales[1]);
+                const float32x4_t sumf_0123  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[0]),
+                                                                       vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[0])));
+                acc_f32[0]                   = vfmaq_f32(acc_f32[0], sb_scale_0123, sumf_0123);
+                // row c4567 blk0 and blk1
+                const int16x4_t   sc_4567_lo = vget_high_s16(q5sb_scales[0]);
+                const int16x4_t   sc_4567_hi = vget_high_s16(q5sb_scales[1]);
+                const float32x4_t sumf_4567  = vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[1]),
+                                                                       vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[1])));
+                acc_f32[1]                   = vfmaq_f32(acc_f32[1], sb_scale_4567, sumf_4567);
+
+                // Bias Correction
+                const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[2 * sb + 0]);
+                const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[2 * sb + 1]);
+
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_lo, vget_low_s16(q5sb_mins[0]));
+                bias_acc[0] = vmlal_s16(bias_acc[0], bsums_vec_hi, vget_low_s16(q5sb_mins[1]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_lo, vget_high_s16(q5sb_mins[0]));
+                bias_acc[1] = vmlal_s16(bias_acc[1], bsums_vec_hi, vget_high_s16(q5sb_mins[1]));
+            }  // for sb
+
+            acc_f32[0] = vmlsq_f32(acc_f32[0], vcvtq_f32_s32(bias_acc[0]), sb_min_0123);
+            acc_f32[1] = vmlsq_f32(acc_f32[1], vcvtq_f32_s32(bias_acc[1]), sb_min_4567);
+        }  // for b
+
+        int base = x * ncols_interleaved;
+        vst1q_f32(s + base, acc_f32[0]);
+        vst1q_f32(s + base + 4, acc_f32[1]);
+    }  // for x
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemv_q5_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemv_q5_K_8x8_q8_K(int                        n,
                             float * GGML_RESTRICT      s,
                             size_t                     bs,
@@ -3205,6 +3364,235 @@ void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
    ggml_gemm_q4_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
 }

+void ggml_gemm_q5_K_8x4_q8_K(int                        n,
+                             float * GGML_RESTRICT      s,
+                             size_t                     bs,
+                             const void * GGML_RESTRICT vx,
+                             const void * GGML_RESTRICT vy,
+                             int                        nr,
+                             int                        nc) {
+    constexpr int qk = QK_K;
+    const int     nb = n / qk;
+
+    constexpr int ncols_interleaved = 8;
+    constexpr int blocklen          = 4;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    constexpr int    q8_k_blocklen = 4;
+    constexpr int    acc_size      = 2 * 4;  // 2 row pairs, 4 col pairs
+    constexpr int    col_groups    = ncols_interleaved / 4;
+    const uint8x16_t m4b           = vdupq_n_u8(0x0f);
+    const uint8x16_t mone          = vdupq_n_u8(1);
+    const uint8x16_t mtwo          = vdupq_n_u8(2);
+
+    // 8 accumulators: 2 row pairs, 4 col pairs
+    float32x4_t acc_f32[acc_size];
+
+    for (int y = 0; y < nr / q8_k_blocklen; y++) {
+        const block_q8_Kx4 * GGML_RESTRICT q8_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * GGML_RESTRICT q5_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+            for (int i = 0; i < acc_size; i++) {
+                acc_f32[i] = vdupq_n_f32(0);
+            }
+
+            for (int b = 0; b < nb; b++) {
+                // d5 0 1 2 3, 4 5 6 7
+                float32x4_t q5_d_0123    = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d));
+                float32x4_t q5_d_4567    = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].d + 4));
+                // d8 0 1 2 3
+                float32x4_t q8_d_0123    = vld1q_f32(q8_ptr[b].d);
+                // mins
+                float32x4_t q5_dmin_0123 = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin));
+                float32x4_t q5_dmin_4567 = vcvt_f32_f16(vld1_f16((const __fp16 *) q5_ptr[b].dmin + 4));
+
+                // Precomputation of scales and mins
+                float32x4_t sbd_scale_0123[q8_k_blocklen];
+                float32x4_t sbd_scale_4567[q8_k_blocklen];
+                float32x4_t sbd_min_0123[q8_k_blocklen];
+                float32x4_t sbd_min_4567[q8_k_blocklen];
+
+                sbd_scale_0123[0] = vmulq_laneq_f32(q5_d_0123, q8_d_0123, 0);
+                sbd_scale_4567[0] = vmulq_laneq_f32(q5_d_4567, q8_d_0123, 0);
+                sbd_min_0123[0]   = vmulq_laneq_f32(q5_dmin_0123, q8_d_0123, 0);
+                sbd_min_4567[0]   = vmulq_laneq_f32(q5_dmin_4567, q8_d_0123, 0);
+
+                sbd_scale_0123[1] = vmulq_laneq_f32(q5_d_0123, q8_d_0123, 1);
+                sbd_scale_4567[1] = vmulq_laneq_f32(q5_d_4567, q8_d_0123, 1);
+                sbd_min_0123[1]   = vmulq_laneq_f32(q5_dmin_0123, q8_d_0123, 1);
+                sbd_min_4567[1]   = vmulq_laneq_f32(q5_dmin_4567, q8_d_0123, 1);
+
+                sbd_scale_0123[2] = vmulq_laneq_f32(q5_d_0123, q8_d_0123, 2);
+                sbd_scale_4567[2] = vmulq_laneq_f32(q5_d_4567, q8_d_0123, 2);
+                sbd_min_0123[2]   = vmulq_laneq_f32(q5_dmin_0123, q8_d_0123, 2);
+                sbd_min_4567[2]   = vmulq_laneq_f32(q5_dmin_4567, q8_d_0123, 2);
+
+                sbd_scale_0123[3] = vmulq_laneq_f32(q5_d_0123, q8_d_0123, 3);
+                sbd_scale_4567[3] = vmulq_laneq_f32(q5_d_4567, q8_d_0123, 3);
+                sbd_min_0123[3]   = vmulq_laneq_f32(q5_dmin_0123, q8_d_0123, 3);
+                sbd_min_4567[3]   = vmulq_laneq_f32(q5_dmin_4567, q8_d_0123, 3);
+
+                // Precomputation of bsums, each vpaddq calcs all the bsums for each row
+                const int16x8_t bsums[q8_k_blocklen] = {
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 0), vld1q_s16(q8_ptr[b].bsums + 16 * 0 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 1), vld1q_s16(q8_ptr[b].bsums + 16 * 1 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 2), vld1q_s16(q8_ptr[b].bsums + 16 * 2 + 8)),
+                    vpaddq_s16(vld1q_s16(q8_ptr[b].bsums + 16 * 3), vld1q_s16(q8_ptr[b].bsums + 16 * 3 + 8)),
+                };
+                int16_t bsums_arr[QK_K / 64][8];
+                for (int q8_row = 0; q8_row < 4; q8_row++) {
+                    vst1q_s16(bsums_arr[q8_row], bsums[q8_row]);
+                }
+
+                // interleaved bias_acc: [0]->r0 0123, [1]->r1 0123, .., [4]->r0 4567, [5]->r1 4567 ..
+                int32x4_t bias_acc[acc_size];
+                for (int i = 0; i < acc_size; i++) {
+                    bias_acc[i] = vdupq_n_s32(0);
+                }
+
+                uint8x16_t qh[col_groups][8];
+                for (int c = 0; c < col_groups; c++) {
+                    for (int i = 0; i < 8; i++) {
+                        qh[c][i] = vld1q_u8(q5_ptr[b].qh + i * 32 + 16 * c);
+                    }
+                }
+
+                for (int sb = 0; sb < QK_K / 64; sb++) {
+                    // Int accumulators for qs vecdot (4 row * 2 col quartets)
+                    int32x4_t acc_lo[acc_size];
+                    int32x4_t acc_hi[acc_size];
+                    for (int i = 0; i < acc_size; i++) {
+                        acc_lo[i] = vdupq_n_s32(0);
+                        acc_hi[i] = vdupq_n_s32(0);
+                    }
+                    // Need scales for the low and high nibbles
+                    // 2 * 12 = 24 bytes per subblock, 4 sbs -> 4 * 24 = 96 bytes total
+                    int16x8_t q5sb_scales[2];
+                    int16x8_t q5sb_mins[2];
+                    for (int i = 0; i < 2; i++) {
+                        int8_t    aux_q5sb[8];
+                        const int offset = sb * 24 + i * 12;
+                        decode_q_Kx8_6bit_scales(&q5_ptr[b].scales[offset], &q5sb_mins[i], aux_q5sb);
+                        q5sb_scales[i] = vmovl_s8(vld1_s8(aux_q5sb));
+                    }
+
+                    constexpr int reads_per_sb = 8;  // 8 * 16 bytes each => 32 qs * 4 rows
+                    for (int k = 0; k < reads_per_sb; k++) {
+                        const int8x16_t q8_blk0 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k);
+                        const int8x16_t q8_blk1 = vld1q_s8(q8_ptr[b].qs + sb * 256 + 16 * k + 128);
+
+                        // 0..3 & 32..35
+                        const uint8x16_t q5_0123 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 32 * k);
+                        const uint8x16_t q5_4567 = vld1q_u8(q5_ptr[b].qs + sb * QK_K + 32 * k + 16);
+
+                        // NOTE: This is the only difference with q4_K
+                        const uint8x16_t hbit_lo_0123 = vandq_u8(qh[0][k], mone);
+                        const uint8x16_t hbit_hi_0123 = vshlq_n_u8(vandq_u8(qh[0][k], mtwo), 3);
+                        qh[0][k]                      = vshrq_n_u8(qh[0][k], 2);
+                        const uint8x16_t hbit_lo_4567 = vandq_u8(qh[1][k], mone);
+                        const uint8x16_t hbit_hi_4567 = vshlq_n_u8(vandq_u8(qh[1][k], mtwo), 3);
+                        qh[1][k]                      = vshrq_n_u8(qh[1][k], 2);
+                        // From here, same as q4_K
+
+                        const int8x16_t q5_0123_lo =
+                            vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q5_0123, m4b), hbit_lo_0123, 4));
+                        const int8x16_t q5_0123_hi =
+                            vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5_0123, 4), hbit_hi_0123));
+
+                        acc_lo[0] = vdotq_laneq_s32(acc_lo[0], q5_0123_lo, q8_blk0, 0);  //  0..3  r0 c0123
+                        acc_lo[1] = vdotq_laneq_s32(acc_lo[1], q5_0123_lo, q8_blk0, 1);  //  0..3  r1 c0123
+                        acc_lo[2] = vdotq_laneq_s32(acc_lo[2], q5_0123_lo, q8_blk0, 2);  //  0..3  r2 c0123
+                        acc_lo[3] = vdotq_laneq_s32(acc_lo[3], q5_0123_lo, q8_blk0, 3);  //  0..3  r3 c0123
+
+                        acc_hi[0] = vdotq_laneq_s32(acc_hi[0], q5_0123_hi, q8_blk1, 0);  // 32..35 r0 c0123
+                        acc_hi[1] = vdotq_laneq_s32(acc_hi[1], q5_0123_hi, q8_blk1, 1);  // 32..35 r1 c0123
+                        acc_hi[2] = vdotq_laneq_s32(acc_hi[2], q5_0123_hi, q8_blk1, 2);  // 32..35 r2 c0123
+                        acc_hi[3] = vdotq_laneq_s32(acc_hi[3], q5_0123_hi, q8_blk1, 3);  // 32..35 r3 c0123
+
+                        const int8x16_t q5_4567_lo =
+                            vreinterpretq_s8_u8(vsliq_n_u8(vandq_u8(q5_4567, m4b), hbit_lo_4567, 4));
+                        const int8x16_t q5_4567_hi =
+                            vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5_4567, 4), hbit_hi_4567));
+
+                        acc_lo[4] = vdotq_laneq_s32(acc_lo[4], q5_4567_lo, q8_blk0, 0);  //  0..3  r0 c4567
+                        acc_lo[5] = vdotq_laneq_s32(acc_lo[5], q5_4567_lo, q8_blk0, 1);  //  0..3  r1 c4567
+                        acc_lo[6] = vdotq_laneq_s32(acc_lo[6], q5_4567_lo, q8_blk0, 2);  //  0..3  r2 c4567
+                        acc_lo[7] = vdotq_laneq_s32(acc_lo[7], q5_4567_lo, q8_blk0, 3);  //  0..3  r3 c4567
+
+                        acc_hi[4] = vdotq_laneq_s32(acc_hi[4], q5_4567_hi, q8_blk1, 0);  // 32..35 r0 c4567
+                        acc_hi[5] = vdotq_laneq_s32(acc_hi[5], q5_4567_hi, q8_blk1, 1);  // 32..35 r1 c4567
+                        acc_hi[6] = vdotq_laneq_s32(acc_hi[6], q5_4567_hi, q8_blk1, 2);  // 32..35 r2 c4567
+                        acc_hi[7] = vdotq_laneq_s32(acc_hi[7], q5_4567_hi, q8_blk1, 3);  // 32..35 r3 c4567
+                    }
+
+                    // Scale and bias application
+                    // acc is stored interleaved to match output layout
+                    const int16x4_t sc_0123_lo = vget_low_s16(q5sb_scales[0]);
+                    const int16x4_t sc_4567_lo = vget_high_s16(q5sb_scales[0]);
+                    const int16x4_t sc_0123_hi = vget_low_s16(q5sb_scales[1]);
+                    const int16x4_t sc_4567_hi = vget_high_s16(q5sb_scales[1]);
+                    for (int row = 0; row < q8_k_blocklen; row++) {
+                        // Bias correction
+                        // row c0123 blk0 and blk1
+                        const float32x4_t sumf_0123 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_0123_lo), acc_lo[row]),
+                                                    vmulq_s32(vmovl_s16(sc_0123_hi), acc_hi[row])));
+                        acc_f32[2 * row] = vfmaq_f32(acc_f32[2 * row], sbd_scale_0123[row], sumf_0123);
+
+                        // row c4567 blk0 and blk1
+                        const float32x4_t sumf_4567 =
+                            vcvtq_f32_s32(vaddq_s32(vmulq_s32(vmovl_s16(sc_4567_lo), acc_lo[row + 4]),
+                                                    vmulq_s32(vmovl_s16(sc_4567_hi), acc_hi[row + 4])));
+                        acc_f32[2 * row + 1] = vfmaq_f32(acc_f32[2 * row + 1], sbd_scale_4567[row], sumf_4567);
+
+                        // Bias
+                        const int16x4_t bsums_vec_lo = vdup_n_s16(bsums_arr[sb][row * 2]);
+                        const int16x4_t bsums_vec_hi = vdup_n_s16(bsums_arr[sb][row * 2 + 1]);
+
+                        // row c0123 blk0 and blk1
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_lo, vget_low_s16(q5sb_mins[0]));
+                        bias_acc[2 * row] = vmlal_s16(bias_acc[2 * row], bsums_vec_hi, vget_low_s16(q5sb_mins[1]));
+
+                        // row c4567 blk0 and blk1
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_lo, vget_high_s16(q5sb_mins[0]));
+                        bias_acc[2 * row + 1] =
+                            vmlal_s16(bias_acc[2 * row + 1], bsums_vec_hi, vget_high_s16(q5sb_mins[1]));
+                    }
+                }  // for sb
+
+                for (int row = 0; row < q8_k_blocklen; row++) {
+                    acc_f32[2 * row] = vmlsq_f32(acc_f32[2 * row], vcvtq_f32_s32(bias_acc[2 * row]), sbd_min_0123[row]);
+                    acc_f32[2 * row + 1] =
+                        vmlsq_f32(acc_f32[2 * row + 1], vcvtq_f32_s32(bias_acc[2 * row + 1]), sbd_min_4567[row]);
+                }
+            }  // for b
+
+            for (int i = 0; i < q8_k_blocklen; i++) {
+                int row = y * q8_k_blocklen + i;
+                for (int j = 0; j < 2; j++) {
+                    int col    = x * ncols_interleaved + j * 4;
+                    int offset = row * bs + col;
+                    vst1q_f32(s + offset, acc_f32[2 * i + j]);
+                }
+            }
+        }  // for x
+    }  // for y
+    return;
+#endif  // defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    ggml_gemm_q5_K_8x4_q8_K_generic(n, s, bs, vx, vy, nr, nc);
+}
+
 void ggml_gemm_q4_K_8x8_q8_K(int                        n,
                             float * GGML_RESTRICT      s,
                             size_t                     bs,
@@ -1954,3 +1954,773 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 #endif
 }

+static const uint8_t sign_gather_indices_arr[64] = {
+    0,0,0,0,0,0,0,0, 1,1,1,1,1,1,1,1, 2,2,2,2,2,2,2,2, 3,3,3,3,3,3,3,3,
+    4,4,4,4,4,4,4,4, 5,5,5,5,5,5,5,5, 6,6,6,6,6,6,6,6, 7,7,7,7,7,7,7,7
+};
+
+static const uint8_t sign_bit_masks_arr[64] = {
+    1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128,
+    1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128, 1,2,4,8,16,32,64,128
+};
+
+static void ggml_vec_dot_iq2_s_q8_K_vl256(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);
+    UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    const uint64_t * grid64 = (const uint64_t *)iq2s_grid;
+
+    // --- Pre-load Constants ---
+    uint16_t gather_qh_arr[8] = {0, 0, 0, 0, 1, 1, 1, 1};
+    vuint16mf2_t v_gather_qh = __riscv_vle16_v_u16mf2(gather_qh_arr, 8);
+    uint16_t shift_qh_arr[8] = {11, 9, 7, 5, 11, 9, 7, 5};
+    vuint16mf2_t v_shift_qh = __riscv_vle16_v_u16mf2(shift_qh_arr, 8);
+
+    // Constants for sign extraction
+    vuint8m2_t v_sign_gather_indices = __riscv_vle8_v_u8m2(sign_gather_indices_arr, 64);
+    vuint8m2_t v_sign_masks = __riscv_vle8_v_u8m2(sign_bit_masks_arr, 64);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float combined_scale = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const uint8_t * signs_ptr = qs + 32;
+
+        float sum_block = 0.0f;
+
+        for (int ib = 0; ib < 4; ++ib) {
+            // Combine low + high bits
+            vuint8mf4_t v_qs_u8 = __riscv_vle8_v_u8mf4(qs, 8);
+            qs += 8;
+            uint16_t qh_val;
+            memcpy(&qh_val, qh, 2);
+            qh += 2;
+            vuint8mf8_t v_qh_raw = __riscv_vle8_v_u8mf8((const uint8_t*)&qh_val, 2);
+            vuint16mf4_t v_qh_u16 = __riscv_vwcvtu_x_x_v_u16mf4(v_qh_raw, 2);
+            vuint16mf2_t v_qh_u16_ext = __riscv_vlmul_ext_v_u16mf4_u16mf2(v_qh_u16);
+            vuint16mf2_t v_qh_expanded = __riscv_vrgather_vv_u16mf2(v_qh_u16_ext, v_gather_qh, 8);
+            v_qh_expanded = __riscv_vsll_vv_u16mf2(v_qh_expanded, v_shift_qh, 8);
+
+            // Mask: We want bits 11-12. 0x1800 = 0001 1000 0000 0000
+            v_qh_expanded = __riscv_vand_vx_u16mf2(v_qh_expanded, 0x1800, 8);
+            vuint16mf2_t v_qs_u16 = __riscv_vwcvtu_x_x_v_u16mf2(v_qs_u8, 8);
+
+            // Multiply by 8 to get byte offset, instead of element offset
+            v_qs_u16 = __riscv_vsll_vx_u16mf2(v_qs_u16, 3, 8);
+            vuint16mf2_t v_grid_offsets = __riscv_vor_vv_u16mf2(v_qs_u16, v_qh_expanded, 8);
+
+            // Lookup Grid using Byte Offsets
+            vuint64m2_t v_grid_vals = __riscv_vluxei16_v_u64m2(grid64, v_grid_offsets, 8);
+
+            vuint8m2_t v_grid_u8 = __riscv_vreinterpret_v_u64m2_u8m2(v_grid_vals);
+            vint8m2_t v_grid_i8 = __riscv_vreinterpret_v_u8m2_i8m2(v_grid_u8);
+
+            // Load signs and generate sign mask
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs_ptr, 8);
+            signs_ptr += 8;
+
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 64);
+
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 64);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 64);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 64);
+            q8 += 64;
+
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative, v_q8, v_q8, 0, 64);
+            vint16m4_t v_dot = __riscv_vwmul_vv_i16m4(v_grid_i8, v_q8_signed, 64);
+
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            int32_t s0 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 0), v_zero, 16));
+            int32_t s1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 1), v_zero, 16));
+            int32_t s2 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 2), v_zero, 16));
+            int32_t s3 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m1_i32m1(
+                __riscv_vget_v_i16m4_i16m1(v_dot, 3), v_zero, 16));
+
+            uint8_t sc0 = scales[0];
+            uint8_t sc1 = scales[1];
+            scales += 2;
+
+            sum_block += s0 * (2 * (sc0 & 0xF) + 1);
+            sum_block += s1 * (2 * (sc0 >> 4)  + 1);
+            sum_block += s2 * (2 * (sc1 & 0xF) + 1);
+            sum_block += s3 * (2 * (sc1 >> 4)  + 1);
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+static void ggml_vec_dot_iq2_s_q8_K_vl128(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);
+    UNUSED(nrc); UNUSED(bx); UNUSED(by); UNUSED(bs);
+
+    const block_iq2_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+    const uint64_t * grid64 = (const uint64_t *)iq2s_grid;
+
+    // Pre-load Constants
+    vuint8m2_t v_ids = __riscv_vid_v_u8m2(32);
+    vuint8m2_t v_sign_gather_indices = __riscv_vsrl_vx_u8m2(v_ids, 3, 32);
+    vuint8m2_t v_ones = __riscv_vmv_v_x_u8m2(1, 32);
+    vuint8m2_t v_shift_amts = __riscv_vand_vx_u8m2(v_ids, 7, 32);
+    vuint8m2_t v_sign_masks = __riscv_vsll_vv_u8m2(v_ones, v_shift_amts, 32);
+    uint16_t shift_qh_arr[4] = {11, 9, 7, 5};
+    vuint16mf2_t v_shift_qh = __riscv_vle16_v_u16mf2(shift_qh_arr, 4);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float combined_scale = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        const uint8_t * signs_ptr = qs + 32;
+        float sum_block = 0.0f;
+
+        for (int ib = 0; ib < 8; ++ib) {
+
+            // Load Low Bits [4 bytes]
+            vuint8mf4_t v_qs_u8 = __riscv_vle8_v_u8mf4(qs, 4);
+            qs += 4;
+
+            // Load 1 byte. It contains bits for 4 mini-blocks.
+            uint8_t qh_val = *qh++;
+
+            // Combine Low + High bits of 10bit indices
+            vuint8mf4_t v_qh_raw = __riscv_vmv_v_x_u8mf4(qh_val, 4);
+            vuint16mf2_t v_qh_u16 = __riscv_vwcvtu_x_x_v_u16mf2(v_qh_raw, 4);
+            vuint16mf2_t v_qh_mf2 = __riscv_vsll_vv_u16mf2(v_qh_u16, v_shift_qh, 4);
+            v_qh_mf2 = __riscv_vand_vx_u16mf2(v_qh_mf2, 0x1800, 4);
+            vuint16mf2_t v_qs_u16_mf2 = __riscv_vwcvtu_x_x_v_u16mf2(v_qs_u8, 4);
+            vuint16mf2_t v_qs_u16 = __riscv_vsll_vx_u16mf2(v_qs_u16_mf2, 3, 4);
+            vuint16mf2_t v_grid_offsets = __riscv_vor_vv_u16mf2(v_qs_u16, v_qh_mf2, 4);
+
+            // Lookup Grid
+            vint8m2_t v_grid_i8 = __riscv_vreinterpret_v_u8m2_i8m2(__riscv_vreinterpret_v_u64m2_u8m2(__riscv_vluxei16_v_u64m2(grid64, v_grid_offsets, 4)));
+
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs_ptr, 4);
+            signs_ptr += 4;
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 32);
+
+            // generating sign mask
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 32);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 32);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 32);
+            q8 += 32;
+
+            // apply signs
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative,v_q8, v_q8, 0, 32);
+            vint16m4_t v_dot = __riscv_vwmul_vv_i16m4(v_grid_i8, v_q8_signed, 32);
+
+            // Reduction
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            // Reduce 0-15 (First Half)
+            int32_t s0 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(
+                __riscv_vget_v_i16m4_i16m2(v_dot, 0), v_zero, 16));
+
+            // Reduce 16-31 (Second Half)
+            int32_t s1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(
+                __riscv_vget_v_i16m4_i16m2(v_dot, 1), v_zero, 16));
+
+            // Apply sub Scales
+            uint8_t sc = *scales++;
+
+            sum_block += s0 * (2 * (sc & 0xF) + 1);
+            sum_block += s1 * (2 * (sc >> 4)  + 1);
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 128:
+            ggml_vec_dot_iq2_s_q8_K_vl128(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        case 256:
+            ggml_vec_dot_iq2_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq3_s_q8_K_vl256(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);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    const uint64_t * grid64 = (const uint64_t *)iq3s_grid;
+
+    // --- Pre-load Constants ---
+    const uint16_t qh_bit_shifts_arr[16] = {
+        0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15
+    };
+    vuint8m2_t v_sign_gather_indices = __riscv_vle8_v_u8m2(sign_gather_indices_arr, 64);
+    vuint8m2_t v_sign_masks = __riscv_vle8_v_u8m2(sign_bit_masks_arr, 64);
+    vuint16m1_t v_qh_shifts = __riscv_vle16_v_u16m1(qh_bit_shifts_arr, 16);
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float combined_scale = d * y[i].d;
+
+        const uint8_t * GGML_RESTRICT qs = x[i].qs;
+        const uint8_t * GGML_RESTRICT qh = x[i].qh;
+        const uint8_t * GGML_RESTRICT scales = x[i].scales;
+        const uint8_t * GGML_RESTRICT signs = x[i].signs;
+        const int8_t  * GGML_RESTRICT q8 = y[i].qs;
+
+        float sum_block = 0.0f;
+
+        // Loop: Process 64 weights (16 mini-blocks of 4) per iteration
+        for (int ib = 0; ib < 4; ++ib) {
+
+            vuint8mf2_t v_qs_u8 = __riscv_vle8_v_u8mf2(qs, 16);
+            qs += 16;
+
+            uint16_t qh_val;
+            memcpy(&qh_val, qh, 2);
+            qh += 2;
+
+            vuint16m1_t v_qh_val = __riscv_vmv_v_x_u16m1(qh_val, 16);
+            // Extract bits: (qh >> i) & 1
+            v_qh_val = __riscv_vsrl_vv_u16m1(v_qh_val, v_qh_shifts, 16);
+            v_qh_val = __riscv_vand_vx_u16m1(v_qh_val, 1, 16);
+
+            vuint16m1_t v_qs_u16 = __riscv_vwcvtu_x_x_v_u16m1(v_qs_u8, 16);
+            v_qs_u16 = __riscv_vsll_vx_u16m1(v_qs_u16, 2, 16);
+            v_qh_val = __riscv_vsll_vx_u16m1(v_qh_val, 10, 16);
+            vuint16m1_t v_grid_offsets = __riscv_vor_vv_u16m1(v_qs_u16, v_qh_val, 16);
+
+            // Grid value is 4xuint8
+            vuint32m2_t v_grid_packed = __riscv_vluxei16_v_u32m2((const uint32_t *)grid64, v_grid_offsets, 16);
+            vuint8m2_t v_grid_u8 = __riscv_vreinterpret_v_u32m2_u8m2(v_grid_packed);
+            vuint8mf4_t v_signs_raw = __riscv_vle8_v_u8mf4(signs, 8);
+            signs += 8;
+
+            // Generate sign mask
+            vuint8m2_t v_signs_source = __riscv_vlmul_ext_v_u8mf4_u8m2(v_signs_raw);
+            vuint8m2_t v_signs_bcast = __riscv_vrgather_vv_u8m2(v_signs_source, v_sign_gather_indices, 64);
+            vuint8m2_t v_sign_bits = __riscv_vand_vv_u8m2(v_signs_bcast, v_sign_masks, 64);
+            vbool4_t m_negative = __riscv_vmsne_vx_u8m2_b4(v_sign_bits, 0, 64);
+
+            vint8m2_t v_q8 = __riscv_vle8_v_i8m2(q8, 64);
+            q8 += 64;
+
+            // Apply Signs
+            vint8m2_t v_q8_signed = __riscv_vrsub_vx_i8m2_mu(m_negative, v_q8, v_q8, 0, 64);
+            vint16m4_t v_dot = __riscv_vwmulsu_vv_i16m4(v_q8_signed, v_grid_u8, 64);
+
+            // Reduction
+            vint16m2_t v_dot_lo = __riscv_vget_v_i16m4_i16m2(v_dot, 0);
+            vint16m2_t v_dot_hi = __riscv_vget_v_i16m4_i16m2(v_dot, 1);
+            vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, 1);
+
+            int32_t s_lo = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(v_dot_lo, v_zero, 32));
+            int32_t s_hi = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(v_dot_hi, v_zero, 32));
+
+            // Apply sub-scales
+            uint8_t sc_byte = *scales++;
+            int sc_lo = (sc_byte & 0xF) * 2 + 1;
+            int sc_hi = (sc_byte >> 4)  * 2 + 1;
+
+            sum_block += s_lo * sc_lo + s_hi * sc_hi;
+        }
+        sumf += sum_block * combined_scale;
+    }
+    *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq3_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_tq1_0_q8_K_vl256(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(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq1_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.0f;
+    uint8_t pow[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
+
+    for (int i = 0; i < nb; i++) {
+        // First loop.
+        vint32m4_t suml1;
+        {
+            const int vl = 32;
+            vuint8m1_t tq = __riscv_vle8_v_u8m1(x[i].qs, vl);
+
+            vuint16m2_t tq0 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(tq, 3, vl), 8, vl);
+            vuint16m2_t tq1 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 3, vl), 3, vl), 8, vl);
+            vuint16m2_t tq2 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 9, vl), 3, vl), 8, vl);
+            vuint16m2_t tq3 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 27, vl), 3, vl), 8, vl);
+            vuint16m2_t tq4 = __riscv_vsrl_vx_u16m2(__riscv_vwmulu_vx_u16m2(__riscv_vmul_vx_u8m1(tq, 81, vl), 3, vl), 8, vl);
+
+            vint16m2_t q80 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 0, vl), vl);
+            vint16m2_t q81 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 32, vl), vl);
+            vint16m2_t q82 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 64, vl), vl);
+            vint16m2_t q83 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 96, vl), vl);
+            vint16m2_t q84 = __riscv_vwcvt_x_x_v_i16m2(__riscv_vle8_v_i8m1(y[i].qs + 128, vl), vl);
+
+            vint16m2_t sum0 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq0, 1, vl)), q80, vl);
+            vint16m2_t sum1 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq1, 1, vl)), q81, vl);
+            vint16m2_t sum2 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq2, 1, vl)), q82, vl);
+            vint16m2_t sum3 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq3, 1, vl)), q83, vl);
+            vint16m2_t sum4 = __riscv_vmul_vv_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vsub_vx_u16m2(tq4, 1, vl)), q84, vl);
+
+            vint32m4_t sumi0 = __riscv_vwadd_vv_i32m4(sum0, sum1, vl);
+            vint32m4_t sumi1 = __riscv_vwadd_vv_i32m4(sum2, sum3, vl);
+            suml1 = __riscv_vadd_vv_i32m4(__riscv_vwcvt_x_x_v_i32m4(sum4, vl), __riscv_vadd_vv_i32m4(sumi0, sumi1, vl), vl);
+        }
+
+        // Second loop.
+        vint32m2_t suml2;
+        {
+            const int vl = 16;
+            vuint8mf2_t tq = __riscv_vle8_v_u8mf2(x[i].qs + 32, vl);
+
+            vuint16m1_t tq0 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(tq, 3 * 1, vl), 8, vl);
+            vuint16m1_t tq1 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 3, vl), 3, vl), 8, vl);
+            vuint16m1_t tq2 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 9, vl), 3, vl), 8, vl);
+            vuint16m1_t tq3 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 27, vl), 3, vl), 8, vl);
+            vuint16m1_t tq4 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vx_u8mf2(tq, 81, vl), 3, vl), 8, vl);
+
+            vint16m1_t q80 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 160, vl), vl);
+            vint16m1_t q81 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 176, vl), vl);
+            vint16m1_t q82 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 192, vl), vl);
+            vint16m1_t q83 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 208, vl), vl);
+            vint16m1_t q84 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 224, vl), vl);
+
+            vint16m1_t sum0 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq0, 1, vl)), q80, vl);
+            vint16m1_t sum1 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq1, 1, vl)), q81, vl);
+            vint16m1_t sum2 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq2, 1, vl)), q82, vl);
+            vint16m1_t sum3 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq3, 1, vl)), q83, vl);
+            vint16m1_t sum4 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq4, 1, vl)), q84, vl);
+
+            vint32m2_t sumi0 = __riscv_vwadd_vv_i32m2(sum0, sum1, vl);
+            vint32m2_t sumi1 = __riscv_vwadd_vv_i32m2(sum2, sum3, vl);
+            suml2 = __riscv_vadd_vv_i32m2(__riscv_vwcvt_x_x_v_i32m2(sum4, vl), __riscv_vadd_vv_i32m2(sumi0, sumi1, vl), vl);
+        }
+
+        // Third loop.
+        vint32m2_t suml3;
+        {
+            const int vl = 16;
+
+            uint32_t qh;
+            memcpy(&qh, &x[i].qh[0], 4);
+            // Prevent fusion with vmv.
+            __asm__ __volatile__("" : "+r"(qh));
+            vuint8mf2_t tq = __riscv_vreinterpret_v_u32mf2_u8mf2(__riscv_vmv_v_x_u32mf2(qh, vl / 4));
+
+            vuint8mf2_t p = __riscv_vle8_v_u8mf2(pow, vl);
+
+            vuint16m1_t tq0 = __riscv_vsrl_vx_u16m1(__riscv_vwmulu_vx_u16m1(__riscv_vmul_vv_u8mf2(tq, p, vl), 3, vl), 8, vl);
+
+            vint16m1_t q80 = __riscv_vwcvt_x_x_v_i16m1(__riscv_vle8_v_i8mf2(y[i].qs + 240, vl), vl);
+
+            vint16m1_t sum0 = __riscv_vmul_vv_i16m1(__riscv_vreinterpret_v_u16m1_i16m1(__riscv_vsub_vx_u16m1(tq0, 1, vl)), q80, vl);
+            suml3 = __riscv_vwcvt_x_x_v_i32m2(sum0, vl);
+        }
+
+        vint32m2_t sumb = __riscv_vadd_vv_i32m2(__riscv_vget_v_i32m4_i32m2(suml1, 0), __riscv_vget_v_i32m4_i32m2(suml1, 1), 16);
+        sumb = __riscv_vadd_vv_i32m2(sumb, suml2, 16);
+        sumb = __riscv_vadd_vv_i32m2(sumb, suml3, 16);
+
+        vint32m1_t sum = __riscv_vredsum_vs_i32m2_i32m1(sumb, __riscv_vmv_v_x_i32m1(0, 1), 16);
+        sumf += __riscv_vmv_x_s_i32m1_i32(sum) * y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_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) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_tq1_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_tq2_0_q8_K_vl256(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);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq2_0 * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0.0f;
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x[0].qs); j += 32) {
+            const int8_t * py0 = &y[i].qs[j * 4 + 0 * 32];
+            const int8_t * py1 = &y[i].qs[j * 4 + 1 * 32];
+            const int8_t * py2 = &y[i].qs[j * 4 + 2 * 32];
+            const int8_t * py3 = &y[i].qs[j * 4 + 3 * 32];
+            const uint8_t* px  = &x[i].qs[j];
+
+            size_t vlmax_16m2 = __riscv_vsetvl_e16m2(32);
+            vint16m2_t vacc16 = __riscv_vmv_v_x_i16m2(0, vlmax_16m2);
+
+            size_t vl = __riscv_vsetvl_e8m1(32);
+
+            vuint8m1_t vx_u8 = __riscv_vle8_v_u8m1(px, vl);
+
+            vint8m1_t vy0 = __riscv_vle8_v_i8m1(py0 , vl);
+            vint8m1_t vy1 = __riscv_vle8_v_i8m1(py1, vl);
+            vint8m1_t vy2 = __riscv_vle8_v_i8m1(py2, vl);
+            vint8m1_t vy3 = __riscv_vle8_v_i8m1(py3, vl);
+
+            // l=0 (bits 1:0)
+            vuint8m1_t t0 = __riscv_vand_vx_u8m1(vx_u8, 0x03, vl);
+            vint8m1_t vq0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t0), 1, vl);
+
+            // l=1 (bits 3:2)
+            vuint8m1_t t1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vx_u8, 2, vl), 0x03, vl);
+            vint8m1_t vq1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t1), 1, vl);
+
+            // l=2 (bits 5:4)
+            vuint8m1_t t2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(vx_u8, 4, vl), 0x03, vl);
+            vint8m1_t vq2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t2), 1, vl);
+
+            // l=3 (bits 7:6)
+            vuint8m1_t t3 = __riscv_vsrl_vx_u8m1(vx_u8, 6, vl); // No final AND needed as vsrl shifts in zeros
+            vint8m1_t vq3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(t3), 1, vl);
+
+            // 4. Multiply and accumulate
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq0, vy0, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq1, vy1, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq2, vy2, vl);
+            vacc16 = __riscv_vwmacc_vv_i16m2(vacc16, vq3, vy3, vl);
+
+            vlmax_16m2 = __riscv_vsetvl_e16m2(32);
+            vint32m1_t vzero32 = __riscv_vmv_v_x_i32m1(0, 1);
+            vint32m1_t vred32 = __riscv_vwredsum_vs_i16m2_i32m1(vacc16, vzero32, vlmax_16m2);
+
+            sumi += __riscv_vmv_x_s_i32m1_i32(vred32);
+        }
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        sumf += (float)sumi * d;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq2_0_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) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_tq2_0_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq1_s_q8_K_vl256(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);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        // Load qh once for the entire superblock.
+        vuint16mf2_t qh = __riscv_vle16_v_u16mf2(x[i].qh, 8);
+
+        // Calculate ls.
+        vuint16mf2_t temp = __riscv_vsrl_vx_u16mf2(qh, 12, 8);
+        temp = __riscv_vand_vx_u16mf2(temp, 7, 8);
+        vint32m1_t ls = __riscv_vreinterpret_v_u32m1_i32m1(__riscv_vwmulu_vx_u32m1(temp, 2, 8));
+        ls = __riscv_vadd_vx_i32m1(ls, 1, 8);
+
+        // Calculate delta.
+        vbool32_t mask = __riscv_vmseq_vx_u16mf2_b32(__riscv_vand_vx_u16mf2(qh, 0x8000, 8), 0, 8);
+        vint32m1_t delta_neg = __riscv_vmv_v_x_i32m1(-1, 8);
+        vint32m1_t delta_pos = __riscv_vmv_v_x_i32m1(1, 8);
+        vint32m1_t delta = __riscv_vmerge_vvm_i32m1(delta_neg, delta_pos, mask, 8);
+
+        // Load qs.
+        vuint8m1_t qs = __riscv_vle8_v_u8m1(x[i].qs, 32);
+
+        // Prepare the indices.
+        const uint64_t shift = 0x0009000600030000;
+        vuint16m2_t qh_shift = __riscv_vreinterpret_v_u64m2_u16m2(__riscv_vmv_v_x_u64m2(shift, 8));
+        vuint16m2_t qh_gather_index = __riscv_vreinterpret_v_i16m2_u16m2(
+            __riscv_vdiv_vx_i16m2(__riscv_vreinterpret_v_u16m2_i16m2(__riscv_vid_v_u16m2(32)), 4, 32));
+        vuint16m2_t qh_ext = __riscv_vlmul_ext_v_u16m1_u16m2(__riscv_vlmul_ext_v_u16mf2_u16m1(qh));
+        vuint16m2_t qh_index = __riscv_vrgather_vv_u16m2(qh_ext, qh_gather_index, 32);
+        qh_index = __riscv_vsrl_vv_u16m2(qh_index, qh_shift, 32);
+        qh_index = __riscv_vand_vx_u16m2(qh_index, 7, 32);
+        qh_index = __riscv_vsll_vx_u16m2(qh_index, 8, 32);
+        qh_index = __riscv_vor_vv_u16m2(qh_index, __riscv_vzext_vf2_u16m2(qs, 32), 32);
+        vuint16m2_t index = __riscv_vsll_vx_u16m2(qh_index, 3, 32);
+
+        // Final lsums.
+        int32_t lsums_s[8];
+        vint32m1_t one_scalar = __riscv_vmv_v_x_i32m1(0, 1);
+
+        // Sub-blocks 1-4
+        {
+            vuint16m1_t grid_index0 = __riscv_vget_v_u16m2_u16m1(index, 0);
+            vint8m4_t grid0 = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vluxei16_v_i64m4((const int64_t*)iq1s_grid, grid_index0, 16));
+            vint8m4_t q80 = __riscv_vle8_v_i8m4(y[i].qs, 128);
+            vint16m8_t lsum0 = __riscv_vwmul_vv_i16m8(grid0, q80, 128);
+            lsums_s[0] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 0), one_scalar, 32));
+            lsums_s[1] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 1), one_scalar, 32));
+            lsums_s[2] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 2), one_scalar, 32));
+            lsums_s[3] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum0, 3), one_scalar, 32));
+        }
+        __asm__ __volatile__("" ::: "memory");
+        // Sub-blocks 5-8
+        {
+            vuint16m1_t grid_index1 = __riscv_vget_v_u16m2_u16m1(index, 1);
+            vint8m4_t grid1 = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vluxei16_v_i64m4((const int64_t*)iq1s_grid, grid_index1, 16));
+            vint8m4_t q81 = __riscv_vle8_v_i8m4(&y[i].qs[128], 128);
+            vint16m8_t lsum1 = __riscv_vwmul_vv_i16m8(grid1, q81, 128);
+            lsums_s[4] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 0), one_scalar, 32));
+            lsums_s[5] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 1), one_scalar, 32));
+            lsums_s[6] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 2), one_scalar, 32));
+            lsums_s[7] = __riscv_vmv_x_s_i32m1_i32(__riscv_vwredsum_vs_i16m2_i32m1(__riscv_vget_v_i16m8_i16m2(lsum1, 3), one_scalar, 32));
+        }
+        __asm__ __volatile__("" ::: "memory");
+        vint32m1_t lsums = __riscv_vle32_v_i32m1(&lsums_s[0], 8);
+
+        // Calculate the bsums.
+        vint16m1_t bsums_0 = __riscv_vle16_v_i16m1(y[i].bsums, 16);
+        const vuint32m1_t bsums_i32 = __riscv_vreinterpret_v_u16m1_u32m1(__riscv_vreinterpret_v_i16m1_u16m1(bsums_0));
+        const vint16mf2_t bsums_i32_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(bsums_i32, 0, 8));
+        const vint16mf2_t bsums_i32_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(bsums_i32, 16, 8));
+        const vint32m1_t bsums = __riscv_vwadd_vv_i32m1(bsums_i32_0, bsums_i32_1, 8);
+
+        // Accumulation.
+        vint32m1_t sumi_v = __riscv_vmul_vv_i32m1(ls, lsums, 8);
+        vint32m1_t sumi1_v = __riscv_vmul_vv_i32m1(__riscv_vmul_vv_i32m1(ls, delta, 8), bsums, 8);
+
+        // Update sumf.
+        int sumi = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m1_i32m1(sumi_v, __riscv_vmv_v_x_i32m1(0.0f, 1), 8));
+        int sumi1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m1_i32m1(sumi1_v, __riscv_vmv_v_x_i32m1(0.0f, 1), 8));
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq1_s_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+static void ggml_vec_dot_iq1_m_q8_K_vl256(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);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * GGML_RESTRICT x = vx;
+    const block_q8_K  * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+    float sumf = 0.0f;
+    for (int i = 0; i < nb; ++i) {
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        // Accumulators.
+        vint32m2_t acc1 = __riscv_vmv_v_x_i32m2(0, 16);
+        vint32m2_t acc2 = __riscv_vmv_v_x_i32m2(0, 16);
+
+        // We process 4 sub-blocks together.
+        for (int ib = 0; ib < QK_K/128; ib++) {
+            // Load qh for 4 sub-blocks.
+            const vuint8mf4_t qh_8 = __riscv_vle8_v_u8mf4(qh, 8);
+            const vuint16mf2_t qh_16_lo = __riscv_vzext_vf2_u16mf2(qh_8, 8);
+            const vuint16mf2_t qh_16_hi = __riscv_vsll_vx_u16mf2(qh_16_lo, 8, 8);
+            const vuint16m1_t qhb = __riscv_vzext_vf2_u16m1(
+                __riscv_vreinterpret_v_u16mf2_u8mf2(__riscv_vor_vv_u16mf2(qh_16_lo, qh_16_hi, 8)), 16);
+            qh += 8;
+
+            // Prepare grid indices.
+            const vuint16m1_t qsb = __riscv_vzext_vf2_u16m1(__riscv_vle8_v_u8mf2(&qs[0], 16), 16);
+            const vuint16m1_t shift = __riscv_vreinterpret_v_u32m1_u16m1(__riscv_vmv_v_x_u32m1(0x00040008, 8));
+            vuint16m1_t index = __riscv_vor_vv_u16m1(qsb, __riscv_vand_vx_u16m1(__riscv_vsll_vv_u16m1(qhb, shift, 16), 0x700, 16), 16);
+            index = __riscv_vsll_vx_u16m1(index, 3, 16);
+            qs += 16;
+
+            // Load the grid.
+            const vint8m4_t iq1b = __riscv_vreinterpret_v_i64m4_i8m4(__riscv_vreinterpret_v_u64m4_i64m4(
+                __riscv_vluxei16_v_u64m4(iq1s_grid, index, 16)));
+
+            // Prepare the deltas.
+            const vbool16_t mask = __riscv_vmsgtu_vx_u16m1_b16(
+                __riscv_vand_vv_u16m1(qhb, __riscv_vreinterpret_v_u32m1_u16m1(__riscv_vmv_v_x_u32m1(0x00800008, 8)), 16), 0, 16);
+            const vint64m4_t delta_pos = __riscv_vmv_v_x_i64m4(0x0101010101010101, 16);
+            const vint64m4_t delta_neg = __riscv_vmv_v_x_i64m4(0xffffffffffffffff, 16);
+            const vint8m4_t delta = __riscv_vreinterpret_v_i64m4_i8m4(
+                __riscv_vmerge_vvm_i64m4(delta_pos, delta_neg, mask, 16));
+
+            // Load q8 for sub-blocks.
+            const vint8m4_t q8b = __riscv_vle8_v_i8m4(q8, 128);
+            q8 += 128;
+
+            // Calculate the lsums.
+            const vint16m8_t lsum1 = __riscv_vwmul_vv_i16m8(iq1b, q8b, 128);
+            const vint16m8_t lsum2 = __riscv_vwmul_vv_i16m8(delta, q8b, 128);
+
+            // Prepare the scales.
+            const int16_t ls_0_0 = 2*((sc[0] >> 0) & 0x7) + 1;
+            const int16_t ls_0_1 = 2*((sc[0] >> 3) & 0x7) + 1;
+            const int16_t ls_1_0 = 2*((sc[0] >> 6) & 0x7) + 1;
+            const int16_t ls_1_1 = 2*((sc[0] >> 9) & 0x7) + 1;
+            const int16_t ls_2_0 = 2*((sc[1] >> 0) & 0x7) + 1;
+            const int16_t ls_2_1 = 2*((sc[1] >> 3) & 0x7) + 1;
+            const int16_t ls_3_0 = 2*((sc[1] >> 6) & 0x7) + 1;
+            const int16_t ls_3_1 = 2*((sc[1] >> 9) & 0x7) + 1;
+            sc += 2;
+
+            // Accumulate in acc0 and acc1 for each sub-block.
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_0_0, __riscv_vget_v_i16m8_i16m1(lsum1, 0), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_0_1, __riscv_vget_v_i16m8_i16m1(lsum1, 1), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_0_0, __riscv_vget_v_i16m8_i16m1(lsum2, 0), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_0_1, __riscv_vget_v_i16m8_i16m1(lsum2, 1), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_1_0, __riscv_vget_v_i16m8_i16m1(lsum1, 2), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_1_1, __riscv_vget_v_i16m8_i16m1(lsum1, 3), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_1_0, __riscv_vget_v_i16m8_i16m1(lsum2, 2), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_1_1, __riscv_vget_v_i16m8_i16m1(lsum2, 3), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_2_0, __riscv_vget_v_i16m8_i16m1(lsum1, 4), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_2_1, __riscv_vget_v_i16m8_i16m1(lsum1, 5), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_2_0, __riscv_vget_v_i16m8_i16m1(lsum2, 4), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_2_1, __riscv_vget_v_i16m8_i16m1(lsum2, 5), 16);
+            //
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_3_0, __riscv_vget_v_i16m8_i16m1(lsum1, 6), 16);
+            acc1 = __riscv_vwmacc_vx_i32m2(acc1, ls_3_1, __riscv_vget_v_i16m8_i16m1(lsum1, 7), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_3_0, __riscv_vget_v_i16m8_i16m1(lsum2, 6), 16);
+            acc2 = __riscv_vwmacc_vx_i32m2(acc2, ls_3_1, __riscv_vget_v_i16m8_i16m1(lsum2, 7), 16);
+        }
+
+        // Reduce and accumulate in `sumf`.
+        vint32m1_t one = __riscv_vmv_v_x_i32m1(0, 1);
+        int sumi1 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m2_i32m1(acc1, one, 16));
+        int sumi2 = __riscv_vmv_x_s_i32m1_i32(__riscv_vredsum_vs_i32m2_i32m1(acc2, one, 16));
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq1_m_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) {
+#if defined __riscv_v_intrinsic
+    switch (__riscv_vlenb() * 8) {
+        case 256:
+            ggml_vec_dot_iq1_m_q8_K_vl256(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+        default:
+            ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+            break;
+    }
+#else
+    ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
@@ -1,333 +0,0 @@
-#pragma once
-
-typedef vector unsigned char vec_t;
-typedef __vector_quad acc_t;
-
-template <typename TA>
-class tinyBLAS_Q0_PPC {
-  public:
-    tinyBLAS_Q0_PPC(int64_t k,
-                    const TA *A, int64_t lda,
-                    const block_q8_0 *B, int64_t ldb,
-                    float *C, int64_t ldc,
-                    int ith, int nth);
-
-    void matmul(int64_t m, int64_t n);
-    void matmul_tiled_q0(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
-        vec_t A_pack[mc*kc*2];
-        vec_t B_pack[nc*kc*2];
-        int comparray[mc*kc];
-        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
-        int64_t ytiles = m / mc;
-        int64_t xtiles = n / nc;
-        int64_t tiles  = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles) {
-            end = tiles;
-        }
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = (job / xtiles) * mc;
-            int64_t jj = (job % xtiles) * nc;
-            for (int64_t kk = 0; kk < k; kk += kc) {
-                if constexpr(is_Ablock_q4) {
-                    packNormalInt4_large(A + ii*lda + kk, lda, mc, 4, (int8_t*)A_pack, comparray);
-                } else {
-                    packNormal_large<int8_t, vector signed char>(A + ii*lda + kk, lda, mc, 8, (int8_t*)A_pack, false, comparray);
-                }
-                packNormal_large<uint8_t, vector unsigned char>(B + jj*ldb + kk, ldb, nc, 8, (uint8_t*)B_pack, true);
-                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack, comparray);
-            }
-        }
-    }
-
-  private:
-    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-        for (int I = 0; I < RM; I++) {
-            for (int J = 0; J < RN; J++) {
-                *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
-            }
-        }
-    }
-
-    inline void add_save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-        for (int I = 0; I < RM; I++) {
-            for (int J = 0; J < RN; J++) {
-                float * c_ptr = (float *)(C+ii+((jj+J)*ldc)+I);
-                *c_ptr += *((float*)&fin_res[idx+I]+J);
-            }
-        }
-    }
-
-    template<typename ArrayType>
-    inline void compute(acc_t* ACC, int c_idx, int s_idx, ArrayType& comparray, vector float* vs, vector float* fin_res) {
-        vector signed int vec_C[4];
-        vector float CA[4] = {0};
-        vector float res[4] = {0};
-        __builtin_mma_disassemble_acc(vec_C, ACC);
-        for (int i = 0; i < 4; i++) {
-            CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
-            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
-            fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
-        }
-    }
-
-    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
-        const vector signed char lowMask = vec_splats((signed char)0xF);
-        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-        const vector signed char v8 = vec_splats((signed char)0x8);
-        vector signed int vsum = {0};
-        vector signed int vsum2 = {0};
-        c[0] = vec_and(c[1], lowMask);
-        c[1] = vec_sr(c[1], v4);
-        c[0] = vec_sub(c[0], v8);
-        c[1] = vec_sub(c[1], v8);
-        vsum = vec_sum4s(c[0], vsum);
-        vsum2 = vec_sum4s(c[1], vsum2);
-        vsum = vec_add(vsum, vsum2);
-        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template <typename V1, typename V2>
-    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
-        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
-        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
-        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
-        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
-        V2 t1, t2, t3, t4, t5, t6, t7, t8;
-        vector unsigned char xor_vector;
-        uint8_t flip_vec = 0x80;
-        xor_vector = vec_splats(flip_vec);
-        t1 = vec_perm(s1, s2, swiz1);
-        t2 = vec_perm(s1, s2, swiz2);
-        t3 = vec_perm(s3, s4, swiz1);
-        t4 = vec_perm(s3, s4, swiz2);
-        t5 = vec_perm(t1, t3, swiz3);
-        t6 = vec_perm(t1, t3, swiz4);
-        t7 = vec_perm(t2, t4, swiz3);
-        t8 = vec_perm(t2, t4, swiz4);
-        if (flip == true) {
-            t5 = vec_xor(t5, xor_vector);
-            t6 = vec_xor(t6, xor_vector);
-            t7 = vec_xor(t7, xor_vector);
-            t8 = vec_xor(t8, xor_vector);
-        }
-        vec_xst(t5, 0, vecOffset);
-        vec_xst(t6, 0, vecOffset+16);
-        vec_xst(t7, 0, vecOffset+32);
-        vec_xst(t8, 0, vecOffset+48);
-    }
-
-    template<int RM, int RN>
-    inline void kernel(int64_t ii, int64_t jj) {
-        if constexpr(RM == 4 && RN == 8) {
-            KERNEL_4x8(ii,jj);
-        } else if constexpr(RM == 8 && RN == 4) {
-            KERNEL_8x4(ii,jj);
-        } else if constexpr(RM == 8 && RN == 8) {
-            KERNEL_8x8(ii,jj);
-        } else {
-            assert(false && "RN/RM values not supported");
-        }
-    }
-    template<int size>
-    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray);
-    template<typename VA, typename VB>
-    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip);
-    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n);
-    void KERNEL_4x8(int64_t ii, int64_t jj);
-    void KERNEL_8x4(int64_t ii, int64_t jj);
-    void KERNEL_8x8(int64_t ii, int64_t jj);
-    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN);
-    template <int RM, int RN>
-    void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n);
-
-    void compute_scale(int64_t ii, int64_t jj, int blk, vector float* vs){
-        for (int I = 0; I<8; I++) {
-            float a_scale = unhalf((A+((ii+I)*lda)+blk)->d);
-            for (int J = 0; J<4; J++) {
-                *((float*)&vs[I]+J) = (a_scale * unhalf((B+((jj+J)*ldb)+blk)->d));
-                *((float*)&vs[I+8]+J) = (a_scale * unhalf((B+((jj+J+4)*ldb)+blk)->d));
-             }
-         }
-    }
-
-    inline void process_q8_elements(const int8_t *qs, int *ca) {
-        vector signed char c1 = vec_xl(0, qs);
-        vector signed char c2 = vec_xl(16, qs);
-        vector signed int vsum1 = {0};
-        vector signed int vsum2 = {0};
-        vsum1 = vec_sum4s(c1, vsum1);
-        vsum2 = vec_sum4s(c2, vsum2);
-        vector signed int vsum = vec_add(vsum1, vsum2);
-        *ca = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template<typename VA, typename VB>
-    void packNormal_large(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip, int* comparray=nullptr) {
-        int64_t i, j;
-        block_q8_0 *aoffset = NULL;
-        VA *vecOffset = NULL;
-        block_q8_0* aoffsets[8];
-        __vector_pair arr[8];
-        VB c[8][2] = {0};
-        VB c1[8] = {0}; VB c2[8] = {0};
-        aoffset = const_cast<block_q8_0*>(a);
-        vecOffset = vec;
-        j = (rows >> 3);
-        int index = 0;
-        if (j > 0) {
-            do {
-                for (int it = 0; it < 8; it++)
-                    aoffsets[it] = aoffset + it*lda;
-                aoffset += 8 * lda;
-                for (int blk = 0; blk < kc; blk++) {
-                    for (int it = 0; it < 8; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)(aoffsets[it]+blk)->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
-                        c1[it] = c[it][0];
-                        c2[it] = c[it][1];
-                        if (comparray){
-                            process_q8_elements((aoffsets[it]+ blk)->qs, &comparray[index + 8*blk + it]);
-                        }
-                    }
-                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
-                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
-                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
-                    vecOffset += 256;
-                }
-                j--;
-                index += 8*kc;
-            } while(j > 0);
-        }
-
-    }
-
-    void packNormalInt4_large(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, int*comparray) {
-        int64_t i, j;
-        TA *aoffset = NULL;
-        int8_t *vecOffset = NULL;
-        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
-        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
-        vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
-        vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
-        aoffset = const_cast<TA*>(a);
-        vecOffset = vec;
-        int index = 0;
-        j = (rows >> 3);
-        if (j > 0) {
-            do {
-                aoffset1 = aoffset;
-                aoffset2 = aoffset1 + lda;
-                aoffset3 = aoffset2 + lda;
-                aoffset4 = aoffset3 + lda;
-                aoffset5 = aoffset4 + lda;
-                aoffset6 = aoffset5 + lda;
-                aoffset7 = aoffset6 + lda;
-                aoffset8 = aoffset7 + lda;
-                aoffset += 8 * lda;
-                for (int blk = 0; blk < kc; blk++) {
-                    c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset1+blk)->qs));
-                    c2[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset2+blk)->qs));
-                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset3+blk)->qs));
-                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset4+blk)->qs));
-                    c5[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset5+blk)->qs));
-                    c6[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset6+blk)->qs));
-                    c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset7+blk)->qs));
-                    c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, (aoffset8+blk)->qs));
-
-                    process_q4_elements(c1, &comparray[index + 8*blk+0]);
-                    process_q4_elements(c2, &comparray[index + 8*blk+1]);
-                    process_q4_elements(c3, &comparray[index + 8*blk+2]);
-                    process_q4_elements(c4, &comparray[index + 8*blk+3]);
-                    process_q4_elements(c5, &comparray[index + 8*blk+4]);
-                    process_q4_elements(c6, &comparray[index + 8*blk+5]);
-                    process_q4_elements(c7, &comparray[index + 8*blk+6]);
-                    process_q4_elements(c8, &comparray[index + 8*blk+7]);
-                    vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
-                    vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
-                    vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
-                    vecOffset += 256;
-                }
-                j--;
-                index += 8*kc;
-            } while (j > 0);
-        }
-    }
-
-    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t *vec_A, vec_t *vec_B, int *comparray) {
-        acc_t acc[8];
-        for (int i = 0; i < mc ; i += 8) {
-            for (int j = 0; j < nc; j += 8) {
-                vector float fin_res[16] = {0};
-                vector float vs[16] = {0};
-                for (int64_t kk = 0; kk < kc; kk+=2) {
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xxsetaccz(&acc[x]);
-                    }
-                    int A_block_idx = (i/8)*(16*kc) + kk*16;
-                    int B_block_idx = (j/8)*(16*kc)+ kk*16;
-                    vec_t *A_block = &vec_A[A_block_idx];
-                    vec_t *B_block = &vec_B[B_block_idx];
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xvi8ger4pp(&acc[0], A_block[x],     B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[1], A_block[x + 8], B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[2], A_block[x],     B_block[x+8]);
-                        __builtin_mma_xvi8ger4pp(&acc[3], A_block[x+8],   B_block[x+8]);
-                    }
-                    compute_scale(ii+i, jj+j, l+kk, vs);
-                    int c_index = (i/8)*(8*kc)+ kk*8;
-                    int* c_block = &comparray[c_index];
-                    compute(&acc[0], 0,  0,  c_block, vs, fin_res);
-                    compute(&acc[1], 4,  4,  c_block, vs, fin_res);
-                    compute(&acc[2], 0,  8,  c_block, vs, fin_res);
-                    compute(&acc[3], 4, 12,  c_block, vs, fin_res);
-
-                    A_block_idx = (i/8)*(16*kc) + (kk+1)*16;
-                    B_block_idx = (j/8)*(16*kc)+ (kk+1)*16;
-                    A_block = &vec_A[A_block_idx];
-                    B_block = &vec_B[B_block_idx];
-                    for (int x = 0; x < 8; x++) {
-                        __builtin_mma_xvi8ger4pp(&acc[4], A_block[x],     B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[5], A_block[x + 8], B_block[x]);
-                        __builtin_mma_xvi8ger4pp(&acc[6], A_block[x],     B_block[x+8]);
-                        __builtin_mma_xvi8ger4pp(&acc[7], A_block[x+8],   B_block[x+8]);
-                    }
-                    compute_scale(ii+i, jj+j, l+kk+1, vs);
-                    c_index = (i/8)*(8*kc)+ (kk+1)*8;
-                    c_block = &comparray[c_index];
-                    compute(&acc[4], 0,  0,  c_block, vs, fin_res);
-                    compute(&acc[5], 4,  4,  c_block, vs, fin_res);
-                    compute(&acc[6], 0,  8,  c_block, vs, fin_res);
-                    compute(&acc[7], 4, 12,  c_block, vs, fin_res);
-
-                }
-                if (l == 0) {
-                    save_res(ii+i,   jj+j,    0,  fin_res);
-                    save_res(ii+i+4, jj+j,    4,  fin_res);
-                    save_res(ii+i,   jj+j+4,  8,  fin_res);
-                    save_res(ii+i+4, jj+j+4, 12,  fin_res);
-                } else {
-                    add_save_res(ii+i,   jj+j,    0,  fin_res);
-                    add_save_res(ii+i+4, jj+j,    4,  fin_res);
-                    add_save_res(ii+i,   jj+j+4,  8,  fin_res);
-                    add_save_res(ii+i+4, jj+j+4, 12,  fin_res);
-                }
-            }
-        }
-    }
-
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *C;
-    const int64_t k;
-    int64_t kc;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
@@ -121,7 +121,8 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
 #endif

 #if defined(__MMA__)
-#include "sgemm-ppc.h"
+typedef vector unsigned char vec_t;
+typedef __vector_quad acc_t;
 #endif
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED FUSED MULTIPLY ADD
@@ -2153,7 +2154,7 @@ class tinyBLAS_HP16_PPC {
            packNormal((B+(jj*ldb)+l), ldb, 8, 4, (uint8_t*)vec_B);
            for (int x = 0; x < 4; x++) {
                mma_instr<TA>::outer_product(&acc_0, vec_A[x], vec_B[x]);
-                mma_instr<TA>::outer_product(&acc_1, vec_A[x], vec_B[x+4]);
+                mma_instr<TA>::outer_product(&acc_1, vec_A[x+4], vec_B[x]);
            }
        }
        SAVE_ACC(&acc_0, ii, jj);
@@ -2301,43 +2302,299 @@ class tinyBLAS_HP16_PPC {
    const int nth;
 };

-    template <typename TA>
-    tinyBLAS_Q0_PPC<TA>::tinyBLAS_Q0_PPC(int64_t k,
-        const TA *A, int64_t lda,
-        const block_q8_0 *B, int64_t ldb,
-        float *C, int64_t ldc,
-        int ith, int nth)
+template <typename TA>
+class tinyBLAS_Q0_PPC {
+  public:
+    tinyBLAS_Q0_PPC(int64_t k,
+             const TA * A, int64_t lda,
+             const block_q8_0 * B, int64_t ldb,
+             float * C, int64_t ldc,
+             int ith, int nth)
        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
-                kc = 64;
    }

-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::matmul(int64_t m, int64_t n) {
-        int mc = 64; int nc = 64;
-        if (n % 8 == 0 && n < nc) {
-                nc = n;
-                mc = 32 ;
-                kc = 32;
+    void matmul(int64_t m, int64_t n) {
+        const int64_t mc = 64;
+        const int64_t kc = 64;
+        int64_t nc = 64;
+        int64_t n_aligned = 0;
+        if (n % 64 == 0) {
+            n_aligned = n;
+        } else if (n == 4) {
+            n_aligned = 4;
+        } else if (n < 64) {
+            n_aligned = (n / 8) * 8;
+        } else {
+            n_aligned = (n / 64) * 64;
        }
-        const bool is_aligned = ((m & (mc - 1)) == 0) & ((n & (nc - 1)) == 0) & ((k & (kc - 1)) == 0);
-        if (is_aligned) {
-            this->matmul_tiled_q0(m, n, mc, nc, kc);
+
+        if (n_aligned > 0) {
+            if (n_aligned % 64 == 0)      nc = 64;
+            else if (n_aligned == n)      nc = n;
+            else if (n_aligned % 32 == 0) nc = 32;
+            else if (n_aligned % 24 == 0) nc = 24;
+            else if (n_aligned % 16 == 0) nc = 16;
+            else                          nc = 8;
+        }
+        bool can_use_tiled = n_aligned > 0 && (m % mc == 0) && (k % kc == 0);
+        if (can_use_tiled) {
+            matmul_tiled(m, n_aligned, mc, nc, kc);
+            if (n > n_aligned) {
+                mnpack(0, m, n_aligned, n);
+            }
        } else {
            mnpack(0, m, 0, n);
        }
    }

-   template<typename TA>
-   template<int size>
-   void tinyBLAS_Q0_PPC<TA>::packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
+  private:
+    inline void save_res(int ii, int jj, int idx, vector float * fin_res, int RM = 4, int RN = 4) {
+        for (int I = 0; I < RM; I++) {
+            for (int J = 0; J < RN; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&fin_res[idx + I] + J);
+            }
+        }
+    }
+
+    inline void save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                *((float *)(C + ii + ((jj + J) * ldc) + I)) = *((float *)&vec_C[I] + J);
+            }
+        }
+    }
+
+    inline void add_save_acc(acc_t * ACC, int64_t ii, int64_t jj) {
+        vec_t vec_C[4];
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int I = 0; I < 4; I++) {
+            for (int J = 0; J < 4; J++) {
+                float * c_ptr = (float *)(C + ii+ ((jj + J) * ldc) + I);
+                *c_ptr += *((float *)&vec_C[I] + J);
+            }
+        }
+    }
+
+    template<typename ArrayType>
+    inline void compute(acc_t * ACC, int c_idx, int s_idx, ArrayType & comparray, vector float * vs, vector float * fin_res) {
+        vector signed int vec_C[4];
+        vector float CA[4] = {0};
+        vector float res[4] = {0};
+        __builtin_mma_disassemble_acc(vec_C, ACC);
+        for (int i = 0; i < 4; i++) {
+            CA[i] = vec_splats((float)(((double)comparray[c_idx + i]) * -128.0));
+            res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
+            fin_res[s_idx + i] = vec_madd(res[i], vs[s_idx + i], fin_res[s_idx + i]);
+        }
+    }
+
+    inline void process_q4_elements(vector signed char (&c)[2], int * ca) {
+        const vector signed char lowMask = vec_splats((signed char)0xF);
+        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+        const vector signed char v8 = vec_splats((signed char)0x8);
+        vector signed int vsum = {0};
+        vector signed int vsum2 = {0};
+        c[0] = vec_and(c[1], lowMask);
+        c[1] = vec_sr(c[1], v4);
+        c[0] = vec_sub(c[0], v8);
+        c[1] = vec_sub(c[1], v8);
+        vsum = vec_sum4s(c[0], vsum);
+        vsum2 = vec_sum4s(c[1], vsum2);
+        vsum = vec_add(vsum, vsum2);
+        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
+    }
+
+    template <typename V1, typename V2>
+    inline void vector_permute_store(V2 & s1, V2 & s2, V2 & s3, V2 & s4, V1 * vecOffset, bool flip) {
+        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
+        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
+        V2 t1, t2, t3, t4, t5, t6, t7, t8;
+        vector unsigned char xor_vector;
+        uint8_t flip_vec = 0x80;
+        xor_vector = vec_splats(flip_vec);
+        t1 = vec_perm(s1, s2, swiz1);
+        t2 = vec_perm(s1, s2, swiz2);
+        t3 = vec_perm(s3, s4, swiz1);
+        t4 = vec_perm(s3, s4, swiz2);
+        t5 = vec_perm(t1, t3, swiz3);
+        t6 = vec_perm(t1, t3, swiz4);
+        t7 = vec_perm(t2, t4, swiz3);
+        t8 = vec_perm(t2, t4, swiz4);
+        if (flip == true) {
+            t5 = vec_xor(t5, xor_vector);
+            t6 = vec_xor(t6, xor_vector);
+            t7 = vec_xor(t7, xor_vector);
+            t8 = vec_xor(t8, xor_vector);
+        }
+        vec_xst(t5, 0, vecOffset);
+        vec_xst(t6, 0, vecOffset + 16);
+        vec_xst(t7, 0, vecOffset + 32);
+        vec_xst(t8, 0, vecOffset + 48);
+    }
+
+    inline void unpack_q4_to_q8(vector signed char packed, vector signed char & lo, vector signed char & hi) {
+        const vector signed char lowMask = vec_splats((signed char)0x0F);
+        const vector signed char v8      = vec_splats((signed char)0x08);
+        const vector unsigned char v4    = vec_splats((unsigned char)4);
+        lo = vec_and(packed, lowMask);
+        hi = vec_sr(packed, v4);
+        lo = vec_sub(lo, v8);
+        hi = vec_sub(hi, v8);
+    }
+
+    inline void vector_permute_store_fp16(vec_t * c, unsigned char * vecOffset) {
+        vec_t t[8], s[8];
+        vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+        for (int i = 0; i < 4; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        for (int i = 4; i < 8; i += 2) {
+            t[i + 0] = vec_perm(c[i + 0], c[i + 1], swiz1);
+            t[i + 1] = vec_perm(c[i + 0], c[i + 1], swiz2);
+        }
+        s[0] = vec_perm(t[0], t[2], swiz3);
+        s[1] = vec_perm(t[0], t[2], swiz4);
+        s[2] = vec_perm(t[1], t[3], swiz3);
+        s[3] = vec_perm(t[1], t[3], swiz4);
+        s[4] = vec_perm(t[4], t[6], swiz3);
+        s[5] = vec_perm(t[4], t[6], swiz4);
+        s[6] = vec_perm(t[5], t[7], swiz3);
+        s[7] = vec_perm(t[5], t[7], swiz4);
+        for (int i = 0; i < 8; ++i) {
+            vec_xst(s[i], 0, (vec_t *)(vecOffset + i * 16));
+        }
+    }
+
+    static inline void convert_and_scale_q8(vector signed char raw, vector float v_scale, vector unsigned short & out_hi, vector unsigned short & out_lo) {
+        vector signed short i16_hi = vec_unpackh(raw);
+        vector signed short i16_lo = vec_unpackl(raw);
+
+        vector float f_hi_h = vec_ctf(vec_unpackh(i16_hi), 0);
+        vector float f_hi_l = vec_ctf(vec_unpackl(i16_hi), 0);
+        vector float f_lo_h = vec_ctf(vec_unpackh(i16_lo), 0);
+        vector float f_lo_l = vec_ctf(vec_unpackl(i16_lo), 0);
+        out_hi = vec_pack_to_short_fp32(vec_mul(f_hi_h, v_scale), vec_mul(f_hi_l, v_scale));
+        out_lo = vec_pack_to_short_fp32(vec_mul(f_lo_h, v_scale), vec_mul(f_lo_l, v_scale));
+    }
+
+    void packNormal_q4_fp16(const block_q4_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        for (int i = 0; i < rows; i += 8) {
+            const block_q4_0 * rows_base[8];
+            for (int r = 0; r < 8; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[8][4];
+                for (int r = 0; r < 8; r++) {
+                    const block_q4_0 * current_blk = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(current_blk->d));
+                    vector signed char v_qs = reinterpret_cast<vector signed char>(vec_xl(0, current_blk->qs));
+                    vector signed char c1, c2;
+                    unpack_q4_to_q8(v_qs, c1, c2);
+                    convert_and_scale_q8(c1, v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c2, v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int c = 0; c < 4; c++) {
+                    vector unsigned char c_arr[8];
+                    for (int r = 0; r < 8; r++) {
+                        c_arr[r] = (vector unsigned char)hp_res[r][c];
+                    }
+                    vector_permute_store_fp16((vec_t *)c_arr, vecOffset);
+                    vecOffset += 128;
+                }
+            }
+        }
+    }
+
+    template <int chunk_size>
+    static inline void pack_q8_block(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        unsigned char * vecOffset = vec;
+        const vec_t swiz1 = {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23};
+        const vec_t swiz2 = {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31};
+        const vec_t swiz3 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
+        const vec_t swiz4 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
+
+        for (int i = 0; i < rows; i += chunk_size) {
+            const block_q8_0 * rows_base[chunk_size];
+            for (int r = 0; r < chunk_size; r++) {
+                rows_base[r] = a + (i + r) * lda;
+            }
+            for (int blk = 0; blk < blocks; blk++) {
+                vector unsigned short hp_res[chunk_size][4];
+                for (int r = 0; r < chunk_size; r++) {
+                    const block_q8_0 * b = rows_base[r] + blk;
+                    vector float v_scale = vec_extract_fp32_from_shorth(vec_splats(b->d));
+                    vector signed char c[2];
+                    __vector_pair pair = __builtin_vsx_lxvp(0, (__vector_pair *)b->qs);
+                    __builtin_vsx_disassemble_pair(c, & pair);
+                    convert_and_scale_q8(c[0], v_scale, hp_res[r][0], hp_res[r][1]);
+                    convert_and_scale_q8(c[1], v_scale, hp_res[r][2], hp_res[r][3]);
+                }
+                for (int col = 0; col < 4; col++) {
+                    if constexpr (chunk_size == 8) {
+                        vec_t t[8];
+                        t[0] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        t[1] = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        t[2] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        t[3] = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+                        t[4] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz1);
+                        t[5] = vec_perm((vec_t)hp_res[4][col], (vec_t)hp_res[5][col], swiz2);
+                        t[6] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz1);
+                        t[7] = vec_perm((vec_t)hp_res[6][col], (vec_t)hp_res[7][col], swiz2);
+
+                        vec_xst(vec_perm(t[0], t[2], swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t[0], t[2], swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t[1], t[3], swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t[1], t[3], swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vec_xst(vec_perm(t[4], t[6], swiz3), 0, (vec_t *)(vecOffset + 64));
+                        vec_xst(vec_perm(t[4], t[6], swiz4), 0, (vec_t *)(vecOffset + 80));
+                        vec_xst(vec_perm(t[5], t[7], swiz3), 0, (vec_t *)(vecOffset + 96));
+                        vec_xst(vec_perm(t[5], t[7], swiz4), 0, (vec_t *)(vecOffset + 112));
+                        vecOffset += 128;
+                    } else {
+                        vec_t t0 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz1);
+                        vec_t t1 = vec_perm((vec_t)hp_res[0][col], (vec_t)hp_res[1][col], swiz2);
+                        vec_t t2 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz1);
+                        vec_t t3 = vec_perm((vec_t)hp_res[2][col], (vec_t)hp_res[3][col], swiz2);
+
+                        vec_xst(vec_perm(t0, t2, swiz3), 0, (vec_t *)(vecOffset + 0));
+                        vec_xst(vec_perm(t0, t2, swiz4), 0, (vec_t *)(vecOffset + 16));
+                        vec_xst(vec_perm(t1, t3, swiz3), 0, (vec_t *)(vecOffset + 32));
+                        vec_xst(vec_perm(t1, t3, swiz4), 0, (vec_t *)(vecOffset + 48));
+                        vecOffset += 64;
+                    }
+                }
+            }
+        }
+    }
+
+    void packNormal_q8_fp16(const block_q8_0 * a, int64_t lda, int rows, int blocks, unsigned char * vec) {
+        if (rows == 4) {
+            pack_q8_block<4>(a, lda, rows, blocks, vec);
+        } else {
+            pack_q8_block<8>(a, lda, rows, blocks, vec);
+        }
+    }
+
+    template<int size>
+    void packNormalInt4(const TA * a, int64_t lda, int rows, int cols, int8_t * vec, std::array<int, size> & comparray) {
        int64_t i, j;
-        TA *aoffset = NULL;
-        int8_t *vecOffset = NULL;
-        TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
-        TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+        TA * aoffset = NULL;
+        int8_t * vecOffset = NULL;
+        TA * aoffset1 = NULL, * aoffset2 = NULL, * aoffset3 = NULL, * aoffset4 = NULL;
+        TA * aoffset5 = NULL, * aoffset6 = NULL, * aoffset7 = NULL, * aoffset8 = NULL;
        vector signed char c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
        vector signed char c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
-        aoffset = const_cast<TA*>(a);
+        aoffset = const_cast<TA *>(a);
        vecOffset = vec;
        j = (rows >> 3);
        if (j > 0) {
@@ -2363,18 +2620,18 @@ class tinyBLAS_HP16_PPC {
                        c7[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset7->qs));
                        c8[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset8->qs));

-                        process_q4_elements(c1, &comparray[0]);
-                        process_q4_elements(c2, &comparray[1]);
-                        process_q4_elements(c3, &comparray[2]);
-                        process_q4_elements(c4, &comparray[3]);
-                        process_q4_elements(c5, &comparray[4]);
-                        process_q4_elements(c6, &comparray[5]);
-                        process_q4_elements(c7, &comparray[6]);
-                        process_q4_elements(c8, &comparray[7]);
+                        process_q4_elements(c1, & comparray[0]);
+                        process_q4_elements(c2, & comparray[1]);
+                        process_q4_elements(c3, & comparray[2]);
+                        process_q4_elements(c4, & comparray[3]);
+                        process_q4_elements(c5, & comparray[4]);
+                        process_q4_elements(c6, & comparray[5]);
+                        process_q4_elements(c7, & comparray[6]);
+                        process_q4_elements(c8, & comparray[7]);
                        vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset+128, false);
-                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset+192, false);
+                        vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[0], c6[0], c7[0], c8[0], vecOffset + 128, false);
+                        vector_permute_store<int8_t, vector signed char>(c5[1], c6[1], c7[1], c8[1], vecOffset + 192, false);
                        aoffset1 += lda;
                        aoffset2 += lda;
                        aoffset3 += lda;
@@ -2405,12 +2662,12 @@ class tinyBLAS_HP16_PPC {
                    c3[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset3->qs));
                    c4[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset4->qs));

-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                    vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                    aoffset1 += lda;
                    aoffset2 += lda;
                    aoffset3 += lda;
@@ -2434,12 +2691,12 @@ class tinyBLAS_HP16_PPC {
                        case 1: c1[1] = reinterpret_cast<vector signed char>(vec_xl(0, aoffset1->qs));
                            break;
                    }
-                    process_q4_elements(c1, &comparray[0]);
-                    process_q4_elements(c2, &comparray[1]);
-                    process_q4_elements(c3, &comparray[2]);
-                    process_q4_elements(c4, &comparray[3]);
+                    process_q4_elements(c1, & comparray[0]);
+                    process_q4_elements(c2, & comparray[1]);
+                    process_q4_elements(c3, & comparray[2]);
+                    process_q4_elements(c4, & comparray[3]);
                    vector_permute_store<int8_t, vector signed char>(c1[0], c2[0], c3[0], c4[0], vecOffset, false);
-                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset+64, false);
+                    vector_permute_store<int8_t, vector signed char>(c1[1], c2[1], c3[1], c4[1], vecOffset + 64, false);
                    aoffset1 += lda;
                    aoffset2 += lda;
                    aoffset3 += lda;
@@ -2450,39 +2707,38 @@ class tinyBLAS_HP16_PPC {
        }
    }

-    template<typename TA>
    template<typename VA, typename VB>
-    void tinyBLAS_Q0_PPC<TA>::packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    void packNormal(const block_q8_0 * a, int64_t lda, int rows, int cols, VA * vec, bool flip) {
        int64_t i, j;
-        block_q8_0 *aoffset = NULL;
-        VA *vecOffset = NULL;
-        block_q8_0* aoffsets[8];
+        block_q8_0 * aoffset = NULL;
+        VA * vecOffset = NULL;
+        block_q8_0 * aoffsets[8];
        __vector_pair arr[8];
        VB c[8][2] = {0};
        VB c1[8] = {0}; VB c2[8] = {0};
-        aoffset = const_cast<block_q8_0*>(a);
+        aoffset = const_cast<block_q8_0 *>(a);
        vecOffset = vec;
        j = (rows >> 3);
        if (j > 0) {
            do {
                aoffsets[0] = aoffset;
                for (int it = 1; it < 8; it++)
-                    aoffsets[it] = aoffsets[it-1] + lda;
+                    aoffsets[it] = aoffsets[it - 1] + lda;
                aoffset += 8 * lda;

                i = (cols >> 3);
                if (i > 0) {
                do {
                    for (int it = 0; it < 8; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                        c1[it] = c[it][0];
                        c2[it] = c[it][1];
                    }
                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
-                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset+128, flip);
-                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset+192, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
+                    vector_permute_store<VA, VB>(c1[4], c1[5], c1[6], c1[7], vecOffset + 128, flip);
+                    vector_permute_store<VA, VB>(c2[4], c2[5], c2[6], c2[7], vecOffset + 192, flip);
                    for (int it = 0; it < 8; it++)
                        aoffsets[it] += lda;
                    vecOffset += 256;
@@ -2501,13 +2757,13 @@ class tinyBLAS_HP16_PPC {
            if (i > 0) {
               do {
                    for (int it = 0; it < 4; it++) {
-                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[it]->qs);
-                        __builtin_vsx_disassemble_pair(c[it], &arr[it]);
+                        arr[it] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[it]->qs);
+                        __builtin_vsx_disassemble_pair(c[it], & arr[it]);
                        c1[it] = c[it][0];
                        c2[it] = c[it][1];
                    }
                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                    for (int it = 0; it < 4; it++) {
                        aoffsets[it] += lda;
                    }
@@ -2520,24 +2776,24 @@ class tinyBLAS_HP16_PPC {
        if (rows & 3) {
            aoffsets[0]  = aoffset;
            for (int it = 1; it < 3; it++ )
-                aoffsets[it] = aoffsets[it-1] + lda;
+                aoffsets[it] = aoffsets[it - 1] + lda;
            i = (cols >> 3);
            if (i > 0) {
                do {
                    switch(rows) {
-                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[2]->qs);
-                                __builtin_vsx_disassemble_pair(c[2], &arr[2]);
+                        case 3: arr[2] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[2]->qs);
+                                __builtin_vsx_disassemble_pair(c[2], & arr[2]);
                                c1[2] = c[2][0]; c2[2] = c[2][1];
-                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[1]->qs);
-                                __builtin_vsx_disassemble_pair(c[1], &arr[1]);
+                        case 2: arr[1] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[1]->qs);
+                                __builtin_vsx_disassemble_pair(c[1], & arr[1]);
                                c1[1] = c[1][0]; c2[1] = c[1][1];
-                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair*)aoffsets[0]->qs);
-                                __builtin_vsx_disassemble_pair(c[0], &arr[0]);
+                        case 1: arr[0] = __builtin_vsx_lxvp(0, (__vector_pair *)aoffsets[0]->qs);
+                                __builtin_vsx_disassemble_pair(c[0], & arr[0]);
                                c1[0] = c[0][0]; c2[0] = c[0][1];
                                break;
                    }
                    vector_permute_store<VA, VB>(c1[0], c1[1], c1[2], c1[3], vecOffset, flip);
-                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset+64, flip);
+                    vector_permute_store<VA, VB>(c2[0], c2[1], c2[2], c2[3], vecOffset + 64, flip);
                    for (int it = 0; it < 3; it++)
                         aoffsets[it] += lda;
                    vecOffset += 128;
@@ -2547,8 +2803,7 @@ class tinyBLAS_HP16_PPC {
        }
    }

-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
        int m_rem = MIN(m - m0, 16);
        int n_rem = MIN(n - n0, 16);

@@ -2585,8 +2840,7 @@ class tinyBLAS_HP16_PPC {
    }


-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_4x8(int64_t ii, int64_t jj) {
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
        vec_t vec_A[8], vec_B[16] = {0};
        acc_t acc_0, acc_1;
        std::array<int, 4> comparray {};
@@ -2594,26 +2848,26 @@ class tinyBLAS_HP16_PPC {
        vector float vs[8] = {0};
        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
            if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<4>((A+(ii*lda)+l), lda, 4, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<4>((A + (ii * lda) + l), lda, 4, 4, (int8_t *)vec_A, comparray);
            } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 4, 8, (int8_t *)vec_A, false);
            }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
            for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x], vec_B[x+8]);
            }
            for (int I = 0; I<4; I++) {
                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 4] + J) = (unhalf((A +((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                }
            }
            if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                for (int i = 0; i < 4; i++) {
                    comparray[i] = 0;
                    int ca = 0;
@@ -2624,15 +2878,14 @@ class tinyBLAS_HP16_PPC {
                    aoffset += lda;
                }
            }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 0, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 0, 4, comparray, vs, fin_res);
        }
        save_res(ii, jj, 0, fin_res);
-        save_res(ii, jj+4, 4, fin_res);
+        save_res(ii, jj + 4, 4, fin_res);
    }

-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x4(int64_t ii, int64_t jj) {
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
        vec_t vec_A[16], vec_B[8] = {0};
        acc_t acc_0, acc_1;
        std::array<int, 8> comparray {};
@@ -2640,25 +2893,25 @@ class tinyBLAS_HP16_PPC {
        vector float vs[8] = {0};
        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
            if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
            } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
            }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 4, 8, (uint8_t *)vec_B, true);
            for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
            }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+            for (int I = 0; I < 8; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                }
            }
            if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                for (int i = 0; i < 8; i++) {
                    comparray[i] = 0;
                    int ca = 0;
@@ -2669,15 +2922,14 @@ class tinyBLAS_HP16_PPC {
                    aoffset += lda;
                }
            }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
        }
        save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
    }

-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x8(int64_t ii, int64_t jj) {
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
        vec_t vec_A[16], vec_B[16] = {0};
        acc_t acc_0, acc_1, acc_2, acc_3;
        acc_t acc_4, acc_5, acc_6, acc_7;
@@ -2686,30 +2938,30 @@ class tinyBLAS_HP16_PPC {
        vector float vs[16] = {0};
        bool isAblock_q4 = std::is_same_v<TA, block_q4_0>;
        for (int l = 0; l < k; l++) {
-            __builtin_mma_xxsetaccz(&acc_0);
-            __builtin_mma_xxsetaccz(&acc_1);
-            __builtin_mma_xxsetaccz(&acc_2);
-            __builtin_mma_xxsetaccz(&acc_3);
+            __builtin_mma_xxsetaccz(& acc_0);
+            __builtin_mma_xxsetaccz(& acc_1);
+            __builtin_mma_xxsetaccz(& acc_2);
+            __builtin_mma_xxsetaccz(& acc_3);
            if (std::is_same_v<TA, block_q4_0>) {
-               packNormalInt4<8>((A+(ii*lda)+l), lda, 8, 4, (int8_t*)vec_A, comparray);
+               packNormalInt4<8>((A + (ii * lda) + l), lda, 8, 4, (int8_t *)vec_A, comparray);
            } else {
-               packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
+               packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, 8, 8, (int8_t *)vec_A, false);
            }
-            packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
+            packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, 8, 8, (uint8_t *)vec_B, true);
            for(int x = 0; x < 8; x++) {
-                __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
-                __builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]);
-                __builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]);
+                __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_1, vec_A[x + 8], vec_B[x]);
+                __builtin_mma_xvi8ger4pp(& acc_2, vec_A[x], vec_B[x + 8]);
+                __builtin_mma_xvi8ger4pp(& acc_3, vec_A[x + 8], vec_B[x + 8]);
            }
-            for (int I = 0; I<8; I++) {
-                for (int J = 0; J<4; J++) {
-                    *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
-                    *((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
+            for (int I = 0; I < 8 ; I++) {
+                for (int J = 0; J < 4; J++) {
+                    *((float *)& vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
+                    *((float *)& vs[I + 8] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J + 4) * ldb) + l)->d));
                }
            }
            if (!isAblock_q4) {
-                auto aoffset = A+(ii*lda)+l;
+                auto aoffset = A + (ii * lda) + l;
                for (int i = 0; i < 8; i++) {
                    comparray[i] = 0;
                    int ca = 0;
@@ -2720,19 +2972,99 @@ class tinyBLAS_HP16_PPC {
                    aoffset += lda;
                }
            }
-            compute(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute(&acc_1, 4, 4, comparray, vs, fin_res);
-            compute(&acc_2, 0, 8, comparray, vs, fin_res);
-            compute(&acc_3, 4, 12, comparray, vs, fin_res);
+            compute(& acc_0, 0, 0, comparray, vs, fin_res);
+            compute(& acc_1, 4, 4, comparray, vs, fin_res);
+            compute(& acc_2, 0, 8, comparray, vs, fin_res);
+            compute(& acc_3, 4, 12, comparray, vs, fin_res);
        }
        save_res(ii, jj, 0, fin_res);
-        save_res(ii+4, jj, 4, fin_res);
-        save_res(ii, jj+4, 8, fin_res);
-        save_res(ii+4, jj+4, 12, fin_res);
+        save_res(ii + 4, jj, 4, fin_res);
+        save_res(ii, jj + 4, 8, fin_res);
+        save_res(ii + 4, jj + 4, 12, fin_res);
    }

-    template<typename TA>
-    void tinyBLAS_Q0_PPC<TA>::gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    void KERNEL_Q0(int64_t ii, int64_t jj, int64_t mc, int64_t nc, int64_t kc, int64_t l, vec_t * vec_A, vec_t * vec_B) {
+        acc_t acc[8];
+        for (int i = 0; i < mc ; i += 16) {
+            for (int j = 0; j < nc; j += 8) {
+                int A0_base = (i / 16) * (2 * 32 * kc);
+                int B0_base = (j / 8) * (32 * kc);
+                for (int x = 0; x < 8; x++) {
+                     __builtin_mma_xxsetaccz(&acc[x]);
+                }
+                for (int64_t kk = 0; kk < kc; kk++) {
+                    int A0_block_idx = A0_base + kk * 32;
+                    int B0_block_idx = B0_base + kk * 32;
+                    int A1_block_idx = A0_block_idx + 32 * kc;
+                    int B1_block_idx = B0_block_idx + 32 * kc;
+                    vec_t * A0_block = & vec_A[A0_block_idx];
+                    vec_t * B0_block = & vec_B[B0_block_idx];
+                    vec_t * A1_block = & vec_A[A1_block_idx];
+                    for (int it = 0; it < 4; it++) {
+                        for (int x = 0; x < 4; x++) {
+                            __builtin_mma_xvf16ger2pp(& acc[0], A0_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[1], A0_block[8 * it + x], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[2], A0_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[3], A0_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[4], A1_block[8 * it + x], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[5], A1_block[8 * it + x], B0_block[8 * it+ x + 4]);
+                            __builtin_mma_xvf16ger2pp(& acc[6], A1_block[8 * it + x + 4], B0_block[8 * it + x]);
+                            __builtin_mma_xvf16ger2pp(& acc[7], A1_block[8 * it + x + 4], B0_block[8 * it + x + 4]);
+                        }
+                    }
+                }
+                if (l == 0) {
+                    save_acc(& acc[0], ii + i, jj + j);
+                    save_acc(& acc[1], ii + i, jj + j + 4);
+                    save_acc(& acc[2], ii + i + 4, jj + j);
+                    save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    save_acc(& acc[4], ii + i + 8, jj + j);
+                    save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    save_acc(& acc[6], ii + i + 12, jj + j);
+                    save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                } else {
+                    add_save_acc(& acc[0], ii + i, jj + j);
+                    add_save_acc(& acc[1], ii + i, jj + j + 4);
+                    add_save_acc(& acc[2], ii + i + 4, jj + j);
+                    add_save_acc(& acc[3], ii + i + 4, jj + j + 4);
+                    add_save_acc(& acc[4], ii + i + 8, jj + j);
+                    add_save_acc(& acc[5], ii + i + 8, jj + j + 4);
+                    add_save_acc(& acc[6], ii + i + 12, jj + j);
+                    add_save_acc(& acc[7], ii + i + 12, jj + j + 4);
+                }
+            }
+        }
+    }
+
+    void matmul_tiled(int64_t m, int64_t n, int64_t mc, int64_t nc, int64_t kc) {
+        vec_t A_pack[mc * kc * 4];
+        vec_t B_pack[nc * kc * 4];
+        constexpr bool is_Ablock_q4 = std::is_same_v<TA, block_q4_0>;
+        int64_t ytiles = m / mc;
+        int64_t xtiles = n / nc;
+        int64_t tiles  = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles) {
+            end = tiles;
+        }
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = (job / xtiles) * mc;
+            int64_t jj = (job % xtiles) * nc;
+            for (int64_t kk = 0; kk < k; kk += kc) {
+                if constexpr(is_Ablock_q4) {
+                    packNormal_q4_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                } else {
+                    packNormal_q8_fp16(A + ii * lda + kk, lda, mc, kc, (uint8_t *)A_pack);
+                }
+                packNormal_q8_fp16(B + jj * ldb + kk, ldb, nc, kc, (uint8_t *)B_pack);
+                KERNEL_Q0(ii, jj, mc, nc, kc, kk, A_pack, B_pack);
+            }
+        }
+    }
+
+    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
        int64_t ytiles = (m - m0) / RM;
        int64_t xtiles = (n - n0) / RN;
        int64_t tiles = xtiles * ytiles;
@@ -2754,32 +3086,32 @@ class tinyBLAS_HP16_PPC {
            vector float fin_res[4] = {0};
            vector float vs[4] = {0};
            vector float CA[4] = {0};
-            __builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value
-            __builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((A + (ii * lda) + 0)->qs, 0, 1); // prefetch first value
+            __builtin_prefetch((B + (jj * ldb) + 0)->qs, 0, 1); // prefetch first value
            for (int l = 0; l < k; l++) {
-                __builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
-                __builtin_mma_xxsetaccz(&acc_0);
+                __builtin_prefetch((A + (ii * lda) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_prefetch((B + (jj * ldb) + (l + 1))->qs, 0, 1); // prefetch one loop ahead
+                __builtin_mma_xxsetaccz(& acc_0);
                if (isAblock_q4) {
-                   packNormalInt4<4>((A+(ii*lda)+l), lda, RM, 4, (int8_t*)vec_A, comparray);
+                    packNormalInt4<4>((A + (ii * lda) + l), lda, RM, 4, (int8_t *)vec_A, comparray);
                } else {
-                   packNormal<int8_t, vector signed char>((const block_q8_0*)(A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
+                    packNormal<int8_t, vector signed char>((const block_q8_0 *)(A + (ii * lda) + l), lda, RM, 8, (int8_t *)vec_A, false);
                }
-                packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
-                for(int x = 0; x < 8; x+=4) {
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]);
-                    __builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]);
+                packNormal<uint8_t, vector unsigned char>((B + (jj * ldb) + l), ldb, RN, 8, (uint8_t *)vec_B, true);
+                for (int x = 0; x < 8; x += 4) {
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x], vec_B[x]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 1], vec_B[x + 1]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 2], vec_B[x + 2]);
+                    __builtin_mma_xvi8ger4pp(& acc_0, vec_A[x + 3], vec_B[x + 3]);
                }
-                for (int I = 0; I<RM; I++) {
-                    for (int J = 0; J<RN; J++) {
-                        *((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
+                for (int I = 0; I < RM; I++) {
+                    for (int J = 0; J < RN; J++) {
+                        *((float*)&vs[I] + J) = (unhalf((A + ((ii + I) * lda) + l)->d) * unhalf((B + ((jj + J) * ldb) + l)->d));
                    }
                }
-                __builtin_mma_disassemble_acc(vec_C, &acc_0);
+                __builtin_mma_disassemble_acc(vec_C, & acc_0);
                if (!isAblock_q4) {
-                    auto aoffset = A+(ii*lda)+l;
+                    auto aoffset = A + (ii * lda) + l;
                    for (int i = 0; i < RM; i++) {
                        comparray[i] = 0;
                        int ca = 0;
@@ -2800,9 +3132,21 @@ class tinyBLAS_HP16_PPC {
        }
    }

-    template<typename TA>
+    template<int RM, int RN>
+    inline void kernel(int64_t ii, int64_t jj) {
+        if constexpr(RM == 4 && RN == 8) {
+            KERNEL_4x8(ii,jj);
+        } else if constexpr(RM == 8 && RN == 4) {
+            KERNEL_8x4(ii,jj);
+        } else if constexpr(RM == 8 && RN == 8) {
+            KERNEL_8x8(ii,jj);
+        } else {
+            assert(false && "RN/RM values not supported");
+        }
+    }
+
    template <int RM, int RN>
-    NOINLINE void tinyBLAS_Q0_PPC<TA>::gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
        int64_t ytiles = (m - m0) / RM;
        int64_t xtiles = (n - n0) / RN;
        int64_t tiles = xtiles * ytiles;
@@ -2814,12 +3158,20 @@ class tinyBLAS_HP16_PPC {
        for (int64_t job = start; job < end; ++job) {
            int64_t ii = m0 + job / xtiles * RM;
            int64_t jj = n0 + job % xtiles * RN;
-            this->kernel<RM, RN>(ii, jj);
+            kernel<RM, RN>(ii, jj);
        }
    }
-
-template class tinyBLAS_Q0_PPC<block_q4_0>;
-template class tinyBLAS_Q0_PPC<block_q8_0>;
+    const TA * const A;
+    const block_q8_0 * const B;
+    float * C;
+    const int64_t k;
+    int64_t kc;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};

 class tinyBLAS_PPC {
  public:
@@ -450,6 +450,208 @@ static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int                        n,
    }
 }

+template <int M, int N>
+static void ggml_gemv_q5_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int         blocklen          = M;
+    constexpr int         ncols_interleaved = N;
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float    sumf[ncols_interleaved];
+    float    sum_minf[ncols_interleaved];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    const block_q8_K * a_ptr = (const block_q8_K *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) {
+            sumf[j]     = 0.0;
+            sum_minf[j] = 0.0;
+        }
+        for (int l = 0; l < nb; l++) {
+            for (int sb = 0; sb < 8; sb++) {
+                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
+                utmp[sb * 4 + 3]      = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                utmp[sb * 4 + 2]      = uaux_0;
+                utmp[sb * 4 + 0] &= kmask1;
+            }
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                constexpr int scale_stride = 32;
+                uint8_t *     scales_0     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
+                uint8_t *     scales_1     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
+
+                const int qh_shift = (k / (32 / blocklen)) * 2;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi1 = 0;
+                    sumi2 = 0;
+                    sumi  = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                        const int qh_idx      = (k * blocklen + i) % 32;
+                        const int qh_chunk    = qh_idx / blocklen;
+                        const int qh_pos      = qh_idx % blocklen;
+                        const int b_qh_offset = qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
+
+                        const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                        const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                        const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                        const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                        const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                        const int q8_offset = (k / (32 / blocklen)) * 64 + (k % (32 / blocklen)) * blocklen + i;
+
+                        sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                        sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
+                        sumi1 = sumi1 * scales_0[j];
+                        sumi2 = sumi2 * scales_1[j];
+                        sumi += sumi1 + sumi2;
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                }
+            }
+            for (int sb = 0; sb < 8; sb++) {
+                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
+                                   GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) {
+            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+        }
+    }
+}
+
+template <int M, int N>
+static void ggml_gemm_q5_K_NxM_q8_K_generic_impl(int                        n,
+                                                 float * GGML_RESTRICT      s,
+                                                 size_t                     bs,
+                                                 const void * GGML_RESTRICT vx,
+                                                 const void * GGML_RESTRICT vy,
+                                                 int                        nr,
+                                                 int                        nc) {
+    constexpr int         blocklen          = M;
+    constexpr int         ncols_interleaved = N;
+    const int             qk                = QK_K;
+    const int             nb                = n / qk;
+    static const uint32_t kmask1            = 0x3f3f3f3f;
+    static const uint32_t kmask2            = 0x0f0f0f0f;
+    static const uint32_t kmask3            = 0x03030303;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float    sumf[4][ncols_interleaved];
+    float    sum_minf[4][ncols_interleaved];
+    uint32_t utmp[32];
+    int      sumi1;
+    int      sumi2;
+    int      sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumf[m][j]     = 0.0;
+                    sum_minf[m][j] = 0.0;
+                }
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int sb = 0; sb < 8; sb++) {
+                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE, K_SCALE_SIZE);
+                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
+                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
+                    utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
+                    utmp[sb * 4 + 2]      = uaux_0;
+                    utmp[sb * 4 + 0] &= kmask1;
+                }
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    constexpr int scale_stride = 32;
+                    uint8_t *     scales_0     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride;
+                    uint8_t *     scales_1     = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16;
+
+                    const int qh_shift = (k / (32 / blocklen)) * 2;
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi1 = 0;
+                            sumi2 = 0;
+                            sumi  = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
+
+                                const int qh_idx   = (k * blocklen + i) % 32;
+                                const int qh_chunk = qh_idx / blocklen;
+                                const int qh_pos   = qh_idx % blocklen;
+                                const int b_qh_offset =
+                                    qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos;
+
+                                const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
+                                const uint8_t h0     = (qh_val >> qh_shift) & 1;
+                                const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
+
+                                const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
+                                const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
+
+                                const int q8_offset = (k / (32 / blocklen)) * 256 +
+                                                      (k % (32 / blocklen)) * 4 * blocklen + m * blocklen + i;
+
+                                sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
+                                sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
+                                sumi1 = sumi1 * scales_0[j];
+                                sumi2 = sumi2 * scales_1[j];
+                                sumi += sumi1 + sumi2;
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+                for (int sb = 0; sb < 8; sb++) {
+                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
+                    for (int m = 0; m < 4; m++) {
+                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
+                                              GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+                }
+            }
+        }
+    }
+}
+
 extern "C" {

 void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -803,98 +1005,12 @@ void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
 }

-void ggml_gemv_q5_K_8x8_q8_K_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int             qk                = QK_K;
-    const int             nb                = n / qk;
-    const int             ncols_interleaved = 8;
-    const int             blocklen          = 8;
-    static const uint32_t kmask1            = 0x3f3f3f3f;
-    static const uint32_t kmask2            = 0x0f0f0f0f;
-    static const uint32_t kmask3            = 0x03030303;
+void ggml_gemv_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}

-    assert(n % qk == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    UNUSED(bs);
-    UNUSED(nr);
-
-    float    sumf[8];
-    float    sum_minf[8];
-    uint32_t utmp[32];
-    int      sumi1;
-    int      sumi2;
-    int      sumi;
-
-    const block_q8_K * a_ptr = (const block_q8_K *) vy;
-    for (int x = 0; x < nc / ncols_interleaved; x++) {
-        const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
-
-        for (int j = 0; j < ncols_interleaved; j++) {
-            sumf[j]     = 0.0;
-            sum_minf[j] = 0.0;
-        }
-        for (int l = 0; l < nb; l++) {
-            for (int sb = 0; sb < 8; sb++) {
-                memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                utmp[sb * 4 + 3]      = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                utmp[sb * 4 + 2]      = uaux_0;
-                utmp[sb * 4 + 0] &= kmask1;
-            }
-            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
-                uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
-
-                const int qh_shift = (k / 4) * 2;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumi1 = 0;
-                    sumi2 = 0;
-                    sumi  = 0;
-                    for (int i = 0; i < blocklen; ++i) {
-                        const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
-
-                        const int qh_idx      = (k * 8 + i) % 32;
-                        const int qh_chunk    = qh_idx / 8;
-                        const int qh_pos      = qh_idx % 8;
-                        const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
-
-                        const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
-                        const uint8_t h0     = (qh_val >> qh_shift) & 1;
-                        const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
-
-                        const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
-                        const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
-
-                        const int q8_offset = (k >> 2) * 64 + (k % 4) * blocklen + i;
-
-                        sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
-                        sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]);
-                        sumi1 = sumi1 * scales_0[j];
-                        sumi2 = sumi2 * scales_1[j];
-                        sumi += sumi1 + sumi2;
-                    }
-                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
-                }
-            }
-            for (int sb = 0; sb < 8; sb++) {
-                uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) *
-                                   GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
-                }
-            }
-        }
-        for (int j = 0; j < ncols_interleaved; j++) {
-            s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
-        }
-    }
+void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemv_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }


@@ -1494,107 +1610,12 @@ void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
    }
 }

-void ggml_gemm_q5_K_8x8_q8_K_generic(int                        n,
-                                     float * GGML_RESTRICT      s,
-                                     size_t                     bs,
-                                     const void * GGML_RESTRICT vx,
-                                     const void * GGML_RESTRICT vy,
-                                     int                        nr,
-                                     int                        nc) {
-    const int qk                = QK_K;
-    const int nb                = n / qk;
-    const int ncols_interleaved = 8;
-    const int blocklen          = 8;
+void ggml_gemm_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc);
+}

-    constexpr uint32_t kmask1 = 0x3f3f3f3f;
-    constexpr uint32_t kmask2 = 0x0f0f0f0f;
-    constexpr uint32_t kmask3 = 0x03030303;
-
-    assert(n % qk == 0);
-    assert(nr % 4 == 0);
-    assert(nc % ncols_interleaved == 0);
-
-    float    sumf[4][8];
-    float    sum_minf[4][8];
-    uint32_t utmp[32];
-    int      sumi1;
-    int      sumi2;
-    int      sumi;
-
-    for (int y = 0; y < nr / 4; y++) {
-        const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb);
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    sumf[m][j]     = 0.0;
-                    sum_minf[m][j] = 0.0;
-                }
-            }
-            for (int l = 0; l < nb; l++) {
-                for (int sb = 0; sb < 8; sb++) {
-                    memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12);
-                    utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4);
-                    const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1;
-                    utmp[sb * 4 + 1]      = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4);
-                    utmp[sb * 4 + 2]      = uaux_0;
-                    utmp[sb * 4 + 0] &= kmask1;
-                }
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    uint8_t * scales_0 = (uint8_t *) utmp + (k / 4) * 32;
-                    uint8_t * scales_1 = (uint8_t *) utmp + (k / 4) * 32 + 16;
-
-                    const int qh_shift = (k / 4) * 2;
-                    for (int m = 0; m < 4; m++) {
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sumi1 = 0;
-                            sumi2 = 0;
-                            sumi  = 0;
-                            for (int i = 0; i < blocklen; ++i) {
-                                const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i;
-
-                                const int qh_idx      = (k * 8 + i) % 32;
-                                const int qh_chunk    = qh_idx / 8;
-                                const int qh_pos      = qh_idx % 8;
-                                const int b_qh_offset = qh_chunk * 64 + j * 8 + qh_pos;
-
-                                const uint8_t qh_val = b_ptr[l].qh[b_qh_offset];
-                                const uint8_t h0     = (qh_val >> qh_shift) & 1;
-                                const uint8_t h1     = (qh_val >> (qh_shift + 1)) & 1;
-
-                                const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4));
-                                const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4));
-
-                                const int q8_offset = (k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i;
-
-                                sumi1 = (v0 * a_ptr[l].qs[q8_offset]);
-                                sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]);
-                                sumi1 = sumi1 * scales_0[j];
-                                sumi2 = sumi2 * scales_1[j];
-                                sumi += sumi1 + sumi2;
-                            }
-                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-                for (int sb = 0; sb < 8; sb++) {
-                    uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16;
-                    for (int m = 0; m < 4; m++) {
-                        const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
-                        for (int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) *
-                                              GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
-                        }
-                    }
-                }
-            }
-            for (int m = 0; m < 4; m++) {
-                for (int j = 0; j < ncols_interleaved; j++) {
-                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
-                }
-            }
-        }
-    }
+void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    ggml_gemm_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc);
 }

 void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
@@ -2029,18 +2050,16 @@ static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_in

    const int end = QK_K * 4 / blck_size_interleave;

-    // Interleave Q5_K quants by taking 8 bytes at a time
+    // Interleave Q5_K quants by taking blck_size_interleave bytes at a time
    for (int i = 0; i < end; ++i) {
        int src_id     = i % 8;
        int src_offset = (i / 8) * blck_size_interleave;
        int dst_offset = i * blck_size_interleave;

-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave);
    }

-    // Repeat for low bits 8 bytes at a time as well, since
+    // Repeat for high bits with the same chunk size, since
    // the high bits are interleaved in Q5_K and the index is
    // qh_idx = (qs_idx % 32);
    // qh_val = qh[qh_idx] >> (qs_idx / 32);
@@ -2049,9 +2068,7 @@ static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_in
        int src_offset = (i / 8) * blck_size_interleave;
        int dst_offset = i * blck_size_interleave;

-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qh[src_offset], sizeof(uint64_t));
-        memcpy(&out.qh[dst_offset], &elems, sizeof(uint64_t));
+        memcpy(&out.qh[dst_offset], &in[src_id].qh[src_offset], blck_size_interleave);
    }

    // The below logic is copied over from Q4_K
@@ -2249,7 +2266,7 @@ static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor *       t,
                                    const void * GGML_RESTRICT data,
                                    size_t                     data_size) {
    GGML_ASSERT(t->type == GGML_TYPE_Q5_K);
-    GGML_ASSERT(interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
    constexpr int nrows_interleaved = 8;

    block_q5_Kx8 *     dst = (block_q5_Kx8 *) t->data;
@@ -2523,6 +2540,10 @@ template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * da
    return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
 }

+template <> int repack<block_q5_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q5_K_to_q5_K_8_bl(t, 4, data, data_size);
+}
+
 template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
    return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size);
 }
@@ -2591,6 +2612,10 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
    ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }

+template <> void gemv<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
    ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -2654,6 +2679,10 @@ template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
    ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }

+template <> void gemm<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
 template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
    ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
 }
@@ -3068,6 +3097,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
    static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;

    // instance for Q5_K
+    static const ggml::cpu::repack::tensor_traits<block_q5_K, 4, 8, GGML_TYPE_Q8_K> q5_K_8x4_q8_K;
    static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K;

    // instance for Q6_K
@@ -3130,6 +3160,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
                return &q5_K_8x8_q8_K;
            }
        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &q5_K_8x4_q8_K;
+            }
+        }
    } else if (cur->type == GGML_TYPE_Q6_K) {
        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
            if (cur->ne[1] % 8 == 0) {
@@ -111,6 +111,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -122,6 +123,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -143,6 +145,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -154,6 +157,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q5_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q5_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x4_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q6_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
@@ -1149,8 +1149,7 @@ struct ggml_cuda_graph {
    size_t num_nodes = 0;
    std::vector<cudaGraphNode_t> nodes;
    bool disable_due_to_gpu_arch = false;
-    bool disable_due_to_too_many_updates = false;
-    int number_consecutive_updates = 0;
+    bool warmup_complete = false;
    std::vector<ggml_cuda_graph_node_properties> props;

    // these are extra tensors (inputs) that participate in the ggml graph but are not nodes
@@ -1159,21 +1158,9 @@ struct ggml_cuda_graph {
    // ref: https://github.com/ggml-org/llama.cpp/pull/19165
    std::vector<ggml_cuda_graph_node_properties> extra;

-    void record_update(bool use_graph, bool update_required) {
-        if (use_graph && update_required) {
-            number_consecutive_updates++;
-        } else {
-            number_consecutive_updates = 0;
-        }
-        if (number_consecutive_updates >= 4) {
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
-            disable_due_to_too_many_updates = true;
-        }
-    }
-
    bool is_enabled() const {
        static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
-        return !(disable_due_to_gpu_arch || disable_cuda_graphs_due_to_env || disable_due_to_too_many_updates);
+        return !(disable_due_to_gpu_arch || disable_cuda_graphs_due_to_env);
    }
 #endif
 };
@@ -1186,8 +1186,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
    GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
    const int gqa_ratio = Q->ne[2] / K->ne[2];

+    // On NVIDIA (Pascal and older) the GQA optimizations seem to be detrimental in some cases.
+    // However, for DKQ == 576, DV == 512 only the kernel variant with GQA optimizations is implemented.
    const bool nvidia = GGML_CUDA_CC_IS_NVIDIA(ggml_cuda_info().devices[ggml_cuda_get_device()].cc);
-    const int gqa_limit = nvidia && gqa_ratio <= 4 ? 16 : INT_MAX;
+    const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
    const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;

    if constexpr (DV == 512) {
@@ -2979,10 +2979,6 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);

-    if (graph->instance == nullptr) {
-        res = true;
-    }
-
    // Check if the graph size has changed
    if (graph->props.size() != (size_t)cgraph->n_nodes) {
        res = true;
@@ -3931,14 +3927,35 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
 #ifdef USE_CUDA_GRAPH
    graph_key = ggml_cuda_graph_get_key(cgraph);

-    use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
+    ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);

    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
    if (graph->is_enabled()) {
-        cuda_graph_update_required = ggml_cuda_graph_update_required(cuda_ctx, cgraph);
-        use_cuda_graph             = ggml_cuda_graph_check_compability(cgraph);
+        const bool graph_compatible = ggml_cuda_graph_check_compability(cgraph);
+        if (graph_compatible) {
+            const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph);

-        graph->record_update(use_cuda_graph, cuda_graph_update_required);
+            if (!graph->warmup_complete) {
+                // Warmup: need at least 2 calls with no property change on the 2nd call
+                if (!properties_changed) {
+                    graph->warmup_complete = true;
+                    GGML_LOG_DEBUG("%s: CUDA graph warmup complete\n", __func__);
+                    use_cuda_graph = true;
+                    cuda_graph_update_required = true;
+                }
+                // else: properties changed or first call - execute directly (use_cuda_graph stays false)
+            } else {
+                // Post-warmup: normal CUDA graph operation
+                if (properties_changed) {
+                    // Properties changed - reset warmup, execute directly until stable again
+                    graph->warmup_complete = false;
+                    GGML_LOG_DEBUG("%s: CUDA graph warmup reset\n", __func__);
+                } else {
+                    use_cuda_graph = true;
+                    cuda_graph_update_required = graph->instance == nullptr;
+                }
+            }
+        }
    }
 #endif // USE_CUDA_GRAPH

@@ -1749,23 +1749,6 @@ static inline bool ggml_backend_buffer_is_hexagon_repack(const struct ggml_backe
    return b->buft->iface.alloc_buffer == ggml_backend_hexagon_repack_buffer_type_alloc_buffer;
 }

-static bool hex_supported_dims2(const struct ggml_tensor * x, const struct ggml_tensor * y) {
-    if (x->ne[0] != y->ne[0]) {
-        return false;
-    }
-    if (x->ne[1] != y->ne[1]) {
-        return false;
-    }
-    if (x->ne[2] != y->ne[2]) {
-        return false;
-    }
-    if (x->ne[3] != y->ne[3]) {
-        return false;
-    }
-
-    return true;
-}
-
 static bool ggml_hexagon_supported_flash_attn_ext(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
    const struct ggml_tensor * src0 = op->src[0];
    const struct ggml_tensor * src1 = op->src[1];
@@ -1797,43 +1780,6 @@ static bool ggml_hexagon_supported_flash_attn_ext(const struct ggml_hexagon_sess
    return opt_experimental;
 }

-static bool hex_supported_src0_type(ggml_type t) {
-    return t == GGML_TYPE_F32;
-}
-
-static bool hex_supported_src1_type(ggml_type t) {
-    return t == GGML_TYPE_F32;
-}
-
-static bool hex_supported_src2_type(ggml_type t) {
-    return t == GGML_TYPE_F32;
-}
-
-static bool hex_supported_src1_type2(ggml_type t) {
-    return t == GGML_TYPE_F16;
-}
-
-static bool hex_supported_src1_type3(ggml_type t) {
-    return t == GGML_TYPE_I32;
-}
-
-static bool hex_supported_dst_type(ggml_type t) {
-    return t == GGML_TYPE_F32;
-}
-
-static bool hex_supported_dims(const struct ggml_tensor * x, const struct ggml_tensor * y) {
-    // TODO: support broadcast for ne[2 and 3]
-    if (x->ne[0] != y->ne[0]) {
-        return false;
-    }
-    if (x->ne[2] != y->ne[2]) {
-        return false;
-    }
-    if (x->ne[3] != y->ne[3]) {
-        return false;
-    }
-    return true;
-}

 static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * sess, const struct ggml_tensor * dst) {
    const struct ggml_tensor * src0 = dst->src[0];
@@ -1919,19 +1865,19 @@ static bool ggml_hexagon_supported_binary(const struct ggml_hexagon_session * se
    const struct ggml_tensor * src1 = op->src[1];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_src1_type(src1->type)) {
+    if (src1->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dims2(src0, dst)) {
+    if (!ggml_are_same_shape(src0, dst)) {
        return false;
    }
-    if (!ggml_can_repeat(src1, src0)) {
+    if (!ggml_can_repeat(src1, src0) || ggml_is_permuted(src1)) {
        return false;
    }

@@ -1943,16 +1889,16 @@ static bool ggml_hexagon_supported_add_id(const struct ggml_hexagon_session * se
    const struct ggml_tensor * src1 = op->src[1];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_src1_type(src1->type)) {
+    if (src1->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dims2(src0, dst)) {
+    if (!ggml_are_same_shape(src0, dst)) {
        return false;
    }

@@ -1968,13 +1914,13 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
    const struct ggml_tensor * src0 = op->src[0];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dims2(src0, dst)) {
+    if (!ggml_are_same_shape(src0, dst)) {
        return false;
    }

@@ -1990,10 +1936,10 @@ static bool ggml_hexagon_supported_sum_rows(const struct ggml_hexagon_session *
    const struct ggml_tensor * src0 = op->src[0];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }

@@ -2011,10 +1957,10 @@ static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session
    const struct ggml_tensor * src1 = op->src[1];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }

@@ -2023,10 +1969,10 @@ static bool ggml_hexagon_supported_activations(const struct ggml_hexagon_session
    }

    if (src1) {
-        if (!hex_supported_src1_type(src1->type)) {
+        if (src1->type != GGML_TYPE_F32) {
            return false;
        }
-        if (!hex_supported_dims2(src0, src1)) {
+        if (!ggml_are_same_shape(src0, src1)) {
            return false;
        }
        if (!ggml_is_contiguous(src1)) {
@@ -2047,15 +1993,15 @@ static bool ggml_hexagon_supported_softmax(const struct ggml_hexagon_session * s
        return false;  // FIXME: add support for sinks
    }

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }

    if (src1) {
-        if (!hex_supported_src1_type(src1->type) && !hex_supported_src1_type2(src1->type)) {
+        if (src1->type != GGML_TYPE_F32 && src1->type != GGML_TYPE_F16) {
            return false;
        }
        if (src0->ne[0] != src1->ne[0]) {
@@ -2162,17 +2108,17 @@ static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess
    const struct ggml_tensor * src2 = op->src[2];
    const struct ggml_tensor * dst  = op;

-    if (!hex_supported_src0_type(src0->type)) {
+    if (src0->type != GGML_TYPE_F32) {
        return false;  // FIXME: add support for GGML_TYPE_F16 for src0
    }
-    if (!hex_supported_dst_type(dst->type)) {
+    if (dst->type != GGML_TYPE_F32) {
        return false;
    }
-    if (!hex_supported_src1_type3(src1->type)) {
+    if (src1->type != GGML_TYPE_I32) {
        return false;
    }
    if (src2) {
-        if (!hex_supported_src2_type(src2->type)) {
+        if (src2->type != GGML_TYPE_F32) {
            return false;
        }
        int n_dims = op_params[1];
@@ -69,27 +69,45 @@
    const uint32_t nb2 = dst->nb[2];   \
    const uint32_t nb3 = dst->nb[3];

-static void glu_swiglu_f32_per_thread(const struct htp_tensor * src0,
-                                       const struct htp_tensor * src1,
-                                       struct htp_tensor *       dst,
-                                       const int32_t *           op_params,
-                                       struct htp_spad *         src0_spad,
-                                       struct htp_spad *         src1_spad,
-                                       struct htp_spad *         dst_spad,
-                                       uint32_t                  nth,
-                                       uint32_t                  ith,
-                                       uint32_t                  src0_nrows_per_thread,
-                                       dma_queue *               dma_queue) {
+struct htp_act_context {
+    struct htp_ops_context *  octx;
+
+    // Precomputed values
+    const uint8_t *           data_src0;
+    const uint8_t *           data_src1;
+    uint8_t *                 data_dst;
+
+    size_t                    src0_row_size;
+    size_t                    src1_row_size;
+    size_t                    dst_row_size;
+
+    size_t                    src0_row_size_aligned;
+    size_t                    src1_row_size_aligned;
+    size_t                    dst_row_size_aligned;
+
+    size_t                    src0_spad_half_size;
+    size_t                    src1_spad_half_size;
+    size_t                    dst_spad_half_size;
+
+    uint32_t                  block;
+    uint32_t                  src0_nrows;
+    uint32_t                  src0_nrows_per_thread;
+    int                       nc;
+};
+
+static void glu_swiglu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_act_context * actx = (struct htp_act_context *) data;
+    const struct htp_tensor * src0 = &actx->octx->src0;
+    const struct htp_tensor * src1 = &actx->octx->src1;
+    const struct htp_tensor * dst  = &actx->octx->dst;
    htp_act_preamble3;

-    size_t src0_row_size = nb01;
-    size_t src1_row_size = nb11;
-    size_t dst_row_size  = nb1;
-
-
-
-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    size_t src0_row_size = actx->src0_row_size;
+    size_t src1_row_size = actx->src1_row_size;
+    size_t dst_row_size  = actx->dst_row_size;

+    const uint32_t src0_nrows = actx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;
    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);

@@ -101,43 +119,34 @@ static void glu_swiglu_f32_per_thread(const struct htp_tensor * src0,
    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
-    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
-    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+    const uint8_t * restrict data_src0 = actx->data_src0;
+    const uint8_t * restrict data_src1 = actx->data_src1;
+    uint8_t * restrict data_dst        = actx->data_dst;

-    const bool src1_valid = src1->ne[0];
-    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
-    if (!src1_valid) {
-        const int32_t swapped = op_params[1];
-        data_src1             = data_src0;
-        src1_row_size         = src0_row_size;
+    const int  nc = actx->nc;

-        const size_t nc_in_bytes = nc * SIZEOF_FP32;
-        data_src0 += swapped ? nc_in_bytes : 0;
-        data_src1 += swapped ? 0 : nc_in_bytes;
-    }
+    const size_t src0_row_size_aligned = actx->src0_row_size_aligned;
+    const size_t src1_row_size_aligned = actx->src1_row_size_aligned;
+    const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;

-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    uint8_t * restrict src0_spad_data = actx->octx->src0_spad.data + (ith * actx->octx->src0_spad.size_per_thread);
+    uint8_t * restrict src1_spad_data = actx->octx->src1_spad.data + (ith * actx->octx->src1_spad.size_per_thread);
+    uint8_t * restrict dst_spad_data  = actx->octx->dst_spad.data + (ith * actx->octx->dst_spad.size_per_thread);

-    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
-    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
-    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+    size_t src0_spad_half_size = actx->src0_spad_half_size;
+    size_t src1_spad_half_size = actx->src1_spad_half_size;
+    size_t dst_spad_half_size  = actx->dst_spad_half_size;

-    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
-    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
-    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
-    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
-
-    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    const int BLOCK = actx->block;
    if (BLOCK == 0) {
        FARF(ERROR,
             "swiglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
-             src0_spad->size_per_thread, src0_row_size_aligned);
+             actx->octx->src0_spad.size_per_thread, src0_row_size_aligned);
        return;
    }

+    dma_queue * dma_queue = actx->octx->ctx->dma[ith];
+
    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
@@ -196,27 +205,22 @@ static void glu_swiglu_f32_per_thread(const struct htp_tensor * src0,
         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void glu_swiglu_oai_f32_per_thread(const struct htp_tensor * src0,
-                                           const struct htp_tensor * src1,
-                                           struct htp_tensor *       dst,
-                                           const int32_t *           op_params,
-                                           struct htp_spad *         src0_spad,
-                                           struct htp_spad *         src1_spad,
-                                           struct htp_spad *         dst_spad,
-                                           uint32_t                  nth,
-                                           uint32_t                  ith,
-                                           uint32_t                  src0_nrows_per_thread,
-                                           dma_queue *               dma_queue) {
+static void glu_swiglu_oai_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_act_context * actx = (struct htp_act_context *) data;
+    const struct htp_tensor * src0 = &actx->octx->src0;
+    const struct htp_tensor * src1 = &actx->octx->src1;
+    const struct htp_tensor * dst  = &actx->octx->dst;
    htp_act_preamble3;

    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    size_t src0_row_size = nb01;
-    size_t src1_row_size = nb11;
-    size_t dst_row_size  = nb1;
+    size_t src0_row_size = actx->src0_row_size;
+    size_t src1_row_size = actx->src1_row_size;
+    size_t dst_row_size  = actx->dst_row_size;

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src0_nrows = actx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -226,45 +230,36 @@ static void glu_swiglu_oai_f32_per_thread(const struct htp_tensor * src0,
        return;
    }

-    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
-    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
-    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+    const uint8_t * restrict data_src0 = actx->data_src0;
+    const uint8_t * restrict data_src1 = actx->data_src1;
+    uint8_t * restrict data_dst        = actx->data_dst;

-    const bool src1_valid = src1->ne[0];
-    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
-    if (!src1_valid) {
-        const int32_t swapped = op_params[1];
-        data_src1             = data_src0;
-        src1_row_size         = src0_row_size;
+    const int nc = actx->nc;

-        const size_t nc_in_bytes = nc * SIZEOF_FP32;
-        data_src0 += swapped ? nc_in_bytes : 0;
-        data_src1 += swapped ? 0 : nc_in_bytes;
-    }
+    const size_t src0_row_size_aligned = actx->src0_row_size_aligned;
+    const size_t src1_row_size_aligned = actx->src1_row_size_aligned;
+    const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;

-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    uint8_t * restrict src0_spad_data = actx->octx->src0_spad.data + (ith * actx->octx->src0_spad.size_per_thread);
+    uint8_t * restrict src1_spad_data = actx->octx->src1_spad.data + (ith * actx->octx->src1_spad.size_per_thread);
+    uint8_t * restrict dst_spad_data  = actx->octx->dst_spad.data + (ith * actx->octx->dst_spad.size_per_thread);

-    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
-    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
-    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+    size_t src0_spad_half_size = actx->src0_spad_half_size;
+    size_t src1_spad_half_size = actx->src1_spad_half_size;
+    size_t dst_spad_half_size  = actx->dst_spad_half_size;

-    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
-    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
-    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
-    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
-
-    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    const int BLOCK = actx->block;
    if (BLOCK == 0) {
        FARF(ERROR,
             "swiglu-oai-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least "
             "%zu\n",
-             src0_spad->size_per_thread, src0_row_size_aligned);
+             actx->octx->src0_spad.size_per_thread, src0_row_size_aligned);
        return;
    }
-    const float alpha = ((const float *) (op_params))[2];
-    const float limit = ((const float *) (op_params))[3];
+    const float alpha = ((const float *) (actx->octx->op_params))[2];
+    const float limit = ((const float *) (actx->octx->op_params))[3];
+
+    dma_queue * dma_queue = actx->octx->ctx->dma[ith];

    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
@@ -335,26 +330,22 @@ static void glu_swiglu_oai_f32_per_thread(const struct htp_tensor * src0,
 }


-static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
-                                       struct htp_tensor *       dst,
-                                       const int32_t *           op_params,
-                                       struct htp_spad *         src0_spad,
-                                       struct htp_spad *         dst_spad,
-                                       uint32_t                  nth,
-                                       uint32_t                  ith,
-                                       uint32_t                  src0_nrows_per_thread,
-                                       dma_queue *               dma_queue) {
+static void unary_gelu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_act_context * actx = (struct htp_act_context *) data;
+    const struct htp_tensor * src0 = &actx->octx->src0;
+    const struct htp_tensor * dst  = &actx->octx->dst;
    htp_act_preamble2;

    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    const size_t src0_row_size = nb01;
-    const size_t dst_row_size  = nb1;
-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    const size_t src0_row_size = actx->src0_row_size;
+    const size_t dst_row_size  = actx->dst_row_size;
+    const size_t src0_row_size_aligned = actx->src0_row_size_aligned;
+    const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+    const uint32_t src0_nrows = actx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -364,25 +355,29 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
        return;
    }

-    const uint8_t * data_src0 = (const uint8_t *) src0->data;
-    uint8_t * data_dst        = (uint8_t *) dst->data;
+    const uint8_t * data_src0 = actx->data_src0;
+    uint8_t * data_dst        = actx->data_dst;

-    uint8_t * src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
-    uint8_t * dst_spad_data  = dst_spad->data  + (ith * dst_spad->size_per_thread);
+    // nc/ne0 matches.
+    const int ne0_val = actx->nc; // == dst->ne[0]

-    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
-    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
-    size_t dst_spad_half_size  = dst_spad->size_per_thread  / 2;
+    uint8_t * src0_spad_data = actx->octx->src0_spad.data + (ith * actx->octx->src0_spad.size_per_thread);
+    uint8_t * dst_spad_data  = actx->octx->dst_spad.data  + (ith * actx->octx->dst_spad.size_per_thread);
+
+    size_t src0_spad_half_size = actx->src0_spad_half_size;
+    size_t dst_spad_half_size  = actx->dst_spad_half_size;

    // In gelu = x*sigmoid(x*1.702)
-    const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block
+    const int BLOCK = actx->block;

    if (BLOCK == 0) {
        FARF(ERROR, "gelu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
-                src0_spad->size_per_thread, src0_row_size_aligned);
+                actx->octx->src0_spad.size_per_thread, src0_row_size_aligned);
        return;
    }

+    dma_queue * dma_queue = actx->octx->ctx->dma[ith];
+
    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
@@ -408,9 +403,9 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
            float* dst_spad_ptr        = dst_spad  + ib * (dst_row_size_aligned  / sizeof(float));

            // gelu = x * sigmoid(1.702 * x) // current implementation
-            hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0);
-            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
-            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_mul_scalar_f32((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (float) 1.702, ne0_val);
+            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0_val);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0_val);
        }

        dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -435,34 +430,23 @@ static void unary_gelu_f32_per_thread(const struct htp_tensor * src0,
         ne03, src0_start_row, src0_end_row, ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void unary_gelu_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-    unary_gelu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
-                               octx->src0_nrows_per_thread, octx->ctx->dma[i]);
-}

-
-
-static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
-                                       struct htp_tensor *       dst,
-                                       const int32_t *           op_params,
-                                       struct htp_spad *         src0_spad,
-                                       struct htp_spad *         dst_spad,
-                                       uint32_t                  nth,
-                                       uint32_t                  ith,
-                                       uint32_t                  src0_nrows_per_thread,
-                                       dma_queue *               dma_queue) {
+static void unary_silu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_act_context * actx = (struct htp_act_context *) data;
+    const struct htp_tensor * src0 = &actx->octx->src0;
+    const struct htp_tensor * dst  = &actx->octx->dst;
    htp_act_preamble2;

    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    const size_t src0_row_size = nb01;
-    const size_t dst_row_size  = nb1;
-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    const size_t src0_row_size = actx->src0_row_size;
+    const size_t dst_row_size  = actx->dst_row_size;
+    const size_t src0_row_size_aligned = actx->src0_row_size_aligned;
+    const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;
+    const uint32_t src0_nrows = actx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -472,24 +456,27 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
        return;
    }

-    const uint8_t * data_src0 = (const uint8_t *) src0->data;
-    uint8_t * data_dst        = (uint8_t *) dst->data;
+    const uint8_t * data_src0 = actx->data_src0;
+    uint8_t * data_dst        = actx->data_dst;

-    uint8_t * src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
-    uint8_t * dst_spad_data  = dst_spad->data  + (ith * dst_spad->size_per_thread);
+    const int ne0_val = actx->nc; // == dst->ne[0]

-    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
-    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
-    size_t dst_spad_half_size  = dst_spad->size_per_thread  / 2;
+    uint8_t * src0_spad_data = actx->octx->src0_spad.data + (ith * actx->octx->src0_spad.size_per_thread);
+    uint8_t * dst_spad_data  = actx->octx->dst_spad.data  + (ith * actx->octx->dst_spad.size_per_thread);

-    const int BLOCK = src0_spad_half_size / src0_row_size_aligned; // How many rows can we process in one block
+    size_t src0_spad_half_size = actx->src0_spad_half_size;
+    size_t dst_spad_half_size  = actx->dst_spad_half_size;
+
+    const int BLOCK = actx->block;

    if (BLOCK == 0) {
        FARF(ERROR, "silu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
-                src0_spad->size_per_thread, src0_row_size_aligned);
+                actx->octx->src0_spad.size_per_thread, src0_row_size_aligned);
        return;
    }

+    dma_queue * dma_queue = actx->octx->ctx->dma[ith];
+
    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
@@ -515,8 +502,8 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
            float* dst_spad_ptr        = dst_spad  + ib * (dst_row_size_aligned  / sizeof(float));

            // silu = x * sigmoid(x)
-            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0);
-            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0);
+            hvx_sigmoid_f32_aa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, ne0_val);
+            hvx_mul_f32_aaa((uint8_t *) dst_spad_ptr, (const uint8_t *) src0_spad_ptr, (const uint8_t *) dst_spad_ptr, ne0_val);
        }

        dma_queue_push_vtcm_to_ddr(dma_queue,
@@ -544,27 +531,22 @@ static void unary_silu_f32_per_thread(const struct htp_tensor * src0,
 static const float GELU_COEF_A     = 0.044715f;
 static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;

-static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
-                                       const struct htp_tensor * src1,
-                                       struct htp_tensor *       dst,
-                                       const int32_t *           op_params,
-                                       struct htp_spad *         src0_spad,
-                                       struct htp_spad *         src1_spad,
-                                       struct htp_spad *         dst_spad,
-                                       uint32_t                  nth,
-                                       uint32_t                  ith,
-                                       uint32_t                  src0_nrows_per_thread,
-                                       dma_queue *               dma_queue) {
+static void glu_geglu_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_act_context * actx = (struct htp_act_context *) data;
+    const struct htp_tensor * src0 = &actx->octx->src0;
+    const struct htp_tensor * src1 = &actx->octx->src1;
+    const struct htp_tensor * dst  = &actx->octx->dst;
    htp_act_preamble3;

-    size_t src0_row_size = nb01;
-    size_t src1_row_size = nb11;
-    size_t dst_row_size  = nb1;
+    size_t src0_row_size = actx->src0_row_size;
+    size_t src1_row_size = actx->src1_row_size;
+    size_t dst_row_size  = actx->dst_row_size;

    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t src0_nrows = actx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = actx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -574,43 +556,34 @@ static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
        return;
    }

-    const uint8_t * restrict data_src0 = (const uint8_t *) src0->data;
-    const uint8_t * restrict data_src1 = (const uint8_t *) src1->data;
-    uint8_t * restrict data_dst        = (uint8_t *) dst->data;
+    const uint8_t * restrict data_src0 = actx->data_src0;
+    const uint8_t * restrict data_src1 = actx->data_src1;
+    uint8_t * restrict data_dst        = actx->data_dst;

-    const bool src1_valid = src1->ne[0];
-    const int  nc         = (src1_valid) ? ne00 : ne00 / 2;
-    if (!src1_valid) {
-        const int32_t swapped = op_params[1];
-        data_src1             = data_src0;
-        src1_row_size         = src0_row_size;
+    const int nc = actx->nc;

-        const size_t nc_in_bytes = nc * SIZEOF_FP32;
-        data_src0 += swapped ? nc_in_bytes : 0;
-        data_src1 += swapped ? 0 : nc_in_bytes;
-    }
+    const size_t src0_row_size_aligned = actx->src0_row_size_aligned;
+    const size_t src1_row_size_aligned = actx->src1_row_size_aligned;
+    const size_t dst_row_size_aligned  = actx->dst_row_size_aligned;

-    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
-    const size_t src1_row_size_aligned = hex_round_up(src1_row_size, VLEN);
-    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);
+    uint8_t * restrict src0_spad_data = actx->octx->src0_spad.data + (ith * actx->octx->src0_spad.size_per_thread);
+    uint8_t * restrict src1_spad_data = actx->octx->src1_spad.data + (ith * actx->octx->src1_spad.size_per_thread);
+    uint8_t * restrict dst_spad_data  = actx->octx->dst_spad.data + (ith * actx->octx->dst_spad.size_per_thread);

-    uint8_t * restrict src0_spad_data = src0_spad->data + (ith * src0_spad->size_per_thread);
-    uint8_t * restrict src1_spad_data = src1_spad->data + (ith * src1_spad->size_per_thread);
-    uint8_t * restrict dst_spad_data  = dst_spad->data + (ith * dst_spad->size_per_thread);
+    size_t src0_spad_half_size = actx->src0_spad_half_size;
+    size_t src1_spad_half_size = actx->src1_spad_half_size;
+    size_t dst_spad_half_size  = actx->dst_spad_half_size;

-    // While given src0_spad->size_per_thread, divide it to two ping-pong buffer for src0
-    size_t src0_spad_half_size = src0_spad->size_per_thread / 2;
-    size_t src1_spad_half_size = src1_spad->size_per_thread / 2;
-    size_t dst_spad_half_size  = dst_spad->size_per_thread / 2;
-
-    const int BLOCK = src0_spad_half_size / src0_row_size_aligned;  // How many rows can we process in one block
+    const int BLOCK = actx->block;
    if (BLOCK == 0) {
        FARF(ERROR,
             "geglu-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
-             src0_spad->size_per_thread, src0_row_size_aligned);
+             actx->octx->src0_spad.size_per_thread, src0_row_size_aligned);
        return;
    }

+    dma_queue * dma_queue = actx->octx->ctx->dma[ith];
+
    // See discussion: https://github.com/ggml-org/llama.cpp/pull/18151#issuecomment-3678235379
    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
@@ -678,33 +651,7 @@ static void glu_geglu_f32_per_thread(const struct htp_tensor * src0,
         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void unary_silu_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-    unary_silu_f32_per_thread(&octx->src0, &octx->dst, octx->op_params, &octx->src0_spad, &octx->dst_spad, n, i,
-                               octx->src0_nrows_per_thread, octx->ctx->dma[i]);
-}
-
-static void glu_swiglu_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-    glu_swiglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
-                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
-}
-
-static void glu_swiglu_oai_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-    glu_swiglu_oai_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
-                                   &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
-}
-
-static void glu_geglu_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-    glu_geglu_f32_per_thread(&octx->src0, &octx->src1, &octx->dst, octx->op_params, &octx->src0_spad,
-                               &octx->src1_spad, &octx->dst_spad, n, i, octx->src0_nrows_per_thread, octx->ctx->dma[i]);
-}
-
 static int execute_op_activations_f32(struct htp_ops_context * octx) {
-    int err = HTP_STATUS_OK;
-
    const struct htp_tensor * src0 = &octx->src0;
    const struct htp_tensor * src1 = &octx->src1;
    struct htp_tensor *       dst  = &octx->dst;
@@ -719,26 +666,26 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {

    switch (octx->op) {
        case HTP_OP_UNARY_SILU:
-            act_op_func = unary_silu_f32;
+            act_op_func = (worker_callback_t)unary_silu_f32_per_thread;
            op_type     = "silu-f32";
            break;

        case HTP_OP_GLU_SWIGLU:
-            act_op_func = glu_swiglu_f32;
+            act_op_func = (worker_callback_t)glu_swiglu_f32_per_thread;
            op_type     = "swiglu-f32";
            break;

        case HTP_OP_GLU_SWIGLU_OAI:
-            act_op_func = glu_swiglu_oai_f32;
+            act_op_func = (worker_callback_t)glu_swiglu_oai_f32_per_thread;
            op_type     = "swiglu-oai-f32";
            break;
        case HTP_OP_UNARY_GELU:
-            act_op_func = unary_gelu_f32;
+            act_op_func = (worker_callback_t)unary_gelu_f32_per_thread;
            op_type     = "gelu-f32";
            break;

        case HTP_OP_GLU_GEGLU:
-            act_op_func = glu_geglu_f32;
+            act_op_func = (worker_callback_t)glu_geglu_f32_per_thread;
            op_type     = "geglu-f32";
            break;
        default:
@@ -797,13 +744,58 @@ static int execute_op_activations_f32(struct htp_ops_context * octx) {
             octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
    }

-    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs = MIN(n_threads, src0_nrows);
-        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, act_op_func, octx, n_jobs);
+    if ((octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
+        return HTP_STATUS_OK;
    }

-    return err;
+    uint32_t n_jobs = MIN(n_threads, src0_nrows);
+
+    // Prepare context
+    struct htp_act_context actx;
+    actx.octx = octx;
+
+    actx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+
+    actx.src0_row_size = src0_row_size;
+    actx.src1_row_size = src1_row_size;
+    actx.dst_row_size  = dst_row_size;
+
+    actx.src0_row_size_aligned = src0_row_size_aligned;
+    actx.src1_row_size_aligned = src1_row_size_aligned;
+    actx.dst_row_size_aligned  = dst_row_size_aligned;
+
+    actx.src0_spad_half_size = octx->src0_spad.size_per_thread / 2;
+    actx.src1_spad_half_size = octx->src1_spad.size_per_thread / 2;
+    actx.dst_spad_half_size  = octx->dst_spad.size_per_thread / 2;
+
+    actx.block = actx.src0_spad_half_size / actx.src0_row_size_aligned;
+    actx.src0_nrows = src0_nrows;
+
+    actx.nc = dst->ne[0];
+
+    // Pointers and GLU logic
+    const uint8_t * data_src0 = (const uint8_t *) src0->data;
+    const uint8_t * data_src1 = (const uint8_t *) src1->data;
+
+    if (!src1_valid && (octx->op == HTP_OP_GLU_SWIGLU || octx->op == HTP_OP_GLU_SWIGLU_OAI || octx->op == HTP_OP_GLU_GEGLU)) {
+         const int32_t swapped = octx->op_params[1];
+         data_src1 = data_src0;
+         actx.src1_row_size = actx.src0_row_size;
+
+         size_t nc_in_bytes = actx.nc * SIZEOF_FP32;
+         if (swapped) {
+             data_src0 += nc_in_bytes;
+         } else {
+             data_src1 += nc_in_bytes;
+         }
+    }
+
+    actx.data_src0 = data_src0;
+    actx.data_src1 = data_src1;
+    actx.data_dst  = (uint8_t *) dst->data;
+
+    worker_pool_run_func(octx->ctx->worker_pool, act_op_func, &actx, n_jobs);
+    return HTP_STATUS_OK;
 }

 int op_activations(struct htp_ops_context * octx) {
@@ -15,6 +15,13 @@
 #include "htp-ops.h"
 #include "hvx-utils.h"

+struct get_rows_context {
+    struct htp_ops_context * octx;
+    uint32_t src1_nrows_per_thread;
+    struct fastdiv_values get_rows_div_ne10;
+    struct fastdiv_values get_rows_div_ne10_ne11;
+};
+
 #define get_rows_preamble \
    const uint32_t ne00 = octx->src0.ne[0]; \
    const uint32_t ne01 = octx->src0.ne[1]; \
@@ -39,20 +46,22 @@
                                            \
    const uint32_t nr = ne10 * ne11 * ne12;

-static int get_rows_thread_f32_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+static void get_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
+    struct get_rows_context * grctx = (struct get_rows_context *)data;
+    struct htp_ops_context * octx = grctx->octx;
    get_rows_preamble;

    // parallelize by src1 elements (which correspond to dst rows)
-    const uint32_t dr  = octx->src1_nrows_per_thread;
+    const uint32_t dr  = grctx->src1_nrows_per_thread;
    const uint32_t ir0 = dr * ith;
    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;

    const bool is_i32 = (octx->src1.type == HTP_TYPE_I32);

    for (uint32_t i = ir0; i < ir1; ++i) {
-        const uint32_t i12 = fastdiv(i, &octx->get_rows_div_ne10_ne11);
+        const uint32_t i12 = fastdiv(i, &grctx->get_rows_div_ne10_ne11);
        const uint32_t rem = i - i12 * ne11 * ne10;
-        const uint32_t i11 = fastdiv(rem, &octx->get_rows_div_ne10);
+        const uint32_t i11 = fastdiv(rem, &grctx->get_rows_div_ne10);
        const uint32_t i10 = rem - i11 * ne10;

        const uintptr_t src1_addr = octx->src1.data + i10*nb10 + i11*nb11 + i12*nb12;
@@ -68,12 +77,6 @@ static int get_rows_thread_f32_f32(struct htp_ops_context * octx, const int nth,
        const uintptr_t dst_ptr  = octx->dst.data  + i10*nb1  + i11*nb2  + i12*nb3;
        hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, ne00);
    }
-
-    return HTP_STATUS_OK;
-}
-
-static void get_rows_work_f32_f32(unsigned int n, unsigned int i, void *data) {
-    get_rows_thread_f32_f32((struct htp_ops_context *) data, n, i);
 }

 int op_get_rows(struct htp_ops_context * octx) {
@@ -95,12 +98,14 @@ int op_get_rows(struct htp_ops_context * octx) {
        return HTP_STATUS_OK;
    }

-    octx->get_rows_div_ne10      = init_fastdiv_values(octx->src1.ne[0]);
-    octx->get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src1.ne[0] * octx->src1.ne[1]);
+    struct get_rows_context grctx;
+    grctx.octx = octx;
+    grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src1.ne[0]);
+    grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src1.ne[0] * octx->src1.ne[1]);

    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    octx->src1_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    grctx.src1_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;

-    worker_pool_run_func(octx->ctx->worker_pool, get_rows_work_f32_f32, octx, n_jobs);
+    worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32, &grctx, n_jobs);
    return HTP_STATUS_OK;
 }
@@ -102,7 +102,7 @@ static inline bool dma_queue_push(dma_queue * q,
    dmlink(q->tail, desc);
    q->tail = desc;

-    // FARF(ERROR, "dma-push: i %u len %u dst %p src %p\n", q->push_idx, len, dst, src);
+    // FARF(ERROR, "dma-push: i %u width %u nrows %d dst %p src %p\n", q->push_idx, width, nrows, dptr.dst, dptr.src);
    q->push_idx = (q->push_idx + 1) & q->idx_mask;
    return true;
 }
@@ -144,11 +144,37 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {

    dptr = q->dptr[q->pop_idx];

-    // FARF(ERROR, "dma-pop: i %u dst %p\n", q->pop_idx, dst);
+    // FARF(ERROR, "dma-pop: i %u dst %p src %p\n", q->pop_idx, dptr.dst, dptr.src);
    q->pop_idx = (q->pop_idx + 1) & q->idx_mask;
    return dptr;
 }

+static inline dma_ptr dma_queue_pop_nowait(dma_queue * q) {
+    dma_ptr dptr  = { NULL };
+
+    if (q->push_idx == q->pop_idx) {
+        return dptr;
+    }
+
+    dptr = q->dptr[q->pop_idx];
+
+    // FARF(ERROR, "dma-pop-nowait: i %u dst %p src %p\n", q->pop_idx, dptr.dst, dptr.src);
+    q->pop_idx = (q->pop_idx + 1) & q->idx_mask;
+    return dptr;
+}
+
+static inline bool dma_queue_empty(dma_queue * q) {
+    return q->push_idx == q->pop_idx;
+}
+
+static inline uint32_t dma_queue_depth(dma_queue * q) {
+    return (q->push_idx - q->pop_idx) & q->idx_mask;
+}
+
+static inline uint32_t dma_queue_capacity(dma_queue * q) {
+    return q->capacity;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif
@@ -44,32 +44,6 @@ struct htp_ops_context {
    uint32_t src0_nrows_per_thread;
    uint32_t src1_nrows_per_thread;

-    struct fastdiv_values src0_div1;  // fastdiv values for ne1
-    struct fastdiv_values src0_div2;  // fastdiv values for ne2
-    struct fastdiv_values src0_div3;  // fastdiv values for ne3
-    struct fastdiv_values src0_div21; // fastdiv values for ne2 * ne1
-
-    struct fastdiv_values src1_div1;  // fastdiv values for ne1
-    struct fastdiv_values src1_div2;  // fastdiv values for ne2
-    struct fastdiv_values src1_div3;  // fastdiv values for ne3
-    struct fastdiv_values src1_div21; // fastdiv values for ne2 * ne1
-
-    struct fastdiv_values src3_div1;  // fastdiv values for ne1
-    struct fastdiv_values src3_div2;  // fastdiv values for ne2
-    struct fastdiv_values src3_div3;  // fastdiv values for ne3
-    struct fastdiv_values src3_div21; // fastdiv values for ne2 * ne1
-
-    struct fastdiv_values broadcast_rk2;
-    struct fastdiv_values broadcast_rk3;
-    struct fastdiv_values broadcast_rv2;
-    struct fastdiv_values broadcast_rv3;
-
-    struct fastdiv_values set_rows_div_ne12; // fastdiv values for ne12
-    struct fastdiv_values set_rows_div_ne11; // fastdiv values for ne11
-
-    struct fastdiv_values get_rows_div_ne10;      // fastdiv values for ne10
-    struct fastdiv_values get_rows_div_ne10_ne11; // fastdiv values for ne10 * ne11
-
    uint32_t flags;
 };

@@ -49,62 +49,6 @@ struct htp_matmul_context {
    struct fastdiv_values mm_div_r3;
 };

-// vdelta control to replicate first 4x fp32 values across lanes
-static const uint8_t __attribute__((aligned(128))) repl_4x_f32[128] = {
-    0x00, 0x00, 0x00, 0x00, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10,
-    0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20,
-    0x20, 0x20, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10, 0x10, 0x04,
-    0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x40, 0x40, 0x40, 0x40,
-    0x44, 0x44, 0x44, 0x44, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04,
-    0x04, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20, 0x20, 0x20, 0x04, 0x04,
-    0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10, 0x10,
-};
-
-// vdelta control to replicate and interleave first 8x fp32 values across lanes
-static const uint8_t __attribute__((aligned(128))) repl_interleave_8x_f32[128] = {
-    0x00, 0x00, 0x00, 0x00, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x00, 0x00, 0x00,
-    0x00, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20,
-    0x20, 0x20, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20, 0x20, 0x20, 0x04,
-    0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x40, 0x40, 0x40, 0x40,
-    0x44, 0x44, 0x44, 0x44, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x40, 0x40, 0x40, 0x40, 0x44, 0x44, 0x44,
-    0x44, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20, 0x20, 0x20, 0x04, 0x04,
-    0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20, 0x20, 0x20,
-};
-
-// vdelta control to replicate first fp32 value across all elements
-static const uint8_t __attribute__((aligned(128))) repl_1x_f32[128] = {
-    0x00, 0x00, 0x00, 0x00, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10,
-    0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x20, 0x20, 0x20, 0x20, 0x04, 0x04,
-    0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08,
-    0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x40, 0x40, 0x40, 0x40, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08,
-    0x04, 0x04, 0x04, 0x04, 0x10, 0x10, 0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04,
-    0x04, 0x20, 0x20, 0x20, 0x20, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04, 0x10, 0x10,
-    0x10, 0x10, 0x04, 0x04, 0x04, 0x04, 0x08, 0x08, 0x08, 0x08, 0x04, 0x04, 0x04, 0x04,
-};
-
-// vdelta control to replicate first fp16 value across all elements
-static const uint8_t __attribute__((aligned(128))) repl_1x_f16[128] = {
-    0x00, 0x00, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x10, 0x10, 0x02,
-    0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x20, 0x20, 0x02, 0x02, 0x04, 0x04,
-    0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08,
-    0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x40, 0x40, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02,
-    0x04, 0x04, 0x02, 0x02, 0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02,
-    0x02, 0x20, 0x20, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x10, 0x10,
-    0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-};
-
-// vdelta control to replicate first fp16 value across all elements
-static const uint8_t __attribute__((aligned(128))) repl_2x_f16[128] = {
-    0x00, 0x00, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x20, 0x20, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x00, 0x00, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x20, 0x20, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-    0x10, 0x10, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02, 0x08, 0x08, 0x02, 0x02, 0x04, 0x04, 0x02, 0x02,
-};
-
 // vdelta control to expand first 32 e8m0 values into 32 uint32 elements
 static const uint8_t __attribute__((aligned(128))) expand_x32_e8m0[128] = {
    0x00, 0x00, 0x00, 0x00, 0x01, 0x04, 0x00, 0x00, 0x02, 0x00, 0x08, 0x08, 0x01, 0x02, 0x00, 0x04, 0x04, 0x00, 0x00,
@@ -2067,10 +2011,10 @@ static inline void quantize_block_f32_q8x1(float * restrict x, uint8_t * restric
    HVX_Vector vx3_qf = Q6_Vqf32_vsub_VsfVsf(vx[3], zero);  // 32 elements

    // Convert to QF32
-    HVX_Vector vmax0_qf = Q6_Vqf32_vsub_VsfVsf(vmax0_sf, zero);
-    HVX_Vector vmax1_qf = Q6_Vqf32_vsub_VsfVsf(vmax1_sf, zero);
-    HVX_Vector vmax2_qf = Q6_Vqf32_vsub_VsfVsf(vmax2_sf, zero);
-    HVX_Vector vmax3_qf = Q6_Vqf32_vsub_VsfVsf(vmax3_sf, zero);
+    HVX_Vector vmax0_qf = Q6_Vqf32_vsub_VsfVsf(vmax0_sf, zero); // replicated over all lanes
+    HVX_Vector vmax1_qf = Q6_Vqf32_vsub_VsfVsf(vmax1_sf, zero); // replicated over all lanes
+    HVX_Vector vmax2_qf = Q6_Vqf32_vsub_VsfVsf(vmax2_sf, zero); // replicated over all lanes
+    HVX_Vector vmax3_qf = Q6_Vqf32_vsub_VsfVsf(vmax3_sf, zero); // replicated over all lanes

    // Combine and convert to fp16
    HVX_Vector vmax01_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vmax1_qf, vmax0_qf)));
@@ -2080,11 +2024,6 @@ static inline void quantize_block_f32_q8x1(float * restrict x, uint8_t * restric
    HVX_Vector vx01_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vx1_qf, vx0_qf)));
    HVX_Vector vx23_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vx3_qf, vx2_qf)));

-    // Replicate first fp16 scale across all lanes
-    HVX_Vector ctrl = *(const HVX_Vector *) repl_2x_f16;
-    vmax01_hf         = Q6_V_vdelta_VV(vmax01_hf, ctrl);
-    vmax23_hf         = Q6_V_vdelta_VV(vmax23_hf, ctrl);
-
    HVX_Vector vd01_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax01_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
    HVX_Vector vd23_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax23_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
    HVX_Vector vd01_hf   = Q6_Vhf_equals_Vqf16(vd01_qf16);
@@ -2130,13 +2069,8 @@ static inline void quantize_block_f32_q8x2(float * restrict x, uint8_t * restric
    HVX_Vector vx23_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vx3_qf, vx2_qf)));

    // Compute max and scale
-    HVX_Vector vmax01_hf = hvx_vec_reduce_max_f16(hvx_vec_abs_f16(vx01_hf));
-    HVX_Vector vmax23_hf = hvx_vec_reduce_max_f16(hvx_vec_abs_f16(vx23_hf));
-
-    // Replicate first fp16 scale across all lanes
-    HVX_Vector ctrl = *(const HVX_Vector *) repl_1x_f16;
-    vmax01_hf         = Q6_V_vdelta_VV(vmax01_hf, ctrl);
-    vmax23_hf         = Q6_V_vdelta_VV(vmax23_hf, ctrl);
+    HVX_Vector vmax01_hf = hvx_vec_reduce_max_f16(hvx_vec_abs_f16(vx01_hf)); // replicated over all lanes
+    HVX_Vector vmax23_hf = hvx_vec_reduce_max_f16(hvx_vec_abs_f16(vx23_hf)); // replicated over all lanes

    HVX_Vector vd01_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax01_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
    HVX_Vector vd23_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax23_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
@@ -2179,11 +2113,7 @@ static inline void quantize_block_f32_q8x4(float * restrict x, uint8_t * restric

    // Compute max and scale
    HVX_Vector vmax_hf = hvx_vec_reduce_max_f16(hvx_vec_abs_f16(vx01_hf));
-    vmax_hf            = hvx_vec_reduce_max2_f16(hvx_vec_abs_f16(vx23_hf), vmax_hf);
-
-    // Replicate first fp16 scale across all lanes
-    HVX_Vector ctrl = *(const HVX_Vector *) repl_1x_f16;
-    vmax_hf         = Q6_V_vdelta_VV(vmax_hf, ctrl);
+    vmax_hf            = hvx_vec_reduce_max2_f16(hvx_vec_abs_f16(vx23_hf), vmax_hf); // replicated over all lanes

    HVX_Vector vd_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
    HVX_Vector vd_hf   = Q6_Vhf_equals_Vqf16(vd_qf16);
@@ -10,6 +10,7 @@

 #include "hex-dma.h"
 #include "hvx-utils.h"
+#include "hex-fastdiv.h"

 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -21,6 +22,9 @@
 #define HTP_ROPE_TYPE_NORMAL 0
 #define HTP_ROPE_TYPE_NEOX   2

+#define HTP_ROPE_SPAD_NROWS  16
+#define HTP_ROPE_SPAD_BLOCK  (HTP_ROPE_SPAD_NROWS/2)
+
 #define htp_rope_preamble              \
    const uint32_t ne00 = src0->ne[0]; \
    const uint32_t ne01 = src0->ne[1]; \
@@ -42,7 +46,7 @@
    const uint32_t nb2 = dst->nb[2];   \
    const uint32_t nb3 = dst->nb[3];

-struct rope_th_ctx {
+struct htp_rope_context {
    int32_t n_dims;
    int32_t mode;
    int32_t n_ctx_orig;
@@ -57,7 +61,19 @@ struct rope_th_ctx {
    float theta_scale;
    float corr_dims[2];

+    uint32_t src0_nrows_per_thread;
+    size_t spad_stride;
+
    struct htp_ops_context * octx;
+
+    size_t src0_row_size;
+    size_t dst_row_size;
+    size_t src0_row_size_aligned;
+    size_t dst_row_size_aligned;
+    size_t theta_cache_offset;
+    uint32_t src0_nrows;
+
+    uint64_t t_start;
 };

 static float rope_yarn_ramp(const float low, const float high, const int i0) {
@@ -117,64 +133,23 @@ static void rope_corr_dims(int     n_dims,
    dims[1]     = MIN(n_dims - 1, end);
 }

-static void init_rope_ctx(struct rope_th_ctx * rope_ctx, struct htp_ops_context * octx) {
-    memset(rope_ctx, 0, sizeof(struct rope_th_ctx));
+static inline void hvx_rope_neox_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
+    const HVX_Vector * restrict vsrc   = (const HVX_Vector *) src0;
+    const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
+    HVX_Vector       * restrict vdst   = (HVX_Vector *) dst;

-    const int32_t * op_params = &octx->op_params[0];
+    uint32_t nvec = (ne / (VLEN_FP32 * 2) * 2); // 2 vecs per loop, step of 2

-    rope_ctx->n_dims     = ((const int32_t *) op_params)[1];
-    rope_ctx->mode       = ((const int32_t *) op_params)[2];
-    rope_ctx->n_ctx_orig = ((const int32_t *) op_params)[4];
+    uint32_t he = ne / 2;         // half_dims offset in elements
+    uint32_t hv = he / VLEN_FP32; // half_dims offset in vectors

-    memcpy(&rope_ctx->freq_base, (int32_t *) op_params + 5, sizeof(float));
-    memcpy(&rope_ctx->freq_scale, (int32_t *) op_params + 6, sizeof(float));
-    memcpy(&rope_ctx->ext_factor, (int32_t *) op_params + 7, sizeof(float));
-    memcpy(&rope_ctx->attn_factor, (int32_t *) op_params + 8, sizeof(float));
-    memcpy(&rope_ctx->beta_fast, (int32_t *) op_params + 9, sizeof(float));
-    memcpy(&rope_ctx->beta_slow, (int32_t *) op_params + 10, sizeof(float));
-    memcpy(&rope_ctx->sections, (int32_t *) op_params + 11, sizeof(int) * 4);
+    #pragma unroll(2)
+    for (uint32_t i = 0; i < nvec; i += 2) {
+        HVX_Vector v0 = vsrc[i/2+0];
+        HVX_Vector v1 = vsrc[i/2+hv];

-    rope_ctx->theta_scale = powf(rope_ctx->freq_base, -2.0f / rope_ctx->n_dims);
-
-    rope_corr_dims(rope_ctx->n_dims, rope_ctx->n_ctx_orig, rope_ctx->freq_base, rope_ctx->beta_fast,
-                   rope_ctx->beta_slow, rope_ctx->corr_dims);
-
-    rope_ctx->octx = octx;
-    FARF(HIGH, "rope-f32 n_dims:%d, ext_factor:%.6f, theta_scale:%.6f, attn_factor:%.6f\n", rope_ctx->n_dims,
-         rope_ctx->ext_factor, rope_ctx->theta_scale, rope_ctx->attn_factor);
-}
-
-static void hvx_calc_rope_neox_f32(const float * restrict src0,
-                                   float * restrict dst,
-                                   const int num_elems,
-                                   const float * restrict theta_cache) {
-    // for (int i = 0; i < num_elems; i += 2) {
-    //const float cos_theta = theta_cache[i + 0];
-    //const float sin_theta = theta_cache[i + 1];
-
-    //const float x0 = src[0];
-    //const float x1 = src[num_elems/2];
-
-    //dst[0] = x0*cos_theta - x1*sin_theta;
-    //dst[num_elems/2] = x0*sin_theta + x1*cos_theta;
-
-    //src += 1;
-    //dst += 1;
-    // }
-
-    const uint8_t * restrict src0_curr  = (const uint8_t *) src0;
-    const uint8_t * restrict theta_curr = (const uint8_t *) theta_cache;
-    uint8_t * restrict dst_curr         = (uint8_t *) dst;
-
-    int step_of_1 = num_elems >> 6;  // 6 because we process two vectors at once
-    int half_size = (sizeof(float) * (num_elems / 2));
-
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector v0 = *(HVX_Vector *) src0_curr;
-        HVX_Vector v1 = *(HVX_Vector *) (src0_curr + half_size);
-
-        HVX_Vector v2 = *(HVX_Vector *) theta_curr;
-        HVX_Vector v3 = *(HVX_Vector *) (theta_curr + VLEN);
+        HVX_Vector v2 = vtheta[i+0];
+        HVX_Vector v3 = vtheta[i+1];

        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta

@@ -186,45 +161,34 @@ static void hvx_calc_rope_neox_f32(const float * restrict src0,
        HVX_Vector v4 = Q6_Vqf32_vsub_Vqf32Vqf32(vx0_c, vx1_s);
        HVX_Vector v5 = Q6_Vqf32_vadd_Vqf32Vqf32(vx0_s, vx1_c);

-        *(HVX_Vector *) dst_curr               = Q6_Vsf_equals_Vqf32(v4);
-        *(HVX_Vector *) (dst_curr + half_size) = Q6_Vsf_equals_Vqf32(v5);
+        vdst[i/2+0]  = Q6_Vsf_equals_Vqf32(v4);
+        vdst[i/2+hv] = Q6_Vsf_equals_Vqf32(v5);
+    }

-        src0_curr += VLEN;
-        theta_curr += 2 * VLEN;
-        dst_curr += VLEN;
+    for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
+        const float cos_theta = theta_cache[i+0];
+        const float sin_theta = theta_cache[i+1];
+        float x0 = src0[i/2];
+        float x1 = src0[i/2 + he];
+        dst[i/2]      = x0 * cos_theta - x1 * sin_theta;
+        dst[i/2 + he] = x0 * sin_theta + x1 * cos_theta;
    }
 }

-static void hvx_calc_rope_f32(const float * restrict src0,
-                              float * restrict dst,
-                              const int num_elems,
-                              const float * restrict theta_cache) {
-    // for (int i = 0; i < num_elems; i += 2) {
-    //const float cos_theta = theta_cache[i + 0];
-    //const float sin_theta = theta_cache[i + 1];
+static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict src0, uint32_t ne, const float * restrict theta_cache) {
+    const HVX_Vector * restrict vsrc   = (const HVX_Vector *) src0;
+    const HVX_Vector * restrict vtheta = (const HVX_Vector *) theta_cache;
+    HVX_Vector       * restrict vdst   = (HVX_Vector *) dst;

-    //const float x0 = src[0];
-    //const float x1 = src[1];
+    uint32_t nvec = (ne / (VLEN_FP32 * 2)) * 2; // 2 vecs per loop, step of two

-    //dst[0] = x0*cos_theta - x1*sin_theta;
-    //dst[1] = x0*sin_theta + x1*cos_theta;
+    #pragma unroll(2)
+    for (uint32_t i = 0; i < nvec; i+=2) {
+        HVX_Vector v0 = vsrc[i+0];
+        HVX_Vector v1 = vsrc[i+1];

-    //src += 2;
-    //dst += 2;
-    // }
-
-    const uint8_t * restrict src0_curr  = (const uint8_t *) src0;
-    const uint8_t * restrict theta_curr = (const uint8_t *) theta_cache;
-    uint8_t * restrict dst_curr         = (uint8_t *) dst;
-
-    int step_of_1 = num_elems >> 6;  // 6 because we process two vectors at once
-
-    for (int i = 0; i < step_of_1; i++) {
-        HVX_Vector v0 = *(HVX_Vector *) src0_curr;
-        HVX_Vector v1 = *(HVX_Vector *) (src0_curr + VLEN);
-
-        HVX_Vector v2 = *(HVX_Vector *) theta_curr;
-        HVX_Vector v3 = *(HVX_Vector *) (theta_curr + VLEN);
+        HVX_Vector v2 = vtheta[i+0];
+        HVX_Vector v3 = vtheta[i+1];

        HVX_VectorPair vx0_x1   = Q6_W_vdeal_VVR(v1, v0, -4);  // vx0_x1[0] = x0, vx0_x1[1] = x1
        HVX_VectorPair vcos_sin = Q6_W_vdeal_VVR(v3, v2, -4);  // vcos_sin[0] = cos_theta, vcos_sin[1] = sin_theta
@@ -239,116 +203,65 @@ static void hvx_calc_rope_f32(const float * restrict src0,

        HVX_VectorPair vstore = Q6_W_vshuff_VVR(Q6_Vsf_equals_Vqf32(v5), Q6_Vsf_equals_Vqf32(v4), -4);

-        *(HVX_Vector *) dst_curr          = Q6_V_lo_W(vstore);
-        *(HVX_Vector *) (dst_curr + VLEN) = Q6_V_hi_W(vstore);
+        vdst[i+0] = Q6_V_lo_W(vstore);
+        vdst[i+1] = Q6_V_hi_W(vstore);
+    }

-        src0_curr += 2 * VLEN;
-        theta_curr += 2 * VLEN;
-        dst_curr += 2 * VLEN;
+    for (uint32_t i = nvec * VLEN_FP32; i < ne; i += 2) {
+        const float cos_theta = theta_cache[i+0];
+        const float sin_theta = theta_cache[i+1];
+        float x0 = src0[i+0];
+        float x1 = src0[i+1];
+        dst[i+0] = x0 * cos_theta - x1 * sin_theta;
+        dst[i+1] = x0 * sin_theta + x1 * cos_theta;
    }
 }

-static void rope_hex_f32(struct rope_th_ctx * rope_ctx,
-                         const uint32_t       ir0,
-                         const uint32_t       ir1,
-                         int                  nth,
-                         int                  ith,
-                         const int            opt_path) {
-    struct htp_ops_context * octx = rope_ctx->octx;
+static void inline rope_basic_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
+                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    #pragma unroll(4)
+    for (uint32_t i = 0; i < nr; i++) {
+        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
+        float * s = (float *) (src + i * rctx->src0_row_size_aligned);
+
+        hvx_rope_f32_aa(d, s, rctx->n_dims, theta_cache);
+
+        // fill the remain channels with data from src tensor
+        if (rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        }
+    }
+}
+
+static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
+                   uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    #pragma unroll(4)
+    for (uint32_t i = 0; i < nr; i++) {
+        float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
+        float * s = (float *) (src + i * rctx->src0_row_size_aligned);
+
+        hvx_rope_neox_f32_aa(d, s, rctx->n_dims, theta_cache);
+
+        // fill the remain channels with data from src tensor
+        if (rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        }
+    }
+}
+
+static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_rope_context * rctx = (struct htp_rope_context *) data;
+    struct htp_ops_context * octx = rctx->octx;

    const struct htp_tensor * src0 = &octx->src0;
    const struct htp_tensor * src1 = &octx->src1;
    const struct htp_tensor * src2 = &octx->src2;
    struct htp_tensor *       dst  = &octx->dst;

-    const int32_t mode    = rope_ctx->mode;
-    const bool    is_neox = mode & HTP_ROPE_TYPE_NEOX;
-
    htp_rope_preamble;

-    const int32_t * pos = (const int32_t *) src1->data;
-
-    float * wp0 = (float *) (octx->src0_spad.data + (ith * nb01));
-
-    const float * freq_factors = NULL;
-    if (src2 != NULL) {
-        freq_factors = (const float *) src2->data;
-    }
-
-    const uint32_t i1_end       = MIN(ir1, ne1);
-    const int32_t  half_dims    = rope_ctx->n_dims / 2;
-    const size_t   remain_bytes = (ne0 - rope_ctx->n_dims) * sizeof(float);
-    for (uint32_t i3 = 0; i3 < ne3; i3++) {      // batch
-        for (uint32_t i2 = 0; i2 < ne2; i2++) {  // seq-len
-            const int32_t p = pos[i2];
-
-            rope_cache_init(p, rope_ctx->freq_scale, freq_factors, rope_ctx->corr_dims, ne0, rope_ctx->ext_factor,
-                            rope_ctx->attn_factor, wp0, rope_ctx->theta_scale);
-
-            for (uint32_t i1 = ir0; i1 < i1_end; i1++) {  // attn-heads
-                const float * src      = (float *) ((char *) src0->data + i3 * nb03 + i2 * nb02 + i1 * nb01);
-                float *       dst_data = (float *) ((char *) dst->data + i3 * nb3 + i2 * nb2 + i1 * nb1);
-
-                const float * src_loc      = src;
-                float *       dst_data_loc = dst_data;
-
-                if (1 == opt_path) {
-                    if (is_neox) {
-                        hvx_calc_rope_neox_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
-                    } else {
-                        hvx_calc_rope_f32(src_loc, dst_data_loc, rope_ctx->n_dims, wp0);
-                    }
-
-                    src_loc += rope_ctx->n_dims;
-                    dst_data_loc += rope_ctx->n_dims;
-                } else {
-                    for (uint32_t i0 = 0; i0 < rope_ctx->n_dims; i0 += 2) {
-                        const float cos_theta = wp0[i0 + 0];
-                        const float sin_theta = wp0[i0 + 1];
-
-                        if (is_neox) {
-                            const float x0 = src_loc[0];
-                            const float x1 = src_loc[half_dims];
-
-                            dst_data_loc[0]         = x0 * cos_theta - x1 * sin_theta;
-                            dst_data_loc[half_dims] = x0 * sin_theta + x1 * cos_theta;
-
-                            src_loc += 1;
-                            dst_data_loc += 1;
-                        } else {
-                            const float x0 = src_loc[0];
-                            const float x1 = src_loc[1];
-
-                            dst_data_loc[0] = x0 * cos_theta - x1 * sin_theta;
-                            dst_data_loc[1] = x0 * sin_theta + x1 * cos_theta;
-
-                            src_loc += 2;
-                            dst_data_loc += 2;
-                        }
-                    }
-
-                    src_loc += (is_neox ? half_dims : 0);
-                    dst_data_loc += (is_neox ? half_dims : 0);
-                }
-
-                // TODO: use simd to speed up the remaining elements copy
-                memcpy(dst_data_loc, src_loc, remain_bytes);
-            }
-        }
-    }
-}
-
-static void rope_job_f32_per_thread(struct rope_th_ctx * rope_ctx, int nth, int ith) {
-    struct htp_ops_context * octx = rope_ctx->octx;
-
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * src1 = &octx->src1;
-    struct htp_tensor *       dst  = &octx->dst;
-
-    htp_rope_preamble;
-
-    const uint32_t src0_nrows            = ne01 * ne02 * ne03;  // src0 rows
-    const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;
+    const uint32_t src0_nrows = rctx->src0_nrows;
+    const uint32_t src0_nrows_per_thread = rctx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -358,32 +271,114 @@ static void rope_job_f32_per_thread(struct rope_th_ctx * rope_ctx, int nth, int
        return;
    }

-    uint64_t t1, t2;
-    t1 = HAP_perf_get_qtimer_count();
+    uint64_t tt = HAP_perf_get_qtimer_count();

-    int is_aligned = 1;
-    int opt_path   = 0;
-    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) || (0 == hex_is_aligned((void *) src1->data, VLEN)) ||
-        (0 == hex_is_aligned((void *) dst->data, VLEN))) {
-        FARF(HIGH, "rope-f32: unaligned addresses in rope op, possibly slower execution\n");
-        is_aligned = 0;
-    }
-    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
-        opt_path = 1;
+    const int32_t mode    = rctx->mode;
+    const bool    is_neox = mode & HTP_ROPE_TYPE_NEOX;
+
+    // VTCM setup
+    uint8_t * src0_spad_base = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
+    float *   theta_cache    = (float *) (src0_spad_base);
+              src0_spad_base = src0_spad_base + rctx->theta_cache_offset;
+    uint8_t * dst_spad_base  = octx->dst_spad.data + (ith * octx->dst_spad.size_per_thread);
+
+    dma_queue * dma_queue = octx->ctx->dma[ith];
+    const int32_t * pos = (const int32_t *) src1->data;
+    const float * freq_factors = src2->data ? (const float *) src2->data : NULL;
+
+    uint32_t ir = 0;
+    uint32_t prev_i2 = (uint32_t) -1;
+
+    for (uint32_t i3 = 0; i3 < ne3; i3++) { // batch
+        for (uint32_t i2 = 0; i2 < ne2; i2++) { // seq-len
+            for (uint32_t i1 = 0; i1 < ne1; ) { // attn-heads
+                if (ir < src0_start_row) { ir++; i1++; continue; }
+                if (ir >= src0_end_row) goto done;
+
+                // Rows in this block
+                const uint32_t nrows = MIN(src0_end_row - ir, ne1 - i1);
+
+                // Depth before prefetch
+                uint32_t dma_depth = dma_queue_depth(dma_queue);
+
+                // FARF(HIGH, "rope-block %u: ir %u n-rows %u dma-depth %u : usec %u", ith, ir, nrows, dma_depth,
+                //             (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
+
+                // Prefetch loop
+                for (uint32_t pnr = 0, pr = 0; pr < nrows && pr < HTP_ROPE_SPAD_NROWS; pr += pnr) {
+                    pnr = MIN(nrows - pr, HTP_ROPE_SPAD_BLOCK);
+
+                    uint32_t pi1 = i1 + pr;
+                    uint32_t pir = ir + pr;
+
+                    // Dummy DMA transaction for sequencing (interleaving dst,src,dst,...)
+                    dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr((void *) dst->data, dst_spad_base + pr * rctx->dst_row_size_aligned), 0, 0, 0);
+
+                    const uint8_t * src_addr = (const uint8_t *) src0->data + i3 * nb03 + i2 * nb02 + pi1 * nb01;
+                          uint8_t * src_spad = src0_spad_base + pr * rctx->src0_row_size_aligned;
+                    dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src_spad, src_addr),
+                        rctx->src0_row_size_aligned, rctx->src0_row_size, pnr);
+
+                    // FARF(HIGH, "rope-prefetch %u: pr %u i1 %u i2 %u i3 %u src-spad %p src-addr %p pnr %u", ith, pir, pi1, i2, i3, src_spad, src_addr, pnr);
+                }
+
+                // Update theta cache
+                if (i2 != prev_i2) {
+                    prev_i2 = i2;
+
+                    const int32_t p = pos[i2];
+                    rope_cache_init(p, rctx->freq_scale, freq_factors, rctx->corr_dims, ne0, rctx->ext_factor, rctx->attn_factor, theta_cache, rctx->theta_scale);
+
+                    // FARF(HIGH, "rope-theta %u: ir %u i1 %u i2 %u i3 %u cache %p : usec %u", ith, ir, i1, i2, i3, theta_cache,
+                    //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
+                }
+
+                // Skip DMA transactions from prev block (if any)
+                // No need to wait for these since the DMA is setup for in-order processing
+                for (uint32_t d=0; d < dma_depth; d++) { dma_queue_pop_nowait(dma_queue); }
+
+                // Compute loop
+                for (uint32_t cnr = 0, cr = 0; cr < nrows; cr += cnr, ir += cnr, i1 += cnr) {
+                    // Number of rows to compute
+                    cnr = MIN(nrows - cr, HTP_ROPE_SPAD_BLOCK);
+
+                    uint8_t * dst_spad = (uint8_t *) dma_queue_pop(dma_queue).src;
+                    uint8_t * src_spad = (uint8_t *) dma_queue_pop(dma_queue).dst;
+
+                    // FARF(HIGH, "rope-compute %u: ir %u i1 %u i2 %u i3 %u src-spad %p cnr %u : usec %u", ith, ir, i1, i2, i3, src_spad, cnr,
+                    //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
+
+                    if (is_neox) {
+                        rope_neox_f32(rctx, dst_spad, src_spad, cnr, ne0, theta_cache);
+                    } else {
+                        rope_basic_f32(rctx, dst_spad, src_spad, cnr, ne0, theta_cache);
+                    }
+
+                    uint8_t * dst_addr = (uint8_t *) dst->data + i3 * nb3 + i2 * nb2 + i1 * nb1;
+                    dma_queue_push_vtcm_to_ddr(dma_queue, dma_make_ptr(dst_addr, dst_spad), rctx->dst_row_size, rctx->dst_row_size_aligned, cnr);
+
+                    // Prefetch more rows (if any)
+                    if ((cr + HTP_ROPE_SPAD_NROWS) < nrows) {
+                        uint32_t pnr = MIN(nrows - (cr + HTP_ROPE_SPAD_NROWS), HTP_ROPE_SPAD_BLOCK);
+                        uint32_t pi1 = i1 + HTP_ROPE_SPAD_NROWS;
+                        uint32_t pir = ir + HTP_ROPE_SPAD_NROWS;
+
+                        const uint8_t * src_addr = (const uint8_t *) src0->data + i3 * nb03 + i2 * nb02 + pi1 * nb01;
+                        dma_queue_push_ddr_to_vtcm(dma_queue, dma_make_ptr(src_spad, src_addr),
+                            rctx->src0_row_size_aligned, rctx->src0_row_size, pnr);
+
+                        // FARF(HIGH, "rope-prefetch %u: pr %u i1 %u i2 %u i3 %u src-spad %p src-addr %p pnr %u", ith, pir, pi1, i2, i3, src_spad, src_addr, pnr);
+                    }
+                }
+            }
+        }
    }

-    rope_hex_f32(rope_ctx, src0_start_row, src0_end_row, nth, ith, opt_path);
+done:
+    dma_queue_flush(dma_queue);
+    tt = HAP_perf_get_qtimer_count() - tt;

-    t2 = HAP_perf_get_qtimer_count();
-
-    FARF(HIGH, "rope-f32: %d/%d/%d: (%u:%u) usec %u\n", ith, nth, opt_path, src0_start_row, src0_end_row,
-         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
-}
-
-static void rope_job_dispatcher_f32(unsigned int n, unsigned int i, void * data) {
-    struct rope_th_ctx * rope_ctx = (struct rope_th_ctx *) data;
-
-    rope_job_f32_per_thread(rope_ctx, n, i);
+    FARF(HIGH, "rope-f32: %d/%d: (%u:%u) usec %u\n", ith, nth, src0_start_row, src0_end_row, (unsigned) HAP_perf_qtimer_count_to_us(tt));
 }

 static int execute_op_rope_f32(struct htp_ops_context * octx) {
@@ -394,17 +389,10 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
    const struct htp_tensor * src2 = &octx->src2;
    struct htp_tensor *       dst  = &octx->dst;

-    worker_callback_t op_func;
-    const char *      op_type = NULL;
-
-    struct rope_th_ctx rope_ctx;
+    const char * op_type = "rope-f32";

    switch (octx->op) {
        case HTP_OP_ROPE:
-            op_func = rope_job_dispatcher_f32;
-            op_type = "rope-f32";
-
-            init_rope_ctx(&rope_ctx, octx);
            break;

        default:
@@ -415,49 +403,79 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
    const uint32_t n_threads = octx->n_threads;

    const size_t src0_row_size = src0->nb[1];
-    const size_t src1_row_size = src0_row_size;
    const size_t dst_row_size  = dst->nb[1];

-    // VTCM scratchpads for all tensors
-    // N rows per thread, padded to HVX vector size
-    octx->dst_spad.size  = hex_round_up(dst_row_size, 128) * n_threads;
-    octx->src0_spad.size = hex_round_up(src0_row_size, 128) * n_threads;
-    octx->src1_spad.size = hex_round_up(src1_row_size, 128) * n_threads;
+    // Aligned row sizes for VTCM
+    const size_t src0_row_size_aligned    = hex_round_up(src0_row_size, VLEN);
+    const size_t dst_row_size_aligned     = hex_round_up(dst_row_size, VLEN);
+    const size_t theta_cache_size_aligned = hex_round_up(src0->ne[0] * sizeof(float), 128);

-    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;
+    // Calculate spad sizes per thread
+    size_t src0_spad_per_thread = theta_cache_size_aligned + HTP_ROPE_SPAD_NROWS * src0_row_size_aligned;
+    size_t dst_spad_per_thread  = HTP_ROPE_SPAD_NROWS * dst_row_size_aligned;
+    size_t spad_per_thread = src0_spad_per_thread + dst_spad_per_thread;

-    if (src2->ne[0]) {
-        FARF(HIGH,
-             "%s: %ux%ux%ux%u (x %ux%ux%ux%u x %ux%ux%ux%u) -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u "
-             "dst-spad-size %u\n",
-             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
-             src1->ne[3], src2->ne[0], src2->ne[1], src2->ne[2], src2->ne[3], dst->ne[0], dst->ne[1], dst->ne[2],
-             dst->ne[3], octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);
-    } else {
-        FARF(HIGH,
-             "%s: %ux%ux%ux%u (%ux%ux%ux%u) -> %ux%ux%ux%u : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n",
-             op_type, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], src1->ne[0], src1->ne[1], src1->ne[2],
-             src1->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], octx->src0_spad.size, octx->src1_spad.size,
-             octx->dst_spad.size);
-    }
-
-    // Make sure the reserved vtcm size is sufficient
-    if (octx->ctx->vtcm_size < spad_size) {
-        FARF(ERROR, "%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
-             spad_size);
+    // Check if we fit in VTCM
+    size_t total_vtcm_needed = spad_per_thread * n_threads;
+    if (octx->ctx->vtcm_size < total_vtcm_needed) {
+        FARF(ERROR, "%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size, total_vtcm_needed);
        return HTP_STATUS_VTCM_TOO_SMALL;
    }

-    octx->src0_spad.data = octx->ctx->vtcm_base;
-    octx->src1_spad.data = octx->src0_spad.data + octx->src0_spad.size;
-    octx->dst_spad.data  = octx->src1_spad.data + octx->src1_spad.size;
+    // Assign sizes
+    octx->src0_spad.size_per_thread = src0_spad_per_thread;
+    octx->dst_spad.size_per_thread  = dst_spad_per_thread;
+    octx->src0_spad.size = n_threads * src0_spad_per_thread;
+    octx->dst_spad.size  = n_threads * dst_spad_per_thread;
+    octx->src1_spad.size = 0;

+    // Assign pointers
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->src1_spad.data = NULL;
+    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size;
+
+    // Fill context
+    struct htp_rope_context rctx;
+    memset(&rctx, 0, sizeof(struct htp_rope_context));
+
+    rctx.t_start = HAP_perf_get_qtimer_count();
+
+    rctx.octx = octx;
+
+    const int32_t * op_params = &octx->op_params[0];
+    rctx.n_dims     = ((const int32_t *) op_params)[1];
+    rctx.mode       = ((const int32_t *) op_params)[2];
+    rctx.n_ctx_orig = ((const int32_t *) op_params)[4];
+
+    memcpy(&rctx.freq_base,   (int32_t *) op_params + 5,  sizeof(float));
+    memcpy(&rctx.freq_scale,  (int32_t *) op_params + 6,  sizeof(float));
+    memcpy(&rctx.ext_factor,  (int32_t *) op_params + 7,  sizeof(float));
+    memcpy(&rctx.attn_factor, (int32_t *) op_params + 8,  sizeof(float));
+    memcpy(&rctx.beta_fast,   (int32_t *) op_params + 9,  sizeof(float));
+    memcpy(&rctx.beta_slow,   (int32_t *) op_params + 10, sizeof(float));
+    memcpy(&rctx.sections,    (int32_t *) op_params + 11, sizeof(int) * 4);
+
+    rctx.theta_scale = powf(rctx.freq_base, -2.0f / rctx.n_dims);
+
+    rope_corr_dims(rctx.n_dims, rctx.n_ctx_orig, rctx.freq_base, rctx.beta_fast, rctx.beta_slow, rctx.corr_dims);
+
+    rctx.src0_row_size = src0_row_size;
+    rctx.dst_row_size  = dst_row_size;
+    rctx.src0_row_size_aligned = src0_row_size_aligned;
+    rctx.dst_row_size_aligned  = dst_row_size_aligned;
+    rctx.theta_cache_offset    = theta_cache_size_aligned;
+
+    uint32_t ne0 = dst->ne[0];
    uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
+    rctx.src0_nrows = src0_nrows;
+
+    FARF(HIGH, "rope-f32 n-rows %u n-dims %d ne0 %u ext-factor %.6f theta-scale %.6f attn-factor %.6f\n", rctx.src0_nrows, rctx.n_dims, ne0,
+         rctx.ext_factor, rctx.theta_scale, rctx.attn_factor);

    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
-        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
-        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, op_func, &rope_ctx, n_jobs);
+        uint32_t n_jobs = MIN(n_threads, src0_nrows);
+        rctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        worker_pool_run_func(octx->ctx->worker_pool, rope_job_f32, &rctx, n_jobs);
    }

    return err;
@@ -43,11 +43,21 @@
                                            \
    const uint32_t nr  = ne01;

-static int set_rows_thread_f32_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+struct htp_set_rows_context {
+    struct htp_ops_context * octx;
+    struct fastdiv_values div_ne12;
+    struct fastdiv_values div_ne11;
+    uint32_t src0_nrows_per_thread;
+};
+
+static void set_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
+    struct htp_set_rows_context * srctx = (struct htp_set_rows_context *)data;
+    struct htp_ops_context * octx = srctx->octx;
+
    set_rows_preamble;

    // parallelize by rows of src0
-    const uint32_t dr  = octx->src0_nrows_per_thread;
+    const uint32_t dr  = srctx->src0_nrows_per_thread;
    const uint32_t ir0 = dr * ith;
    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;

@@ -56,8 +66,8 @@ static int set_rows_thread_f32_f32(struct htp_ops_context * octx, const int nth,
    for (uint32_t i03 = 0; i03 < ne03; ++i03) {
        for (uint32_t i02 = 0; i02 < ne02; ++i02) {
            for (uint32_t i = ir0; i < ir1; ++i) {
-                const uint32_t i12 = fastmodulo(i03, ne12, &octx->set_rows_div_ne12);
-                const uint32_t i11 = fastmodulo(i02, ne11, &octx->set_rows_div_ne11);
+                const uint32_t i12 = fastmodulo(i03, ne12, &srctx->div_ne12);
+                const uint32_t i11 = fastmodulo(i02, ne11, &srctx->div_ne11);
                const uint32_t i10 = i;

                const uintptr_t src1_addr = octx->src1.data + i10*nb10 + i11*nb11 + i12*nb12;
@@ -76,15 +86,16 @@ static int set_rows_thread_f32_f32(struct htp_ops_context * octx, const int nth,
            }
        }
    }
-
-    return HTP_STATUS_OK;
 }

-static int set_rows_thread_f16_f32(struct htp_ops_context * octx, const int nth, const int ith) {
+static void set_rows_thread_f16_f32(unsigned int nth, unsigned int ith, void *data) {
+    struct htp_set_rows_context * srctx = (struct htp_set_rows_context *)data;
+    struct htp_ops_context * octx = srctx->octx;
+
    set_rows_preamble;

    // parallelize by rows of src0
-    const uint32_t dr  = octx->src0_nrows_per_thread;
+    const uint32_t dr  = srctx->src0_nrows_per_thread;
    const uint32_t ir0 = dr * ith;
    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;

@@ -93,8 +104,8 @@ static int set_rows_thread_f16_f32(struct htp_ops_context * octx, const int nth,
    for (uint32_t i03 = 0; i03 < ne03; ++i03) {
        for (uint32_t i02 = 0; i02 < ne02; ++i02) {
            for (uint32_t i = ir0; i < ir1; ++i) {
-                const uint32_t i12 = fastmodulo(i03, ne12, &octx->set_rows_div_ne12);
-                const uint32_t i11 = fastmodulo(i02, ne11, &octx->set_rows_div_ne11);
+                const uint32_t i12 = fastmodulo(i03, ne12, &srctx->div_ne12);
+                const uint32_t i11 = fastmodulo(i02, ne11, &srctx->div_ne11);
                const uint32_t i10 = i;

                const uintptr_t src1_addr = octx->src1.data + i10*nb10 + i11*nb11 + i12*nb12;
@@ -112,16 +123,6 @@ static int set_rows_thread_f16_f32(struct htp_ops_context * octx, const int nth,
            }
        }
    }
-
-    return HTP_STATUS_OK;
-}
-
-static void set_rows_work_f16_f32(unsigned int n, unsigned int i, void *data) {
-    set_rows_thread_f16_f32((struct htp_ops_context *) data, n, i);
-}
-
-static void set_rows_work_f32_f32(unsigned int n, unsigned int i, void *data) {
-    set_rows_thread_f32_f32((struct htp_ops_context *) data, n, i);
 }

 int op_set_rows(struct htp_ops_context * octx) {
@@ -143,18 +144,20 @@ int op_set_rows(struct htp_ops_context * octx) {
        return HTP_STATUS_OK;
    }

-    octx->set_rows_div_ne12 = init_fastdiv_values(ne12);
-    octx->set_rows_div_ne11 = init_fastdiv_values(ne11);
+    struct htp_set_rows_context srctx;
+    srctx.octx = octx;
+    srctx.div_ne12 = init_fastdiv_values(ne12);
+    srctx.div_ne11 = init_fastdiv_values(ne11);

    const uint32_t n_jobs = MIN(nr, octx->n_threads);
-    octx->src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+    srctx.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;

    switch(octx->dst.type) {
    case HTP_TYPE_F32:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_work_f32_f32, octx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_jobs);
        break;
    case HTP_TYPE_F16:
-        worker_pool_run_func(octx->ctx->worker_pool, set_rows_work_f16_f32, octx, n_jobs);
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_jobs);
        break;
    default:
        return HTP_STATUS_NO_SUPPORT;
@@ -10,6 +10,7 @@

 #include "hex-dma.h"
 #include "hvx-utils.h"
+#include "hex-fastdiv.h"

 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
@@ -48,7 +49,7 @@
    const uint32_t nb2 = dst->nb[2];                       \
    const uint32_t nb3 = dst->nb[3];

-struct softmax_th_ctx {
+struct htp_softmax_context {
    bool     use_f16;
    bool     use_src1;
    uint32_t n_head;
@@ -59,28 +60,48 @@ struct softmax_th_ctx {
    float m0;
    float m1;

+    uint32_t src0_nrows_per_thread;
+    struct fastdiv_values fastdiv_ne01;
+    struct fastdiv_values fastdiv_ne02;
+    struct fastdiv_values fastdiv_ne12; // For mask broadcasting
+    struct fastdiv_values fastdiv_ne13; // For mask broadcasting
+    size_t spad_stride;
+
    struct htp_ops_context * octx;
 };

-static void init_softmax_ctx(struct softmax_th_ctx * softmax_ctx, struct htp_ops_context * octx) {
+static void init_softmax_ctx(struct htp_softmax_context * smctx, struct htp_ops_context * octx) {
    const struct htp_tensor * src0 = &octx->src0;
    const struct htp_tensor * src1 = &octx->src1;

-    memset(softmax_ctx, 0, sizeof(struct softmax_th_ctx));
+    memset(smctx, 0, sizeof(struct htp_softmax_context));

-    memcpy(&softmax_ctx->scale, (float *) octx->op_params, sizeof(float));
-    memcpy(&softmax_ctx->max_bias, (float *) octx->op_params + 1, sizeof(float));
+    memcpy(&smctx->scale, (float *) octx->op_params, sizeof(float));
+    memcpy(&smctx->max_bias, (float *) octx->op_params + 1, sizeof(float));

-    softmax_ctx->n_head      = src0->ne[2];
-    softmax_ctx->n_head_log2 = 1u << (uint32_t) floor(log2(softmax_ctx->n_head));
+    smctx->n_head      = src0->ne[2];
+    smctx->n_head_log2 = 1u << (uint32_t) floor(log2(smctx->n_head));

-    softmax_ctx->m0 = powf(2.0f, -(softmax_ctx->max_bias) / softmax_ctx->n_head_log2);
-    softmax_ctx->m1 = powf(2.0f, -(softmax_ctx->max_bias / 2.0f) / softmax_ctx->n_head_log2);
+    smctx->m0 = powf(2.0f, -(smctx->max_bias) / smctx->n_head_log2);
+    smctx->m1 = powf(2.0f, -(smctx->max_bias / 2.0f) / smctx->n_head_log2);

-    softmax_ctx->use_src1 = (src1->ne[0] != 0);
-    softmax_ctx->use_f16  = (src1->ne[0] != 0) && (src1->type == HTP_TYPE_F16);
+    smctx->use_src1 = (src1->ne[0] != 0);
+    smctx->use_f16  = (src1->ne[0] != 0) && (src1->type == HTP_TYPE_F16);

-    softmax_ctx->octx = octx;
+    smctx->octx = octx;
+
+    // Initialize fastdiv values
+    const uint32_t ne01 = src0->ne[1];
+    const uint32_t ne02 = src0->ne[2];
+
+    if (ne01 > 0) smctx->fastdiv_ne01 = init_fastdiv_values(ne01);
+    if (ne02 > 0) smctx->fastdiv_ne02 = init_fastdiv_values(ne02);
+
+    const uint32_t ne12 = (src1->ne[0]) ? src1->ne[2] : 1;
+    const uint32_t ne13 = (src1->ne[0]) ? src1->ne[3] : 1;
+
+    if (ne12 > 0) smctx->fastdiv_ne12 = init_fastdiv_values(ne12);
+    if (ne13 > 0) smctx->fastdiv_ne13 = init_fastdiv_values(ne13);
 }

 static void hvx_fast_softmax_prep_f32(const uint8_t * restrict src,
@@ -139,8 +160,7 @@ static void hvx_fast_softmax_f32(const uint8_t * restrict src,
        max_vec       = Q6_Vsf_vmax_VsfVsf(max_vec, v1);
    }

-    HVX_Vector v = hvx_vec_reduce_max_f32(max_vec);
-    max_vec      = hvx_vec_repl4(v);
+    max_vec = hvx_vec_reduce_max_f32(max_vec); // replicated over all lanes

    #pragma unroll(4)
    for (int i = 0; i < step_of_1; i++) {
@@ -154,8 +174,7 @@ static void hvx_fast_softmax_f32(const uint8_t * restrict src,
        v_pad[i] = v3;
    }

-    v       = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_vec));
-    sum_vec = hvx_vec_repl4(v);
+    sum_vec = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_vec)); // replicated over all lanes

    HVX_VectorPred pos_sum   = Q6_Q_vcmp_gt_VwVw(sum_vec, zero_v);
    HVX_Vector     v4        = hvx_vec_inverse_f32(sum_vec);
@@ -183,83 +202,9 @@ static float hvx_softmax_f32(const uint8_t * restrict src,
    return sum;
 }

-static void softmax_htp_f32(int nth, int ith, struct softmax_th_ctx * softmax_ctx, int opt_path) {
-    struct htp_ops_context * octx = softmax_ctx->octx;
-
-    const struct htp_tensor * src0 = &octx->src0;
-    const struct htp_tensor * src1 = &octx->src1;
-    const struct htp_tensor * dst  = &octx->dst;
-
-    htp_softmax_preamble3;
-
-    uint8_t * src0_spad_data = octx->src0_spad.data + (ith * nb01);
-    uint8_t * src1_spad_data = octx->src1_spad.data + (ith * nb01);
-    uint8_t * dst_spad_data  = octx->dst_spad.data + (ith * nb1);
-
-    float * wp0 = (float *) src0_spad_data;
-    float * wp1 = (float *) src1_spad_data;
-    float * wp2 = (float *) dst_spad_data;
-
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            for (uint32_t i01 = ith; i01 < ne01; i01 += nth) {
-                const uint32_t i11 = i01;
-                const uint32_t i12 = i02 % ne12;
-                const uint32_t i13 = i03 % ne13;
-
-                // ALiBi
-                const uint32_t h = i02;  // head
-
-                const float slope = (softmax_ctx->max_bias > 0.0f) ?
-                                        h < softmax_ctx->n_head_log2 ?
-                                        powf(softmax_ctx->m0, h + 1) :
-                                        powf(softmax_ctx->m1, 2 * (h - softmax_ctx->n_head_log2) + 1) :
-                                        1.0f;
-
-                float * sp = (float *) ((char *) octx->src0.data + i01 * nb01 + i02 * nb02 + i03 * nb03);
-                float * dp = (float *) ((char *) octx->dst.data + i01 * nb1 + i02 * nb2 + i03 * nb3);
-
-                // broadcast the mask across rows
-                __fp16 * mp_f16 = (softmax_ctx->use_src1) ?
-                                      (__fp16 *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
-                                      NULL;
-                float *  mp_f32 = (softmax_ctx->use_src1) ?
-                                      (float *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
-                                      NULL;
-
-                if ((1 == opt_path) && (mp_f32) && !(softmax_ctx->use_f16)) {
-                    hvx_fast_softmax_prep_f32((const uint8_t *) sp, (uint8_t *) wp0, ne00, softmax_ctx->scale,
-                                              (const uint8_t *) mp_f32, slope);
-                } else {
-                    hvx_scale_f32((uint8_t *) wp0, (const uint8_t *) sp, ne00, softmax_ctx->scale);
-                    if (mp_f32) {
-                        if (softmax_ctx->use_f16) {
-                            for (int i = 0; i < ne00; ++i) {
-                                wp0[i] += slope * (float) mp_f16[i];
-                            }
-                        } else {
-                            for (int i = 0; i < ne00; ++i) {
-                                wp0[i] += slope * mp_f32[i];
-                            }
-                        }
-                    }
-                }
-
-                if (1 == opt_path) {
-                    hvx_fast_softmax_f32((const uint8_t *) wp0, (uint8_t *) dp, (uint8_t *) wp1, ne00);
-                } else {
-                    float max = hvx_reduce_max_f32((const uint8_t *) wp0, ne00);
-                    float sum = hvx_softmax_f32((const uint8_t *) wp0, (uint8_t *) wp2, (uint8_t *) wp1, ne00, max);
-                    sum       = sum > 0.0 ? (1.0 / sum) : 1;
-                    hvx_scale_f32((uint8_t *) dp, (const uint8_t *) wp2, ne00, sum);
-                }
-            }
-        }
-    }
-}
-
-static void softmax_job_f32_per_thread(struct softmax_th_ctx * softmax_ctx, int nth, int ith) {
-    struct htp_ops_context * octx = softmax_ctx->octx;
+static void softmax_job_f32(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_softmax_context * smctx = (struct htp_softmax_context *) data;
+    struct htp_ops_context * octx = smctx->octx;

    const struct htp_tensor * src0 = &octx->src0;
    const struct htp_tensor * src1 = &octx->src1;
@@ -268,7 +213,7 @@ static void softmax_job_f32_per_thread(struct softmax_th_ctx * softmax_ctx, int
    htp_softmax_preamble3;

    const uint32_t src0_nrows            = ne01 * ne02 * ne03;  // src0 rows
-    const uint32_t src0_nrows_per_thread = octx->src0_nrows_per_thread;
+    const uint32_t src0_nrows_per_thread = smctx->src0_nrows_per_thread;

    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
@@ -291,20 +236,103 @@ static void softmax_job_f32_per_thread(struct softmax_th_ctx * softmax_ctx, int
        opt_path = 1;
    }

-    softmax_htp_f32(nth, ith, softmax_ctx, opt_path);
+    uint8_t * src0_spad_data = octx->src0_spad.data + (ith * smctx->spad_stride);
+    uint8_t * src1_spad_data = octx->src1_spad.data + (ith * smctx->spad_stride);
+    uint8_t * dst_spad_data  = octx->dst_spad.data + (ith * smctx->spad_stride);
+
+    float * wp0 = (float *) src0_spad_data;
+    float * wp1 = (float *) src1_spad_data;
+    float * wp2 = (float *) dst_spad_data;
+
+    uint32_t prev_i2 = (uint32_t)-1;
+    float slope = 1.0f;
+
+    for (uint32_t r = src0_start_row; r < src0_end_row; ++r) {
+        uint32_t i1 = fastmodulo(r, ne01, &smctx->fastdiv_ne01);
+        uint32_t r_div_ne01 = fastdiv(r, &smctx->fastdiv_ne01);
+        uint32_t i2 = fastmodulo(r_div_ne01, ne02, &smctx->fastdiv_ne02);
+        uint32_t i3 = fastdiv(r_div_ne01, &smctx->fastdiv_ne02);
+
+        // Map to original logic indices
+        // i01 = i1
+        // i02 = i2
+        // i03 = i3
+
+        const uint32_t i11 = i1;
+        // const uint32_t i12 = i2 % ne12;
+        // const uint32_t i13 = i3 % ne13;
+
+        uint32_t i12, i13;
+        if (ne12 == ne02) {
+             i12 = i2;
+        } else {
+             i12 = fastmodulo(i2, ne12, &smctx->fastdiv_ne12);
+        }
+
+        if (ne13 == ne03) {
+             i13 = i3;
+        } else {
+             i13 = fastmodulo(i3, ne13, &smctx->fastdiv_ne13);
+        }
+
+        // ALiBi
+        if (i2 != prev_i2) {
+            const uint32_t h = i2;  // head
+
+            slope = (smctx->max_bias > 0.0f) ?
+                        h < smctx->n_head_log2 ?
+                        powf(smctx->m0, h + 1) :
+                        powf(smctx->m1, 2 * (h - smctx->n_head_log2) + 1) :
+                        1.0f;
+            prev_i2 = i2;
+        }
+
+        float * sp = (float *) ((char *) octx->src0.data + i1 * nb01 + i2 * nb02 + i3 * nb03);
+        float * dp = (float *) ((char *) octx->dst.data + i1 * nb1 + i2 * nb2 + i3 * nb3);
+
+        // broadcast the mask across rows
+        __fp16 * mp_f16 = (smctx->use_src1) ?
+                              (__fp16 *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
+                              NULL;
+        float *  mp_f32 = (smctx->use_src1) ?
+                              (float *) ((char *) octx->src1.data + i11 * nb11 + i12 * nb12 + i13 * nb13) :
+                              NULL;
+
+        if ((1 == opt_path) && (mp_f32) && !(smctx->use_f16)) {
+            hvx_fast_softmax_prep_f32((const uint8_t *) sp, (uint8_t *) wp0, ne00, smctx->scale,
+                                      (const uint8_t *) mp_f32, slope);
+        } else {
+            hvx_scale_f32((uint8_t *) wp0, (const uint8_t *) sp, ne00, smctx->scale);
+            if (mp_f32) {
+                if (smctx->use_f16) {
+                    for (int i = 0; i < ne00; ++i) {
+                        wp0[i] += slope * (float) mp_f16[i];
+                    }
+                } else {
+                    for (int i = 0; i < ne00; ++i) {
+                        wp0[i] += slope * mp_f32[i];
+                    }
+                }
+            }
+        }
+
+        if (1 == opt_path) {
+            hvx_fast_softmax_f32((const uint8_t *) wp0, (uint8_t *) dp, (uint8_t *) wp1, ne00);
+        } else {
+            float max = hvx_reduce_max_f32((const uint8_t *) wp0, ne00);
+            float sum = hvx_softmax_f32((const uint8_t *) wp0, (uint8_t *) wp2, (uint8_t *) wp1, ne00, max);
+            sum       = sum > 0.0 ? (1.0 / sum) : 1;
+            hvx_scale_f32((uint8_t *) dp, (const uint8_t *) wp2, ne00, sum);
+        }
+    }

    t2 = HAP_perf_get_qtimer_count();

    FARF(HIGH, "softmax-f32 %d/%d/%d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
-         softmax_ctx->use_f16, opt_path, ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13,
+         smctx->use_f16, opt_path, ne00, ne01, ne02, ne03, src0_start_row, src0_end_row, ne10, ne11, ne12, ne13,
         ne0, ne1, ne2, ne3, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void softmax_job_dispatcher_f32(unsigned int n, unsigned int i, void * p_data) {
-    struct softmax_th_ctx * p_softmax_ctx = (struct softmax_th_ctx *) p_data;
-    softmax_job_f32_per_thread(p_softmax_ctx, n, i);
-}
-
 static int execute_op_softmax_f32(struct htp_ops_context * octx) {
    int err = HTP_STATUS_OK;

@@ -312,17 +340,12 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
    const struct htp_tensor * src1 = &octx->src1;
    struct htp_tensor *       dst  = &octx->dst;

-    worker_callback_t op_func;
-    const char *      op_type = NULL;
-
-    struct softmax_th_ctx softmax_ctx;
+    struct htp_softmax_context smctx;
+    const char * op_type = "softmax-f32";

    switch (octx->op) {
        case HTP_OP_SOFTMAX:
-            op_func = softmax_job_dispatcher_f32;
-            op_type = "softmax-f32";
-
-            init_softmax_ctx(&softmax_ctx, octx);
+            init_softmax_ctx(&smctx, octx);
            break;

        default:
@@ -342,6 +365,9 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {
    octx->src0_spad.size = hex_round_up(src0_row_size, 128) * n_threads;
    octx->src1_spad.size = hex_round_up(src1_row_size, 128) * n_threads;

+    // Use stride for calculating offset
+    smctx.spad_stride = hex_round_up(src0_row_size, 128);
+
    size_t spad_size = octx->src0_spad.size + octx->src1_spad.size + octx->dst_spad.size;

    if (src1->ne[0]) {
@@ -371,8 +397,8 @@ static int execute_op_softmax_f32(struct htp_ops_context * octx) {

    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
        uint32_t n_jobs             = MIN(n_threads, src0_nrows);
-        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
-        worker_pool_run_func(octx->ctx->worker_pool, op_func, &softmax_ctx, n_jobs);
+        smctx.src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        worker_pool_run_func(octx->ctx->worker_pool, softmax_job_f32, &smctx, n_jobs);
    }

    return err;
@@ -17,7 +17,6 @@
 #include "htp-msg.h"
 #include "htp-ops.h"

-
 #define sum_rows_preamble                       \
    struct htp_tensor *src0 =  &octx->src0;\
    struct htp_tensor *dst  = &octx->dst;  \
@@ -42,53 +41,54 @@
    const uint32_t  nb2 = dst->nb[2];      \
    const uint32_t  nb3 = dst->nb[3];      \

-static int sum_rows_thread_f32(struct htp_ops_context * octx, const int nth, const int ith) {
-    sum_rows_preamble;
+struct sum_rows_context {
+    const uint8_t * src_data;
+    uint8_t       * dst_data;
+    uint32_t        ne00;
+    size_t          src_stride;
+    size_t          dst_stride;
+    uint32_t        rows_per_thread;
+    uint32_t        total_rows;
+    bool            opt_path;
+};

-    const uint32_t src0_nrows_per_thread  = octx->src0_nrows_per_thread;
-    const size_t src0_row_size = nb01;
-    const size_t dst_row_size  = nb1;
+static void sum_rows_thread_f32(unsigned int nth, unsigned int ith, void *data) {
+    const struct sum_rows_context * smctx = (const struct sum_rows_context *) data;

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    const uint32_t rows_per_thread = smctx->rows_per_thread;
+    const uint32_t total_rows      = smctx->total_rows;

-    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
-    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);
+    const uint32_t start_row = rows_per_thread * ith;
+    const uint32_t end_row   = MIN(start_row + rows_per_thread, total_rows);

-    // no work for this thread
-    if (src0_start_row >= src0_end_row) {
-        return HTP_STATUS_OK;
+    if (start_row >= end_row) {
+        return;
    }

-    int opt_path   = 0;
-    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
-        opt_path = 1;
-    }
+    const size_t   src_stride = smctx->src_stride;
+    const size_t   dst_stride = smctx->dst_stride;
+    const uint32_t ne00       = smctx->ne00;
+    const bool     opt_path   = smctx->opt_path;

-    const uint8_t * restrict data_src = (const uint8_t *) src0->data;
-    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+    const float * restrict src_th = (const float *) (smctx->src_data + (start_row * src_stride));
+    float       * restrict dst_th = (float *)       (smctx->dst_data + (start_row * dst_stride));

-    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
-    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
+    // Calculate actual number of rows for this thread
+    const uint32_t n_rows = end_row - start_row;

-    for (uint32_t ir = 0; ir < src0_nrows_per_thread; ir++) {
-        const float * restrict src_local = src_th + (ir * ne00);
+    for (uint32_t ir = 0; ir < n_rows; ir++) {
+        const float * restrict src_local = src_th + (ir * (src_stride / sizeof(float)));

-        if (ir + 1 < src0_nrows_per_thread) {
-            hex_l2fetch(src_local + ne00, src0_row_size, src0_row_size, 1);
+        if (ir + 1 < n_rows) {
+            hex_l2fetch(src_local + (src_stride / sizeof(float)), src_stride, src_stride, 1);
        }

-        if (1 == opt_path) {
+        if (opt_path) {
            dst_th[ir] = hvx_reduce_sum_f32_a((const uint8_t *) src_local, ne00);
        } else {
            dst_th[ir] = hvx_reduce_sum_f32((const uint8_t *) src_local, ne00);
        }
    }
-
-    return HTP_STATUS_OK;
-}
-
-static void sum_rows_work_f32(unsigned int n, unsigned int i, void *data) {
-    sum_rows_thread_f32((struct htp_ops_context *) data, n, i);
 }

 int op_sum_rows(struct htp_ops_context * octx) {
@@ -106,10 +106,25 @@ int op_sum_rows(struct htp_ops_context * octx) {
    const uint32_t src0_nrows = ne01 * ne02 * ne03;

    uint32_t n_jobs = MIN(n_threads, src0_nrows);
-    octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+    uint32_t rows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;

-    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_work_f32, octx, n_jobs);
+    bool opt_path = false;
+    if ((0 == hex_is_aligned((void *) src0->data, VLEN)) && !(nb01 & (VLEN - 1))) {
+        opt_path = true;
+    }
+
+    struct sum_rows_context smctx = {
+        .src_data        = (const uint8_t *) src0->data,
+        .dst_data        = (uint8_t *) dst->data,
+        .ne00            = ne00,
+        .src_stride      = nb01,
+        .dst_stride      = nb1,
+        .rows_per_thread = rows_per_thread,
+        .total_rows      = src0_nrows,
+        .opt_path        = opt_path,
+    };
+
+    worker_pool_run_func(octx->ctx->worker_pool, sum_rows_thread_f32, &smctx, n_jobs);

    return HTP_STATUS_OK;
 }
-
@@ -17,6 +17,28 @@
 #include "htp-msg.h"
 #include "htp-ops.h"

+struct htp_unary_context {
+    struct htp_ops_context * octx;
+
+    // Precomputed values
+    const uint8_t *           data_src0;
+    uint8_t *                 data_dst;
+
+    size_t                    src0_row_size;
+    size_t                    dst_row_size;
+
+    size_t                    src0_row_size_aligned;
+    size_t                    dst_row_size_aligned;
+
+    size_t                    src0_spad_half_size;
+    size_t                    dst_spad_half_size;
+
+    uint32_t                  block;
+    uint32_t                  src0_nrows;
+    uint32_t                  src0_nrows_per_thread;
+    uint32_t                  nc;
+};
+
 #define htp_unary_preamble            \
    const uint32_t ne00 = src->ne[0]; \
    const uint32_t ne01 = src->ne[1]; \
@@ -57,8 +79,7 @@ static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
    }

-    HVX_Vector reduced_sum = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
-    sum_v                  = hvx_vec_repl4(reduced_sum);
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v)); // replicated over all lanes

    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
@@ -75,128 +96,95 @@ static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
    }
 }

-static void scale_htp_f32(const float * restrict src,
-                          float * restrict dst,
-                          uint8_t * restrict spad,
-                          const uint32_t num_rows,
-                          const uint32_t row_elems,
-                          const size_t   row_size,
-                          int32_t *      op_params,
-                          int            opt_path) {
+static void scale_f32(const float * restrict src,
+                      float * restrict dst,
+                      uint8_t * restrict spad,
+                      const uint32_t num_rows,
+                      const uint32_t row_elems,
+                      const size_t   row_size,
+                      int32_t *      op_params) {
    float scale = 0.f;
    float bias  = 0.f;
    memcpy(&scale, &op_params[0], sizeof(float));
    memcpy(&bias,  &op_params[1], sizeof(float));

    for (uint32_t ir = 0; ir < num_rows; ir++) {
-        const float * restrict src_local = src + (ir * row_elems);
-        float * restrict dst_local       = dst + (ir * row_elems);
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);

-        if (ir + 1 < num_rows) {
-            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
-        }
-
-        hvx_scale_offset_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems, scale, bias);
+        hvx_scale_offset_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems, scale, bias);
    }
 }

-static void rms_norm_htp_f32(const float * restrict src,
-                             float * restrict dst,
-                             uint8_t * restrict spad,
-                             const uint32_t num_rows,
-                             const uint32_t row_elems,
-                             const size_t   row_size,
-                             int32_t *      op_params,
-                             int            opt_path) {
+static void rms_norm_f32(const float * restrict src,
+                         float * restrict dst,
+                         uint8_t * restrict spad,
+                         const uint32_t num_rows,
+                         const uint32_t row_elems,
+                         const size_t   row_size,
+                         int32_t *      op_params) {
    float epsilon = 0.f;
    memcpy(&epsilon, op_params, sizeof(float));

    for (uint32_t ir = 0; ir < num_rows; ir++) {
-        const float * restrict src_local = src + (ir * row_elems);
-        float * restrict dst_local       = dst + (ir * row_elems);
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);

-        if (ir + 1 < num_rows) {
-            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
-        }
-
-        if (1 == opt_path) {
-            hvx_fast_rms_norm_f32((const uint8_t *) src_local, (uint8_t *) dst_local, spad, row_elems, epsilon);
-        } else {
-            float sum = hvx_sum_of_squares_f32((const uint8_t *) src_local, row_elems);
-
-            const float mean  = sum / row_elems;
-            const float scale = 1.0f / sqrtf(mean + epsilon);
-
-            hvx_scale_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems, scale);
-        }
+        hvx_fast_rms_norm_f32((const uint8_t *) src_local, (uint8_t *) dst_local, spad, row_elems, epsilon);
    }
 }

-static void sqr_htp_f32(const float * restrict src,
-                          float * restrict dst,
-                          uint8_t * restrict spad,
-                          const uint32_t num_rows,
-                          const uint32_t row_elems,
-                          const size_t   row_size,
-                          int32_t *      op_params,
-                          int            opt_path) {
+static void sqr_f32(const float * restrict src,
+                    float * restrict dst,
+                    uint8_t * restrict spad,
+                    const uint32_t num_rows,
+                    const uint32_t row_elems,
+                    const size_t   row_size,
+                    int32_t *      op_params) {

    for (uint32_t ir = 0; ir < num_rows; ir++) {
-        const float * restrict src_local = src + (ir * row_elems);
-        float * restrict dst_local       = dst + (ir * row_elems);
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);

-        if (ir + 1 < num_rows) {
-            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
-        }
-
-        if (1 == opt_path) {
-            hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
-        } else {
-            hvx_sqr_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
-        }
+        hvx_sqr_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
    }
 }

-static void sqrt_htp_f32(const float * restrict src,
-                          float * restrict dst,
-                          uint8_t * restrict spad,
-                          const uint32_t num_rows,
-                          const uint32_t row_elems,
-                          const size_t   row_size,
-                          int32_t *      op_params,
-                          int            opt_path) {
+static void sqrt_f32(const float * restrict src,
+                     float * restrict dst,
+                     uint8_t * restrict spad,
+                     const uint32_t num_rows,
+                     const uint32_t row_elems,
+                     const size_t   row_size,
+                     int32_t *      op_params) {

    for (uint32_t ir = 0; ir < num_rows; ir++) {
-        const float * restrict src_local = src + (ir * row_elems);
-        float * restrict dst_local       = dst + (ir * row_elems);
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * row_size);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * row_size);

-        if (ir + 1 < num_rows) {
-            hex_l2fetch(src_local + row_elems, row_size, row_size, 1);
-        }
-
-        if (1 == opt_path) {
-            hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
-        } else {
-            hvx_sqrt_f32((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
-        }
+        hvx_sqrt_f32_aa((uint8_t *) dst_local, (const uint8_t *) src_local, row_elems);
    }
 }

-static void unary_job_f32_per_thread(const struct htp_tensor * src,
-                                     struct htp_tensor *       dst,
-                                     uint8_t *                 spad,
-                                     int                       htp_op,
-                                     int32_t *                 op_params,
-                                     uint32_t                  nth,
-                                     uint32_t                  ith,
-                                     uint32_t                  src0_nrows_per_thread) {
+static void unary_job_f32_per_thread(unsigned int nth, unsigned int ith, void * data) {
+    const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;
+    struct htp_ops_context * octx = uctx->octx;
+    const struct htp_tensor * src = &octx->src0;
+    const struct htp_tensor * dst = &octx->dst;
+
    htp_unary_preamble;

-    const size_t src0_row_size = nb01;
-    const size_t dst_row_size  = nb1;
+    int                       htp_op = octx->op;
+    int32_t *                 op_params = octx->op_params;
+    uint32_t                  src0_nrows_per_thread = uctx->src0_nrows_per_thread;

-    const uint32_t src0_nrows = ne01 * ne02 * ne03;  // src0 rows
+    const size_t src0_row_size = uctx->src0_row_size;
+    const size_t dst_row_size  = uctx->dst_row_size;

+    const size_t src0_row_size_aligned = uctx->src0_row_size_aligned;
+    const size_t dst_row_size_aligned  = uctx->dst_row_size_aligned;
+
+    const uint32_t src0_nrows = uctx->src0_nrows;
    const uint32_t src0_start_row = src0_nrows_per_thread * ith;
    const uint32_t src0_end_row   = MIN(src0_start_row + src0_nrows_per_thread, src0_nrows);

@@ -208,79 +196,104 @@ static void unary_job_f32_per_thread(const struct htp_tensor * src,
    uint64_t t1, t2;
    t1 = HAP_perf_get_qtimer_count();

-    int is_aligned = 1;
-    int opt_path   = 0;
-    if ((0 == hex_is_aligned((void *) src->data, VLEN)) || (0 == hex_is_aligned((void *) dst->data, VLEN))) {
-        is_aligned = 0;
-    }
-    if ((1 == is_aligned) && !(nb01 & (VLEN - 1))) {
-        opt_path = 1;
+    const uint8_t * restrict data_src = uctx->data_src0;
+    uint8_t * restrict       data_dst = uctx->data_dst;
+
+    uint8_t * src0_spad_data = octx->src0_spad.data + (ith * octx->src0_spad.size_per_thread);
+    uint8_t * dst_spad_data  = octx->dst_spad.data  + (ith * octx->dst_spad.size_per_thread);
+
+    size_t src0_spad_half_size = uctx->src0_spad_half_size;
+    size_t dst_spad_half_size  = uctx->dst_spad_half_size;
+
+    const int BLOCK = uctx->block;
+    if (BLOCK == 0) {
+        FARF(ERROR, "unary-f32 : current VTCM reservation %zu is too small for even 1 row per thread, needed at least %zu\n",
+             octx->src0_spad.size_per_thread, src0_row_size_aligned);
+        return;
    }

-    const uint8_t * restrict data_src = (const uint8_t *) src->data;
-    uint8_t * restrict data_dst       = (uint8_t *) dst->data;
+    dma_queue * dma_queue = octx->ctx->dma[ith];

-    const float * restrict src_th = (float *) (data_src + (src0_start_row * src0_row_size));
-    float * restrict dst_th       = (float *) (data_dst + (src0_start_row * dst_row_size));
-    uint8_t * restrict spad_th    = (uint8_t *) spad + (ith * nb01);
+    for (uint32_t ir = src0_start_row, spad_idx = 0; ir < src0_end_row && spad_idx < 2; ir += BLOCK, spad_idx++) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);

-    switch (htp_op) {
-        case HTP_OP_RMS_NORM:
-            rms_norm_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
-            break;
-        case HTP_OP_SCALE:
-            scale_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
-            break;
-        case HTP_OP_SQR:
-            sqr_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
-            break;
-        case HTP_OP_SQRT:
-            sqrt_htp_f32(src_th, dst_th, spad_th, src0_end_row - src0_start_row, ne0, nb1, op_params, opt_path);
-            break;
+        // Dummy DMA transation for sequencing (interleaving dst,src,dst,...)
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst, dst_spad_data + (spad_idx * dst_spad_half_size)),
+            dst_row_size, dst_row_size_aligned, 0);

-        default:
-            break;
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+            dma_make_ptr(src0_spad_data + (spad_idx * src0_spad_half_size), data_src + (ir * src0_row_size)),
+            src0_row_size_aligned, src0_row_size, block_size);
    }

+    for (uint32_t ir = src0_start_row; ir < src0_end_row; ir += BLOCK) {
+        const uint32_t block_size = MIN(BLOCK, src0_end_row - ir);
+
+        float * dst_spad  = (float *) dma_queue_pop(dma_queue).src;
+        float * src0_spad = (float *) dma_queue_pop(dma_queue).dst;
+
+        // Process block in VTCM
+        switch (htp_op) {
+            case HTP_OP_RMS_NORM:
+                rms_norm_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SCALE:
+                scale_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SQR:
+                sqr_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            case HTP_OP_SQRT:
+                sqrt_f32(src0_spad, dst_spad, NULL, block_size, ne0, src0_row_size_aligned, op_params);
+                break;
+            default:
+                break;
+        }
+
+        dma_queue_push_vtcm_to_ddr(dma_queue,
+            dma_make_ptr(data_dst + (ir * dst_row_size), dst_spad),
+            dst_row_size, dst_row_size_aligned, block_size);
+
+        // prefetch N+2 loop iteration if any
+        const uint32_t pref_block = (ir + BLOCK * 2);
+        if (pref_block < src0_end_row) {
+            const uint32_t pref_block_size = MIN(BLOCK, src0_end_row - pref_block);
+            dma_queue_push_ddr_to_vtcm(dma_queue,
+                dma_make_ptr(src0_spad, data_src + (pref_block * src0_row_size)),
+                src0_row_size_aligned, src0_row_size, pref_block_size);
+        }
+    }
+
+    dma_queue_flush(dma_queue);
+
    t2 = HAP_perf_get_qtimer_count();

-    FARF(HIGH, "unary-f32 %d/%d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, opt_path, src->ne[0],
+    FARF(HIGH, "unary-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n", ith, nth, src->ne[0],
         src->ne[1], src->ne[2], src->ne[3], src0_start_row, src0_end_row, dst->ne[0], dst->ne[1], dst->ne[2],
         dst->ne[3], (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
 }

-static void unary_job_dispatcher_f32(unsigned int n, unsigned int i, void * data) {
-    struct htp_ops_context * octx = (struct htp_ops_context *) data;
-
-    unary_job_f32_per_thread(&octx->src0, &octx->dst, octx->src0_spad.data, octx->op, octx->op_params, n, i,
-                             octx->src0_nrows_per_thread);
-}
-
 static int execute_op_unary_f32(struct htp_ops_context * octx) {
    int err = HTP_STATUS_OK;

    const struct htp_tensor * src0 = &octx->src0;
    struct htp_tensor *       dst  = &octx->dst;

-    worker_callback_t unary_op_func;
-    const char *      op_type = NULL;
+    const char * op_type = NULL;

    switch (octx->op) {
        case HTP_OP_RMS_NORM:
-            unary_op_func = unary_job_dispatcher_f32;
-            op_type       = "rmsnorm-f32";
+            op_type = "rmsnorm-f32";
            break;
        case HTP_OP_SCALE:
-            unary_op_func = unary_job_dispatcher_f32;
-            op_type       = "scale-f32";
+            op_type = "scale-f32";
            break;
        case HTP_OP_SQR:
-            unary_op_func = unary_job_dispatcher_f32;
-            op_type       = "sqr-f32";
+            op_type = "sqr-f32";
            break;
        case HTP_OP_SQRT:
-            unary_op_func = unary_job_dispatcher_f32;
-            op_type       = "sqrt-f32";
+            op_type = "sqrt-f32";
            break;

        default:
@@ -294,32 +307,61 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
    const size_t src0_row_size = src0->nb[1];
    const size_t dst_row_size  = dst->nb[1];

-    // VTCM scratchpads for all tensors
-    octx->dst_spad.size  = hex_round_up(dst_row_size, 128) * n_threads;
-    octx->src0_spad.size = hex_round_up(src0_row_size, 128) * n_threads;
+    const size_t src0_row_size_aligned = hex_round_up(src0_row_size, VLEN);
+    const size_t dst_row_size_aligned  = hex_round_up(dst_row_size, VLEN);

-    size_t spad_size = octx->src0_spad.size + octx->dst_spad.size;
+    // VTCM scratchpads for all tensors
+    // N rows per thread, padded to HVX vector size
+    // Double buffering requires 2x size per buffer
+
+    size_t spad_size_per_row   = 2 * (src0_row_size_aligned + dst_row_size_aligned);
+    size_t vtcm_row_per_thread = (octx->ctx->vtcm_size)/ (n_threads * spad_size_per_row);
+
+    // Make sure the reserved vtcm size is sufficient
+    if (vtcm_row_per_thread == 0) {
+        FARF(ERROR, "unary-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
+             spad_size_per_row * n_threads);
+        return HTP_STATUS_VTCM_TOO_SMALL;
+    }
+
+    octx->src0_spad.size_per_thread = src0_row_size_aligned * vtcm_row_per_thread * 2;
+    octx->dst_spad.size_per_thread  = dst_row_size_aligned * vtcm_row_per_thread * 2;
+
+    octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
+    octx->dst_spad.size  = n_threads * octx->dst_spad.size_per_thread;
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;
+    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size;

    FARF(HIGH, "%s: (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-spad-size %u src1-spad-size %u dst-spad-size %u\n", op_type,
         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
         octx->src0_spad.size, octx->src1_spad.size, octx->dst_spad.size);

-    // Make sure the reserved vtcm size is sufficient
-    if (octx->ctx->vtcm_size < spad_size) {
-        FARF(ERROR, "unary-%s : current VTCM reservation %zu is too small, needed %zu\n", op_type, octx->ctx->vtcm_size,
-             spad_size);
-        return HTP_STATUS_VTCM_TOO_SMALL;
-    }
-
-    octx->src0_spad.data = octx->ctx->vtcm_base;
-    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size;
-
    if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
        uint32_t n_jobs = MIN(n_threads, src0_nrows);

-        octx->src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs;
+        struct htp_unary_context uctx = {
+            .octx                  = octx,
+            .src0_nrows_per_thread = (src0_nrows + n_jobs - 1) / n_jobs,
+            .src0_nrows            = src0_nrows,

-        worker_pool_run_func(octx->ctx->worker_pool, unary_op_func, octx, n_jobs);
+            .data_src0             = (const uint8_t *)src0->data,
+            .data_dst              = (uint8_t *)dst->data,
+
+            .src0_row_size         = src0_row_size,
+            .dst_row_size          = dst_row_size,
+
+            .src0_row_size_aligned = src0_row_size_aligned,
+            .dst_row_size_aligned  = dst_row_size_aligned,
+
+            .src0_spad_half_size   = octx->src0_spad.size_per_thread / 2,
+            .dst_spad_half_size    = octx->dst_spad.size_per_thread / 2,
+
+            .block                 = (octx->src0_spad.size_per_thread / 2) / src0_row_size_aligned,
+            .nc                    = src0->ne[0],
+        };
+
+        worker_pool_run_func(octx->ctx->worker_pool, unary_job_f32_per_thread, &uctx, n_jobs);
    }

    return err;
@@ -403,19 +403,20 @@ enum FaCodePath {
 };

 struct vk_fa_pipeline_state {
-    vk_fa_pipeline_state(uint32_t HSK, uint32_t HSV, bool small_rows, bool small_cache, FaCodePath path, bool aligned, bool f32acc, uint32_t flags)
-        : HSK(HSK), HSV(HSV), small_rows(small_rows), small_cache(small_cache), path(path), aligned(aligned), f32acc(f32acc), flags(flags) {}
-
    uint32_t HSK, HSV;
-    bool small_rows, small_cache;
+    uint32_t Br, Bc;
+    uint32_t D_split, row_split;
+    bool shmem_staging;
    FaCodePath path;
+    uint32_t workgroup_size, subgroup_size;
    bool aligned;
    bool f32acc;
    uint32_t flags;
+    uint32_t limit_occupancy_shmem;

    bool operator<(const vk_fa_pipeline_state &b) const {
-        return std::tie(HSK, HSV, small_rows, small_cache, path, aligned, f32acc, flags) <
-               std::tie(b.HSK, b.HSV, b.small_rows, b.small_cache, b.path, b.aligned, b.f32acc, b.flags);
+        return std::tie(HSK, HSV, Br, Bc, D_split, row_split, shmem_staging, path, workgroup_size, subgroup_size, aligned, f32acc, flags, limit_occupancy_shmem) <
+               std::tie(b.HSK, b.HSV, b.Br, b.Bc, b.D_split, b.row_split, b.shmem_staging, b.path, b.workgroup_size, b.subgroup_size, b.aligned, b.f32acc, b.flags, b.limit_occupancy_shmem);
    }
 };

@@ -623,6 +624,8 @@ struct vk_device_struct {
    // floor(log2(maxComputeWorkGroupInvocations))
    uint32_t max_workgroup_size_log2 {};

+    bool flash_attention_fp16;
+
    bool coopmat_support;
    bool coopmat_acc_f32_support {};
    bool coopmat_acc_f16_support {};
@@ -944,6 +947,7 @@ struct vk_mat_mat_push_constants {
    uint32_t M; uint32_t N; uint32_t K;
    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t base_work_group_z; uint32_t num_batches;
    uint32_t k_split;
    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
    uint32_t padded_N;
@@ -963,6 +967,7 @@ struct vk_mat_vec_push_constants {
    uint32_t batch_stride_b;
    uint32_t batch_stride_d;
    uint32_t fusion_flags;
+    uint32_t base_work_group_y;
    uint32_t ne02;
    uint32_t ne12;
    uint32_t broadcast2;
@@ -1654,6 +1659,7 @@ static bool vk_perf_logger_concurrent = false;
 static bool vk_enable_sync_logger = false;
 // number of calls between perf logger prints
 static uint32_t vk_perf_logger_frequency = 1;
+static std::string vk_pipeline_stats_filter;

 class vk_perf_logger {
  public:
@@ -2170,7 +2176,32 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
        executableInfo.pipeline = pipeline->pipeline;

        auto statistics = device->device.getPipelineExecutableStatisticsKHR(executableInfo);
+
+        bool print_stats = !vk_pipeline_stats_filter.empty() &&
+                           pipeline->name.find(vk_pipeline_stats_filter) != std::string::npos;
+        if (print_stats) {
+            std::cerr << "ggml_vulkan: pipeline stats for " << pipeline->name << ":" << std::endl;
+        }
+
        for (auto & s : statistics) {
+            if (print_stats) {
+                std::cerr << "ggml_vulkan:   " << s.name.data() << ": ";
+                switch (s.format) {
+                    case vk::PipelineExecutableStatisticFormatKHR::eBool32:
+                        std::cerr << (s.value.b32 ? "true" : "false");
+                        break;
+                    case vk::PipelineExecutableStatisticFormatKHR::eInt64:
+                        std::cerr << s.value.i64;
+                        break;
+                    case vk::PipelineExecutableStatisticFormatKHR::eUint64:
+                        std::cerr << s.value.u64;
+                        break;
+                    case vk::PipelineExecutableStatisticFormatKHR::eFloat64:
+                        std::cerr << s.value.f64;
+                        break;
+                }
+                std::cerr << std::endl;
+            }
            // "Register Count" is reported by NVIDIA drivers.
            if (strcmp(s.name, "Register Count") == 0) {
                VK_LOG_DEBUG(pipeline->name << " " << s.name << ": " << s.value.u64 << " registers");
@@ -2753,78 +2784,214 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
    );
 }

-// 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;
+struct vk_fa_tuning_params {
+    FaCodePath path;
+    uint32_t workgroup_size;
+    uint32_t subgroup_size;
+    uint32_t block_rows;
+    uint32_t block_cols;
+    uint32_t d_split;
+    uint32_t row_split;
+    bool shmem_staging;
+    bool disable_subgroups;
+    uint32_t limit_occupancy_shmem;

-static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv, bool small_cache) {
-    if (hsv >= 192) {
-        return 2;
-    } else if ((hsv | hsk) & 8 || small_cache) {
-        return 4;
-    } else {
-        return 8;
+    void print() const {
+        std::cerr << "path=" << path << " workgroup_size=" << workgroup_size << " subgroup_size=" << subgroup_size <<
+                     " block_rows=" << block_rows << " block_cols=" << block_cols << " d_split=" << d_split <<
+                     " row_split=" << row_split << " shmem_staging=" << shmem_staging << " disable_subgroups=" << disable_subgroups <<
+                     " limit_occupancy_shmem=" << limit_occupancy_shmem << std::endl;
    }
-}
+};

-// 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 bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc);
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc);

-static uint32_t get_fa_num_small_rows(FaCodePath path) {
-    if (path == FA_COOPMAT2) {
-        return flash_attention_num_small_rows;
+static vk_fa_tuning_params get_fa_tuning_params_scalar(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type kv_type, bool f32acc) {
+    GGML_UNUSED(kv_type);
+
+    vk_fa_tuning_params result{};
+    result.path = FA_SCALAR;
+
+    if (device->vendor_id == VK_VENDOR_ID_INTEL) {
+        // Disable subgroup use due to performance issues when enforcing subgroup sizes
+        result.subgroup_size = 32;
+        result.disable_subgroups = true;
+    } else if (device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != AMD_GCN) {
+        result.subgroup_size = n_rows < 4 ? 32 : device->subgroup_size;
    } else {
-        return scalar_flash_attention_num_small_rows;
+        result.subgroup_size = device->subgroup_size;
    }
-}

-static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) {
-    GGML_UNUSED(clamp);
+    // Row split splits the workgroup so that synchronization only has to happen within subgroups, which avoids barriers
+    uint32_t row_split_max_hsk = 64;
+    if (device->vendor_id == VK_VENDOR_ID_AMD && device->architecture != AMD_GCN && !device->uma) {
+        row_split_max_hsk = n_rows <= 8 ? 64 : 128;
+    }
+    result.row_split = (n_rows < 4 || hsk <= row_split_max_hsk) ? 1 : 4;

-    if (path == FA_SCALAR) {
-        if (small_rows) {
-            return {scalar_flash_attention_num_small_rows, 64};
+    if (result.subgroup_size > 32 && (n_rows < 4 || hsk < (result.row_split == 1 ? 128 : 64))) {
+        result.workgroup_size = result.subgroup_size * 2;
+    } else {
+        result.workgroup_size = result.subgroup_size * 4;
+    }
+
+    const uint32_t D = hsk | hsv;
+
+    const bool reduce_block_rows = D & 8 || n_kv < 1024 || device->vendor_id == VK_VENDOR_ID_INTEL;
+
+    if (n_rows == 1) {
+        result.block_rows = 1;
+        result.block_cols = 64;
+    } else {
+        // row_split 1 means higher register use per row, so block size has to be adjusted
+        if (result.row_split == 1) {
+            result.block_rows = n_rows == 2 ? 2 : ((n_rows <= 4 || reduce_block_rows) ? 4 : 8);
        } else {
-            if ((hsv | hsk) & 8) {
-                // HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter
-                // larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not.
-                return {get_fa_scalar_num_large_rows(hsk, hsv, small_cache), 64};
-            } else {
-                return {get_fa_scalar_num_large_rows(hsk, hsv, small_cache), 32};
-            }
+            result.block_rows = n_rows <= 4 ? 4 : ((n_rows <= 8 || reduce_block_rows) ? 8 : 16);
        }
+
+        result.block_cols = (D & 8) ? 64 : 32;
+    }
+
+    const uint32_t D_lsb = D ^ (D & (D-1));  // extract lowest set bit
+
+    result.d_split = std::min(std::min(result.subgroup_size, 8u), D_lsb / 4);
+
+    result.shmem_staging = (device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256 && hsv < 256) ? 1 : 0;
+
+    if (!reduce_block_rows && !ggml_vk_flash_attn_scalar_shmem_support(device, result, hsk, hsv, f32acc)) {
+        result.block_rows /= 2;
+    }
+
+    // On AMD RDNA, for small head sizes and big batch size the shader uses few registers, so too many subgroups get scheduled
+    // at once and end up thrashing the cache. Fix this by setting a large (unused) shmem buffer that reduces occupancy.
+    // This targets an occupancy of 4 subgroups per SIMD.
+    if (device->vendor_id == VK_VENDOR_ID_AMD && device->properties.limits.maxComputeSharedMemorySize == 65536) {
+        if (device->architecture != AMD_GCN && n_rows >= 64 && hsk <= 128) {
+            // 30kb target for hsk > 64, 26kb for <= 64 due to smaller workgroup size
+            // Values are guessed, tested on RDNA2
+            result.limit_occupancy_shmem = (hsk <= 64 ? 26 : 30) * 1024 / 4 / 4;
+        } else if (device->architecture == AMD_GCN && n_rows <= 8 && hsk >= 256) {
+            // Same thing for GCN, with an occupancy target of 2 subgroups per SIMD.
+            // Here low-batch FA with large head size is affected.
+            // n_rows < 4 switch because workgroup size switches from 128 to 256 there.
+            result.limit_occupancy_shmem = (n_rows < 4 ? 14 : 26) * 1024 / 4 / 4;
+        }
+    }
+
+    return result;
+}
+
+static vk_fa_tuning_params get_fa_tuning_params_coopmat1(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type kv_type, bool f32acc) {
+    GGML_UNUSED(n_rows);
+    GGML_UNUSED(n_kv);
+    GGML_UNUSED(kv_type);
+    GGML_UNUSED(f32acc);
+
+    vk_fa_tuning_params result{};
+    result.path = FA_COOPMAT1;
+
+    const uint32_t D = hsk | hsv;
+
+    const uint32_t coopmat_block_rows = 16;
+    const uint32_t coopmat_block_cols = 16;
+
+    const uint32_t num_subgroups = 4;
+
+    result.block_rows = coopmat_block_rows;
+    result.block_cols = coopmat_block_cols * num_subgroups;
+    result.row_split = num_subgroups;
+    result.subgroup_size = device->subgroup_size;
+    result.workgroup_size = num_subgroups * result.subgroup_size;
+
+    const uint32_t D_lsb = D ^ (D & (D-1));  // extract lowest set bit
+    result.d_split = std::min(std::min(result.subgroup_size, 8u), D_lsb / 4);
+
+    result.shmem_staging = (device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256 && hsv < 256) ? 1 : 0;
+
+    return result;
+}
+
+static vk_fa_tuning_params get_fa_tuning_params_coopmat2(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type kv_type, bool f32acc) {
+    GGML_UNUSED(n_kv);
+    GGML_UNUSED(f32acc);
+
+    vk_fa_tuning_params result{};
+    result.path = FA_COOPMAT2;
+
+    const uint32_t D = hsk | hsv;
+
+    const bool small_rows = n_rows < 32;
+
+    if (small_rows) {
+        result.block_rows = 32;
+        result.block_cols = 32;
+    } else if (ggml_is_quantized(kv_type) || hsk >= 256 || hsv >= 256) {
+        result.block_rows = (hsk >= 512 || hsv >= 512) ? 32 : 64;
+        result.block_cols = 32;
+    } else {
+        result.block_rows = 64;
+        result.block_cols = 64;
+    }
+
+    result.subgroup_size = device->subgroup_size;
+    result.workgroup_size = (small_rows && (D % 32) == 0) ? 256 : 128;
+
+    return result;
+}
+
+static vk_fa_tuning_params get_fa_tuning_params(const vk_device& device, uint32_t hsk, uint32_t hsv, uint32_t n_rows, uint32_t n_kv, ggml_type kv_type, bool f32acc) {
+    FaCodePath path = device->coopmat2 ? FA_COOPMAT2 :
+                      device->coopmat1_fa_support ? FA_COOPMAT1 : FA_SCALAR;
+
+    if (path == FA_COOPMAT1 && device->architecture == vk_device_architecture::NVIDIA_TURING) {
+        // Nvidia compiler bug, see https://github.com/ggml-org/llama.cpp/pull/19075#issuecomment-3820716090
+        path = FA_SCALAR;
    }

    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};
+        bool shape_ok = (f32acc && device->coopmat_support_16x16x16_f32acc) ||
+                        (!f32acc && device->coopmat_support_16x16x16_f16acc);
+        const vk_fa_tuning_params params = get_fa_tuning_params_coopmat1(device, hsk, hsv, n_rows, n_kv, kv_type, f32acc);
+        bool shmem_ok = ggml_vk_flash_attn_coopmat_shmem_support(device, params, hsk, hsv, f32acc);
+
+        if (!shape_ok || !shmem_ok) {
+            path = FA_SCALAR;
        }
    }

-    // small rows, large cols
-    if (small_rows) {
-        return {get_fa_num_small_rows(FA_COOPMAT2), 32};
+    // scalar is faster than coopmat when N==1
+    if (n_rows == 1 && (path == FA_COOPMAT1 || path == FA_COOPMAT2)) {
+        path = FA_SCALAR;
    }

-    // small cols to reduce register count
-    if (ggml_is_quantized(type) || hsk >= 256 || hsv >= 256) {
-        if (hsk >= 512 || hsv >= 512) {
-            return {32, 32};
-        } else {
-            return {64, 32};
-        }
+    switch (path) {
+    case FA_SCALAR:
+        return get_fa_tuning_params_scalar(device, hsk, hsv, n_rows, n_kv, kv_type, f32acc);
+    case FA_COOPMAT1:
+        return get_fa_tuning_params_coopmat1(device, hsk, hsv, n_rows, n_kv, kv_type, f32acc);
+    case FA_COOPMAT2:
+        return get_fa_tuning_params_coopmat2(device, hsk, hsv, n_rows, n_kv, kv_type, f32acc);
+    default:
+        throw std::runtime_error("unsupported FaCodePath");
    }
-    return {64, 64};
 }

-static uint32_t fa_align(FaCodePath path, uint32_t hsk, uint32_t hsv, ggml_type type, bool small_rows, bool small_cache) {
-    return fa_rows_cols(path, hsk, hsv, 0, type, small_rows, small_cache)[1];
+static vk_fa_pipeline_state get_fa_pipeline_state(const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool aligned, bool f32acc,
+                                                  bool use_mask, bool use_mask_opt, bool use_logit_softcap) {
+    uint32_t flags = (use_mask_opt      ? 1 : 0) |
+                     (use_mask          ? 2 : 0) |
+                     (use_logit_softcap ? 4 : 0);
+
+    const uint32_t subgroup_size = params.disable_subgroups ? 0 : params.subgroup_size;
+
+    return vk_fa_pipeline_state{hsk, hsv, params.block_rows, params.block_cols, params.d_split, params.row_split, params.shmem_staging, params.path, params.workgroup_size, subgroup_size, aligned, f32acc, flags, params.limit_occupancy_shmem};
+}
+
+static std::vector<uint32_t> get_fa_spec_constants(const vk_fa_pipeline_state& state) {
+    return {state.workgroup_size, state.Br, state.Bc, state.HSK, state.HSV, !state.aligned, state.D_split,
+            state.row_split, state.subgroup_size, state.shmem_staging ? 1u : 0u, state.flags, state.limit_occupancy_shmem};
 }

 static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) {
@@ -3191,76 +3358,43 @@ static void ggml_vk_load_shaders(vk_device& device) {
                                       align, disable_robustness, require_full_subgroups, required_subgroup_size);
    };

-    auto const &fa_wg_denoms = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache) -> std::array<uint32_t, 3> {
-        return {fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache)[0], 1, 1};
-    };
-
-    auto const &fa_spec_constants = [&](FaCodePath path, uint32_t hsk, uint32_t hsv, uint32_t clamp, ggml_type type, bool small_rows, bool small_cache, uint32_t flags) -> 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)
-        // The same D_split value is used for both HSK and HSV, so just base it on the union of the LSBs.
-        const uint32_t D = (hsk|hsv);
-        auto rows_cols = fa_rows_cols(path, hsk, hsv, clamp, type, small_rows, small_cache);
-
-        uint32_t wg_size;
-        switch (path) {
-        case FA_COOPMAT2:
-            wg_size = ((small_rows && (D % 32) == 0) ? 256 : 128);
-            break;
-        case FA_COOPMAT1:
-            wg_size = (rows_cols[1] / 16) * device->subgroup_size; // enough subgroups for Bc/MatBc
-            break;
-        default:
-            wg_size = scalar_flash_attention_workgroup_size;
-            break;
-        }
-
-        // 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.
-        const uint32_t D_lsb = D ^ (D & (D-1));
-        uint32_t D_split = std::min(std::min(device->subgroup_size, 8u), D_lsb / 4);
-
-        // Nvidia prefers shared memory use to load large tiles of K.
-        // Switch to loading from global memory when it would use too much shared memory.
-        // AMD prefers loading K directly from global memory
-        const uint32_t k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256 ? 1 : 0;
-
-        return {wg_size, rows_cols[0], rows_cols[1], hsk, hsv, clamp, D_split, device->subgroup_size, k_load_shmem, flags};
-    };
-
 #define CREATE_FA(TYPE, NAMELC, FAPATH, SUFFIX) \
        for (auto &fa : device->pipeline_flash_attn_f32_f16[TYPE]) { \
-            uint32_t HSK = fa.first.HSK; \
-            uint32_t HSV = fa.first.HSV; \
-            bool small_rows = fa.first.small_rows; \
-            bool small_cache = fa.first.small_cache; \
            FaCodePath path = fa.first.path; \
+            uint32_t Br = fa.first.Br; \
+            uint32_t Bc = fa.first.Bc; \
            bool aligned = fa.first.aligned; \
            bool f32acc = fa.first.f32acc; \
-            uint32_t flags = fa.first.flags; \
+            uint32_t fa_sgs = fa.first.subgroup_size; \
+            bool fa_ds = fa.first.subgroup_size == 0; \
            if (path == FAPATH) { \
                if (aligned) { \
                    if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), {Br, 1, 1}, get_fa_spec_constants(fa.first), Bc, true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? fa_sgs : 0));     \
                    } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,small_rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,small_rows,small_cache), true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), {Br, 1, 1}, get_fa_spec_constants(fa.first), Bc, true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? fa_sgs : 0));     \
                    } \
                } else { \
                    if (f32acc) { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache,flags), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ##            SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), {Br, 1, 1}, get_fa_spec_constants(fa.first), 1,  true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? fa_sgs : 0));     \
                    } else { \
-                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,small_rows,small_cache,flags), 1,                                        true, FAPATH==FA_COOPMAT1, (FAPATH==FA_COOPMAT1 ? device->subgroup_size : 0));     \
+                        ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data,  "main", 7, sizeof(vk_flash_attn_push_constants), {Br, 1, 1}, get_fa_spec_constants(fa.first), 1,  true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? fa_sgs : 0));     \
                    } \
                } \
            } \
        }

-    CREATE_FA(GGML_TYPE_F32, f32, FA_SCALAR, )
-    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 (device->flash_attention_fp16) {
+        CREATE_FA(GGML_TYPE_F32, f32, FA_SCALAR, )
+        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, )
+    } else {
+        CREATE_FA(GGML_TYPE_F32, f32, FA_SCALAR, _fp32)
+        CREATE_FA(GGML_TYPE_F16, f16, FA_SCALAR, _fp32)
+        CREATE_FA(GGML_TYPE_Q4_0, q4_0, FA_SCALAR, _fp32)
+        CREATE_FA(GGML_TYPE_Q8_0, q8_0, FA_SCALAR, _fp32)
+    }
 #if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
    if (device->coopmat1_fa_support) {
        CREATE_FA(GGML_TYPE_F32, f32, FA_COOPMAT1, _cm1)
@@ -3778,10 +3912,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
        && !device->coopmat_bf16_support
 #endif
        ) {
+        const uint32_t s_warptile_wm = device->subgroup_size == 8 ? 8 : 32;
+
        // use scalar tile sizes
        l_warptile = { 128, 128, 128, 16, subgroup_size_8 * 2, 64, 2, 4, 4, 1, subgroup_size_8 };
        m_warptile = { 128,  64,  64, 16, subgroup_size_8, 32, 2, 4, 2, 1, subgroup_size_8 };
-        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, 2, 2, 1, subgroup_size_8 };
+        s_warptile = { subgroup_size_32, 32, 32, 16, s_warptile_wm, 32, 2, 2, 2, 1, subgroup_size_8 };

        l_wg_denoms = {128, 128, 1 };
        m_wg_denoms = { 64,  64, 1 };
@@ -4531,6 +4667,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
 }

 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch);
+static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev);

 static vk_device ggml_vk_get_device(size_t idx) {
    VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")");
@@ -4747,6 +4884,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
            device->shader_core_count = sm_props.shaderSMCount;
        } else if (amd_shader_core_properties2) {
            device->shader_core_count = amd_shader_core_properties2_props.activeComputeUnitCount;
+        } else if (device->vendor_id == VK_VENDOR_ID_INTEL) {
+            device->shader_core_count = ggml_vk_intel_shader_core_count(device->physical_device);
        } else {
            device->shader_core_count = 0;
        }
@@ -4966,11 +5105,7 @@ 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;
+        device->coopmat1_fa_support = device->coopmat_support && device->subgroup_require_full_support;
 #endif

        if (coopmat2_support) {
@@ -5288,6 +5423,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
            device->mmvq_mode = 1;
        }

+        // Driver issues with older AMD GPUs on Windows, see https://github.com/ggml-org/llama.cpp/pull/19625#issuecomment-3940840613
+        const bool is_amd_proprietary_gcn = device->vendor_id == VK_VENDOR_ID_AMD && device->architecture == AMD_GCN && device->driver_id == vk::DriverId::eAmdProprietary;
+        device->flash_attention_fp16 = device->fp16 && !is_amd_proprietary_gcn;
+
        return device;
    }

@@ -5538,6 +5677,10 @@ static void ggml_vk_instance_init() {
    vk_perf_logger_concurrent = getenv("GGML_VK_PERF_LOGGER_CONCURRENT") != nullptr;
    vk_enable_sync_logger = getenv("GGML_VK_SYNC_LOGGER") != nullptr;
    vk_memory_logger_enabled = getenv("GGML_VK_MEMORY_LOGGER") != nullptr;
+    const char* GGML_VK_PIPELINE_STATS = getenv("GGML_VK_PIPELINE_STATS");
+    if (GGML_VK_PIPELINE_STATS != nullptr) {
+        vk_pipeline_stats_filter = GGML_VK_PIPELINE_STATS;
+    }
    const char* GGML_VK_PERF_LOGGER_FREQUENCY = getenv("GGML_VK_PERF_LOGGER_FREQUENCY");

    if (GGML_VK_PERF_LOGGER_FREQUENCY != nullptr) {
@@ -6773,8 +6916,16 @@ static void ggml_vk_matmul(
        uint32_t padded_n) {
        VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")");
    if (split_k == 1) {
-        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, CEIL_DIV(batch, ctx->device->properties.limits.maxComputeWorkGroupCount[2]));
+
+        uint32_t base_work_group_z = 0;
+        while (base_work_group_z < batch) {
+            uint32_t groups_z = std::min(batch - base_work_group_z, ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+
+            const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, base_work_group_z, batch, k, ne02, ne12, broadcast2, broadcast3, padded_n };
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, groups_z });
+            base_work_group_z += groups_z;
+        }
        return;
    }

@@ -6788,9 +6939,17 @@ static void ggml_vk_matmul(
    uint32_t k_split = CEIL_DIV(k, split_k);
    k_split = ROUNDUP_POW2(k_split, 256);

-    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
-    // Make sure enough workgroups get assigned for split k to work
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, CEIL_DIV(batch, ctx->device->properties.limits.maxComputeWorkGroupCount[2]));
+
+    uint32_t base_work_group_z = 0;
+    while (base_work_group_z < batch) {
+        uint32_t groups_z = std::min(batch - base_work_group_z, ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+
+        const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, base_work_group_z, batch, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
+        // Make sure enough workgroups get assigned for split k to work
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, groups_z });
+        base_work_group_z += groups_z;
+    }
    ggml_vk_sync_buffers(ctx, subctx);
    const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
@@ -7186,7 +7345,6 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
        }

        // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
        if (qx_needs_dequant) {
            ggml_pipeline_request_descriptor_sets(ctx, to_fp16_vk_0, 1);
        }
@@ -7484,7 +7642,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
        if (quantize_y) {
            ggml_pipeline_request_descriptor_sets(ctx, to_q8_1, 1);
        }
-        ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
    }

    vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
@@ -7579,22 +7736,29 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
        fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS1;
    }

-    // compute
-    const vk_mat_vec_push_constants pc = {
-        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        stride_batch_x, stride_batch_y, stride_batch_d,
-        fusion_flags,
-        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
-    };
-    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
-                              {
-                                d_X,
-                                d_Y,
-                                d_D,
-                                d_F0,
-                                d_F1,
-                              },
-                              pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+    ggml_pipeline_request_descriptor_sets(ctx, dmmv, CEIL_DIV(ne12 * ne13, ctx->device->properties.limits.maxComputeWorkGroupCount[1]));
+
+    uint32_t base_work_group_y = 0;
+    while (base_work_group_y < ne12 * ne13) {
+
+        uint32_t groups_y = std::min((uint32_t)(ne12 * ne13) - base_work_group_y, ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+        const vk_mat_vec_push_constants pc = {
+            (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
+            stride_batch_x, stride_batch_y, stride_batch_d,
+            fusion_flags, base_work_group_y,
+            (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
+        };
+        ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+                                  {
+                                    d_X,
+                                    d_Y,
+                                    d_D,
+                                    d_F0,
+                                    d_F1,
+                                  },
+                                  pc, { groups_x, groups_y, groups_z });
+        base_work_group_y += groups_y;
+    }

    if (x_non_contig) {
        ctx->prealloc_x_need_sync = true;
@@ -7832,10 +7996,15 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
        src1->nb[2] <= src1->nb[1] &&
        src1->nb[1] <= src1->nb[3] &&
        src0->ne[3] == 1 &&
-        src1->ne[3] == 1) {
+        src1->ne[3] == 1 &&
+        src0->ne[1] <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
+        src1->ne[2] <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]) {
        ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, cgraph, node_idx);
    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
-               !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
+               !ggml_is_permuted(src0) && !ggml_is_permuted(src1) &&
+               src0->ne[3] <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] &&
+               src0->ne[1] <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] &&
+               src1->ne[2] <= ctx->device->properties.limits.maxComputeWorkGroupCount[2]) {
        ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, cgraph, node_idx);
    // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
    // when ne12 and ne13 are one.
@@ -8391,21 +8560,27 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
    }
 }

-static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool small_cache) {
+static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc) {
+    GGML_UNUSED(f32acc);
    // Needs to be kept up to date on shader changes
-    GGML_UNUSED(hsv);
-    const uint32_t wg_size = scalar_flash_attention_workgroup_size;
-    const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv, small_cache);
-    const uint32_t Bc = scalar_flash_attention_Bc;
+    const uint32_t wg_size = params.workgroup_size;
+    const uint32_t Br = params.block_rows;
+    const uint32_t Bc = params.block_cols;

+    const uint32_t float_type_size = device->flash_attention_fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
+
+    // tmpsh is overestimated slightly
    const uint32_t tmpsh = wg_size * sizeof(float);
-    const uint32_t tmpshv4 = wg_size * 4 * sizeof(float);
+    const uint32_t tmpshv4 = wg_size * 4 * float_type_size;

-    const uint32_t masksh = Bc * Br * sizeof(float);
+    const uint32_t masksh = Bc * (Br + 1) * float_type_size;

-    const uint32_t Qf = Br * (hsk / 4 + 2) * 4 * sizeof(float);
+    const uint32_t Qf = Br * (hsk / 4 + 1) * 4 * float_type_size;

-    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf;
+    const uint32_t D = std::max(hsk, hsv);
+    const uint32_t kvsh = params.shmem_staging ? Bc * (D / 4 + 1) * 4 * float_type_size : 4 * float_type_size;
+
+    const uint32_t total_size = tmpsh + tmpshv4 + masksh + Qf + kvsh;
    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;

    VK_LOG_DEBUG("ggml_vk_flash_attn_scalar_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", total_size=" << total_size << ", supported=" << supported);
@@ -8413,18 +8588,17 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con
    return supported;
 }

-static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const uint32_t hsk, uint32_t hsv, bool f32acc, ggml_type kv_type) {
+static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, const vk_fa_tuning_params& params, uint32_t hsk, uint32_t hsv, bool f32acc) {
    // Needs to be kept up to date on shader changes
-    GGML_UNUSED(hsv);
-    const auto rows_cols = fa_rows_cols(FA_COOPMAT1, hsk, hsv, 0, kv_type, false, false);
-    const uint32_t Br = rows_cols[0];
-    const uint32_t Bc = rows_cols[1];
+    const uint32_t Br = params.block_rows;
+    const uint32_t Bc = params.block_cols;

    const uint32_t MatBr = 16, MatBc = 16;

    const uint32_t row_split = Bc / MatBc;

    const uint32_t hsk_pad = ROUNDUP_POW2(hsk, 16);
+    const uint32_t hsv_pad = ROUNDUP_POW2(hsv, 16);

    const uint32_t acctype = f32acc ? 4 : 2;
    const uint32_t f16vec4 = 8;
@@ -8440,17 +8614,19 @@ static bool ggml_vk_flash_attn_coopmat_shmem_support(const vk_device& device, co
    const uint32_t sfshstride = (hsk <= 128) ? (Br + 8) : Br;
    const uint32_t sfsh = Bc * sfshstride * acctype;

-    const bool k_load_shmem = device->vendor_id == VK_VENDOR_ID_NVIDIA && hsk < 256;
-    const uint32_t kshstride = (k_load_shmem ? hsk_pad : MatBr) / 4 + 2;
+    const uint32_t kvshstride = (params.shmem_staging ? std::max(hsk_pad, hsv_pad) : MatBr) / 4 + 2;
    const uint32_t vsh_stride = MatBc / 4 * row_split;
-    const uint32_t ksh = ((kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)) * f16vec4;
+    const uint32_t ksh = ((kvshstride >= vsh_stride) ? (Bc * kvshstride) : (Bc * vsh_stride)) * f16vec4;
+
+    const uint32_t osh_stride = params.row_split * MatBr / 4;
+    const uint32_t pvsh = MatBc * osh_stride * f16vec4;

    const uint32_t slope = Br * acctype;

-    const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + slope;
+    const uint32_t total_size = tmpsh + Qf + Psh + sfsh + ksh + pvsh + slope;
    const bool supported = total_size <= device->properties.limits.maxComputeSharedMemorySize;

-    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", kv_type=" << kv_type << ", total_size=" << total_size << ", supported=" << supported);
+    VK_LOG_DEBUG("ggml_vk_flash_attn_coopmat_shmem_support(HSK=" << hsk << ", HSV=" << hsv << ", f32acc=" << f32acc << ", total_size=" << total_size << ", supported=" << supported);

    return supported;
 }
@@ -8508,48 +8684,18 @@ 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);

-    FaCodePath path = ctx->device->coopmat2 ? FA_COOPMAT2 :
-                      ctx->device->coopmat1_fa_support ? FA_COOPMAT1 : FA_SCALAR;
-
-    if (path == FA_COOPMAT1 && ctx->device->architecture == vk_device_architecture::NVIDIA_TURING) {
-        // Nvidia compiler bug, see https://github.com/ggml-org/llama.cpp/pull/19075#issuecomment-3820716090
-        path = 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, HSK, HSV, dst->op_params[3] == GGML_PREC_F32, k->type);
-
-        if (!coopmat_shape_supported || !coopmat_shmem_supported) {
-            path = FA_SCALAR;
-        }
-    }
-
    uint32_t gqa_ratio = 1;
    uint32_t qk_ratio = neq2 / nek2;
    uint32_t workgroups_x = (uint32_t)neq1;
    uint32_t workgroups_y = (uint32_t)neq2;
    uint32_t workgroups_z = (uint32_t)neq3;

-    const bool small_cache = nek1 < 1024;
+    const bool f32acc = !ctx->device->flash_attention_fp16 || dst->op_params[3] == GGML_PREC_F32;

    // 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 = get_fa_scalar_num_large_rows(HSK, HSV, small_cache);
-        break;
-    case FA_COOPMAT2:
-        max_gqa = get_fa_num_small_rows(FA_COOPMAT2);
-        break;
-    default:
-        GGML_ASSERT(0);
-    }
+    vk_fa_tuning_params tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, 512, KV, k->type, f32acc);
+    const uint32_t max_gqa = std::min(tuning_params.block_rows, 32u);

    if (N <= 8 && qk_ratio > 1 && qk_ratio <= max_gqa &&
        qk_ratio * nek2 == neq2 && nek2 == nev2 && nem2 <= 1) {
@@ -8561,24 +8707,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        workgroups_y /= gqa_ratio;
    }

-    bool small_rows = N <= get_fa_num_small_rows(path);
-
-    // 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;
-    }
-
-    // with large hsk/hsv, scalar path may need to use small_rows to fit in shared memory
-    if (path == FA_SCALAR &&
-        !ggml_vk_flash_attn_scalar_shmem_support(ctx->device, HSK, HSV, small_cache)) {
-        small_rows = true;
-    }
+    tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, N, KV, k->type, f32acc);

    const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
    uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
@@ -8592,18 +8721,16 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        v_stride /= 4;
    }

-    uint32_t alignment = fa_align(path, HSK, HSV, k->type, small_rows, small_cache);
+    const uint32_t alignment = tuning_params.block_cols;
    bool aligned = (KV % alignment) == 0 &&
                   // the "aligned" shader variant will forcibly align strides, for performance
                   (q_stride & 7) == 0 && (k_stride & 7) == 0 && (v_stride & 7) == 0;

    // Need to use the coopmat2 variant that clamps loads when HSK/HSV aren't sufficiently aligned.
-    if (((HSK | HSV) % 16) != 0 && path == FA_COOPMAT2) {
+    if (((HSK | HSV) % 16) != 0 && tuning_params.path == FA_COOPMAT2) {
        aligned = false;
    }

-    bool f32acc = path == FA_SCALAR || dst->op_params[3] == GGML_PREC_F32;
-
    float scale         = 1.0f;
    float max_bias      = 0.0f;
    float logit_softcap = 0.0f;
@@ -8618,12 +8745,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx

    // Only use mask opt when the mask is fairly large. This hasn't been tuned extensively.
    bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768;
-
-    uint32_t flags = (use_mask_opt       ? 1 : 0) |
-                     (mask != nullptr    ? 2 : 0) |
-                     (logit_softcap != 0 ? 4 : 0);
-
-    vk_fa_pipeline_state fa_pipeline_state(HSK, HSV, small_rows, small_cache, path, aligned, f32acc, flags);
+    vk_fa_pipeline_state fa_pipeline_state = get_fa_pipeline_state(tuning_params, HSK, HSV, aligned, f32acc,
+                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0);

    vk_pipeline pipeline = nullptr;

@@ -8645,22 +8768,35 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
    uint32_t split_kv = KV;
    uint32_t split_k = 1;

+    // Intel Alchemist prefers more workgroups
+    const uint32_t shader_core_count_multiplier = (ctx->device->vendor_id == VK_VENDOR_ID_INTEL && ctx->device->architecture != INTEL_XE2) ? 2 : 1;
+
    // Use a placeholder core count if one isn't available. split_k is a big help for perf.
-    const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count : 16;
+    const uint32_t shader_core_count = ctx->device->shader_core_count ? ctx->device->shader_core_count * shader_core_count_multiplier : 16;
+
+    const uint32_t Br = fa_pipeline_state.Br;
+    const uint32_t Bc = fa_pipeline_state.Bc;
+
+    GGML_ASSERT(Br == pipeline->wg_denoms[0]);
+    const uint32_t Tr = CEIL_DIV(N, Br);

    // Try to use split_k when KV is large enough to be worth the overhead.
-    // Must either be a single batch or be using gqa, we can't mix the two.
-    if (workgroups_x <= pipeline->wg_denoms[0] && (workgroups_x == 1 || gqa_ratio > 1)) {
-        // Try to run two workgroups per SM.
+    if (gqa_ratio > 1 && workgroups_x <= Br) {
        split_k = shader_core_count * 2 / (workgroups_x * workgroups_y * workgroups_z);
-        if (split_k > 1) {
-            // Try to evenly split KV into split_k chunks, but it needs to be a multiple
-            // of "align", so recompute split_k based on that.
-            split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), alignment);
-            split_k = CEIL_DIV(KV, split_kv);
+    } else if (gqa_ratio <= 1) {
+        uint32_t total_wgs_no_split = Tr * workgroups_y * workgroups_z;
+        if (total_wgs_no_split < shader_core_count * 2) {
+            split_k = shader_core_count * 2 / total_wgs_no_split;
        }
    }

+    if (split_k > 1) {
+        // Try to evenly split KV into split_k chunks, but it needs to be a multiple
+        // of "align", so recompute split_k based on that.
+        split_kv = ROUNDUP_POW2(std::max(1u, KV / split_k), alignment);
+        split_k = CEIL_DIV(KV, split_kv);
+    }
+
    // Reserve space for split_k temporaries. For each split x batch, we need to store the O matrix (D x ne1)
    // and the per-row m and L values (ne1 rows). We store all the matrices first, followed by the rows.
    // For matrices, the order is (inner to outer) [HSV, ne1, k, ne2, ne3].
@@ -8674,10 +8810,6 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        ggml_vk_preallocate_buffers(ctx, subctx);
    }

-    auto rows_cols = fa_rows_cols(path, HSK, HSV, !aligned, k->type, small_rows, small_cache);
-    const uint32_t Br = rows_cols[0];
-    const uint32_t Bc = rows_cols[1];
-
    const uint32_t mask_opt_num_dwords = CEIL_DIV(nem0, 16 * Bc);
    const uint64_t mask_opt_size = sizeof(uint32_t) * mask_opt_num_dwords * CEIL_DIV(nem1, Br) * nem2 * nem3;

@@ -8757,15 +8889,21 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
        if (ctx->prealloc_split_k_need_sync) {
            ggml_vk_sync_buffers(ctx, subctx);
        }
-        workgroups_x *= pipeline->wg_denoms[0];
+
+        // We reuse workgroups_x to mean the number of splits, so we need to
+        // cancel out the divide by wg_denoms[0].
+        uint32_t dispatch_x;
+        if (gqa_ratio > 1) {
+            workgroups_x *= pipeline->wg_denoms[0];
+            dispatch_x = split_k * workgroups_x;
+        } else {
+            dispatch_x = Tr * split_k * pipeline->wg_denoms[0];
+        }
+
        vk_subbuffer split_k_buf = ggml_vk_subbuffer(ctx, ctx->prealloc_split_k, 0);
        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                    {q_buf, k_buf, v_buf, mask_buf, sinks_buf, split_k_buf, mask_opt_buf},
-                                    // We only use split_k when group query attention is enabled, which means
-                                    // there's no more than one tile of rows (i.e. workgroups_x would have been
-                                    // one). We reuse workgroups_x to mean the number of splits, so we need to
-                                    // cancel out the divide by wg_denoms[0].
-                                    pc, { split_k * workgroups_x, workgroups_y, workgroups_z });
+                                    pc, { dispatch_x, workgroups_y, workgroups_z });

        ggml_vk_sync_buffers(ctx, subctx);
        const vk_op_flash_attn_split_k_reduce_push_constants pc2 = { HSV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, split_k, (sinks != nullptr) };
@@ -11560,7 +11698,6 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
        }
    }

-    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
    if (split_k > 1) {
        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);

@@ -12069,7 +12206,6 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
        // y[i] = i % k;
    }

-    ggml_pipeline_request_descriptor_sets(ctx, p, num_it);
    if (split_k > 1) {
        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_matmul_split_k_reduce, num_it);

@@ -15392,6 +15528,46 @@ static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDevicePrope
    }
 }

+static uint32_t ggml_vk_intel_shader_core_count(const vk::PhysicalDevice& vkdev) {
+    VkPhysicalDeviceProperties2 props = vkdev.getProperties2();
+
+    if (props.properties.vendorID != VK_VENDOR_ID_INTEL) {
+        return 0;
+    }
+
+    const uint32_t device_id = props.properties.deviceID;
+
+    switch (device_id) {
+    case 0x56A6:  // A310
+        return 6;
+    case 0x5693:  // A370M
+    case 0x56A5:  // A380
+    case 0x56B1:  // Pro A40/A50
+        return 8;
+    case 0x5697:  // A530M
+        return 12;
+    case 0x5692:  // A550M
+    case 0x56B3:  // Pro A60
+        return 16;
+    case 0x56A2:  // A580
+        return 24;
+    case 0x5691:  // A730M
+    case 0x56A1:  // A750
+        return 28;
+    case 0x56A0:  // A770
+    case 0x5690:  // A770M
+        return 32;
+    case 0xE212:  // Pro B50
+        return 16;
+    case 0xE20C:  // B570
+        return 18;
+    case 0xE20B:  // B580
+        return 20;
+    default:
+        return 0;
+    }
+}
+
 // checks

 #ifdef GGML_VULKAN_CHECK_RESULTS
@@ -16068,7 +16244,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_cgraph *
        ggml_vk_print_graph_origin(tensor, done);
    }

-    if (avg_err > 0.5 || std::isnan(avg_err)) {
+    if (avg_err > 0.01 || std::isnan(avg_err)) {
        std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
        if (src0 != nullptr) {
@@ -3,9 +3,13 @@
 #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

+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_subgroup_extended_types_float16 : require
+#endif
+
 #extension GL_KHR_shader_subgroup_shuffle : enable
 #extension GL_KHR_shader_subgroup_vote : enable

@@ -15,8 +19,10 @@
 const uint32_t HSK_per_thread = HSK / D_split;
 const uint32_t HSV_per_thread = HSV / D_split;

-const uint32_t cols_per_iter = WorkGroupSize / D_split;
+const uint32_t rows_per_thread = Br / row_split;
+const uint32_t cols_per_iter = WorkGroupSize / D_split / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
+const uint32_t num_subgroups = SubGroupSize == 0 ? 0 : WorkGroupSize / SubGroupSize;


 layout (binding = 0) readonly buffer Q {float data_q[];};
@@ -27,20 +33,22 @@ 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) * HSV + c;
-    data_o[o_offset + offset] = D_TYPE(elem);
-    return elem;
-}
+// If SubGroupSize is set to 0 then only use shmem reductions
+const uint32_t tmpsh_size = (SubGroupSize > 0) ? (row_split == 1 ? num_subgroups * D_split : num_subgroups) : WorkGroupSize;
+shared float tmpsh[tmpsh_size];
+shared FLOAT_TYPEV4 tmpshv4[tmpsh_size];

-shared FLOAT_TYPE tmpsh[WorkGroupSize];
-shared vec4 tmpshv4[WorkGroupSize];
+const uint32_t masksh_stride = Br + 1;
+shared FLOAT_TYPE masksh[Bc * masksh_stride];

-shared float masksh[Bc][Br];
-shared vec4 Qf[Br][HSK / 4];
+const uint32_t qf_stride = HSK / 4 + 1;
+shared FLOAT_TYPEV4 Qf[Br * qf_stride];
+
+const uint32_t D = HSK > HSV ? HSK : HSV;
+const uint32_t kvsh_stride = D / 4 + 1;
+shared FLOAT_TYPEV4 kvsh[SHMEM_STAGING != 0 ? Bc * kvsh_stride : 1];
+
+shared vec4 occupancy_limiter[LIMIT_OCCUPANCY_SHMEM > 0 ? LIMIT_OCCUPANCY_SHMEM : 1];

 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -50,8 +58,24 @@ void main() {
    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 rowgroup_tid = gl_LocalInvocationIndex % threads_per_rowgroup;
    const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
-    const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
+    const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
+
+    if (LIMIT_OCCUPANCY_SHMEM > 0) {
+        // This just exists to avoid the occupancy_limiter array getting optimized out
+        occupancy_limiter[tid] = vec4(tid);
+
+        barrier();
+
+        if (occupancy_limiter[tid] == vec4(99999.0)) {
+            data_ov4[0] = D_TYPEV4(occupancy_limiter[tid]);
+        }
+    }
+
+#define tile_row(r) (row_tid * rows_per_thread + (r))

    uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02 + iq3*p.nb03) / 4;

@@ -60,37 +84,37 @@ void main() {
        uint32_t r = (idx + tid) / (HSK / 4);
        if (r < Br && d < HSK / 4 &&
            i * Br + r < N) {
-            Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
+            Qf[r * qf_stride + d] = FLOAT_TYPEV4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
        }
    }
    barrier();

-    vec4 Of[Br][HSV_per_thread / 4];
+    FLOAT_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            Of[r][d] = vec4(0.0);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] = FLOAT_TYPEV4(0.0);
        }
    }

-    float Lf[Br], Mf[Br];
+    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 < Br; ++r) {
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
        Lf[r] = 0;
        Mf[r] = NEG_FLT_MAX_OVER_2;
    }

-    float slope[Br];
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        slope[r] = 1.0;
+    ACC_TYPE slope[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        slope[r] = ACC_TYPE(1.0);
    }

    // ALiBi
    if (p.max_bias > 0.0f) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            slope[r] = perElemOpComputeSlope(tile_row(r), col_tid, ACC_TYPE(0), iq2);
        }
    }

@@ -113,75 +137,141 @@ void main() {

    uint32_t mask_opt = 0;
    uint32_t mask_opt_idx = ~0;
+    uint32_t mask_opt_bits = 0;

    [[dont_unroll]]
    for (uint32_t j = start_j; j < end_j; ++j) {
+        if (MASK_ENABLE) {
+            if (USE_MASK_OPT && mask_opt_idx != j / 16) {
+                mask_opt_idx = j / 16;
+                mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
+            }
+            mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
+            if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
+                // skip this block
+                continue;
+            }
+            // Only load if the block is not all zeros
+            if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
+                bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;

-        if (USE_MASK_OPT && mask_opt_idx != j / 16) {
-            mask_opt_idx = j / 16;
-            mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
+                float max_mask = NEG_FLT_MAX_OVER_2;
+                barrier();
+                [[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) {
+                        if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
+                            FLOAT_TYPE m = FLOAT_TYPE(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
+                            masksh[c * masksh_stride + r] = m;
+                            max_mask = max(max_mask, float(m));
+                        } else {
+                            masksh[c * masksh_stride + r] = FLOAT_TYPE(0);
+                        }
+                    }
+                }
+                // skip the block if the mask is entirely -inf
+                bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
+                barrier();
+                if (gl_SubgroupInvocationID == 0) {
+                    tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
+                }
+                barrier();
+                [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
+                    max_mask = max(max_mask, tmpsh[s]);
+                }
+                if (max_mask <= NEG_FLT_MAX_OVER_2) {
+                    continue;
+                }
+            }
        }
-        uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
-        if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
-            // skip this block
-            continue;
-        }
-        // Only load if the block is not all zeros
-        if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
-            bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;

-            float max_mask = NEG_FLT_MAX_OVER_2;
-            [[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) {
-                    if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
-                        float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
-                        masksh[c][r] = m;
-                        max_mask = max(max_mask, m);
+        ACC_TYPE Sf[rows_per_thread][cols_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                Sf[r][c] = ACC_TYPE(0.0);
+            }
+        }
+
+        if (SHMEM_STAGING != 0) {
+            barrier();
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSK / 4);
+                uint32_t c = (idx + tid) / (HSK / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSK / 4 || c < Bc) {
+                    FLOAT_TYPEV4 K_Tf = FLOAT_TYPEV4(0);
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+#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);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+#else
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = K_Tf;
+                }
+            }
+            barrier();
+        }
+
+        // More d iterations means Q register caching becomes relevant
+        // Few iterations means the additional registers needed are worse than the speed-up from caching
+        if (HSK_per_thread / 4 > 4) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                FLOAT_TYPEV4 Q_cache[rows_per_thread];
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    Q_cache[r] = Qf[tile_row(r) * qf_stride + d * D_split + d_tid];
+                }
+
+                [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
+                    if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                        continue;
+                    }
+
+                    FLOAT_TYPEV4 K_Tf;
+                    if (SHMEM_STAGING != 0) {
+                        K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
                    } else {
-                        masksh[c][r] = float(0);
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+#else
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
+#endif
+                    }
+                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                        Sf[r][c] += ACC_TYPE(dot(Q_cache[r], K_Tf));
                    }
                }
            }
-            // skip the block if the mask is entirely -inf
-            bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
-            barrier();
-            if (gl_SubgroupInvocationID == 0) {
-                tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
-            }
-            barrier();
-            [[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
-                max_mask = max(max_mask, tmpsh[s]);
-            }
-            if (max_mask <= NEG_FLT_MAX_OVER_2) {
-                continue;
-            }
-        }
-
-        float Sf[Br][cols_per_thread];
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+        } else {
            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                Sf[r][c] = 0.0;
-            }
-        }
+                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
+                    continue;
+                }

-
-        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
-                continue;
-            }
-            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
+                    FLOAT_TYPEV4 K_Tf;
+                    if (SHMEM_STAGING != 0) {
+                        K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
+                    } else {
 #if BLOCK_SIZE > 1
-                uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
-                uint ib = coord / BLOCK_SIZE;
-                uint iqs = (coord % BLOCK_SIZE);
-                vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
+                        uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
-                vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
+                        K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
 #endif
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                    Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf);
+                    }
+                    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                        Sf[r][c] += ACC_TYPE(dot(Qf[tile_row(r) * qf_stride + d * D_split + d_tid], K_Tf));
+                    }
                }
            }
        }
@@ -189,89 +279,109 @@ void main() {
        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
            // Compute sum across the D_split
            [[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) {
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                    Sf[r][c] += subgroupShuffleXor(Sf[r][c], s);
                }
            }
        }

        if (LOGIT_SOFTCAP) {
-            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
                [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                    Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]);
+                    Sf[r][c] = ACC_TYPE(p.logit_softcap * tanh(Sf[r][c]));
                }
            }
        }

        if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                    float mvf = masksh[c * cols_per_iter + col_tid][r];
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    FLOAT_TYPE mvf = masksh[(c * cols_per_iter + col_tid) * masksh_stride + tile_row(r)];

                    Sf[r][c] += slope[r]*mvf;
                }
            }
-            barrier();
        }

-        float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            rowmaxf[r] = NEG_FLT_MAX_OVER_2;
+        float eMf[rows_per_thread];
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float rowmaxf = NEG_FLT_MAX_OVER_2;
            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
                    continue;
                }
-                rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
+                rowmaxf = max(rowmaxf, float(Sf[r][c]));
            }
-            Moldf[r] = Mf[r];
+            float Moldf = Mf[r];

            // M = max(rowmax, Mold)
            // P = e^(S - M)
            // eM = e^(Mold - M)
-            Mf[r] = max(rowmaxf[r], Moldf[r]);
-            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                Pf[r][c] = exp(Sf[r][c] - Mf[r]);
-            }
-            eMf[r] = exp(Moldf[r] - Mf[r]);
-
-            // Compute sum across row of P
-            rowsumf[r] = 0.0;
-            [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-                if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
-                    continue;
-                }
-                rowsumf[r] += Pf[r][c];
-            }
-
-            Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
+            Mf[r] = max(rowmaxf, Moldf);
+            eMf[r] = exp(Moldf - Mf[r]);
+            Lf[r] = eMf[r]*Lf[r];
        }

        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-            [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-                Of[r][d] = eMf[r] * Of[r][d];
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Of[r][d] = FLOAT_TYPE(eMf[r]) * Of[r][d];
            }
        }

+        if (SHMEM_STAGING != 0) {
+            barrier();
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSV / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSV / 4);
+                uint32_t c = (idx + tid) / (HSV / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSV / 4 || c < Bc) {
+                    FLOAT_TYPEV4 V_Tf = FLOAT_TYPEV4(0);
+                    if (!KV_bounds_check || j * Bc + c < KV) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                        V_Tf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = V_Tf;
+                }
+            }
+            barrier();
+        }
+
        [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
            if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
                continue;
            }
+
+            FLOAT_TYPE Pf[rows_per_thread];
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                Pf[r] = FLOAT_TYPE(exp(float(Sf[r][c]) - Mf[r]));
+                Lf[r] += Pf[r];
+            }
+
            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                FLOAT_TYPEV4 Vf;
+                if (SHMEM_STAGING != 0) {
+                    Vf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
+                } else {
 #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);
+                    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);
+                    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]);
+                    Vf = FLOAT_TYPEV4(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 < Br; ++r) {
-                    Of[r][d] += Pf[r][c] * Vf;
+                }
+                [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                    Of[r][d] += FLOAT_TYPEV4(Pf[r] * Vf);
                }
            }
        }
-
-        barrier();
    }

    // prevent race on tmpsh
@@ -279,58 +389,108 @@ void main() {

    // reduce across threads

-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-        float rowmaxf, eMf;
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+        float rowmaxf = Mf[r];

-        tmpsh[tid] = Mf[r];
        // Compute max across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-            if (tid < s) {
-                tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]);
+        if (SubGroupSize > 0) {
+            [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                rowmaxf = max(rowmaxf, subgroupShuffleXor(rowmaxf, s));
            }
+            if (row_split == 1) {
+                // Reduce inside workgroup with shmem
+                barrier();
+                if (gl_SubgroupInvocationID == d_tid) {
+                    tmpsh[gl_SubgroupID * D_split + d_tid] = rowmaxf;
+                }
+                barrier();
+                rowmaxf = tmpsh[d_tid];
+                [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                    rowmaxf = max(rowmaxf, tmpsh[s * D_split + d_tid]);
+                }
+            }
+        } else {
            barrier();
+            tmpsh[tid] = rowmaxf;
+            barrier();
+            [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                if (rowgroup_tid < s) {
+                    tmpsh[tid] = max(tmpsh[tid], tmpsh[tid ^ s]);
+                }
+                barrier();
+            }
+            rowmaxf = tmpsh[row_tid * threads_per_rowgroup + d_tid];
        }
-        rowmaxf = tmpsh[d_tid];
-        barrier();

        float Moldf = Mf[r];

        // M = max(rowmax, Mold)
        // eM = e^(Mold - M)
        Mf[r] = max(rowmaxf, Moldf);
-        eMf = exp(Moldf - Mf[r]);
+        float eMf = exp(Moldf - Mf[r]);

        Lf[r] = eMf*Lf[r];

-        tmpsh[tid] = Lf[r];
-
        // Compute sum across the row
-        barrier();
-        [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-            if (tid < s) {
-                tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s];
+        if (SubGroupSize > 0) {
+            [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                Lf[r] += subgroupShuffleXor(Lf[r], s);
            }
+            if (row_split == 1) {
+                barrier();
+                if (gl_SubgroupInvocationID == d_tid) {
+                    tmpsh[gl_SubgroupID * D_split + d_tid] = Lf[r];
+                }
+                barrier();
+                Lf[r] = tmpsh[d_tid];
+                [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                    Lf[r] += tmpsh[s * D_split + d_tid];
+                }
+            }
+        } else {
            barrier();
-        }
-        Lf[r] = tmpsh[d_tid];
-        barrier();
-
-        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-
-            Of[r][d] = eMf * Of[r][d];
-            tmpshv4[tid] = Of[r][d];
-
+            tmpsh[tid] = Lf[r];
            barrier();
-            [[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
-                if (tid < s) {
-                    Of[r][d] += tmpshv4[tid + s];
-                    tmpshv4[tid] = Of[r][d];
+            [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                if (rowgroup_tid < s) {
+                    tmpsh[tid] = tmpsh[tid] + tmpsh[tid ^ s];
                }
                barrier();
            }
-            Of[r][d] = tmpshv4[d_tid];
-            barrier();
+            Lf[r] = tmpsh[row_tid * threads_per_rowgroup + d_tid];
+        }
+
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+            Of[r][d] = FLOAT_TYPE(eMf) * Of[r][d];
+
+            if (SubGroupSize > 0) {
+                [[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
+                    Of[r][d] += subgroupShuffleXor(Of[r][d], s);
+                }
+                if (row_split == 1) {
+                    barrier();
+                    if (gl_SubgroupInvocationID == d_tid) {
+                        tmpshv4[gl_SubgroupID * D_split + d_tid] = Of[r][d];
+                    }
+                    barrier();
+                    Of[r][d] = tmpshv4[d_tid];
+                    [[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
+                        Of[r][d] += tmpshv4[s * D_split + d_tid];
+                    }
+                }
+            } else {
+                barrier();
+                tmpshv4[tid] = Of[r][d];
+                barrier();
+                [[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
+                    if (rowgroup_tid < s) {
+                        Of[r][d] += tmpshv4[tid ^ s];
+                        tmpshv4[tid] = Of[r][d];
+                    }
+                    barrier();
+                }
+                Of[r][d] = tmpshv4[row_tid * threads_per_rowgroup + d_tid];
+            }
        }
    }

@@ -338,33 +498,53 @@ void main() {
    // 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) {
-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;

-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                if (row < N) {
+                    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                        gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
                    }
                }
            }
-        }

-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
-                perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
-                perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                if (row < N) {
+                    perElemOpStoreCol0(row, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
+                    perElemOpStoreCol0(row, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
+                }
+            }
+        } else {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                const uint global_row = i * Br + row;
+
+                if (global_row < N) {
+                    uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
+
+                    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
+                        data_ov4[o_offset + iq2 * HSV/4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
+                    }
+                }
+
+                if (global_row < N && d_tid == 0 && col_tid == 0) {
+                    uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+                    data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
+                    data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
+                }
            }
        }
-
        return;
    }

    if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            float sink = perElemOpGetSink(tile_row(r), 0u, ACC_TYPE(0), iq2);

            float ms = 1.0f;
            float vs = 1.0f;
@@ -373,7 +553,7 @@ void main() {
                ms = exp(Mf[r] - sink);

                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    Of[r][d] *= ms;
+                    Of[r][d] *= FLOAT_TYPE(ms);
                }
            } else {
                vs = exp(sink - Mf[r]);
@@ -383,39 +563,37 @@ void main() {
        }
    }

-    float Lfrcp[Br];
-    [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
+    float Lfrcp[rows_per_thread];
+    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
        Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
    }

    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            Of[r][d] *= Lfrcp[r];
-#if defined(ACC_TYPE_MAX)
-            Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            Of[r][d] *= FLOAT_TYPE(Lfrcp[r]);
+#if defined(FLOAT_TYPE_MAX)
+            Of[r][d] = clamp(Of[r][d], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
 #endif
        }
    }

-    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;

    if (p.gqa_ratio > 1) {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (r < N) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            const uint row = tile_row(r);
+            if (row < N) {
                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
-                    }
+                    gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
                }
            }
        }
    } else {
-        [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
-            if (i * Br + r < N) {
+        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+            const uint row = tile_row(r);
+            if (i * Br + row < N) {
                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
-                    }
+                    data_ov4[o_offset + (iq2 * HSV + (i * Br + row) * p.ne1 * HSV) / 4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
                }
            }
        }
@@ -1,16 +1,18 @@

 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 HSK = 32;
-layout (constant_id = 4) const uint32_t HSV = 32;
-layout (constant_id = 5) const uint32_t Clamp = 0;
-layout (constant_id = 6) const uint32_t D_split = 16;
-layout (constant_id = 7) const uint32_t SubGroupSize = 32;
-layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0;
-layout (constant_id = 9) const uint32_t Flags = 0;
+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 HSK = 32;
+layout (constant_id =  4) const uint32_t HSV = 32;
+layout (constant_id =  5) const uint32_t Clamp = 0;
+layout (constant_id =  6) const uint32_t D_split = 16;
+layout (constant_id =  7) const uint32_t row_split = 1;
+layout (constant_id =  8) const uint32_t SubGroupSize = 32;
+layout (constant_id =  9) const uint32_t SHMEM_STAGING = 0;
+layout (constant_id = 10) const uint32_t Flags = 0;
+layout (constant_id = 11) const uint32_t LIMIT_OCCUPANCY_SHMEM = 0;

 const bool USE_MASK_OPT  = (Flags & 1) != 0;
 const bool MASK_ENABLE   = (Flags & 2) != 0;
@@ -69,6 +71,7 @@ layout (push_constant) uniform parameter {
 layout (binding = 4) readonly buffer S {float data_s[];};

 layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
+layout (binding = 5) writeonly buffer OV4 {D_TYPEV4 data_ov4[];};

 layout (binding = 6) readonly buffer MO {uint32_t data_mask_opt[];};

@@ -94,12 +97,12 @@ layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16
 #define BLOCK_SIZE 4
 #define BLOCK_BYTE_SIZE 16

-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
    // iqs is currently always zero in the flash attention shaders
    if (binding_idx == BINDING_IDX_K) {
-        return k_packed.k_data_packed[a_offset + ib];
+        return FLOAT_TYPEV4(k_packed.k_data_packed[a_offset + ib]);
    } else {
-        return v_packed.v_data_packed[a_offset + ib];
+        return FLOAT_TYPEV4(v_packed.v_data_packed[a_offset + ib]);
    }
 }
 #endif
@@ -107,7 +110,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
 #if defined(DATA_A_Q4_0)
 #define BLOCK_BYTE_SIZE 18

-vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
    if (binding_idx == BINDING_IDX_K) {
        uint vui_lo = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
        uint vui_hi = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -115,7 +118,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
        vui_lo >>= shift;
        vui_hi >>= shift;

-        return float(k_packed.k_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+        return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(8.0f));
    } else {
        uint vui_lo = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
        uint vui_hi = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -123,24 +126,24 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
        vui_lo >>= shift;
        vui_hi >>= shift;

-        return float(v_packed.v_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
+        return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(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) {
+FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
    if (binding_idx == BINDING_IDX_K) {
        const i8vec2 v0 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
        const i8vec2 v1 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;

-        return float(k_packed.k_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
    } else {
        const i8vec2 v0 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
        const i8vec2 v1 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;

-        return float(v_packed.v_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
+        return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
    }
 }
 #endif
@@ -189,10 +192,16 @@ void init_indices()
    KV = p.KV;

    if (p.k_num > 1) {
-        i = 0;
-        // batch and split_k share gl_WorkGroupID.x
-        gqa_iq1 = gl_WorkGroupID.x / p.k_num;
-        split_k_index = gl_WorkGroupID.x % p.k_num;
+        if (p.gqa_ratio > 1) {
+            i = 0;
+            // batch and split_k share gl_WorkGroupID.x
+            gqa_iq1 = gl_WorkGroupID.x / p.k_num;
+            split_k_index = gl_WorkGroupID.x % p.k_num;
+        } else {
+            gqa_iq1 = 0;
+            split_k_index = gl_WorkGroupID.x % p.k_num;
+            i = gl_WorkGroupID.x / p.k_num;
+        }
    } else if (p.gqa_ratio > 1) {
        i = 0;
        gqa_iq1 = gl_WorkGroupID.x;
@@ -244,3 +253,11 @@ void init_indices()
 // Bias applied to softmax to stay in fp16 range.
 // Based on ggml-cuda issue https://github.com/ggml-org/llama.cpp/issues/18606
 const float FATTN_KQ_MAX_OFFSET = 3.0f*0.6931f;
+
+// Store the output when doing grouped query attention.
+// Rows index by Q's dimension 2, and the first N rows are valid.
+void gqaStore(const in uint32_t r, const in uint32_t c, const in FLOAT_TYPEV4 elems, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
+{
+    uint32_t offset = (iq2 + r) * HSV / 4 + c;
+    data_ov4[o_offset + offset] = D_TYPEV4(elems);
+}
@@ -19,7 +19,6 @@
 const uint32_t MatBr = 16;
 const uint32_t MatBc = 16;

-const uint32_t row_split = Bc / MatBc;
 const uint32_t rows_per_thread = Br / row_split;
 const uint32_t cols_per_iter = gl_WorkGroupSize.x / row_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -33,15 +32,6 @@ 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) * HSV + c;
-    data_o[o_offset + offset] = D_TYPE(elem);
-    return elem;
-}
-
 shared float tmpsh[row_split];

 const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
@@ -54,10 +44,14 @@ shared f16vec4 Psh[Bc * psh_stride];
 const uint32_t sfshstride = (HSK <= 128) ? (Br / 4 + 2) : Br / 4;
 shared ACC_TYPEV4 sfsh[Bc * sfshstride];

-const uint32_t kshstride = (K_LOAD_SHMEM != 0 ? HSK_pad : MatBr) / 4 + 2; // in units of f16vec4
+const uint32_t D_pad = HSK_pad > HSV_pad ? HSK_pad : HSV_pad;
+const uint32_t kvsh_stride = (SHMEM_STAGING != 0 ? D_pad : MatBr) / 4 + 2; // in units of f16vec4
 const uint v_cols = MatBc / 4 * row_split; // total cols, 4 vec4s per MatBc * number of subgroups
 const uint vsh_stride = v_cols;
-shared f16vec4 ksh[(kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)];
+shared f16vec4 kvsh[(kvsh_stride >= vsh_stride) ? (Bc * kvsh_stride) : (Bc * vsh_stride)];
+
+const uint32_t osh_stride = row_split * MatBr / 4;
+shared f16vec4 pvsh[MatBc * osh_stride];

 shared ACC_TYPE slope[Br];

@@ -84,11 +78,6 @@ void main() {
                Qf[i + tid] = f16vec4(0);
            }
        }
-        [[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
-            if (i + tid < Bc * kshstride) {
-                ksh[i + tid] = f16vec4(0);
-            }
-        }
        barrier();
    }

@@ -104,10 +93,10 @@ void main() {
    }
    barrier();

-    ACC_TYPEV4 Of[rows_per_thread][d_per_thread];
+    f16vec4 Of[rows_per_thread][d_per_thread];
    [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
        [[unroll]] for (uint32_t d = 0; d < d_per_thread; ++d) {
-            Of[r][d] = ACC_TYPEV4(0.0);
+            Of[r][d] = f16vec4(0.0);
        }
    }

@@ -153,22 +142,22 @@ void main() {

    uint32_t mask_opt = 0;
    uint32_t mask_opt_idx = ~0;
+    uint32_t mask_opt_bits = 0;
+    f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];

    [[dont_unroll]]
    for (uint32_t j = start_j; j < end_j; ++j) {

-        f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
        [[unroll]] for (uint32_t idx = 0; idx < mask_cache.length(); ++idx) {
            mask_cache[idx] = f16vec4(0);
        }

        if (MASK_ENABLE) {
-
            if (USE_MASK_OPT && mask_opt_idx != j / 16) {
                mask_opt_idx = j / 16;
                mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
            }
-            uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
+            mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
            if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
                // skip this block
                continue;
@@ -231,24 +220,24 @@ void main() {
            }
        }

-        if (K_LOAD_SHMEM != 0) {
-            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
-                uint32_t d = (idx + tid) % (HSK / 4);
-                uint32_t c = (idx + tid) / (HSK / 4);
-                if (c < Bc && d < HSK / 4) {
+        if (SHMEM_STAGING != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSK_pad / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSK_pad / 4);
+                uint32_t c = (idx + tid) / (HSK_pad / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSK_pad / 4 || c < Bc) {
                    f16vec4 K_Tf = f16vec4(0);
-                    if (!KV_bounds_check || j * Bc + c < KV) {
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (HSK == HSK_pad || d < HSK / 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);
-                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
                        K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
 #endif
                    }

-                    ksh[c * kshstride + d] = K_Tf;
+                    kvsh[c * kvsh_stride + d] = K_Tf;
                }
            }
            barrier();
@@ -262,7 +251,11 @@ void main() {
        coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;

        [[unroll]] for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
-            if (K_LOAD_SHMEM == 0) {
+            // If SHMEM_STAGING is set, a Bc * HSK_pad size tile of K is loaded to shmem
+            // If not, f16 K is loaded directly from global memory if aligned, otherwise
+            // staged through a Bc * MatBr size staging buffer.
+            // If K is not type f16, then it is always staged for dequantization.
+            if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
            if (KV_bounds_check || d * 16 + 16 > HSK) {
 #endif
@@ -277,13 +270,13 @@ void main() {
                        uint coord = (j * Bc + row) * k_stride * BLOCK_SIZE + d * 16 + col_vec * 4;
                        uint ib = coord / BLOCK_SIZE;
                        uint iqs = (coord % BLOCK_SIZE);
-                        K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
+                        K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
 #else
                        K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + row) * k_stride / 4 + d * 16 / 4 + col_vec]);
 #endif
                    }

-                    ksh[row * kshstride + col_vec] = K_Tf;
+                    kvsh[row * kvsh_stride + col_vec] = K_Tf;
                }
            }
            barrier();
@@ -295,8 +288,8 @@ void main() {
            if (KV_bounds_check || d * 16 + 16 > HSK)
 #endif
            {
-                uint coord = (gl_SubgroupID * MatBc) * kshstride;
-                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+                uint coord = (gl_SubgroupID * MatBc) * kvsh_stride;
+                coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
            }
 #if BLOCK_SIZE == 1
            else {
@@ -305,8 +298,8 @@ void main() {
            }
 #endif
            } else {
-                uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
-                coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
+                uint coord = (gl_SubgroupID * MatBc) * kvsh_stride + d * 16 / 4;
+                coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
            }

            coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
@@ -329,7 +322,7 @@ void main() {
            barrier();
        }

-        if (MASK_ENABLE) {
+        if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
            [[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
                uint32_t c = (idx + tid) / (Br / 4);
                uint32_t r = (idx + tid) % (Br / 4);
@@ -374,7 +367,7 @@ void main() {
        [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
            const uint d_local = d0 / threads_per_rowgroup;
            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d_local] = ACC_TYPE(eMf[r]) * Of[r][d_local];
+                Of[r][d_local] = float16_t(eMf[r]) * Of[r][d_local];
            }
        }

@@ -397,19 +390,47 @@ void main() {
            }
        }

+        if (SHMEM_STAGING != 0) {
+            [[unroll]] for (uint32_t idx = 0; idx < Bc * HSV_pad / 4; idx += gl_WorkGroupSize.x) {
+                uint32_t d = (idx + tid) % (HSV_pad / 4);
+                uint32_t c = (idx + tid) / (HSV_pad / 4);
+                if (idx + gl_WorkGroupSize.x <= Bc * HSV_pad / 4 || c < Bc) {
+                    f16vec4 V_Tf = f16vec4(0);
+                    if ((!KV_bounds_check || j * Bc + c < KV) && (HSV == HSV_pad || d < HSV / 4)) {
+#if BLOCK_SIZE > 1
+                        uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
+                        uint ib = coord / BLOCK_SIZE;
+                        uint iqs = (coord % BLOCK_SIZE);
+                        V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
+#else
+                        V_Tf = f16vec4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
+#endif
+                    }
+
+                    kvsh[c * kvsh_stride + d] = V_Tf;
+                }
+            }
+        }
+        barrier();
+
        const uint num_hsv_tiles = (HSV + MatBc * row_split - 1) / (MatBc * row_split); // round up

        // Each subgroup handles HSV/4 columns
        [[unroll]] for (uint32_t hsv_tile = 0; hsv_tile < num_hsv_tiles; ++hsv_tile) {
            const uint hsv_offset = (hsv_tile * row_split + gl_SubgroupID) * 16;

-            SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
+            coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> PVMat = coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);

            // Preload V tiles for [Bc, 16 * num subgroups]
            const uint v_rows = Bc;
            const uint v_total = v_rows * v_cols;
            const uint v_loads_per_thread = v_total / gl_WorkGroupSize.x;

+            // If SHMEM_STAGING is set, a Bc * HSV_pad size tile of V is loaded to shmem.
+            // If not, f16 V is loaded directly from global memory if aligned, otherwise
+            // staged through a Bc * MatBr size staging buffer.
+            // If V is not type f16, then it is always staged for dequantization.
+            if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
            // For f16, only preload if not aligned
            if (KV_bounds_check) {
@@ -428,44 +449,52 @@ void main() {

                if (!KV_bounds_check || (v_row < KV && v_col < HSV)) {
 #if BLOCK_SIZE > 1
-                    ksh[row * vsh_stride + col] = f16vec4(dequantize4(ib, iqs, v_offset, BINDING_IDX_V));
+                    kvsh[row * vsh_stride + col] = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
 #else
-                    ksh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
+                    kvsh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
 #endif
                } else {
-                    ksh[row * vsh_stride + col] = f16vec4(0.0f);
+                    kvsh[row * vsh_stride + col] = f16vec4(0.0f);
                }
            }
+
 #if BLOCK_SIZE == 1
            }
 #endif
-
+            }
            barrier();

-            [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
-                coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
+            const uint o_offset = gl_SubgroupID * MatBr / 4;

+            if (hsv_offset < HSV_pad) {
+                [[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
+                    coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
+
+                    if (SHMEM_STAGING == 0) {
 #if BLOCK_SIZE == 1
-                if (!KV_bounds_check) {
-                    // F16 values can be loaded directly from global memory
-                    const uint v_tile_row = j * Bc + bc_chunk * MatBc;
-                    const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
-                    coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
-                } else
+                    if (!KV_bounds_check) {
+                        // F16 values can be loaded directly from global memory
+                        const uint v_tile_row = j * Bc + bc_chunk * MatBc;
+                        const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
+                        coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
+                    } else
 #endif
-                {
-                    const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
-                    coopMatLoad(QMat, ksh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    {
+                        const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
+                        coopMatLoad(QMat, kvsh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    }
+                    } else {
+                        const uint v_tile_offset = bc_chunk * MatBc * kvsh_stride + (hsv_tile * row_split + gl_SubgroupID) * (MatBc / 4);
+                        coopMatLoad(QMat, kvsh, v_tile_offset, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
+                    }
+
+                    PVMat = coopMatMulAdd(KMat, QMat, PVMat);
                }

-                SfMat = coopMatMulAdd(KMat, QMat, SfMat);
+                // Store PVMat to pvsh and load into Of
+                coopMatStore(PVMat, pvsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
            }

-            // Store SfMat to sfsh and load into Of
-            const uint osh_stride = row_split * MatBc / 4;
-            const uint o_offset = gl_SubgroupID * MatBc / 4;
-            coopMatStore(SfMat, sfsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
-
            barrier();

            const uint hsv_per_tile = row_split * MatBc;
@@ -484,7 +513,7 @@ void main() {

                    if (hsv_col >= hsv_base && hsv_col < hsv_base + hsv_per_tile && hsv_col < HSV) {
                        const uint local_hsv = (hsv_col - hsv_base) / 4;
-                        Of[r][d_local] += ACC_TYPEV4(sfsh[row * osh_stride + local_hsv]);
+                        Of[r][d_local] += pvsh[row * osh_stride + local_hsv];
                    }
                }
            }
@@ -500,27 +529,48 @@ void main() {
    // 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) {
-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;

-        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            if (tile_row(r) < N) {
-                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
-                    const uint d = d0 + col_tid;
-                    if (d >= HSV/4) break;
-                    const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                if (tile_row(r) < N) {
+                    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                        const uint d = d0 + col_tid;
+                        if (d >= HSV/4) break;
+                        const uint d_local = d0 / threads_per_rowgroup;
+                        gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
                    }
                }
            }
-        }

-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        [[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);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            [[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);
+                }
+            }
+        } else {
+            [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
+                const uint row = tile_row(r);
+                const uint global_row = i * Br + row;
+
+                if (global_row < N) {
+                    uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
+
+                    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
+                        const uint d = d0 + col_tid;
+                        if (d >= HSV/4) break;
+                        data_ov4[o_offset + iq2 * HSV/4 + d] = D_TYPEV4(Of[r][d/threads_per_rowgroup]);
+                    }
+                }
+
+                if (global_row < N && col_tid == 0) {
+                    uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+                    data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
+                    data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
+                }
            }
        }

@@ -539,7 +589,7 @@ void main() {

                [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
                    const uint d_local = d0 / threads_per_rowgroup;
-                    Of[r][d_local] *= ACC_TYPE(ms);
+                    Of[r][d_local] *= float16_t(ms);
                }
            } else {
                vs = exp(sink - Mf[r]);
@@ -557,14 +607,14 @@ void main() {
    [[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
        const uint d_local = d0 / threads_per_rowgroup;
        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d_local] *= ACC_TYPE(Lfrcp[r]);
-#if defined(ACC_TYPE_MAX)
-            Of[r][d_local] = clamp(Of[r][d_local], -ACC_TYPE_MAX, ACC_TYPE_MAX);
+            Of[r][d_local] *= float16_t(Lfrcp[r]);
+#if defined(FLOAT_TYPE_MAX)
+            Of[r][d_local] = clamp(Of[r][d_local], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
 #endif
        }
    }

-    uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
+    uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;

    if (p.gqa_ratio > 1) {
        [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -573,9 +623,7 @@ void main() {
                    const uint d = d0 + col_tid;
                    if (d >= HSV / 4) break;
                    const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
-                    }
+                    gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
                }
            }
        }
@@ -586,9 +634,7 @@ void main() {
                    const uint d = d0 + col_tid;
                    if (d >= HSV / 4) break;
                    const uint d_local = d0 / threads_per_rowgroup;
-                    [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4 * d + comp] = D_TYPE(Of[r][d_local][comp]);
-                    }
+                    data_ov4[o_offset + (iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV) / 4 + d] = D_TYPEV4(Of[r][d_local]);
                }
            }
        }
@@ -72,6 +72,28 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
    return elem;
 }

+// Store O values for non-GQA split_k. Rows are tokens, not heads.
+D_TYPE perElemOpNonGqaSplitKStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t unused, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c < HSV) {
+        uint32_t o_off = HSV * p.ne1
+            * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[o_off + iq2 * HSV + c] = D_TYPE(elem);
+    }
+    return elem;
+}
+
+// Store L/M values for non-GQA split_k.
+ACC_TYPE perElemOpNonGqaSplitKStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t lm_base, const in uint32_t iq2, const in uint32_t N) {
+    uint32_t global_row = i * Br + r;
+    if (global_row < N && c == 0) {
+        uint32_t lm_off = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3
+            + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
+        data_o[lm_off + lm_base + iq2] = D_TYPE(elem);
+    }
+    return elem;
+}
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
    init_iq_shmem(gl_WorkGroupSize);
@@ -290,13 +312,19 @@ void main() {
    if (p.k_num > 1) {
        coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);

-        // note: O and Q have swapped coord 1,2.
-        uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
+        if (p.gqa_ratio > 1) {
+            // note: O and Q have swapped coord 1,2.
+            uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);

-        o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
-        coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
-        coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+            o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
+            coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
+        } else {
+            coopMatPerElementNV(O_D, O_D, perElemOpNonGqaSplitKStore, 0u, iq2, N);
+            coopMatPerElementNV(L, L, perElemOpNonGqaSplitKStoreCol0, 0u, iq2, N);
+            coopMatPerElementNV(M, M, perElemOpNonGqaSplitKStoreCol0, p.ne1, iq2, N);
+        }
        return;
    }

@@ -32,6 +32,7 @@ layout (push_constant) uniform parameter
    uint expert_i1;
    uint nbi1;
 #else
+    uint base_work_group_y;
    uint ne02;
    uint ne12;
    uint broadcast2;
@@ -45,9 +46,9 @@ uint expert_id;

 void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #ifdef MUL_MAT_ID
-    const uint expert_i0 = gl_GlobalInvocationID.y;
+    const uint expert_i0 = gl_WorkGroupID.y;
 #else
-    const uint batch_idx = gl_GlobalInvocationID.y;
+    const uint batch_idx = gl_WorkGroupID.y + p.base_work_group_y;
 #endif

 #ifndef MUL_MAT_ID
@@ -90,6 +90,8 @@ layout (push_constant) uniform parameter
    uint nbi1;
    uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
    uint k_split;
    uint ne02;
    uint ne12;
@@ -139,7 +141,7 @@ void main() {
    const uint ic = gl_WorkGroupID.y;

 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
    if (ic * BN >= data_expert_count[expert_idx]) {
        return;
    }
@@ -149,7 +151,7 @@ void main() {
 #endif

 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;

    const uint i13 = batch_idx / p.ne12;
    const uint i12 = batch_idx % p.ne12;
@@ -366,7 +368,7 @@ void main() {
    const uint dc = ic * BN + warp_c * WN;

 #ifndef MUL_MAT_ID
-    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif

 #ifdef COOPMAT
@@ -53,6 +53,8 @@ layout (push_constant) uniform parameter
    uint nbi1;
    uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
    uint k_split;
    uint ne02;
    uint ne12;
@@ -165,7 +167,9 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
        uint id = ids[iter++];
        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);

-        ballots_sh[gl_SubgroupID] = ballot;
+        if (gl_SubgroupInvocationID == 0) {
+            ballots_sh[gl_SubgroupID] = ballot;
+        }
        barrier();

        uint subgroup_base = 0;
@@ -197,7 +201,7 @@ void main() {
    const uint ic = gl_WorkGroupID.y;

 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
    if (ic * BN >= data_expert_count[expert_idx]) {
        return;
    }
@@ -215,7 +219,7 @@ void main() {
 #endif

 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;

    const uint i13 = batch_idx / p.ne12;
    const uint i12 = batch_idx % p.ne12;
@@ -255,7 +259,7 @@ void main() {
 #else
    uint pos_a = batch_idx_a * (p.batch_stride_a / QUANT_K);
    uint pos_b = batch_idx * p.batch_stride_b;
-    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif

    uint stride_a = p.stride_a / QUANT_K;
@@ -43,7 +43,9 @@ void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
        uint id = ids[iter++];
        uvec4 ballot = subgroupBallot(in_range && id == expert_idx);

-        ballots_sh[gl_SubgroupID] = ballot;
+        if (gl_SubgroupInvocationID == 0) {
+            ballots_sh[gl_SubgroupID] = ballot;
+        }
        barrier();

        uint subgroup_base = 0;
@@ -57,6 +57,8 @@ layout (push_constant) uniform parameter
    uint nbi1;
    uint ne11;
 #else
+    uint base_work_group_z;
+    uint num_batches;
    uint k_split;
    uint ne02;
    uint ne12;
@@ -108,7 +110,7 @@ void main() {
    const uint ic = gl_WorkGroupID.y;

 #ifdef MUL_MAT_ID
-    const uint expert_idx = gl_GlobalInvocationID.z;
+    const uint expert_idx = gl_WorkGroupID.z;
    if (ic * BN >= data_expert_count[expert_idx]) {
        return;
    }
@@ -118,7 +120,7 @@ void main() {
 #endif

 #ifndef MUL_MAT_ID
-    const uint batch_idx = gl_GlobalInvocationID.z;
+    const uint batch_idx = gl_WorkGroupID.z + p.base_work_group_z;

    const uint i13 = batch_idx / p.ne12;
    const uint i12 = batch_idx % p.ne12;
@@ -276,7 +278,7 @@ void main() {
    const uint dc = ic * BN + warp_c * WN;

 #ifndef MUL_MAT_ID
-    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * p.num_batches;
 #endif

    [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
@@ -595,8 +595,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 }

 void process_shaders() {
-    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
-
    // matmul
    for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
        // No coopmats
@@ -622,49 +620,63 @@ void process_shaders() {
        }
    }

-    // flash attention
-    for (const auto& f16acc : {false, true}) {
-        std::map<std::string, std::string> fa_base_dict = base_dict;
-        fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
-        fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
-        if (f16acc) {
-            fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
+    for (const bool& fp16 : {false, true}) {
+        std::map<std::string, std::string> base_dict;
+        if (fp16) {
+            base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
+        } else {
+            base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV4", "vec4"}};
        }

-        for (const auto& tname : type_names) {
-            if (tname == "bf16") continue;
-
-#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
-            if (tname == "f16") {
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
-                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, true, f16acc);
-            } else {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
+        // flash attention
+        for (const bool& f16acc : {false, true}) {
+            std::map<std::string, std::string> fa_base_dict = base_dict;
+            fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
+            fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
+            if (fp16 && f16acc) {
+                fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
            }
+
+            for (const auto& tname : type_names) {
+                if (tname == "bf16") continue;
+
+                if (fp16) {
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
+                } else {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, 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(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"COOPMAT", "1"}}), true, true, false, f16acc);
-            } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
-            }
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
+                } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
+                }
 #endif
-            if (tname == "f16") {
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
-                    merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, false, f16acc);
-            } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
-                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
-                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
-                    merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
+                }
+
+                if (tname == "f16") {
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                        merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
+                } else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
+                    std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                    string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
+                        merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, false, f16acc);
+                }
            }
        }
    }

+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
+
    for (const auto& tname : type_names) {
        // mul mat vec
        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
@@ -1,5 +1,4 @@
-#decl(BYTE_HELPERS)
-
+#ifdef BYTE_HELPERS
 fn get_byte(value: u32, index: u32) -> u32 {
    return (value >> (index * 8)) & 0xFF;
 }
@@ -7,76 +6,74 @@ fn get_byte(value: u32, index: u32) -> u32 {
 fn get_byte_i32(value: u32, index: u32) -> i32 {
    return bitcast<i32>(((value >> (index * 8)) & 0xFF) << 24) >> 24;
 }
+#endif

-#enddecl(BYTE_HELPERS)
-
-#decl(Q4_0_T)
+#ifdef Q4_0_T
 struct q4_0 {
    d: f16,
    qs: array<f16, 8>
 };
-#enddecl(Q4_0_T)
+#endif

-#decl(Q4_1_T)
+#ifdef Q4_1_T
 struct q4_1 {
    d: f16,
    m: f16,
    qs: array<u32, 4>
 };
-#enddecl(Q4_1_T)
+#endif

-#decl(Q5_0_T)
+#ifdef Q5_0_T
 struct q5_0 {
    d: f16,
    qh: array<f16, 2>,
    qs: array<f16, 8>
 };
-#enddecl(Q5_0_T)
+#endif

-#decl(Q5_1_T)
+#ifdef Q5_1_T
 struct q5_1 {
    d: f16,
    m: f16,
    qh: u32,
    qs: array<u32, 4>
 };
-#enddecl(Q5_1_T)
+#endif

-#decl(Q8_0_T)
+#ifdef Q8_0_T
 struct q8_0 {
    d: f16,
    qs: array<f16, 16>
 };
-#enddecl(Q8_0_T)
+#endif

-#decl(Q8_1_T)
+#ifdef Q8_1_T
 struct q8_1 {
    d: f16,
    m: f16,
    qs: array<u32, 8>
 };
-#enddecl(Q8_1_T)
+#endif

-#decl(Q2_K_T)
-struct q2_k {
+#ifdef Q2_K_T
+struct q2_K {
    scales: array<u32, 4>,
    qs: array<u32, 16>,
    d: f16,
    dmin: f16
 };
-#enddecl(Q2_K_T)
+#endif

-#decl(Q3_K_T)
-struct q3_k {
+#ifdef Q3_K_T
+struct q3_K {
    hmask: array<f16, 16>,
    qs: array<f16, 32>,
    scales: array<f16, 6>,
    d: f16
 };
-#enddecl(Q3_K_T)
-
-#decl(Q45_K_SCALE_MIN)
+#endif

+#if defined(Q4_K_SCALE_MIN) || defined(Q5_K_SCALE_MIN)
 fn get_scale_min(is: u32, scales: array<u32, 3>) -> vec2<f32> {
    if (is < 4) {
        let sc_byte = get_byte(scales[is / 4], is % 4);
@@ -91,69 +88,67 @@ fn get_scale_min(is: u32, scales: array<u32, 3>) -> vec2<f32> {
        return vec2(f32(sc), f32(m));
    }
 }
-
-#enddecl(Q45_K_SCALE_MIN)
-
-#decl(Q4_K_T)
-struct q4_k {
+#endif
+#ifdef Q4_K_T
+struct q4_K {
    d: f16,
    dmin: f16,
    scales: array<u32, 3>,
    qs: array<u32, 32>
 };
-#enddecl(Q4_K_T)
+#endif

-#decl(Q5_K_T)
-struct q5_k {
+#ifdef Q5_K_T
+struct q5_K {
    d: f16,
    dmin: f16,
    scales: array<u32, 3>,
    qh: array<u32, 8>,
    qs: array<u32, 32>
 };
-#enddecl(Q5_K_T)
+#endif

-#decl(Q6_K_T)
-struct q6_k {
+#ifdef Q6_K_T
+struct q6_K {
    ql: array<f16, 64>,
    qh: array<f16, 32>,
    scales: array<f16, 8>,
    d: f16
 };
-#enddecl(Q6_K_T)
+#endif

-#decl(IQ2_XXS_T)
+#ifdef IQ2_XXS_T
 struct iq2_xxs {
    d: f16,
    qs: array<f16, 32>
 };
-#enddecl(IQ2_XXS_T)
+#endif

-#decl(IQ2_XS_T)
+#ifdef IQ2_XS_T
 struct iq2_xs {
    d: f16,
    qs: array<f16, 32>,
    scales: array<f16, 4>
 };
-#enddecl(IQ2_XS_T)
+#endif

-#decl(IQ2_S_T)
+#ifdef IQ2_S_T
 struct iq2_s {
    d: f16,
    qs: array<f16, 32>,
    qh: array<f16, 4>,
    scales: array<f16, 4>
 };
-#enddecl(IQ2_S_T)
+#endif

-#decl(IQ3_XSS_T)
+#ifdef IQ3_XXS_T
 struct iq3_xxs {
    d: f16,
    qs: array<f16, 48>
 };
-#enddecl(IQ3_XSS_T)
+#endif

-#decl(IQ3_S_T)
+#ifdef IQ3_S_T
 struct iq3_s {
    d: f16,
    qs: array<f16, 32>,
@@ -161,41 +156,41 @@ struct iq3_s {
    signs: array<f16, 16>,
    scales: array<f16, 2>
 };
-#enddecl(IQ3_S_T)
+#endif

-#decl(IQ1_S_T)
+#ifdef IQ1_S_T
 struct iq1_s {
    d: f16,
    qs: array<f16, 16>,
    qh: array<f16, 8>
 };
-#enddecl(IQ1_S_T)
+#endif

-#decl(IQ1_M_T)
+#ifdef IQ1_M_T
 struct iq1_m {
    qs: array<u32, 8>,
    qh: array<u32, 4>,
    scales: array<u32, 2>
 };
-#enddecl(IQ1_M_T)
+#endif

-#decl(IQ4_NL_T)
+#ifdef IQ4_NL_T
 struct iq4_nl {
    d: f16,
    qs: array<f16, 8>,
 };
-#enddecl(IQ4_NL_T)
+#endif

-#decl(IQ4_XS_T)
+#ifdef IQ4_XS_T
 struct iq4_xs {
    d: f16,
    scales_h: f16,
    scales_l: u32,
    qs: array<u32, 32>
 };
-#enddecl(IQ4_XS_T)
+#endif

-#decl(IQ23_TABLES)
+#if defined(IQ2_XXS_TABLES) || defined(IQ2_XS_TABLES) || defined(IQ2_S_TABLES) || defined(IQ3_XXS_TABLES) || defined(IQ3_S_TABLES)
 const kmask_iq2xs : array<u32, 2> = array<u32, 2>(
    0x08040201u, // 1, 2, 4, 8
    0x80402010u  // 16, 32, 64, 128
@@ -211,9 +206,9 @@ const ksigns_iq2xs: array<u32, 32> = array<u32, 32>(
    0x63e2e160,0xe76665e4,0xeb6a69e8,0x6feeed6c,
    0xf37271f0,0x77f6f574,0x7bfaf978,0xff7e7dfc
 );
-#enddecl(IQ23_TABLES)
+#endif

-#decl(IQ2_XXS_GRID)
+#ifdef IQ2_XXS_GRID
 const iq2xxs_grid = array<u32, 512>(
    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x082b0808, 0x08080808,
@@ -280,9 +275,9 @@ const iq2xxs_grid = array<u32, 512>(
    0x0808082b, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b081919, 0x2b2b0808, 0x08082b19, 0x2b2b0819,
    0x08080808, 0x2b2b082b, 0x08192b08, 0x2b2b1908, 0x19190808, 0x2b2b2b08, 0x08081908, 0x2b2b2b19
 );
-#enddecl(IQ2_XXS_GRID)
+#endif

-#decl(IQ2_XS_GRID)
+#ifdef IQ2_XS_GRID
 const iq2xs_grid = array<u32, 1024>(
    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
@@ -413,9 +408,9 @@ const iq2xs_grid = array<u32, 1024>(
    0x2b2b2b08, 0x2b2b2b08, 0x08081908, 0x2b2b2b19, 0x2b081908, 0x2b2b2b19, 0x2b08192b, 0x2b2b2b19,
    0x082b2b08, 0x2b2b2b2b, 0x082b2b2b, 0x2b2b2b2b, 0x2b190819, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
 );
-#enddecl(IQ2_XS_GRID)
+#endif

-#decl(IQ2_S_GRID)
+#ifdef IQ2_S_GRID
 const iq2s_grid = array<u32, 2048>(
    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
@@ -674,10 +669,9 @@ const iq2s_grid = array<u32, 2048>(
    0x2b08192b, 0x2b2b2b19, 0x08082b08, 0x2b2b2b2b, 0x08082b2b, 0x2b2b2b2b, 0x082b0808, 0x2b2b2b2b,
    0x082b082b, 0x2b2b2b2b, 0x082b2b08, 0x2b2b2b2b, 0x2b082b08, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
 );
-#enddecl(IQ2_S_GRID)
-
-#decl(IQ3_XSS_GRID)
+#endif

+#ifdef IQ3_XXS_GRID
 const iq3xxs_grid = array<u32, 256>(
    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
@@ -712,10 +706,9 @@ const iq3xxs_grid = array<u32, 256>(
    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04
 );
-#enddecl(IQ3_XSS_GRID)
-
-#decl(IQ3_S_GRID)
+#endif

+#ifdef IQ3_S_GRID
 const iq3s_grid = array<u32, 512>(
    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
@@ -782,9 +775,9 @@ const iq3s_grid = array<u32, 512>(
    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101
 );
-#enddecl(IQ3_S_GRID)
+#endif

-#decl(IQ1_GRID)
+#if defined(IQ1_S_GRID) || defined(IQ1_M_GRID)

 const IQ1_DELTA: f32 = 0.125;

@@ -919,12 +912,12 @@ const iq1_grid = array<u32, 1024>(
    0x55dd55df, 0x55d555d7, 0x5503550c, 0x557f5501, 0x5577557d, 0x55405575, 0x555d555f, 0x55555557
 );

-#enddecl(IQ1_GRID)
+#endif

-#decl(IQ4_GRID)
+#if defined(IQ4_NL_GRID) || defined(IQ4_XS_GRID)

 const kvalues_iq4nl = array<i32, 16>(
    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113
 );

-#enddecl(IQ4_GRID)
+#endif
@@ -56,12 +56,46 @@ def expand_includes(shader, input_dir):
    return include_pattern.sub(replacer, shader)


-def write_shader(shader_name, shader_code, output_dir, outfile):
+def chunk_shader(shader_code, max_chunk_len=60000):
+    """Split shader_code into safe raw-string sized chunks."""
+    return [shader_code[i : i + max_chunk_len] for i in range(0, len(shader_code), max_chunk_len)]
+
+
+def raw_delim(shader_code):
+    """Pick a raw-string delimiter that does not appear in the shader."""
+    delim = "wgsl"
+    while f"){delim}\"" in shader_code:
+        delim += "_x"
+    return delim
+
+
+def write_shader(shader_name, shader_code, output_dir, outfile, input_dir):
+    shader_code = expand_includes(shader_code, input_dir)
+
    if output_dir:
        wgsl_filename = os.path.join(output_dir, f"{shader_name}.wgsl")
        with open(wgsl_filename, "w", encoding="utf-8") as f_out:
            f_out.write(shader_code)
-    outfile.write(f'const char* wgsl_{shader_name} = R"({shader_code})";\n\n')
+
+    delim = raw_delim(shader_code)
+    chunks = chunk_shader(shader_code)
+
+    if len(chunks) == 1:
+        outfile.write(f'const char* wgsl_{shader_name} = R"{delim}({shader_code}){delim}";\n\n')
+    else:
+        for idx, chunk in enumerate(chunks):
+            outfile.write(f'static const char wgsl_{shader_name}_part{idx}[] = R"{delim}({chunk}){delim}";\n\n')
+        outfile.write(f'static const std::string& wgsl_{shader_name}_str() {{\n')
+        outfile.write('    static const std::string s = []{\n')
+        outfile.write('        std::string tmp;\n')
+        outfile.write(f'        tmp.reserve({len(shader_code)});\n')
+        for idx in range(len(chunks)):
+            outfile.write(f'        tmp.append(wgsl_{shader_name}_part{idx});\n')
+        outfile.write('        return tmp;\n')
+        outfile.write('    }();\n')
+        outfile.write('    return s;\n')
+        outfile.write('}\n')
+        outfile.write(f'const char* wgsl_{shader_name} = wgsl_{shader_name}_str().c_str();\n\n')


 def generate_variants(fname, input_dir, output_dir, outfile):
@@ -74,7 +108,7 @@ def generate_variants(fname, input_dir, output_dir, outfile):
    try:
        variants = ast.literal_eval(extract_block(text, "VARIANTS"))
    except ValueError:
-        write_shader(shader_base_name, text, output_dir, outfile)
+        write_shader(shader_base_name, text, output_dir, outfile, input_dir)
    else:
        try:
            decls_map = parse_decls(extract_block(text, "DECLS"))
@@ -123,7 +157,7 @@ def generate_variants(fname, input_dir, output_dir, outfile):
                output_name = f"{shader_base_name}_" + variant["REPLS"]["TYPE"]
            else:
                output_name = shader_base_name
-            write_shader(output_name, final_shader, output_dir, outfile)
+            write_shader(output_name, final_shader, output_dir, outfile, input_dir)


 def main():
@@ -137,7 +171,8 @@ def main():
        os.makedirs(args.output_dir, exist_ok=True)

    with open(args.output_file, "w", encoding="utf-8") as out:
-        out.write("// Auto-generated shader embedding\n\n")
+        out.write("// Auto-generated shader embedding\n")
+        out.write("#include <string>\n\n")
        for fname in sorted(os.listdir(args.input_dir)):
            if fname.endswith(".wgsl"):
                generate_variants(fname, args.input_dir, args.output_dir, out)
@@ -1,222 +1,31 @@
-#define(VARIANTS)
+enable f16;
+#include "common_decls.tmpl"

-[
-  {
-    "SHADER_SUFFIX": "f32_vec",
-    "REPLS": {
-      "TYPE" : "vec4<f32>",
-      "DST_TYPE": "vec4<f32>",
-      "BLOCK_SIZE": 4
-    },
-    "DECLS": ["F32_VEC"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f32",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 1
-    },
-    "DECLS": ["F32"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "f16",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 1
-    },
-    "DECLS": ["F16"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "i32",
-      "DST_TYPE": "i32",
-      "BLOCK_SIZE": 1
-    },
-    "DECLS": ["I32"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q4_0",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q4_0_T", "Q4_0"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q4_1",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q4_1_T", "Q4_1"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q5_0",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q5_0_T", "Q5_0"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q5_1",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q5_1_T", "Q5_1"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q8_0",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q8_0_T", "Q8_0"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q2_k",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q2_K_T", "Q2_K"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q3_k",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q3_K_T", "Q3_K"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q4_k",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K_T", "Q4_K"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q5_k",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K_T", "Q5_K"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "q6_k",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q6_K_T", "Q6_K"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "iq2_xxs",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS_GRID", "IQ2_XXS_T", "IQ2_XXS"]
-  },
-  {
-    "REPLS": {
-      "TYPE" : "iq2_xs",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS_GRID", "IQ2_XS_T", "IQ2_XS"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq2_s",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S_GRID", "IQ2_S_T", "IQ2_S"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq3_xxs",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS_GRID", "IQ3_XSS_T", "IQ3_XSS"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq3_s",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S_GRID", "IQ3_S_T", "IQ3_S"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq1_s",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_S_T", "IQ1_S"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq1_m",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_M_T", "IQ1_M"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq4_nl",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 32,
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_NL_T", "IQ4_NL"]
-  },
-  {
-    "REPLS": {
-      "TYPE": "iq4_xs",
-      "DST_TYPE": "f32",
-      "BLOCK_SIZE": 256,
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_XS_T", "IQ4_XS"]
-  }
-]
-
-#end(VARIANTS)
-
-#define(DECLS)
-
-#decl(F32_VEC)
+#ifdef F32_VEC
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    dst[(dst_base / 4) + offset] = src[(src_base / 4) + offset];
 }
-#enddecl(F32_VEC)
+#endif

-#decl(F32)
+#ifdef F32
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    dst[dst_base + offset] = src[src_base + offset];
 }
-#enddecl(F32)
+#endif

-#decl(F16)
+#ifdef F16
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    dst[dst_base + offset] = f32(src[src_base + offset]);
 }
-#enddecl(F16)
+#endif

-#decl(I32)
+#ifdef I32
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    dst[dst_base + offset] = src[src_base + offset];
 }
-#enddecl(I32)
+#endif

-#decl(Q4_0)
+#ifdef Q4_0
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block_q4_0 = src[src_base + offset];
    let d = f32(block_q4_0.d);
@@ -232,9 +41,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q4_0)
+#endif

-#decl(Q4_1)
+#ifdef Q4_1
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block_q4_1 = src[src_base + offset];
    let d = f32(block_q4_1.d);
@@ -251,9 +60,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q4_1)
+#endif

-#decl(Q5_0)
+#ifdef Q5_0
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block_q5_0 = src[src_base + offset];
    let d = f32(block_q5_0.d);
@@ -272,10 +81,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
+#endif

-#enddecl(Q5_0)
-
-#decl(Q5_1)
+#ifdef Q5_1
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block_q5_1 = src[src_base + offset];
    let d = f32(block_q5_1.d);
@@ -294,9 +102,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q5_1)
+#endif

-#decl(Q8_0)
+#ifdef Q8_0
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block_q8_0 = src[src_base + offset];
    let d = f32(block_q8_0.d);
@@ -310,9 +118,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q8_0)
+#endif

-#decl(Q2_K)
+#ifdef Q2_K
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -340,9 +148,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q2_K)
+#endif

-#decl(Q3_K)
+#ifdef Q3_K
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -398,9 +206,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q3_K)
+#endif

-#decl(Q4_K)
+#ifdef Q4_K
 // 8 blocks of 32 elements each
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
@@ -425,9 +233,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q4_K)
+#endif

-#decl(Q5_K)
+#ifdef Q5_K
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -455,9 +263,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(Q5_K)
+#endif

-#decl(Q6_K)
+#ifdef Q6_K
 // 16 blocks of 16 elements each
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
@@ -511,10 +319,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        sc_b_idx += 8;
    }
 }
+#endif

-#enddecl(Q6_K)
-
-#decl(IQ2_XXS)
+#ifdef IQ2_XXS
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -536,9 +343,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(IQ2_XXS)
+#endif

-#decl(IQ2_XS)
+#ifdef IQ2_XS
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -568,9 +375,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(IQ2_XS)
+#endif

-#decl(IQ2_S)
+#ifdef IQ2_S
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -608,10 +415,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
+#endif

-#enddecl(IQ2_S)
-
-#decl(IQ3_XSS)
+#ifdef IQ3_XXS
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -638,9 +444,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(IQ3_XSS)
+#endif

-#decl(IQ3_S)
+#ifdef IQ3_S
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -683,9 +489,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
-#enddecl(IQ3_S)
+#endif

-#decl(IQ1_S)
+#ifdef IQ1_S
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -707,10 +513,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
+#endif

-#enddecl(IQ1_S)
-
-#decl(IQ1_M)
+#ifdef IQ1_M
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];

@@ -751,10 +556,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        }
    }
 }
+#endif

-#enddecl(IQ1_M)
-
-#decl(IQ4_NL)
+#ifdef IQ4_NL
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -770,9 +574,9 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        dst_i++;
    }
 }
-#enddecl(IQ4_NL)
+#endif

-#decl(IQ4_XS)
+#ifdef IQ4_XS
 fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
    let block = src[src_base + offset];
    let d = f32(block.d);
@@ -791,24 +595,16 @@ fn copy_elements(src_base: u32, dst_base: u32, offset: u32) {
        dst_i += 16;
    }
 }
-#enddecl(IQ4_XS)
-
-#end(DECLS)
-
-#define(SHADER)
-
-enable f16;
-
-DECLS
+#endif

@group(0) @binding(0)
-var<storage, read_write> src: array<{{TYPE}}>;
+var<storage, read_write> src: array<SRC_TYPE>;

@group(0) @binding(1)
 var<storage, read_write> idx: array<i32>;

@group(0) @binding(2)
-var<storage, read_write> dst: array<{{DST_TYPE}}>;
+var<storage, read_write> dst: array<DST_TYPE>;

 struct Params {
    offset_src: u32, // in elements
@@ -842,8 +638,7 @@ struct Params {
@group(0) @binding(3)
 var<uniform> params: Params;

-override wg_size: u32;
-@compute @workgroup_size(wg_size)
+@compute @workgroup_size(WG_SIZE)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    if (gid.x >= params.n_rows * params.ne2 * params.ne3) {
        return;
@@ -866,9 +661,8 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    let i_src_row = params.offset_src + idx_val * params.stride_src1 + i_dst2 * params.stride_src2 + i_dst3 * params.stride_src3;
    let i_dst_row = params.offset_dst + i_dst1 * params.stride_dst1 + i_dst2 * params.stride_dst2 + i_dst3 * params.stride_dst3;

-    for (var i: u32 = 0; i < params.ne0/{{BLOCK_SIZE}}; i++) {
+    for (var i: u32 = 0; i < params.ne0/BLOCK_SIZE; i++) {
      copy_elements(i_src_row, i_dst_row, i);
    }
 }

-#end(SHADER)
@@ -1,195 +1,24 @@
-#define(VARIANTS)
+enable f16;

-[
-  {
-    "REPLS": {
-      "SRC0_TYPE" : "f32",
-      "SRC1_TYPE" : "f32",
-      "BLOCK_SIZE" : 1
-    },
-    "DECLS" : ["FLOAT"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f16",
-      "BLOCK_SIZE" : 1
-    },
-    "DECLS" : ["FLOAT"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "BLOCK_SIZE" : 1
-    },
-    "DECLS" : ["FLOAT"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q4_0",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q4_0_T", "Q4_0"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q4_1",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q4_1_T", "Q4_1"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q5_0",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q5_0_T", "Q5_0"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q5_1",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q5_1_T", "Q5_1"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q8_0",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32
-    },
-    "DECLS": ["BYTE_HELPERS", "Q8_0_T", "Q8_0"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q2_k",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q2_K_T", "Q2_K"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q3_k",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q3_K_T", "Q3_K"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q4_k",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q4_K_T", "Q4_K"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q5_k",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["Q45_K_SCALE_MIN", "BYTE_HELPERS", "Q5_K_T", "Q5_K"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "q6_k",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "Q6_K_T", "Q6_K"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq2_xxs",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XXS_GRID", "IQ2_XXS_T", "IQ2_XXS"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq2_xs",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_XS_GRID", "IQ2_XS_T", "IQ2_XS"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq2_s",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ2_S_GRID", "IQ2_S_T", "IQ2_S"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq3_xxs",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_XSS_GRID", "IQ3_XSS_T", "IQ3_XSS"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq3_s",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ23_TABLES", "IQ3_S_GRID", "IQ3_S_T", "IQ3_S"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq1_s",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_S_T", "IQ1_S"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq1_m",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ1_GRID", "IQ1_M_T", "IQ1_M"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq4_nl",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 32,
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_NL_T", "IQ4_NL"]
-  },
-  {
-    "REPLS": {
-      "SRC0_TYPE": "iq4_xs",
-      "SRC1_TYPE": "f32",
-      "BLOCK_SIZE": 256,
-    },
-    "DECLS": ["BYTE_HELPERS", "IQ4_GRID", "IQ4_XS_T", "IQ4_XS"]
-  }
-]
+#include "common_decls.tmpl"

-#end(VARIANTS)
+#ifdef FLOAT
+const BLOCK_SIZE = 1u;

-#define(DECLS)
+#elif defined(Q4_0) || defined(Q4_1) || defined(Q5_0) || defined(Q5_1) || defined(Q8_0) || defined(Q8_1) || defined(IQ4_NL)
+const BLOCK_SIZE = 32u;

-#decl(FLOAT)
+#elif defined(Q2_K) || defined(Q3_K) || defined(Q4_K) || defined(Q5_K) || defined(Q6_K) || defined(IQ2_XXS) || defined(IQ2_XS) || defined(IQ2_S) || defined(IQ3_XXS) || defined(IQ3_S) || defined(IQ1_S) || defined(IQ1_M) || defined(IQ4_XS)
+const BLOCK_SIZE = 256u;
+#endif
+
+#ifdef FLOAT
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    return f32(src0[src0_idx_base + offset]) * f32(src1[src1_idx_base + offset]);
 }
-#enddecl(FLOAT)
+#endif

-#decl(Q4_0)
+#ifdef Q4_0
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q4_0 = src0[src0_idx_base + offset];
    let d = f32(block_q4_0.d);
@@ -207,9 +36,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q4_0)
+#endif

-#decl(Q4_1)
+#ifdef Q4_1
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q4_1 = src0[src0_idx_base + offset];
    let d = f32(block_q4_1.d);
@@ -228,9 +57,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q4_1)
+#endif

-#decl(Q5_0)
+#ifdef Q5_0
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q5_0 = src0[src0_idx_base + offset];
    let d = f32(block_q5_0.d);
@@ -251,9 +80,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q5_0)
+#endif

-#decl(Q5_1)
+#ifdef Q5_1
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q5_1 = src0[src0_idx_base + offset];
    let d = f32(block_q5_1.d);
@@ -274,9 +103,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q5_1)
+#endif

-#decl(Q8_0)
+#ifdef Q8_0
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q8_0 = src0[src0_idx_base + offset];
    let d = f32(block_q8_0.d);
@@ -292,9 +121,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q8_0)
+#endif

-#decl(Q8_1)
+#ifdef Q8_1
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block_q8_1 = src0[src0_idx_base + offset];
    let d = f32(block_q8_1.d);
@@ -311,9 +140,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(Q8_1)
+#endif

-#decl(Q2_K)
+#ifdef Q2_K
 // 16 blocks of 16 elements each
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
@@ -344,10 +173,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(Q2_K)
-
-#decl(Q3_K)
+#ifdef Q3_K
 // 16 blocks of 16 elements each
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
@@ -406,10 +234,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(Q3_K)
-
-#decl(Q4_K)
+#ifdef Q4_K
 // 8 blocks of 32 elements each
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
@@ -436,10 +263,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(Q4_K)
-
-#decl(Q5_K)
+#ifdef Q5_K
 // 8 blocks of 32 elements each
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
@@ -470,10 +296,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(Q5_K)
-
-#decl(Q6_K)
+#ifdef Q6_K
 // 16 blocks of 16 elements each
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
@@ -529,10 +354,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(Q6_K)
-
-#decl(IQ2_XXS)
+#ifdef IQ2_XXS
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -556,10 +380,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ2_XXS)
-
-#decl(IQ2_XS)
+#ifdef IQ2_XS
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -591,10 +414,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ2_XS)
-
-#decl(IQ2_S)
+#ifdef IQ2_S
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -634,11 +456,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-
-#enddecl(IQ2_S)
-
-#decl(IQ3_XSS)
+#ifdef IQ3_XXS
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -667,10 +487,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ3_XSS)
-
-#decl(IQ3_S)
+#ifdef IQ3_S
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -715,9 +534,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-#enddecl(IQ3_S)
+#endif

-#decl(IQ1_S)
+#ifdef IQ1_S
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -741,10 +560,10 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ1_S)

-#decl(IQ1_M)
+#ifdef IQ1_M
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];

@@ -787,10 +606,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ1_M)
-
-#decl(IQ4_NL)
+#ifdef IQ4_NL
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -808,10 +626,9 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
+#endif

-#enddecl(IQ4_NL)
-
-#decl(IQ4_XS)
+#ifdef IQ4_XS
 fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    let block = src0[src0_idx_base + offset];
    let d = f32(block.d);
@@ -832,16 +649,7 @@ fn multiply_add(src0_idx_base: u32, src1_idx_base: u32, offset: u32) -> f32 {
    }
    return sum;
 }
-
-#enddecl(IQ4_XS)
-
-#end(DECLS)
-
-#define(SHADER)
-
-enable f16;
-
-DECLS
+#endif

 struct MulMatParams {
    offset_src0: u32, // in elements/blocks
@@ -864,8 +672,8 @@ struct MulMatParams {
    broadcast3: u32
 };

-@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
@group(0) @binding(2) var<storage, read_write> dst: array<f32>; // M rows, N columns

@group(0) @binding(3) var<uniform> params: MulMatParams;
@@ -898,10 +706,8 @@ fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
    let src1_idx_base = params.offset_src1 + src13_idx * params.stride_13 + src12_idx * params.stride_12 + row * params.stride_11;

    var sum = 0.0;
-    for (var i: u32 = 0u; i < params.k/{{BLOCK_SIZE}}; i = i + 1u) {
+    for (var i: u32 = 0u; i < params.k/BLOCK_SIZE; i = i + 1u) {
        sum += multiply_add(src0_idx_base, src1_idx_base, i);
    }
    dst[params.offset_dst + dst3_idx * dst3_stride + dst2_idx * dst2_stride + row * params.m + col] = sum;
 }
-
-#end(SHADER)
@@ -1,58 +1,65 @@
-#decl(SHMEM_VEC)
+#ifdef VEC
+#define VEC_SIZE 4
+#define SHMEM_TYPE vec4<f16>
+#define DST_TYPE vec4<f32>
+#define SRC0_TYPE vec4<SRC0_INNER_TYPE>
+#define SRC1_TYPE vec4<SRC1_INNER_TYPE>
+
 fn store_shmem(val: vec4<f16>, idx: u32) {
    shmem[idx] = val.x;
    shmem[idx + 1] = val.y;
    shmem[idx + 2] = val.z;
    shmem[idx + 3] = val.w;
 }
-#enddecl(SHMEM_VEC)
+#endif
+
+#ifdef SCALAR
+#define VEC_SIZE 1
+#define SHMEM_TYPE f16
+#define DST_TYPE f32
+#define SRC0_TYPE SRC0_INNER_TYPE
+#define SRC1_TYPE SRC1_INNER_TYPE

-#decl(SHMEM_SCALAR)
 fn store_shmem(val: f16, idx: u32) {
    shmem[idx] = val;
 }
-#enddecl(SHMEM_SCALAR)
-
-#decl(INIT_SRC0_SHMEM_FLOAT)
+#endif

+#ifdef INIT_SRC0_SHMEM_FLOAT
 fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u32) {
-    for (var elem_idx = thread_id * {{VEC_SIZE}}; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+    for (var elem_idx = thread_id * VEC_SIZE; elem_idx < TILE_SRC0_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * VEC_SIZE) {
        let tile_m = elem_idx / TILE_K;
        let tile_k = elem_idx % TILE_K;
        let global_m = offset_m + tile_m;
        let global_k = k_outer + tile_k;
        let src0_idx = batch_offset + global_m * params.stride_01 + global_k;
        let src0_val = select( // taking a slight performance hit to avoid oob
-            {{SRC0_TYPE}}(0.0),
-            src0[src0_idx/{{VEC_SIZE}}],
+            SRC0_TYPE(0.0),
+            src0[src0_idx/VEC_SIZE],
            global_m < params.m && global_k < params.k);
-        store_shmem({{SHMEM_TYPE}}(src0_val), elem_idx);
+        store_shmem(SHMEM_TYPE(src0_val), elem_idx);
    }
 }
+#endif

-#enddecl(INIT_SRC0_SHMEM_FLOAT)
-
-#decl(INIT_SRC1_SHMEM)
-
+#ifdef INIT_SRC1_SHMEM_FLOAT
 fn init_shmem_src1(thread_id: u32, batch_offset: u32, offset_n: u32, k_outer: u32) {
-    for (var elem_idx = thread_id * {{VEC_SIZE}}; elem_idx < TILE_SRC1_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+    for (var elem_idx = thread_id * VEC_SIZE; elem_idx < TILE_SRC1_SHMEM; elem_idx += TOTAL_WORKGROUP_SIZE * VEC_SIZE) {
        let tile_n = elem_idx / TILE_K;
        let tile_k = elem_idx % TILE_K;
        let global_n = offset_n + tile_n;
        let global_k = k_outer + tile_k;
        let src1_idx = batch_offset + global_n * params.stride_11 + global_k;
        let src1_val = select(
-            {{SRC1_TYPE}}(0.0),
-            src1[src1_idx/{{VEC_SIZE}}],
+            SRC1_TYPE(0.0),
+            src1[src1_idx/VEC_SIZE],
            global_n < params.n && global_k < params.k);
-        store_shmem({{SHMEM_TYPE}}(src1_val), TILE_SRC0_SHMEM + elem_idx);
+        store_shmem(SHMEM_TYPE(src1_val), TILE_SRC0_SHMEM + elem_idx);
    }
 }
+#endif

-#enddecl(INIT_SRC1_SHMEM)
-
-#decl(INIT_SRC0_SHMEM_Q4_0)
-
+#ifdef INIT_SRC0_SHMEM_Q4_0
 const BLOCK_SIZE = 32u;
 // the number of blocks per k-tile. Note that this currently only works if TILE_K is a multiple of BLOCK_SIZE, which may need to be rethought for larger quantized types.
 override BLOCKS_K = TILE_K/BLOCK_SIZE;
@@ -93,5 +100,4 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
        }
    }
 }
-
-#enddecl(INIT_SRC0_SHMEM_Q4_0)
+#endif
@@ -1,115 +1,19 @@
-#define(VARIANTS)
-[
-  {
-    "SHADER_SUFFIX": "f32_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f32>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f32_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f32",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f16>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f16",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "q4_0_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["BYTE_HELPERS", "VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "q4_0_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["BYTE_HELPERS", "SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
-  }
-]
+enable f16;

-#end(VARIANTS)
+#include "common_decls.tmpl"
+#include "mul_mat_decls.tmpl"

-#define(DECLS)
-
-#decl(VEC)
+#ifdef VEC
 fn store_val(acc: array<array<f16, TILE_N>, TILE_M>, tn: u32, tm: u32) -> vec4<f32> {
    return vec4<f32>(f32(acc[tm][tn]), f32(acc[tm + 1][tn]), f32(acc[tm + 2][tn]), f32(acc[tm + 3][tn]));
 }
-#enddecl(VEC)
+#endif

-#decl(SCALAR)
+#ifdef SCALAR
 fn store_val(acc: array<array<f16, TILE_N>, TILE_M>, tn: u32, tm: u32) -> f32 {
    return f32(acc[tm][tn]);
 }
-#enddecl(SCALAR)
-
-#end(DECLS)
-
-#define(SHADER)
-enable f16;
+#endif

 struct MulMatParams {
    offset_src0: u32,
@@ -130,14 +34,12 @@ struct MulMatParams {
    broadcast3: u32
 };

-@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>; // M rows, N columns (transposed)
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)

@group(0) @binding(3) var<uniform> params: MulMatParams;

-DECLS
-
 fn get_local_n(thread_id: u32) -> u32 {
    return thread_id / WORKGROUP_SIZE_M;
 }
@@ -145,18 +47,9 @@ fn get_local_m(thread_id: u32) -> u32 {
    return thread_id % WORKGROUP_SIZE_M;
 }

-// TILE_M must be multiple of 4 for vec4 loads
-const TILE_M = {{WEBGPU_TILE_M}}u;
-const TILE_N = {{WEBGPU_TILE_N}}u;
-
-override WORKGROUP_SIZE_M: u32;
-override WORKGROUP_SIZE_N: u32;
-override TILE_K: u32;
-
-override TOTAL_WORKGROUP_SIZE = WORKGROUP_SIZE_M * WORKGROUP_SIZE_N;
-override TILE_SRC0_SHMEM = TILE_K * WORKGROUP_SIZE_M * TILE_M;
-override TILE_SRC1_SHMEM = TILE_K * WORKGROUP_SIZE_N * TILE_N;
-
+const TOTAL_WORKGROUP_SIZE = WORKGROUP_SIZE_M * WORKGROUP_SIZE_N;
+const TILE_SRC0_SHMEM = TILE_K * WORKGROUP_SIZE_M * TILE_M;
+const TILE_SRC1_SHMEM = TILE_K * WORKGROUP_SIZE_N * TILE_N;
 var<workgroup> shmem: array<f16, TILE_SRC0_SHMEM + TILE_SRC1_SHMEM>;

@compute @workgroup_size(TOTAL_WORKGROUP_SIZE)
@@ -233,15 +126,13 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
    for (var tn = 0u; tn < TILE_N; tn++) {
        let global_col = output_col_base + tn;
        if (global_col < params.n) {
-            for (var tm = 0u; tm < TILE_M; tm += {{VEC_SIZE}}) {
+            for (var tm = 0u; tm < TILE_M; tm += VEC_SIZE) {
                let global_row = output_row_base + tm;
                if (global_row < params.m) {
                    let dst_idx = dst_batch_offset + global_col * params.m + global_row;
-                    dst[dst_idx/{{VEC_SIZE}}] = store_val(acc, tn, tm);
+                    dst[dst_idx/VEC_SIZE] = store_val(acc, tn, tm);
                }
            }
        }
    }
 }
-
-#end(SHADER)
@@ -1,100 +1,12 @@
-#define(VARIANTS)
-[
-  {
-    "SHADER_SUFFIX": "f32_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f32>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f32_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f32",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f16>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f16",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_FLOAT", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "q4_0_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE" : "vec4<f32>",
-      "SHMEM_TYPE" : "vec4<f16>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["BYTE_HELPERS", "VEC", "SHMEM_VEC", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
-  },
-  {
-    "SHADER_SUFFIX": "q4_0_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE" : "f32",
-      "SHMEM_TYPE" : "f16",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["BYTE_HELPERS", "SCALAR", "SHMEM_SCALAR", "INIT_SRC0_SHMEM_Q4_0", "INIT_SRC1_SHMEM"]
-  }
-]
+diagnostic(off, chromium.subgroup_matrix_uniformity);
+enable f16;
+enable subgroups;
+enable chromium_experimental_subgroup_matrix;

-#end(VARIANTS)
+#include "common_decls.tmpl"
+#include "mul_mat_decls.tmpl"

-#define(DECLS)
-
-#decl(VEC)
+#ifdef VEC
 fn store_dst(shmem_idx: u32, dst_idx: u32) {
    dst[dst_idx] = vec4<f32>(
        f32(shmem[shmem_idx]),
@@ -103,21 +15,13 @@ fn store_dst(shmem_idx: u32, dst_idx: u32) {
        f32(shmem[shmem_idx + 3])
    );
 }
-#enddecl(VEC)
+#endif

-#decl(SCALAR)
+#ifdef SCALAR
 fn store_dst(shmem_idx: u32, dst_idx: u32) {
    dst[dst_idx] = f32(shmem[shmem_idx]);
 }
-#enddecl(SCALAR)
-
-#end(DECLS)
-
-#define(SHADER)
-diagnostic(off, chromium.subgroup_matrix_uniformity);
-enable f16;
-enable subgroups;
-enable chromium_experimental_subgroup_matrix;
+#endif

 struct MulMatParams {
    offset_src0: u32,
@@ -138,36 +42,19 @@ struct MulMatParams {
    broadcast3: u32
 };

-@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // M rows, K columns
-@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>; // M rows, N columns (transposed)
+// SRC0_TYPE and SRC1_TYPE are defined in mul_mat_decls, which is included
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)

@group(0) @binding(3) var<uniform> params: MulMatParams;

-DECLS
-
-// Note: These are string interpolated at build time, cannot use override constants due to limitations in
-// current Dawn version type definitions/matrix load requirements for constant memory sizes.
-const SUBGROUP_M = {{WEBGPU_SUBGROUP_M}}u;
-const SUBGROUP_N = {{WEBGPU_SUBGROUP_N}}u;
-// For portability we assume the max subgroup size, meaning some subgroups will be masked out if the
-// runtime subgroup size is smaller.
-const MAX_SUBGROUP_SIZE = {{WEBGPU_MAX_SUBGROUP_SIZE}}u;
-
-const EXPECTED_SUBGROUPS = SUBGROUP_M * SUBGROUP_N;
-
-const SUBGROUP_MATRIX_M_SIZE = {{WEBGPU_SG_MAT_M_SIZE}}u;
-const SUBGROUP_MATRIX_N_SIZE = {{WEBGPU_SG_MAT_N_SIZE}}u;
-const SUBGROUP_MATRIX_K_SIZE = {{WEBGPU_SG_MAT_K_SIZE}}u;
-
-const SUBGROUP_MATRIX_M = {{WEBGPU_SUBGROUP_MATRIX_M}}u;
-const SUBGROUP_MATRIX_N = {{WEBGPU_SUBGROUP_MATRIX_N}}u;
-
-const TILE_K = {{WEBGPU_TILE_K}}u;
-
 const WG_M_SG_TILE_SIZE = SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
 const WG_N_SG_TILE_SIZE = SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;

+// For portability we assume the max subgroup size, meaning some subgroups will be masked out if the
+// runtime subgroup size is smaller.
+const EXPECTED_SUBGROUPS = SUBGROUP_M * SUBGROUP_N;
 const TOTAL_WORKGROUP_SIZE = SUBGROUP_M * SUBGROUP_N * MAX_SUBGROUP_SIZE;
 const TILE_SRC0_SHMEM = TILE_K * SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
 const TILE_SRC1_SHMEM = TILE_K * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
@@ -285,7 +172,7 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,
    let tile_dst_row_base = wg_m * SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
    let tile_dst_col_base = wg_n * SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;

-    for (var idx = thread_id * {{VEC_SIZE}}; idx < total_tile_elems; idx += TOTAL_WORKGROUP_SIZE * {{VEC_SIZE}}) {
+    for (var idx = thread_id * VEC_SIZE; idx < total_tile_elems; idx += TOTAL_WORKGROUP_SIZE * VEC_SIZE) {
        let local_row = idx % WG_TILE_STRIDE;
        let local_col = idx / WG_TILE_STRIDE;

@@ -294,9 +181,8 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>,

        if (global_col < params.n && global_row < params.m) {
            let dst_idx = dst_batch_offset + global_col * params.m + global_row;
-            store_dst(idx, dst_idx/{{VEC_SIZE}});
+            store_dst(idx, dst_idx/VEC_SIZE);
        }
    }
 }

-#end(SHADER)
@@ -1,84 +1,17 @@
-#define(VARIANTS)
-[
-  {
-    "SHADER_SUFFIX": "f32_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f32>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE": "vec4<f32>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "f32_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f32",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE": "f32",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f32>",
-      "DST_TYPE": "vec4<f32>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE": "f32",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16_vec",
-    "REPLS": {
-      "SRC0_TYPE" : "vec4<f16>",
-      "SRC1_TYPE" : "vec4<f16>",
-      "DST_TYPE": "vec4<f32>",
-      "VEC_SIZE" : 4,
-    },
-    "DECLS": ["VEC", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "f16_f16",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f16",
-      "DST_TYPE": "f32",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["SCALAR", "MUL_ACC_FLOAT"]
-  },
-  {
-    "SHADER_SUFFIX": "q4_0_f32",
-    "REPLS": {
-      "SRC0_TYPE" : "f16",
-      "SRC1_TYPE" : "f32",
-      "DST_TYPE": "f32",
-      "VEC_SIZE" : 1,
-    },
-    "DECLS": ["BYTE_HELPERS", "SCALAR", "MUL_ACC_Q4_0"]
-  }
-]

-#end(VARIANTS)
+enable f16;

-#define(DECLS)
+#include "common_decls.tmpl"

-#decl(VEC)
-fn inner_dot(src0_val: {{SRC0_TYPE}}, src1_val: {{SRC1_TYPE}}) -> f32 {
-    return f32(dot({{SRC1_TYPE}}(src0_val), src1_val));
+#ifdef VEC
+
+#define VEC_SIZE 4
+#define DST_TYPE vec4<f32>
+#define SRC0_TYPE vec4<SRC0_INNER_TYPE>
+#define SRC1_TYPE vec4<SRC1_INNER_TYPE>
+
+fn inner_dot(src0_val: SRC0_TYPE, src1_val: SRC1_TYPE) -> f32 {
+    return f32(dot(SRC1_TYPE(src0_val), src1_val));
 }

 fn store_val(group_base: u32) -> vec4<f32> {
@@ -87,33 +20,37 @@ fn store_val(group_base: u32) -> vec4<f32> {
                     partial_sums[group_base + THREADS_PER_OUTPUT * 2],
                     partial_sums[group_base + THREADS_PER_OUTPUT * 3]);
 }
-#enddecl(VEC)
+#endif

-#decl(SCALAR)
-fn inner_dot(src0_val: {{SRC0_TYPE}}, src1_val: {{SRC1_TYPE}}) -> f32 {
+#ifdef SCALAR
+
+#define VEC_SIZE 1
+#define DST_TYPE f32
+#define SRC0_TYPE SRC0_INNER_TYPE
+#define SRC1_TYPE SRC1_INNER_TYPE
+
+fn inner_dot(src0_val: SRC0_TYPE, src1_val: SRC1_TYPE) -> f32 {
    return f32(src0_val) * f32(src1_val);
 }

 fn store_val(group_base: u32) -> f32 {
    return partial_sums[group_base];
 }
-#enddecl(SCALAR)
-
-#decl(MUL_ACC_FLOAT)
+#endif

+#ifdef MUL_ACC_FLOAT
 fn mul_acc(tig:u32, tile_size: u32, idx_base: u32, k_outer: u32) -> f32 {
    var local_sum = 0.0;
-    for (var i = tig * {{VEC_SIZE}}; i < tile_size; i += THREADS_PER_OUTPUT * {{VEC_SIZE}}) {
-        let a = src0[(idx_base + k_outer + i) / {{VEC_SIZE}}];
-        let b = shared_vector[i / {{VEC_SIZE}}];
+    for (var i = tig * VEC_SIZE; i < tile_size; i += THREADS_PER_OUTPUT * VEC_SIZE) {
+        let a = src0[(idx_base + k_outer + i) / VEC_SIZE];
+        let b = shared_vector[i / VEC_SIZE];
        local_sum += inner_dot(a, b);
    }
    return local_sum;
 }
+#endif

-#enddecl(MUL_ACC_FLOAT)
-
-#decl(MUL_ACC_Q4_0)
+#ifdef MUL_ACC_Q4_0

 const BLOCK_SIZE = 32;
 const NQ = 16u; // number of weights per thread
@@ -145,15 +82,7 @@ fn mul_acc(tig:u32, tile_size: u32, idx_base: u32, k_outer: u32) -> f32 {
    }
    return local_sum;
 }
-
-#enddecl(MUL_ACC_Q4_0)
-
-#end(DECLS)
-
-#define(SHADER)
-enable f16;
-
-DECLS
+#endif

 struct MulMatParams {
    offset_src0: u32,
@@ -174,22 +103,20 @@ struct MulMatParams {
    broadcast3: u32
 };

-@group(0) @binding(0) var<storage, read_write> src0: array<{{SRC0_TYPE}}>; // Matrix (M x K)
-@group(0) @binding(1) var<storage, read_write> src1: array<{{SRC1_TYPE}}>; // Vector (K x 1, transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<{{DST_TYPE}}>;  // Result vector (transposed)
+// SRC0_TYPE and SRC1_TYPE are defined in mul_mat_decls, which is included
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
+@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
+@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)

@group(0) @binding(3) var<uniform> params: MulMatParams;

-override WORKGROUP_SIZE: u32;
-override TILE_K: u32;
-override OUTPUTS_PER_WG: u32;
-override THREADS_PER_OUTPUT = WORKGROUP_SIZE / OUTPUTS_PER_WG;
+const THREADS_PER_OUTPUT = WG_SIZE / OUTPUTS_PER_WG;

 // Shared memory for collaborative loading and reduction
-var<workgroup> shared_vector: array<{{SRC1_TYPE}}, TILE_K/{{VEC_SIZE}}>;  // Cache vector tile
-var<workgroup> partial_sums: array<f32, WORKGROUP_SIZE>;   // For reduction
+var<workgroup> shared_vector: array<SRC1_TYPE, TILE_K/VEC_SIZE>;  // Cache vector tile
+var<workgroup> partial_sums: array<f32, WG_SIZE>;   // For reduction

-@compute @workgroup_size(WORKGROUP_SIZE)
+@compute @workgroup_size(WG_SIZE)
 fn main(
    @builtin(local_invocation_id) local_id: vec3<u32>,
    @builtin(workgroup_id) wg_id: vec3<u32>,
@@ -232,8 +159,8 @@ fn main(
        let tile_size = min(TILE_K, params.k - k_tile);

        // Cooperatively load vector tile into shared memory (all threads)
-        for (var i = thread_id * {{VEC_SIZE}}; i < tile_size; i += WORKGROUP_SIZE * {{VEC_SIZE}}) {
-            shared_vector[i / {{VEC_SIZE}}] = src1[(src1_idx_base + k_tile + i) / {{VEC_SIZE}}];
+        for (var i = thread_id * VEC_SIZE; i < tile_size; i += WG_SIZE * VEC_SIZE) {
+            shared_vector[i / VEC_SIZE] = src1[(src1_idx_base + k_tile + i) / VEC_SIZE];
        }

        workgroupBarrier();
@@ -250,7 +177,7 @@ fn main(
    workgroupBarrier();
    let group_base = thread_group * THREADS_PER_OUTPUT;
    let thread_base = group_base + thread_in_group;
-    var offset = THREADS_PER_OUTPUT / 2;
+    var offset: u32 = THREADS_PER_OUTPUT / 2;
    while (offset > 0) {
        if (thread_in_group < offset) {
            partial_sums[thread_base] += partial_sums[thread_base + offset];
@@ -260,8 +187,8 @@ fn main(
    }

    // Store back to global memory
-    if (output_row < params.m && thread_group % {{VEC_SIZE}} == 0 && thread_in_group == 0) {
-        dst[dst_idx / {{VEC_SIZE}}] = store_val(group_base);
+    if (output_row < params.m && thread_group % VEC_SIZE == 0 && thread_in_group == 0) {
+        dst[dst_idx / VEC_SIZE] = store_val(group_base);
    }
 }
-#end(SHADER)
+
@@ -1,21 +1,11 @@
-#define(VARIANTS)
+#ifdef INPLACE
+@group(0) @binding(1)
+var<uniform> params: Params;

-[
-  {
-    "SHADER_NAME": "scale_f32",
-    "DECLS": ["NOT_INPLACE"]
-  },
-  {
-    "SHADER_NAME": "scale_f32_inplace",
-    "DECLS": ["INPLACE"]
-  }
-]
-
-#end(VARIANTS)
-
-#define(DECLS)
-
-#decl(NOT_INPLACE)
+fn store_scale(val: f32, offset: u32) {
+    src[offset] = val;
+}
+#else
@group(0) @binding(1)
 var<storage, read_write> dst: array<f32>;

@@ -25,20 +15,7 @@ var<uniform> params: Params;
 fn store_scale(val: f32, offset: u32) {
    dst[offset] = val;
 }
-#enddecl(NOT_INPLACE)
-
-#decl(INPLACE)
-@group(0) @binding(1)
-var<uniform> params: Params;
-
-fn store_scale(val: f32, offset: u32) {
-    src[offset] = val;
-}
-#enddecl(INPLACE)
-
-#end(DECLS)
-
-#define(SHADER)
+#endif

 struct Params {
    offset_src: u32,
@@ -65,10 +42,7 @@ struct Params {
@group(0) @binding(0)
 var<storage, read_write> src: array<f32>;

-DECLS
-
-override wg_size: u32;
-@compute @workgroup_size(wg_size)
+@compute @workgroup_size(WG_SIZE)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
    if (gid.x >= params.ne) {
        return;
@@ -87,4 +61,3 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {

    store_scale(src[i_src] * params.scale + params.bias, i_dst);
 }
-#end(SHADER)
@@ -170,6 +170,20 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 #ifdef TRUNC
    let res = trunc(src[params.offset_src + src_idx]);
 #endif
+#ifdef SQR
+    let res = src[params.offset_src + src_idx] * src[params.offset_src + src_idx];
+#endif
+#ifdef SQRT
+    let res = sqrt(src[params.offset_src + src_idx]);
+#endif
+#ifdef SIN
+    let res_f32 = sin(f32(src[params.offset_src + src_idx]));
+    let res = TYPE(res_f32);
+#endif
+#ifdef COS
+    let res_f32 = cos(f32(src[params.offset_src + src_idx]));
+    let res = TYPE(res_f32);
+#endif

 #ifdef INPLACE
    src[params.offset_src + src_idx] = res;
@@ -899,7 +899,8 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
 };

 const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
-    GGML_ASSERT(type < GGML_TYPE_COUNT);
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
    return &type_traits[type];
 }

@@ -1265,27 +1266,33 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
 }

 int64_t ggml_blck_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
    return type_traits[type].blck_size;
 }

 size_t ggml_type_size(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
    return type_traits[type].type_size;
 }

 size_t ggml_row_size(enum ggml_type type, int64_t ne) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
    assert(ne % ggml_blck_size(type) == 0);
    return ggml_type_size(type)*ne/ggml_blck_size(type);
 }

-double ggml_type_sizef(enum ggml_type type) {
-    return ((double)(type_traits[type].type_size))/type_traits[type].blck_size;
-}
-
 const char * ggml_type_name(enum ggml_type type) {
-    return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
+    return type_traits[type].type_name;
 }

 bool ggml_is_quantized(enum ggml_type type) {
+    assert(type >= 0);
+    assert(type < GGML_TYPE_COUNT);
    return type_traits[type].is_quantized;
 }

@@ -1629,11 +1636,23 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
    const size_t cur_end  = cur_offs + cur_size;

    // align to GGML_MEM_ALIGN
+    GGML_ASSERT(size <= SIZE_MAX - (GGML_MEM_ALIGN - 1));
    size_t size_needed = GGML_PAD(size, GGML_MEM_ALIGN);

    char * const mem_buffer = ctx->mem_buffer;
    struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);

+    // integer overflow checks
+    if (cur_end > SIZE_MAX - size_needed) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu)\n", __func__, cur_end, size_needed);
+        return NULL;
+    }
+    if (cur_end + size_needed > SIZE_MAX - GGML_OBJECT_SIZE) {
+        GGML_LOG_WARN("%s: overflow detected in cur_end (%zu) + size_needed (%zu) + GGML_OBJECT_SIZE (%zu)\n", __func__,
+                cur_end, size_needed, (size_t) GGML_OBJECT_SIZE);
+        return NULL;
+    }
+
    if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
        GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
                __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
@@ -1702,6 +1721,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
        obj_alloc_size = data_size;
    }

+    GGML_ASSERT(GGML_TENSOR_SIZE <= SIZE_MAX - obj_alloc_size);
+
    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
    GGML_ASSERT(obj_new);

@@ -15,6 +15,17 @@
 #include <string>
 #include <vector>

+#define GGUF_MAX_STRING_LENGTH  (1024*1024*1024)
+#define GGUF_MAX_ARRAY_ELEMENTS (1024*1024*1024)
+
+#ifdef _WIN32
+#    define gguf_ftell _ftelli64
+#    define gguf_fseek _fseeki64
+#else
+#    define gguf_ftell ftello
+#    define gguf_fseek fseeko
+#endif
+
 template <typename T>
 struct type_to_gguf_type;

@@ -228,6 +239,26 @@ struct gguf_reader {

    template <typename T>
    bool read(std::vector<T> & dst, const size_t n) const {
+        if (n > GGUF_MAX_ARRAY_ELEMENTS) {
+            return false;
+        }
+        const uint64_t nbytes = nbytes_remain();
+        if constexpr (std::is_same<T, std::string>::value) {
+            // strings are prefixed with their length, so we need to account for that
+            if (n > SIZE_MAX / sizeof(uint64_t)) {
+                return false;
+            }
+            if (nbytes < n * sizeof(uint64_t)) {
+                return false;
+            }
+        } else {
+            if (n > SIZE_MAX / sizeof(T)) {
+                return false;
+            }
+            if (nbytes < n * sizeof(T)) {
+                return false;
+            }
+        }
        dst.resize(n);
        for (size_t i = 0; i < dst.size(); ++i) {
            if constexpr (std::is_same<T, bool>::value) {
@@ -277,13 +308,43 @@ struct gguf_reader {
        if (!read(size)) {
            return false;
        }
-        dst.resize(size);
+        if (size > GGUF_MAX_STRING_LENGTH) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds maximum %" PRIu64 "\n", __func__, size, (uint64_t) GGUF_MAX_STRING_LENGTH);
+            return false;
+        }
+        const uint64_t nbytes = nbytes_remain();
+        if (size > nbytes) {
+            GGML_LOG_ERROR("%s: string length %" PRIu64 " exceeds remaining file size %" PRIu64 " bytes\n", __func__, size, nbytes);
+            return false;
+        }
+        dst.resize(static_cast<size_t>(size));
        return fread(dst.data(), 1, dst.length(), file) == dst.length();
    }

    bool read(void * dst, const size_t size) const {
        return fread(dst, 1, size, file) == size;
    }
+
+    // remaining bytes in the file
+    uint64_t nbytes_remain() const {
+        const int64_t cur = gguf_ftell(file);
+        if (cur < 0) {
+            return 0;
+        }
+        if (gguf_fseek(file, 0, SEEK_END) != 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        const int64_t end = gguf_ftell(file);
+        if (end < 0) {
+            gguf_fseek(file, cur, SEEK_SET);
+
+            return 0;
+        }
+        gguf_fseek(file, cur, SEEK_SET);
+        return static_cast<uint64_t>(end - cur);
+    }
 };

 struct gguf_context * gguf_init_empty(void) {
@@ -568,8 +629,8 @@ 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) {
-                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));
+                GGML_LOG_ERROR("%s: tensor '%s' has invalid ggml type %d. should be in [0, %d)\n",
+                    __func__, info.t.name, info.t.type, GGML_TYPE_COUNT);
                ok = false;
                break;
            }
@@ -618,14 +679,14 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
    GGML_ASSERT(int64_t(ctx->info.size()) == n_tensors);

    // 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) {
+    if (gguf_fseek(file, GGML_PAD(gguf_ftell(file), ctx->alignment), SEEK_SET) != 0) {
        GGML_LOG_ERROR("%s: failed to seek to beginning of data section\n", __func__);
        gguf_free(ctx);
        return nullptr;
    }

    // store the current file offset - this is where the data section starts
-    ctx->offset = ftell(file);
+    ctx->offset = gguf_ftell(file);

    // compute the total size of the data section, taking into account the alignment
    {
@@ -657,10 +718,34 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
        //   the ggml_tensor structs to the appropriate locations in the binary blob

        // compute the exact size needed for the new ggml_context
-        const size_t mem_size =
-            params.no_alloc ?
-            (n_tensors    )*ggml_tensor_overhead() :
-            (n_tensors + 1)*ggml_tensor_overhead() + ctx->size;
+        size_t mem_size = 0;
+        if (params.no_alloc) {
+            if (n_tensors != 0 && SIZE_MAX / n_tensors < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = n_tensors * ggml_tensor_overhead();
+
+            mem_size = overhead;
+        } else {
+            if ((n_tensors + 1) != 0 && SIZE_MAX / (n_tensors + 1) < ggml_tensor_overhead()) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            const size_t overhead = (n_tensors + 1) * ggml_tensor_overhead();
+
+            if (SIZE_MAX - overhead < ctx->size) {
+                GGML_LOG_ERROR("%s: memory size overflow while allocating ggml context\n", __func__);
+                gguf_free(ctx);
+                return nullptr;
+            }
+
+            mem_size = overhead + ctx->size;
+        }

        struct ggml_init_params pdata = {
            /*mem_size   =*/ mem_size,
@@ -435,6 +435,7 @@ class MODEL_ARCH(IntEnum):
    T5               = auto()
    T5ENCODER        = auto()
    JAIS             = auto()
+    JAIS2            = auto()
    NEMOTRON         = auto()
    NEMOTRON_H       = auto()
    NEMOTRON_H_MOE   = auto()
@@ -472,6 +473,7 @@ class MODEL_ARCH(IntEnum):
    RND1             = auto()
    PANGU_EMBED      = auto()
    MISTRAL3         = auto()
+    PADDLEOCR        = auto()
    MIMO2            = auto()
    STEP35           = auto()
    LLAMA_EMBED      = auto()
@@ -652,6 +654,7 @@ class MODEL_TENSOR(IntEnum):
    ENC_OUTPUT_NORM      = auto()
    CLS                  = auto() # classifier
    CLS_OUT              = auto() # classifier output projection
+    CLS_NORM             = auto()
    CONV1D               = auto()
    CONVNEXT_DW          = auto()
    CONVNEXT_NORM        = auto()
@@ -873,6 +876,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.T5:               "t5",
    MODEL_ARCH.T5ENCODER:        "t5encoder",
    MODEL_ARCH.JAIS:             "jais",
+    MODEL_ARCH.JAIS2:            "jais2",
    MODEL_ARCH.NEMOTRON:         "nemotron",
    MODEL_ARCH.NEMOTRON_H:       "nemotron_h",
    MODEL_ARCH.NEMOTRON_H_MOE:   "nemotron_h_moe",
@@ -911,6 +915,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.RND1:             "rnd1",
    MODEL_ARCH.PANGU_EMBED:      "pangu-embedded",
    MODEL_ARCH.MISTRAL3:         "mistral3",
+    MODEL_ARCH.PADDLEOCR:        "paddleocr",
    MODEL_ARCH.MIMO2:            "mimo2",
    MODEL_ARCH.STEP35:           "step35",
    MODEL_ARCH.LLAMA_EMBED:      "llama-embed",
@@ -1088,6 +1093,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.ENC_OUTPUT_NORM:           "enc.output_norm",
    MODEL_TENSOR.CLS:                       "cls",
    MODEL_TENSOR.CLS_OUT:                   "cls.output",
+    MODEL_TENSOR.CLS_NORM:                  "cls.norm",
    MODEL_TENSOR.CONV1D:                    "conv1d",
    MODEL_TENSOR.CONVNEXT_DW:               "convnext.{bid}.dw",
    MODEL_TENSOR.CONVNEXT_NORM:             "convnext.{bid}.norm",
@@ -1507,6 +1513,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_NORM,
        MODEL_TENSOR.CLS,
        MODEL_TENSOR.CLS_OUT,
+        MODEL_TENSOR.CLS_NORM,
    ],
    MODEL_ARCH.NOMIC_BERT: [
        MODEL_TENSOR.TOKEN_EMBD,
@@ -2660,6 +2667,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_UP,
        MODEL_TENSOR.ATTN_POST_NORM,
        MODEL_TENSOR.FFN_POST_NORM,
+        # NextN/MTP tensors - preserved but unused
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
    ],
    MODEL_ARCH.GLM4_MOE: [
        MODEL_TENSOR.TOKEN_EMBD,
@@ -2807,6 +2821,19 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_GATE,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.JAIS2: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    MODEL_ARCH.NEMOTRON: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -3161,6 +3188,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.PADDLEOCR: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
    MODEL_ARCH.FALCON_H1: [
        # Token embedding
        MODEL_TENSOR.TOKEN_EMBD,
@@ -3822,6 +3863,7 @@ class VisionProjectorType:
    VOXTRAL = "voxtral"
    LFM2 = "lfm2"
    KIMIVL = "kimivl"
+    PADDLEOCR = "paddleocr"
    KIMIK25 = "kimik25"
    LIGHTONOCR = "lightonocr"
    COGVLM = "cogvlm"
@@ -175,6 +175,9 @@ class GGUFReader:
            if new_align.types != [GGUFValueType.UINT32]:
                raise ValueError('Bad type for general.alignment field')
            self.alignment = new_align.parts[-1][0]
+            # Ensure alignment is a non-zero power of two
+            if self.alignment == 0 or (self.alignment & (self.alignment - 1)) != 0:
+                raise ValueError('Invalid alignment: must be a non-zero power of two')
        padding = offs % self.alignment
        if padding != 0:
            offs += self.alignment - padding
@@ -202,11 +205,11 @@ class GGUFReader:

    def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int:
        if field.name in self.fields:
-            # TODO: add option to generate error on duplicate keys
-            # raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')
+            # TODO: add option to make this a warning and accept duplicate keys like below
+            raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}')

-            logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
-            self.fields[field.name + '_{}'.format(field.offset)] = field
+            # logger.warning(f'Duplicate key {field.name} at offset {field.offset}')
+            # self.fields[field.name + '_{}'.format(field.offset)] = field
        else:
            self.fields[field.name] = field
        return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts)
@@ -501,6 +501,8 @@ class GGUFWriter:
        self.add_uint32(Keys.General.QUANTIZATION_VERSION, quantization_version)

    def add_custom_alignment(self, alignment: int) -> None:
+        if alignment <= 0 or (alignment & (alignment - 1)) != 0:
+            raise ValueError('Invalid alignment: must be a non-zero power of two')
        self.data_alignment = alignment
        self.add_uint32(Keys.General.ALIGNMENT, alignment)

@@ -1240,6 +1240,10 @@ class TensorNameMap:
        MODEL_TENSOR.CLS_OUT: (
            "classifier.out_proj", # roberta
        ),
+
+        MODEL_TENSOR.CLS_NORM: (
+            "head.norm", # modern-bert
+        ),
        #############################################################################

        MODEL_TENSOR.CONVNEXT_DW: (
@@ -1321,6 +1325,7 @@ class TensorNameMap:
            "multi_modal_projector.linear_{bid}",
            "mm_projector.proj.linear_{bid}", # Kimi-K2.5
            "visual.merger.mlp.{bid}", # qwen2vl
+            "mlp_AR.linear_{bid}", # PaddleOCR-VL
            "merger.mlp.{bid}",
        ),

@@ -1404,6 +1409,7 @@ class TensorNameMap:
        MODEL_TENSOR.V_ENC_ATTN_Q_NORM: (
            "vision_tower.vision_model.encoder.layers.{bid}.attn.q_norm", # InternVL
            "model.vision_tower.encoder.layer.{bid}.attention.q_norm", # Intern-S1
+            "visual.blocks.{bid}.attn.q_norm", # GLM-OCR
        ),

        MODEL_TENSOR.V_ENC_ATTN_K: (
@@ -1422,6 +1428,7 @@ class TensorNameMap:
        MODEL_TENSOR.V_ENC_ATTN_K_NORM: (
            "vision_tower.vision_model.encoder.layers.{bid}.attn.k_norm", # InternVL
            "model.vision_tower.encoder.layer.{bid}.attention.k_norm", # Intern-S1
+            "visual.blocks.{bid}.attn.k_norm", # GLM-OCR
        ),

        MODEL_TENSOR.V_ENC_ATTN_V: (
@@ -1568,6 +1575,7 @@ class TensorNameMap:
            "mm_projector.pre_norm", # Kimi-K2.5
            "pre_mm_projector_norm",
            "model.vision.linear_proj.norm1", # cogvlm
+            "mlp_AR.pre_norm", # PaddleOCR-VL
            "merger.ln_q",
        ),

@@ -1593,6 +1601,7 @@ class TensorNameMap:

        MODEL_TENSOR.V_RESMPL_ATTN_OUT: (
            "resampler.attn.out_proj",
+            "model.vision_model.head.attention.out_proj",
        ),

        MODEL_TENSOR.V_RESMPL_KV: (
@@ -389,6 +389,7 @@ extern "C" {
        bool only_copy;                       // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
        bool pure;                            // quantize all tensors to the default type
        bool keep_split;                      // quantize to the same number of shards
+        bool dry_run;                         // calculate and show the final quantization size without performing quantization
        void * imatrix;                       // pointer to importance matrix data
        void * kv_overrides;                  // pointer to vector containing overrides
        void * tensor_types;                  // pointer to vector containing tensor types
@@ -0,0 +1,80 @@
+{% macro render_content(content) %}{% if content is none %}{{- '' }}{% elif content is string %}{{- content }}{% elif content is mapping %}{{- content['value'] if 'value' in content else content['text'] }}{% elif content is iterable %}{% for item in content %}{% if item.type == 'text' %}{{- item['value'] if 'value' in item else item['text'] }}{% elif item.type == 'image' %}<im_patch>{% endif %}{% endfor %}{% endif %}{% endmacro %}
+{{bos_token}}{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- render_content(messages[0].content) + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou have access to the following functions in JSONSchema format:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson(ensure_ascii=False) }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...>\n...\n</function> block must be nested within <tool_call>\n...\n</tool_call> XML tags\n- Required parameters MUST be specified\n</IMPORTANT><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + render_content(messages[0].content) + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and render_content(message.content) is string and not(render_content(message.content).startswith('<tool_response>') and render_content(message.content).endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content) %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {%- set role_name = 'observation' if (message.role == "system" and not loop.first and message.name == 'observation') else message.role %}
+        {{- '<|im_start|>' + role_name + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = render_content(message.reasoning_content) %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- else %}
+                {%- set reasoning_content = '' %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- if tool_call.arguments is defined %}
+                    {%- set arguments = tool_call.arguments %}
+                    {%- for args_name, args_value in arguments|items %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson(ensure_ascii=False) | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>tool_response\n' }}
+        {%- endif %}
+        {{- '<tool_response>' }}
+        {{- content }}
+        {{- '</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n<think>\n' }}
+{%- endif %}
@@ -25,16 +25,12 @@ Example usage:
 """


-def generate_input_prompt(length: int) -> list[str]:
-    CORPUS = """
-    You are an advanced AI assistant capable of using tools to gather information, perform calculations, or execute tasks. Always think step by step before responding. If a user's query requires external data, computation, or actions beyond your internal knowledge, use the appropriate tools via function calls.
-
-    ### Tool Call Format:
-    When you need to use a tool, output the call in this exact XML format. Include the opening and closing tags. Do not escape arguments; they will be parsed as plain text.
-
-    You can make multiple calls in one go by placing them one after another.
-    """
-    words = [w.strip() for w in CORPUS.strip().split(" ")]
+def get_remote_corpus(url: str, length: int) -> list[str]:
+    response = requests.get(url)
+    response.raise_for_status()
+    corpus = response.text
+    words = [w.strip() for w in corpus.strip().split(" ")]
+    words = [w for w in words if "<" not in w] # make sure nothing looks like special tokens
    words = [w for w in words if len(w) > 0]  # filter out empty strings
    while len(words) < length:
        words += words
@@ -226,9 +222,9 @@ def parse_args() -> argparse.Namespace:
    )
    parser_dump.add_argument(
        "--file",
-        type=Path,
-        default=None,
-        help="File containing prompt to use instead of the default",
+        type=str,
+        default="https://raw.githubusercontent.com/ggml-org/llama.cpp/eaba92c3dcc980ebe753348855d4a5d75c069997/tools/server/README.md",
+        help="File containing prompt to use instead of the default (can also be an URL)",
    )
    parser_dump.add_argument(
        "--pattern",
@@ -259,17 +255,19 @@ def main():

    if args.verb == "dump":
        pattern = parse_pattern(args.pattern)
-        input_length = sum(n for _, n in pattern)
-        input_words = generate_input_prompt(input_length)
-        if args.file is not None:
-            with args.file.open("r") as f:
+        required_words = sum(n for _, n in pattern)
+        if args.file.startswith("http"):
+            input_words = get_remote_corpus(args.file, required_words)
+            logger.info(f"Fetched {len(input_words)} words from remote {args.file}")
+        else:
+            with open(args.file, "r") as f:
                input_words = f.read().strip().split(" ")
-                if input_length < sum(n for _, n in pattern):
+                input_words = [w for w in input_words if len(w) > 0]  # filter out empty strings
+                if len(input_words) < required_words:
                    raise ValueError(
-                        f"Input file has only {input_length} words, but pattern requires at least {input_length} words."
+                        f"Input file has only {len(input_words)} words, but pattern requires at least {required_words} words."
                    )
-                input_length = len(input_words)
-        logger.info(f"Using {input_length} words")
+        logger.info(f"Using {len(input_words)} words")
        dump_logits(args.endpoint, args.output, input_words, pattern, args.api_key)
    elif args.verb == "compare":
        compare_logits(args.input1, args.input2, args.output)
@@ -54,6 +54,6 @@ adb $adbserial $adbhost shell " \
    $verbose $experimental $sched $opmask $profile $nhvx $ndev $hb \
      ./$branch/bin/llama-cli --no-mmap -m $basedir/../gguf/$model \
         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1           \
-         --ctx-size 8192 --batch-size 128 -fa on \
-         -ngl 99 --device $device $cli_opts $@   \
+         --ctx-size 8192 --ubatch-size 256 -fa on                  \
+         -ngl 99 --device $device $cli_opts $@                     \
 "
@@ -54,6 +54,6 @@ adb $adbserial $adbhost shell " \
    $verbose $experimental $sched $opmask $profile $nhvx $ndev $hb        \
      ./$branch/bin/llama-completion --no-mmap -m $basedir/../gguf/$model \
         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                  \
-         --ctx-size 8192 --batch-size 128 -fa on \
-         -ngl 99 -no-cnv --device $device $cli_opts $@   \
+         --ctx-size 8192 --ubatch-size 256 -fa on                         \
+         -ngl 99 -no-cnv --device $device $cli_opts $@                    \
 "
@@ -58,11 +58,11 @@ adb $adbserial $adbhost shell " \
  cd $basedir; ulimit -c unlimited;        \
    LD_LIBRARY_PATH=$basedir/$branch/lib   \
    ADSP_LIBRARY_PATH=$basedir/$branch/lib \
-    $verbose $experimental $sched $opmask $profile $nhvx $ndev $mtmd_backend       \
-      ./$branch/bin/llama-mtmd-cli --no-mmap -m $basedir/../gguf/$model   \
-         --mmproj $basedir/../gguf/$mmproj \
-         --image $basedir/../gguf/$image \
-         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1             \
-         --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on \
-         -ngl 99 --device $device -v $cli_opts $@ \
+    $verbose $experimental $sched $opmask $profile $nhvx $ndev $mtmd_backend \
+      ./$branch/bin/llama-mtmd-cli --no-mmap -m $basedir/../gguf/$model      \
+         --mmproj $basedir/../gguf/$mmproj                                   \
+         --image $basedir/../gguf/$image                                     \
+         --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1                     \
+         --ctx-size 8192 --ubatch-size 256 -fa on                            \
+         -ngl 99 --device $device -v $cli_opts $@                            \
 "
@@ -49,5 +49,5 @@ $env:ADSP_LIBRARY_PATH="$basedir\lib"
 & "$basedir\bin\llama-completion.exe" `
    --no-mmap -no-cnv -m $basedir\..\..\gguf\$model `
    --poll 1000 -t 6 --cpu-mask 0xfc --cpu-strict 1 `
-    --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on `
+    --ctx-size 8192 --ubatch-size 128 -fa on `
    -ngl 99 --device $device $cli_opts
@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess

-HTTPLIB_VERSION = "d4180e923f846b44a3d30acd938438d6e64fc9f6"
+HTTPLIB_VERSION = "refs/tags/v0.34.0"

 vendor = {
    "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -84,6 +84,7 @@ add_library(llama
            models/hunyuan-moe.cpp
            models/internlm2.cpp
            models/jais.cpp
+            models/jais2.cpp
            models/jamba.cpp
            models/kimi-linear.cpp
            models/lfm2.cpp
@@ -109,6 +110,7 @@ add_library(llama
            models/openai-moe-iswa.cpp
            models/openelm.cpp
            models/orion.cpp
+            models/paddleocr.cpp
            models/pangu-embedded.cpp
            models/phi2.cpp
            models/phi3.cpp
@@ -79,6 +79,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_T5,               "t5"               },
    { LLM_ARCH_T5ENCODER,        "t5encoder"        },
    { LLM_ARCH_JAIS,             "jais"             },
+    { LLM_ARCH_JAIS2,            "jais2"            },
    { LLM_ARCH_NEMOTRON,         "nemotron"         },
    { LLM_ARCH_NEMOTRON_H,       "nemotron_h"       },
    { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
@@ -120,6 +121,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_RND1,             "rnd1"             },
    { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
    { LLM_ARCH_MISTRAL3,         "mistral3"         },
+    { LLM_ARCH_PADDLEOCR,        "paddleocr"        },
    { LLM_ARCH_MIMO2,            "mimo2"            },
    { LLM_ARCH_STEP35,           "step35"           },
    { LLM_ARCH_LLAMA_EMBED,      "llama-embed"      },
@@ -367,6 +369,7 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
    { LLM_TENSOR_TOKEN_TYPES,                            "token_types" },
    { LLM_TENSOR_CLS,                                    "cls" },
    { LLM_TENSOR_CLS_OUT,                                "cls.output" },
+    { LLM_TENSOR_CLS_NORM,                               "cls.norm" },
    { LLM_TENSOR_ENC_OUTPUT_NORM,                        "enc.output_norm" },
    { LLM_TENSOR_FFN_GATE_INP_SHEXP,                     "blk.%d.ffn_gate_inp_shexp" },
    { LLM_TENSOR_SSM_A_NOSCAN,                           "blk.%d.ssm_a" },
@@ -737,6 +740,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
        case LLM_ARCH_INTERNLM2:
        case LLM_ARCH_GRANITE:
        case LLM_ARCH_ERNIE4_5:
+        case LLM_ARCH_PADDLEOCR:
        case LLM_ARCH_SMOLLM3:
        case LLM_ARCH_DREAM:
        case LLM_ARCH_LLADA:
@@ -828,6 +832,7 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                LLM_TENSOR_FFN_NORM,
                LLM_TENSOR_CLS,
                LLM_TENSOR_CLS_OUT,
+                LLM_TENSOR_CLS_NORM,
            };
        case LLM_ARCH_JINA_BERT_V2:
            return {
@@ -1633,6 +1638,12 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                LLM_TENSOR_FFN_DOWN,
                LLM_TENSOR_ATTN_POST_NORM,
                LLM_TENSOR_FFN_POST_NORM,
+                LLM_TENSOR_NEXTN_EH_PROJ,
+                LLM_TENSOR_NEXTN_EMBED_TOKENS,
+                LLM_TENSOR_NEXTN_ENORM,
+                LLM_TENSOR_NEXTN_HNORM,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
            };
        case LLM_ARCH_GLM4_MOE:
            return {
@@ -1783,6 +1794,20 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                LLM_TENSOR_FFN_GATE,
                LLM_TENSOR_FFN_DOWN,
            };
+        case LLM_ARCH_JAIS2:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_DOWN,
+            };
        case LLM_ARCH_NEMOTRON_H:
            return {
                LLM_TENSOR_TOKEN_EMBD,
@@ -2512,6 +2537,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
    {LLM_TENSOR_OUTPUT,                     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
    {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
    {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
+    {LLM_TENSOR_CLS_NORM,                   {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
    {LLM_TENSOR_DENSE_2_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
    {LLM_TENSOR_DENSE_3_OUT,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, // Dense layer output
    {LLM_TENSOR_OUTPUT_NORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
@@ -83,6 +83,7 @@ enum llm_arch {
    LLM_ARCH_T5,
    LLM_ARCH_T5ENCODER,
    LLM_ARCH_JAIS,
+    LLM_ARCH_JAIS2,
    LLM_ARCH_NEMOTRON,
    LLM_ARCH_NEMOTRON_H,
    LLM_ARCH_NEMOTRON_H_MOE,
@@ -124,6 +125,7 @@ enum llm_arch {
    LLM_ARCH_RND1,
    LLM_ARCH_PANGU_EMBED,
    LLM_ARCH_MISTRAL3,
+    LLM_ARCH_PADDLEOCR,
    LLM_ARCH_MIMO2,
    LLM_ARCH_STEP35,
    LLM_ARCH_LLAMA_EMBED,
@@ -497,6 +499,7 @@ enum llm_tensor {
    LLM_TENSOR_ENC_OUTPUT_NORM,
    LLM_TENSOR_CLS,
    LLM_TENSOR_CLS_OUT,
+    LLM_TENSOR_CLS_NORM,
    LLM_TENSOR_CONV1D,
    LLM_TENSOR_CONVNEXT_DW,
    LLM_TENSOR_CONVNEXT_NORM,
@@ -712,8 +712,6 @@ int64_t llama_context::output_resolve_row(int32_t i) const {
 }

 float * llama_context::get_logits_ith(int32_t i) {
-    int64_t j = -1;
-
    output_reorder();

    try {
@@ -721,26 +719,7 @@ float * llama_context::get_logits_ith(int32_t i) {
            throw std::runtime_error("no logits");
        }

-        // TODO: use output_resolve_row()
-        if (i < 0) {
-            j = n_outputs + i;
-            if (j < 0) {
-                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
-            }
-        } else if ((size_t) i >= output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
-        } else {
-            j = output_ids[i];
-        }
-
-        if (j < 0) {
-            throw std::runtime_error(format("batch.logits[%d] != true", i));
-        }
-        if (j >= n_outputs) {
-            // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
-        }
-
+        const int64_t j = output_resolve_row(i);
        return logits.data + j*model.vocab.n_tokens();
    } catch (const std::exception & err) {
        LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
@@ -763,8 +742,6 @@ llama_token * llama_context::get_sampled_tokens()  const{
 }

 float * llama_context::get_embeddings_ith(int32_t i) {
-    int64_t j = -1;
-
    output_reorder();

    try {
@@ -772,26 +749,7 @@ float * llama_context::get_embeddings_ith(int32_t i) {
            throw std::runtime_error("no embeddings");
        }

-        // TODO: use output_resolve_row()
-        if (i < 0) {
-            j = n_outputs + i;
-            if (j < 0) {
-                throw std::runtime_error(format("negative index out of range [0, %d)", n_outputs));
-            }
-        } else if ((size_t) i >= output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %zu)", output_ids.size()));
-        } else {
-            j = output_ids[i];
-        }
-
-        if (j < 0) {
-            throw std::runtime_error(format("batch.logits[%d] != true", i));
-        }
-        if (j >= n_outputs) {
-            // This should not happen
-            throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs));
-        }
-
+        const int64_t j = output_resolve_row(i);
        const uint32_t n_embd_out = model.hparams.n_embd_out();
        return embd.data + j*n_embd_out;
    } catch (const std::exception & err) {
@@ -2482,64 +2440,6 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
        // TODO: add more model-specific info which should prevent loading the session file if not identical
    }

-    // write output ids
-    {
-        LLAMA_LOG_DEBUG("%s: - writing output ids\n", __func__);
-
-        const auto n_outputs    = this->n_outputs;
-        const auto & output_ids = this->output_ids;
-
-        std::vector<int32_t> w_output_pos;
-
-        w_output_pos.resize(n_outputs);
-
-        // build a more compact representation of the output ids
-        for (size_t i = 0; i < n_batch(); ++i) {
-            // map an output id to a position in the batch
-            int64_t pos = output_ids[i];
-            if (pos >= 0) {
-                GGML_ASSERT(pos < n_outputs);
-                w_output_pos[pos] = i;
-            }
-        }
-
-        io.write(&n_outputs, sizeof(n_outputs));
-
-        if (n_outputs) {
-            io.write(w_output_pos.data(), n_outputs * sizeof(int32_t));
-        }
-    }
-
-    // [TAG_CONTEXT_STATE_LOGITS]
-    // write logits
-    {
-        LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
-
-        const uint64_t logits_size = std::min((uint64_t) this->logits.size, (uint64_t) n_outputs * model.vocab.n_tokens());
-
-        io.write(&logits_size, sizeof(logits_size));
-
-        if (logits_size) {
-            io.write(logits.data, logits_size * sizeof(float));
-        }
-    }
-
-    // write embeddings
-    {
-        LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__);
-
-        const uint64_t embd_size = std::min((uint64_t) this->embd.size, (uint64_t) n_outputs * model.hparams.n_embd);
-
-        io.write(&embd_size, sizeof(embd_size));
-
-        if (embd_size) {
-            io.write(embd.data, embd_size * sizeof(float));
-        }
-    }
-
-    // TODO: handle sampling buffers and samplers state ?
-    //       https://github.com/ggml-org/llama.cpp/pull/17004
-
    if (memory != nullptr) {
        LLAMA_LOG_DEBUG("%s: - writing memory module\n", __func__);
        memory->state_write(io);
@@ -2565,70 +2465,6 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
        // TODO: add more info which needs to be identical but which is not verified otherwise
    }

-    // read output ids
-    {
-        LLAMA_LOG_DEBUG("%s: - reading output ids\n", __func__);
-
-        auto n_outputs = this->n_outputs;
-        io.read_to(&n_outputs, sizeof(n_outputs));
-
-        if (n_outputs > output_reserve(n_outputs)) {
-            throw std::runtime_error("could not reserve outputs");
-        }
-
-        std::vector<int32_t> output_pos;
-
-        if (n_outputs) {
-            output_pos.resize(n_outputs);
-            io.read_to(output_pos.data(), n_outputs * sizeof(int32_t));
-
-            for (int32_t i = 0; i < (int32_t) output_pos.size(); ++i) {
-                int32_t id = output_pos[i];
-                if ((uint32_t) id >= n_batch()) {
-                    throw std::runtime_error(format("invalid output id, %d does not fit in batch size of %u", id, n_batch()));
-                }
-                this->output_ids[id] = i;
-            }
-
-            this->n_outputs = n_outputs;
-        }
-    }
-
-    // read logits
-    {
-        LLAMA_LOG_DEBUG("%s: - reading logits\n", __func__);
-
-        uint64_t logits_size;
-        io.read_to(&logits_size, sizeof(logits_size));
-
-        if (this->logits.size < logits_size) {
-            throw std::runtime_error("logits buffer too small");
-        }
-
-        if (logits_size) {
-            io.read_to(this->logits.data, logits_size * sizeof(float));
-        }
-    }
-
-    // read embeddings
-    {
-        LLAMA_LOG_DEBUG("%s: - reading embeddings\n", __func__);
-
-        uint64_t embd_size;
-        io.read_to(&embd_size, sizeof(embd_size));
-
-        if (this->embd.size < embd_size) {
-            throw std::runtime_error("embeddings buffer too small");
-        }
-
-        if (embd_size) {
-            io.read_to(this->embd.data, embd_size * sizeof(float));
-        }
-    }
-
-    // TODO: handle sampling buffers and samplers state ?
-    //       https://github.com/ggml-org/llama.cpp/pull/17004
-
    if (memory) {
        LLAMA_LOG_DEBUG("%s: - reading memory module\n", __func__);

@@ -2761,6 +2597,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
    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);
+    llama_set_param(model->cls_norm,        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) {
@@ -185,7 +185,10 @@ bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) {
 }

 void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) {
-    if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) {
+    if (cparams.embeddings   &&
+       (cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN ||
+        cparams.pooling_type == LLAMA_POOLING_TYPE_RANK )) {
+
        const int64_t n_tokens     = ubatch->n_tokens;
        const int64_t n_seq_tokens = ubatch->n_seq_tokens;
        const int64_t n_seqs_unq   = ubatch->n_seqs_unq;
@@ -1125,8 +1128,8 @@ ggml_tensor * llm_graph_context::build_ffn(

    if (down) {
        cur = build_lora_mm(down, cur);
-        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
-            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
+            // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
        }
    }
@@ -1721,7 +1724,8 @@ ggml_tensor * llm_graph_context::build_attn_mha(

    ggml_tensor * cur;

-    if (cparams.flash_attn && kq_b == nullptr) {
+    const bool use_flash_attn = cparams.flash_attn && kq_b == nullptr;
+    if (use_flash_attn) {
        GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");

        if (v_trans) {
@@ -1981,8 +1985,8 @@ ggml_tensor * llm_graph_context::build_attn(

    if (wo) {
        cur = build_lora_mm(wo, cur);
-        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) {
-            // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators
+        if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE || arch == LLM_ARCH_JAIS2) {
+            // GLM4, GLM4_MOE, and JAIS2 seem to have numerical issues with half-precision accumulators
            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
        }
    }
@@ -2414,8 +2418,9 @@ llm_graph_input_mem_hybrid_iswa * llm_graph_context::build_inp_mem_hybrid_iswa()

 void llm_graph_context::build_dense_out(
    ggml_tensor * dense_2,
+    ggml_tensor * dense_2_b,
    ggml_tensor * dense_3) const {
-    if (!cparams.embeddings || !(dense_2 || dense_3)) {
+    if (!cparams.embeddings || !(dense_2 || dense_2_b || dense_3)) {
        return;
    }
    ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd;
@@ -2424,6 +2429,9 @@ void llm_graph_context::build_dense_out(
    if (dense_2) {
        cur = ggml_mul_mat(ctx0, dense_2, cur);
    }
+    if (dense_2_b) {
+        cur = ggml_add(ctx0, cur, dense_2_b);
+    }
    if (dense_3) {
        cur = ggml_mul_mat(ctx0, dense_3, cur);
    }
@@ -2437,7 +2445,8 @@ void llm_graph_context::build_pooling(
        ggml_tensor * cls,
        ggml_tensor * cls_b,
        ggml_tensor * cls_out,
-        ggml_tensor * cls_out_b) const {
+        ggml_tensor * cls_out_b,
+        ggml_tensor * cls_norm) const {
    if (!cparams.embeddings) {
        return;
    }
@@ -2476,8 +2485,15 @@ void llm_graph_context::build_pooling(
            } break;
        case LLAMA_POOLING_TYPE_RANK:
            {
-                ggml_tensor * inp_cls = build_inp_cls();
-                cur = ggml_get_rows(ctx0, inp, inp_cls);
+                if (arch == LLM_ARCH_MODERN_BERT) {
+                    // modern bert gte reranker builds mean first then applies prediction head and classifier
+                    // https://github.com/huggingface/transformers/blob/main/src/transformers/models/modernbert/modular_modernbert.py#L1404-1411
+                    ggml_tensor * inp_mean = build_inp_mean();
+                    cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, inp)), inp_mean);
+                } else {
+                    ggml_tensor * inp_cls = build_inp_cls();
+                    cur = ggml_get_rows(ctx0, inp, inp_cls);
+                }

                // classification head
                // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566
@@ -2486,7 +2502,15 @@ void llm_graph_context::build_pooling(
                    if (cls_b) {
                        cur = ggml_add(ctx0, cur, cls_b);
                    }
-                    cur = ggml_tanh(ctx0, cur);
+                    if (arch == LLM_ARCH_MODERN_BERT) {
+                        cur = ggml_gelu(ctx0, cur);
+                    } else {
+                        cur = ggml_tanh(ctx0, cur);
+                    }
+                    if (cls_norm) {
+                        // head norm
+                        cur = build_norm(cur, cls_norm, NULL, LLM_NORM, -1);
+                    }
                }

                // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
@@ -1000,7 +1000,8 @@ struct llm_graph_context {
            ggml_tensor * cls,
            ggml_tensor * cls_b,
            ggml_tensor * cls_out,
-            ggml_tensor * cls_out_b) const;
+            ggml_tensor * cls_out_b,
+            ggml_tensor * cls_norm) const;

    //
    // sampling (backend sampling)
@@ -1014,6 +1015,7 @@ struct llm_graph_context {

    void build_dense_out(
            ggml_tensor * dense_2,
+            ggml_tensor * dense_2_b,
            ggml_tensor * dense_3) const;
 };

@@ -109,9 +109,9 @@ std::string llama_format_tensor_shape(const std::vector<int64_t> & ne) {

 std::string llama_format_tensor_shape(const struct ggml_tensor * t) {
    char buf[256];
-    snprintf(buf, sizeof(buf), "%5" PRId64, t->ne[0]);
+    snprintf(buf, sizeof(buf), "%6" PRId64, t->ne[0]);
    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %5" PRId64, t->ne[i]);
+        snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf), ", %6" PRId64, t->ne[i]);
    }
    return buf;
 }
@@ -163,7 +163,7 @@ bool llama_memory_recurrent::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos
            const auto & cell = cells[tail_id];
            // partial intersection is invalid if it includes the final pos
            if (0 < p0 && p0 <= cell.pos && p1 > cell.pos) {
-                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false\n");
+                //printf("[DEBUG] inside `llama_memory_recurrent::seq_rm`: partial intersection is invalid, so returning false, p0 = %d, cell.pos = %d, p1 = %d\n", p0, cell.pos, p1);
                return false;
            }
            // invalidate tails which will be cleared
@@ -271,6 +271,7 @@ void llama_model_saver::add_tensors_from_model() {
    add_tensor(model.cls_b);
    add_tensor(model.cls_out);
    add_tensor(model.cls_out_b);
+    add_tensor(model.cls_norm);

    for (const struct llama_layer & layer : model.layers) {
        for (size_t i = 0; i < sizeof(layer)/sizeof(struct ggml_tensor *); ++i) {
@@ -123,6 +123,7 @@ const char * llm_type_name(llm_type type) {
        case LLM_TYPE_8B_A1B:        return "8B.A1B";
        case LLM_TYPE_16B_A1B:       return "16B.A1B";
        case LLM_TYPE_21B_A3B:       return "21B.A3B";
+        case LLM_TYPE_24B_A2B:       return "24B.A2B";
        case LLM_TYPE_30B_A3B:       return "30B.A3B";
        case LLM_TYPE_31B_A3_5B:     return "31B.A3.5B";
        case LLM_TYPE_35B_A3B:       return "35B.A3B";
@@ -908,7 +909,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {

                    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa);
                    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
-                    hparams.set_swa_pattern(swa_period);
+                    hparams.set_swa_pattern(swa_period, true);
                } else {
                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
                }
@@ -1703,8 +1704,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            } break;
        case LLM_ARCH_DEEPSEEK2:
            {
-                // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
-                const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26);
+                // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B, Kanana-2-30B-A3B
+                const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26 || (hparams.n_layer == 48 && n_vocab == 128256));

                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead);
@@ -1784,7 +1785,15 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+                // NextN/MTP parameters (GLM-OCR)
+                ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
+
+                // TODO: when MTP is implemented, this should probably be updated if needed
+                hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
                switch (hparams.n_layer) {
+                    case 17: type = LLM_TYPE_1B; break; // GLM-OCR
                    case 40: type = LLM_TYPE_9B; break;
                    case 61: type = LLM_TYPE_32B; break;
                    default: type = LLM_TYPE_UNKNOWN;
@@ -1929,6 +1938,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                    default: type = LLM_TYPE_UNKNOWN;
                }
            } break;
+        case LLM_ARCH_JAIS2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+                switch (hparams.n_layer) {
+                    case 32: type = LLM_TYPE_8B; break;
+                    case 68: type = LLM_TYPE_70B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
        case LLM_ARCH_NEMOTRON:
            {
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
@@ -2226,7 +2245,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
            } break;
        case LLM_ARCH_ERNIE4_5:
        case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_PADDLEOCR:
            {
+                // paddleocr need mrope_section
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                if (arch == LLM_ARCH_ERNIE4_5_MOE) {
                    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
@@ -2340,6 +2363,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                    case 10752: type = LLM_TYPE_2_6B; break;
                    default:    type = LLM_TYPE_UNKNOWN;
                }
+                if (const auto is_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); is_swa && hparams.n_swa > 0) {
+                    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+                    for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+                        hparams.swa_layers[il] = !hparams.recurrent_layer_arr[il];
+                    }
+                }
            } break;
        case LLM_ARCH_LFM2MOE:
            {
@@ -2353,7 +2382,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                    hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
                }

-                type = LLM_TYPE_8B_A1B;
+                switch (hparams.n_layer) {
+                    case 24: type = LLM_TYPE_8B_A1B;  break;
+                    case 40: type = LLM_TYPE_24B_A2B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
            } break;
        case LLM_ARCH_SMALLTHINKER:
            {
@@ -3505,9 +3538,10 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
                    }

-                    cls       = create_tensor(tn(LLM_TENSOR_CLS,     "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
-                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
-                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+                    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+                    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+                    cls       = create_tensor(tn(LLM_TENSOR_CLS,      "weight"), {n_embd, n_embd},            TENSOR_NOT_REQUIRED);
+                    cls_norm  = create_tensor(tn(LLM_TENSOR_CLS_NORM, "weight"), {n_embd},                    TENSOR_NOT_REQUIRED);

                } break;
            case LLM_ARCH_NEO_BERT:
@@ -5360,6 +5394,45 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
                    }
                } break;
+            case LLM_ARCH_JAIS2:
+                {
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+                    if (!output) {
+                        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+                        // attention biases - all have shape n_embd (output dimension of projections)
+                        layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
+                        layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
+                        layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
+                        layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+                        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+                        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+                        // Jais-2 uses simple MLP (no gate) with biases
+                        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+                        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+                        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+                        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+                    }
+                } break;
            case LLM_ARCH_CHATGLM:
                {
                    tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
@@ -5410,30 +5483,48 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                    }

                    for (int i = 0; i < n_layer; ++i) {
-                        auto & layer = layers[i];
-
-                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
-
-                        if (layer.wqkv == nullptr) {
-                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
-                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
-                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
-                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
-                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
-                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED);
+                        int flags = 0;
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            // skip all tensors in the NextN layers
+                            flags |= TENSOR_SKIP;
                        }

-                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        auto & layer = layers[i];

-                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, flags | TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, flags | TENSOR_NOT_REQUIRED);

-                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
-                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
-                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, flags);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, flags);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, flags);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, flags | TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, flags | TENSOR_NOT_REQUIRED);
+                        }

-                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
+
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
+
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, flags);
+
+                        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
+
+                        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+                        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+                            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+                            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+                            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+                            // Optional tensors
+                            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+                            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+                            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+                        }
                    }
                } break;
            case LLM_ARCH_GLM4_MOE:
@@ -6549,6 +6640,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                } break;
            case LLM_ARCH_ERNIE4_5:
            case LLM_ARCH_ERNIE4_5_MOE:
+            case LLM_ARCH_PADDLEOCR:
                {
                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

@@ -6869,7 +6961,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                    }

                    // for LFM2-ColBert-350M
-                    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers   = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+                    dense_2_out_layers_b = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "bias"),   {hparams.n_embd_out()        }, TENSOR_NOT_REQUIRED);
                } break;
            case LLM_ARCH_SMALLTHINKER:
                {
@@ -8527,6 +8620,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
            {
                llm = std::make_unique<llm_build_jais>(*this, params);
            } break;
+        case LLM_ARCH_JAIS2:
+            {
+                llm = std::make_unique<llm_build_jais2>(*this, params);
+            } break;
        case LLM_ARCH_NEMOTRON:
            {
                llm = std::make_unique<llm_build_nemotron>(*this, params);
@@ -8622,6 +8719,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
            {
                llm = std::make_unique<llm_build_ernie4_5_moe>(*this, params);
            } break;
+        case LLM_ARCH_PADDLEOCR:
+            {
+                llm = std::make_unique<llm_build_paddleocr>(*this, params);
+            } break;
        case LLM_ARCH_HUNYUAN_MOE:
            {
                llm = std::make_unique<llm_build_hunyuan_moe>(*this, params);
@@ -8645,7 +8746,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
        case LLM_ARCH_LFM2:
        case LLM_ARCH_LFM2MOE:
            {
-                llm = std::make_unique<llm_build_lfm2>(*this, params);
+                if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+                    llm = std::make_unique<llm_build_lfm2<true>>(*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_lfm2<false>>(*this, params);
+                }
            } break;
        case LLM_ARCH_SMALLTHINKER:
            {
@@ -8708,7 +8813,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
    }

    // add on pooling layer
-    llm->build_pooling(cls, cls_b, cls_out, cls_out_b);
+    llm->build_pooling(cls, cls_b, cls_out, cls_out_b, cls_norm);

    // add backend sampling layers (if any)
    llm->build_sampling();
@@ -8717,7 +8822,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
    // there will be two additional dense projection layers
    // dense linear projections are applied after pooling
    // TODO: move reranking logic here and generalize
-    llm->build_dense_out(dense_2_out_layers, dense_3_out_layers);
+    llm->build_dense_out(dense_2_out_layers, dense_2_out_layers_b, dense_3_out_layers);

    llm->res->set_outputs();

@@ -8935,6 +9040,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_BAILINGMOE2:
        case LLM_ARCH_DOTS1:
        case LLM_ARCH_HUNYUAN_MOE:
+        case LLM_ARCH_JAIS2:
        case LLM_ARCH_OPENAI_MOE:
        case LLM_ARCH_HUNYUAN_DENSE:
        case LLM_ARCH_LFM2:
@@ -8953,6 +9059,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
            return LLAMA_ROPE_TYPE_NEOX;

        case LLM_ARCH_QWEN2VL:
+        case LLM_ARCH_PADDLEOCR:
            return LLAMA_ROPE_TYPE_MROPE;
        case LLM_ARCH_QWEN3VL:
        case LLM_ARCH_QWEN3VLMOE:
@@ -116,6 +116,7 @@ enum llm_type {
    LLM_TYPE_8B_A1B, // lfm2moe
    LLM_TYPE_16B_A1B,
    LLM_TYPE_21B_A3B, // Ernie MoE small
+    LLM_TYPE_24B_A2B, // lfm2moe
    LLM_TYPE_30B_A3B,
    LLM_TYPE_31B_A3_5B,
    LLM_TYPE_35B_A3B, // Qwen3.5
@@ -475,6 +476,7 @@ struct llama_model {
    struct ggml_tensor * cls_b     = nullptr;
    struct ggml_tensor * cls_out   = nullptr;
    struct ggml_tensor * cls_out_b = nullptr;
+    struct ggml_tensor * cls_norm  = nullptr;

    struct ggml_tensor * conv1d   = nullptr;
    struct ggml_tensor * conv1d_b = nullptr;
@@ -491,8 +493,9 @@ struct llama_model {
    //Dense linear projections for SentenceTransformers models like embeddinggemma
    // For Sentence Transformers models structure see
    // https://sbert.net/docs/sentence_transformer/usage/custom_models.html#structure-of-sentence-transformer-models
-    struct ggml_tensor * dense_2_out_layers = nullptr;
-    struct ggml_tensor * dense_3_out_layers = nullptr;
+    struct ggml_tensor * dense_2_out_layers   = nullptr;
+    struct ggml_tensor * dense_2_out_layers_b = nullptr;
+    struct ggml_tensor * dense_3_out_layers   = nullptr;

    // gguf metadata
    std::unordered_map<std::string, std::string> gguf_kv;
@@ -479,6 +479,17 @@ static size_t llama_tensor_quantize_impl(enum ggml_type new_type, const float *
    return new_size;
 }

+static bool tensor_type_requires_imatrix(const ggml_tensor * t, const ggml_type dst_type, const llama_ftype ftype) {
+    return (
+        dst_type == GGML_TYPE_IQ2_XXS || dst_type == GGML_TYPE_IQ2_XS ||
+        dst_type == GGML_TYPE_IQ3_XXS || dst_type == GGML_TYPE_IQ1_S  ||
+        dst_type == GGML_TYPE_IQ2_S   || dst_type == GGML_TYPE_IQ1_M  ||
+        (   // Q2_K_S is the worst k-quant type - only allow it without imatrix for token embeddings
+            dst_type == GGML_TYPE_Q2_K && ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(t->name, "token_embd.weight") != 0
+        )
+    );
+}
+
 static void llama_model_quantize_impl(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
    ggml_type default_type;
    llama_ftype ftype = params->ftype;
@@ -735,24 +746,36 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
    };

    const auto tn = LLM_TN(model.arch);
-    new_ofstream(0);
+
+    // no output file for --dry-run
+    if (!params->dry_run) {
+        new_ofstream(0);
+    }
+
+    // flag for `--dry-run`, to let the user know if imatrix will be required for a real
+    // quantization, as a courtesy
+    bool will_require_imatrix = false;
+
    for (const auto * it : tensors) {
        const auto & weight = *it;
        ggml_tensor * tensor = weight.tensor;
-        if (weight.idx != cur_split && params->keep_split) {
+        if (!params->dry_run && (weight.idx != cur_split && params->keep_split)) {
            close_ofstream();
            new_ofstream(weight.idx);
        }

        const std::string name = ggml_get_name(tensor);
+        const size_t tensor_size = ggml_nbytes(tensor);

-        if (!ml.use_mmap) {
-            if (read_data.size() < ggml_nbytes(tensor)) {
-                read_data.resize(ggml_nbytes(tensor));
+        if (!params->dry_run) {
+            if (!ml.use_mmap) {
+                if (read_data.size() < tensor_size) {
+                    read_data.resize(tensor_size);
+                }
+                tensor->data = read_data.data();
            }
-            tensor->data = read_data.data();
+            ml.load_data_for(tensor);
        }
-        ml.load_data_for(tensor);

        LLAMA_LOG_INFO("[%4d/%4d] %36s - [%s], type = %6s, ",
               ++idx, ml.n_tensors,
@@ -900,129 +923,155 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
            quantize = tensor->type != new_type;
        }

-        if (!quantize) {
-            new_type = tensor->type;
-            new_data = tensor->data;
-            new_size = ggml_nbytes(tensor);
-            LLAMA_LOG_INFO("size = %8.3f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0);
-        } else {
-            const int64_t nelements = ggml_nelements(tensor);
-
-            const float * imatrix = nullptr;
-            if (imatrix_data) {
-                auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
-                if (it == imatrix_data->end()) {
-                    LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
-                } else {
-                    if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
-                        imatrix = it->second.data();
-                    } else {
-                        LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
-                                int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);
-
-                        // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
-                        // this is a significant error and it may be good idea to abort the process if this happens,
-                        // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
-                        // tok_embd should be ignored in this case, since it always causes this warning
-                        if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
-                            throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
-                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
-                        }
-                    }
+        // we have now decided on the target type for this tensor
+        if (params->dry_run) {
+            // the --dry-run option calculates the final quantization size without quantizting
+            if (quantize) {
+                new_size = ggml_nrows(tensor) * ggml_row_size(new_type, tensor->ne[0]);
+                LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB (%s)\n",
+                               tensor_size/1024.0/1024.0,
+                               new_size/1024.0/1024.0,
+                               ggml_type_name(new_type));
+                if (!will_require_imatrix && tensor_type_requires_imatrix(tensor, new_type, params->ftype)) {
+                    will_require_imatrix = true;
                }
-            }
-            if ((new_type == GGML_TYPE_IQ2_XXS ||
-                 new_type == GGML_TYPE_IQ2_XS  ||
-                 new_type == GGML_TYPE_IQ2_S   ||
-                 new_type == GGML_TYPE_IQ1_S   ||
-                (new_type == GGML_TYPE_IQ1_M && strcmp(tensor->name, "token_embd.weight") && strcmp(tensor->name, "output.weight"))  ||
-                (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
-                LLAMA_LOG_ERROR("\n\n============================================================\n");
-                LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
-                LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
-                LLAMA_LOG_ERROR("============================================================\n\n");
-                throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
-            }
-
-            float * f32_data;
-
-            if (tensor->type == GGML_TYPE_F32) {
-                f32_data = (float *) tensor->data;
-            } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
-                throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
            } else {
-                llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
-                f32_data = (float *) f32_conv_buf.data();
+                new_size = tensor_size;
+                LLAMA_LOG_INFO("size = %8.3f MiB\n", new_size/1024.0/1024.0);
            }
+            total_size_org += tensor_size;
+            total_size_new += new_size;
+            continue;
+        } else {
+            // no --dry-run, perform quantization
+            if (!quantize) {
+                new_type = tensor->type;
+                new_data = tensor->data;
+                new_size = tensor_size;
+                LLAMA_LOG_INFO("size = %8.3f MiB\n", tensor_size/1024.0/1024.0);
+            } else {
+                const int64_t nelements = ggml_nelements(tensor);

-            LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
-            fflush(stdout);
+                const float * imatrix = nullptr;
+                if (imatrix_data) {
+                    auto it = imatrix_data->find(remap_imatrix(tensor->name, mapped));
+                    if (it == imatrix_data->end()) {
+                        LLAMA_LOG_INFO("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
+                    } else {
+                        if (it->second.size() == (size_t)tensor->ne[0]*tensor->ne[2]) {
+                            imatrix = it->second.data();
+                        } else {
+                            LLAMA_LOG_INFO("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
+                                    int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name);

-            if (work.size() < (size_t)nelements * 4) {
-                work.resize(nelements * 4); // upper bound on size
-            }
-            new_data = work.data();
-
-            const int64_t n_per_row = tensor->ne[0];
-            const int64_t nrows = tensor->ne[1];
-
-            static const int64_t min_chunk_size = 32 * 512;
-            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
-
-            const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
-            const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
-            const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
-
-            // quantize each expert separately since they have different importance matrices
-            new_size = 0;
-            for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
-                const float * f32_data_03 = f32_data + i03 * nelements_matrix;
-                void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
-                const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
-
-                new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
-
-                // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
-#if 0
-                if (new_type == GGML_TYPE_MXFP4) {
-                    auto * x = f32_data_03;
-
-                    //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
-                    std::vector<float> deq(nrows*n_per_row);
-                    const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
-                    qtype->to_float(new_data_03, deq.data(), deq.size());
-
-                    double err = 0.0f;
-                    for (int i = 0; i < (int) deq.size(); ++i) {
-                        err += fabsf(deq[i] - x[i]);
-                        //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
-                        if (deq[i] != x[i]) {
-                            LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                            // this can happen when quantizing an old mixtral model with split tensors with a new incompatible imatrix
+                            // this is a significant error and it may be good idea to abort the process if this happens,
+                            // since many people will miss the error and not realize that most of the model is being quantized without an imatrix
+                            // tok_embd should be ignored in this case, since it always causes this warning
+                            if (name != tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
+                                throw std::runtime_error(format("imatrix size %d is different from tensor size %d for %s",
+                                        int(it->second.size()), int(tensor->ne[0]*tensor->ne[2]), tensor->name));
+                            }
                        }
                    }
-                    //LLAMA_LOG_INFO("err = %f\n", err);
-                    GGML_ASSERT(err == 0.00000);
                }
+                if (!imatrix && tensor_type_requires_imatrix(tensor, new_type, params->ftype)) {
+                    LLAMA_LOG_ERROR("\n\n============================================================\n");
+                    LLAMA_LOG_ERROR("Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
+                    LLAMA_LOG_ERROR("The result will be garbage, so bailing out\n");
+                    LLAMA_LOG_ERROR("============================================================\n\n");
+                    throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
+                }
+
+                float * f32_data;
+
+                if (tensor->type == GGML_TYPE_F32) {
+                    f32_data = (float *) tensor->data;
+                } else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
+                    throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
+                } else {
+                    llama_tensor_dequantize_impl(tensor, f32_conv_buf, workers, nelements, nthread);
+                    f32_data = (float *) f32_conv_buf.data();
+                }
+
+                LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
+                fflush(stdout);
+
+                if (work.size() < (size_t)nelements * 4) {
+                    work.resize(nelements * 4); // upper bound on size
+                }
+                new_data = work.data();
+
+                const int64_t n_per_row = tensor->ne[0];
+                const int64_t nrows = tensor->ne[1];
+
+                static const int64_t min_chunk_size = 32 * 512;
+                const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
+
+                const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
+                const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
+                const int64_t nthread_use = nthread > 1 ? std::max((int64_t)1, std::min((int64_t)nthread, nchunk)) : 1;
+
+                // quantize each expert separately since they have different importance matrices
+                new_size = 0;
+                for (int64_t i03 = 0; i03 < tensor->ne[2]; ++i03) {
+                    const float * f32_data_03 = f32_data + i03 * nelements_matrix;
+                    void * new_data_03 = (char *)new_data + ggml_row_size(new_type, n_per_row) * i03 * nrows;
+                    const float * imatrix_03 = imatrix ? imatrix + i03 * n_per_row : nullptr;
+
+                    new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
+
+                    // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
+#if 0
+                    if (new_type == GGML_TYPE_MXFP4) {
+                        auto * x = f32_data_03;
+
+                        //LLAMA_LOG_INFO("nrows = %d, n_per_row = %d\n", nrows, n_per_row);
+                        std::vector<float> deq(nrows*n_per_row);
+                        const ggml_type_traits * qtype = ggml_get_type_traits(new_type);
+                        qtype->to_float(new_data_03, deq.data(), deq.size());
+
+                        double err = 0.0f;
+                        for (int i = 0; i < (int) deq.size(); ++i) {
+                            err += fabsf(deq[i] - x[i]);
+                            //if (fabsf(deq[i] - x[i]) > 0.00001 && i < 256) {
+                            if (deq[i] != x[i]) {
+                                LLAMA_LOG_INFO("deq[%d] = %f, x[%d] = %f\n", i, deq[i], i, x[i]);
+                            }
+                        }
+                        //LLAMA_LOG_INFO("err = %f\n", err);
+                        GGML_ASSERT(err == 0.00000);
+                    }
 #endif
+                }
+                LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", tensor_size/1024.0/1024.0, new_size/1024.0/1024.0);
            }
-            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB\n", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
-        }
-        total_size_org += ggml_nbytes(tensor);
-        total_size_new += new_size;
+            total_size_org += tensor_size;
+            total_size_new += new_size;

-        // update the gguf meta data as we go
-        gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
-        GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
-        gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);
+            // update the gguf meta data as we go
+            gguf_set_tensor_type(ctx_outs[cur_split].get(), name.c_str(), new_type);
+            GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), name.c_str())) == new_size);
+            gguf_set_tensor_data(ctx_outs[cur_split].get(), name.c_str(), new_data);

-        // write tensor data + padding
-        fout.write((const char *) new_data, new_size);
-        zeros(fout, GGML_PAD(new_size, align) - new_size);
+            // write tensor data + padding
+            fout.write((const char *) new_data, new_size);
+            zeros(fout, GGML_PAD(new_size, align) - new_size);
+        } // no --dry-run
+    } // iterate over tensors
+
+    if (!params->dry_run) {
+        close_ofstream();
    }
-    close_ofstream();

-    LLAMA_LOG_INFO("%s: model size  = %8.2f MiB\n", __func__, total_size_org/1024.0/1024.0);
-    LLAMA_LOG_INFO("%s: quant size  = %8.2f MiB\n", __func__, total_size_new/1024.0/1024.0);
+    LLAMA_LOG_INFO("%s: model size  = %8.2f MiB (%.2f BPW)\n", __func__, total_size_org/1024.0/1024.0, total_size_org*8.0/ml.n_elements);
+    LLAMA_LOG_INFO("%s: quant size  = %8.2f MiB (%.2f BPW)\n", __func__, total_size_new/1024.0/1024.0, total_size_new*8.0/ml.n_elements);
+
+    if (!params->imatrix && params->dry_run && will_require_imatrix) {
+        LLAMA_LOG_WARN("%s: WARNING: dry run completed successfully, but actually completing this quantization will require an imatrix!\n",
+                       __func__
+        );
+    }

    if (qs.n_fallback > 0) {
        LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) required fallback quantization\n",
@@ -1045,6 +1094,7 @@ llama_model_quantize_params llama_model_quantize_default_params() {
        /*.only_copy                   =*/ false,
        /*.pure                        =*/ false,
        /*.keep_split                  =*/ false,
+        /*.dry_run                     =*/ false,
        /*.imatrix                     =*/ nullptr,
        /*.kv_overrides                =*/ nullptr,
        /*.tensor_type                 =*/ nullptr,
@@ -289,6 +289,15 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
                };
                break;
+            case LLAMA_VOCAB_PRE_TYPE_JAIS2:
+                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}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
+
+                    // adapted: same as llama3 but with cascading whitespace pattern
+                    "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s{512}(?!\\S)|\\s{256}(?!\\S)|\\s{128}(?!\\S)|\\s{64}(?!\\S)|\\s{32}(?!\\S)|\\s{16}(?!\\S)|\\s{8}(?!\\S)|\\s{4}(?!\\S)|\\s{1,2}(?!\\S)|\\s{1}",
+                };
+                break;
            case LLAMA_VOCAB_PRE_TYPE_DBRX:
            case LLAMA_VOCAB_PRE_TYPE_SMAUG:
                regex_exprs = {
@@ -1921,8 +1930,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                    tokenizer_pre == "jina-v2-de" ||
                    tokenizer_pre == "a.x-4.0" ||
                    tokenizer_pre == "mellum"  ||
-                    tokenizer_pre == "modern-bert" ) {
+                    tokenizer_pre == "modern-bert") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT2;
+            } else if (
+                    tokenizer_pre == "jais-2") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_JAIS2;
            } else if (
                    tokenizer_pre == "jina-v1-en" ||
                    tokenizer_pre == "jina-v2-code" ||
@@ -2015,7 +2027,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
            } else if (
                tokenizer_pre == "gpt-4o" ||
-                tokenizer_pre == "llama4") {
+                tokenizer_pre == "llama4" ||
+                tokenizer_pre == "kanana2") {
                pre_type = LLAMA_VOCAB_PRE_TYPE_GPT4O;
                clean_spaces = false;
            } else if (
@@ -2458,6 +2471,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                    || t.first == "<|calls|>"  // solar-open
                    || t.first == "<end_of_turn>"
                    || t.first == "<|endoftext|>"
+                    || t.first == "</s>"      // paddleocr
                    || t.first == "<|eom_id|>"
                    || t.first == "<EOT>"
                    || t.first == "_<EOT>"
@@ -57,6 +57,7 @@ enum llama_vocab_pre_type {
    LLAMA_VOCAB_PRE_TYPE_QWEN35          = 46,
    LLAMA_VOCAB_PRE_TYPE_TINY_AYA        = 47,
    LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM       = 48,
+    LLAMA_VOCAB_PRE_TYPE_JAIS2           = 49,
 };

 struct LLM_KV;
@@ -27,6 +27,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne

    const int64_t S_v = v->ne[0];
    const int64_t H_v = v->ne[1];
+    const bool kda = (g->ne[0] == S_k && g->ne[1] == H_k);

    GGML_ASSERT(S_k == S_v);
    GGML_ASSERT(H_v % H_k == 0);
@@ -35,9 +36,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);

-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
-    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+    GGML_ASSERT(g->ne[0] == 1   || g->ne[0] == S_v);
+    GGML_ASSERT(                   g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
+    GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);

    const float scale = 1.0f / sqrtf(S_k);

@@ -52,8 +54,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
    k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs]
    v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs]
-    g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [  1, n_tokens, H_v, n_seqs]
-    b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [  1, n_tokens, H_v, n_seqs]
+    g = ggml_permute(ctx0, g, 0, 2, 1, 3); // [g_0, n_tokens, H_v, n_seqs]
+    b = ggml_permute(ctx0, b, 0, 2, 1, 3); // [  1, n_tokens, H_v, n_seqs]

    const int CS = CHUNK_SIZE;

@@ -78,33 +80,76 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    v   = ggml_reshape_4d(ctx0, v,   S_v, CS, n_chunks, H_v * n_seqs);
    v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs);

-    g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs);
-    b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs);
+    g = ggml_reshape_4d(ctx0, g, g->ne[0], CS, n_chunks, H_v * n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1,        CS, n_chunks, H_v * n_seqs);

-    // [CS, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_cs = ggml_cumsum(ctx0, g);
+    // [CS, g_0, n_chunks, H_v * n_seqs]
+    // TODO: extend ggml_cumsum with axis parameter to avoid transpose
+    ggml_tensor * g_cs = ggml_cumsum(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, g)));
    cb(g_cs, "g_cs", il);

-    ggml_tensor * g_cs_i = g_cs;
-    ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
+    ggml_tensor * kb = nullptr;
+    ggml_tensor * kq = nullptr;
+    if (kda) {
+        const int64_t CHB = n_chunks * H_k * n_seqs;

-    g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
+        ggml_tensor * g_cs_i = ggml_reshape_4d(ctx0, g_cs, CS, 1, S_k, CHB);  // [chunk_size, 1, S_k, CHB]
+        ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, S_k, CHB);  // [1, chunk_size, S_k, CHB]

-    // [CS, CS, n_chunks, H_v * n_seqs]
-    ggml_tensor * decay_mask;
-    decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
-    decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
-    decay_mask = ggml_exp(ctx0, decay_mask);
-    cb(decay_mask, "decay_mask", il);
+        g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, S_k, CHB);  // [1, chunk_size, S_k, CHB] -> [chunk_size, chunk_size, S_k, CHB]

-    // [CS, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * kb;
-    kb = ggml_mul_mat(ctx0, k,  k_b);
-    kb = ggml_mul    (ctx0, kb, decay_mask);
+        // decay_mask [chunk_size,chunk_size,S_k,CHB]
+        ggml_tensor * decay_mask;
+        decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+        decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
+        decay_mask = ggml_exp(ctx0, decay_mask);
+        cb(decay_mask, "decay_mask", il);
+
+        // decay_mask [S_k,BT_j,BT_i,CHB] *Note* second and third chunk_sizes are switched
+        decay_mask = ggml_cont_4d(ctx0, ggml_permute(ctx0, decay_mask, 2, 1, 0, 3), S_k, CS, CS, CHB);
+
+        ggml_tensor * k_b_i = ggml_reshape_4d(ctx0, k_b, S_k, CS,  1, CHB);
+        ggml_tensor * k_j   = ggml_reshape_4d(ctx0, k,   S_k,  1, CS, CHB);
+        ggml_tensor * q_i   = ggml_reshape_4d(ctx0, q,   S_k, CS,  1, CHB);
+
+        ggml_tensor * decay_k_b_i = ggml_mul(ctx0, decay_mask, k_b_i);
+        ggml_tensor * decay_q_i   = ggml_mul(ctx0, decay_mask, q_i);
+
+        // decay_k_b_i [S,BT,BT,CHB] @ k_j [S,1,BT,CHB] = Akk [BT,1,BT,CHB]
+        kb = ggml_mul_mat(ctx0, decay_k_b_i, k_j);
+        kq = ggml_mul_mat(ctx0, decay_q_i,   k_j);
+
+        kb = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kb, CS, CS, n_chunks, H_v * n_seqs)));
+        kq = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_4d(ctx0, kq, CS, CS, n_chunks, H_v * n_seqs)));
+    } else {
+        ggml_tensor * g_cs_i = g_cs;
+        ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs);
+
+        g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs);
+
+        // [CS, CS, n_chunks, H_v * n_seqs]
+        ggml_tensor * decay_mask;
+        decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i);
+        decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG);
+        decay_mask = ggml_exp(ctx0, decay_mask);
+        cb(decay_mask, "decay_mask", il);
+
+        // [CS, CS, n_chunks, H_k * n_seqs]
+        kb = ggml_mul_mat(ctx0, k,  k_b);
+        kb = ggml_mul    (ctx0, kb, decay_mask);
+
+        // [CS, CS, n_chunks, H_k * n_seqs]
+        kq = ggml_mul_mat(ctx0, k, q);
+        kq = ggml_mul(ctx0, kq, decay_mask);
+    }
+
+    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
+    cb(kq, "kq", il);

    // [CS, CS, n_chunks, H_k * n_seqs]
    ggml_tensor * attn;
    attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER);
+    cb(attn, "attn", il);

    ggml_tensor * identity;
    identity = ggml_view_1d(ctx0, attn, CS, 0);
@@ -115,6 +160,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    cb(lhs, "dnet_add_ch_lhs", il);

    attn = ggml_neg(ctx0, attn);
+    cb(attn, "attn_pre_solve", il);

    ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false);
    attn = ggml_add(ctx0, lin_solve, identity);
@@ -123,7 +169,7 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    // [S_v, CS, n_chunks, H_v * n_seqs]
    v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn);

-    // [CS, 1, n_chunks, H_v * n_seqs]
+    // [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
    ggml_tensor * g_exp = ggml_exp(ctx0, g_cs);

    k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b));
@@ -136,16 +182,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn);
    cb(k_cd, "k_cumdecay", il);

-    // [S_k, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp);
+    // [1, CS, n_chunks, H_k * n_seqs] KDA: [S_k, CS, n_chunks, H_k * n_seqs]
+    ggml_tensor * g_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_exp));
    ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t);

-    // [CS, CS, n_chunks, H_k * n_seqs]
-    ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
-    kq = ggml_mul(ctx0, kq, decay_mask);
-    kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG);
-    cb(kq, "kq", il);
-
    // vectorized calculation of key_gdiff
    // improved from the chunked version:
    //   g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1)
@@ -156,8 +196,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne

    // get last element in g_cumsum along CS dimension (ne0)
    // example: [[x, y, z, ..., last], ...] -> [[last], ...]
-    // [1, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3],
+    // [1, 1, n_chunks, H_v * n_seqs] KDA: [1, S_k, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, g_cs->ne[1], g_cs->ne[2], g_cs->ne[3],
            g_cs->nb[1],
            g_cs->nb[2],
            g_cs->nb[3],
@@ -167,16 +207,15 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    // TODO: remove this cont when CUDA supports non-cont unary ops
    g_last = ggml_cont(ctx0, g_last);

-    // [1, 1, n_chunks, H_v * n_seqs]
-    ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last);
-    cb(g_last_exp, "g_last_exp", il);
+    // [1, 1, n_chunks, H_v * n_seqs] KDA: [S_k, 1, n_chunks, H_v * n_seqs]
+    ggml_tensor * g_last_exp_t = ggml_transpose(ctx0, ggml_exp(ctx0, g_last));
+    cb(g_last_exp_t, "g_last_exp_t", il);

-    // [CS, 1, n_chunks, H_v * n_seqs]
+    // [CS, 1, n_chunks, H_v * n_seqs] KDA: [CS, S_k, n_chunks, H_v * n_seqs]
    ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last));
    cb(g_diff, "g_diff", il);

-    ggml_tensor * g_diff_exp   = ggml_exp(ctx0, g_diff);
-    ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp);
+    ggml_tensor * g_diff_exp_t = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_exp(ctx0, g_diff)));

    // [S_k, CS, n_chunks, H_v * n_seqs]
    ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t);
@@ -227,8 +266,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
        ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs]

        // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew
-        ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk);
-        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp);
+        ggml_tensor * ch_g_last_exp_t = get_slice_2d(ctx0, g_last_exp_t, chunk);
+
+        s_t = ggml_mul(ctx0, s_t, ch_g_last_exp_t);
        s_t = ggml_add(ctx0, s_t, kgv);
        cb(s_t, "dnet_add_ch_state", il);
    }
@@ -241,9 +281,9 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
            ggml_row_size(v->type, S_v),
            ggml_row_size(v->type, S_v * CS * n_chunks),
            ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0);
-
    o = ggml_permute  (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs]
-    s = ggml_transpose(ctx0, s_t);           // [S_v, S_v, H_v, n_seqs]
+    s = ggml_transpose(ctx0, s_t);
+    cb(s, "output_state", il);

    return {o, s};
 }
@@ -273,9 +313,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);

-    GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs);
-    GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
-    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
+    GGML_ASSERT(g->ne[0] == 1   || g->ne[0] == S_v);
+    GGML_ASSERT(                   g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
+    GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);

    const float scale = 1.0f / sqrtf(S_k);

@@ -291,8 +332,10 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
    cb(b, "b_in", il);
    cb(g, "g_in", il);

-    g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs);
-    b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs);
+    // GDA: [1,  1,  H_v, n_seqs]
+    // KDA: [1, S_k, H_v, n_seqs]
+    g = ggml_reshape_4d(ctx0, g, 1, g->ne[0], H_v, n_seqs);
+    b = ggml_reshape_4d(ctx0, b, 1,        1, H_v, n_seqs);

    // [S_v, S_v, H_v, n_seqs]
    g = ggml_exp(ctx0, g);
--- a/Show More
+++ b/Show More