llama : reuse graph result instances across scheduler reserves (#32089 )

Allocate gf_res_reserve and both gf_res_prev entries once in the llama_context constructor. Add llm_graph_result::set_max_nodes() and reset the existing instances in sched_reserve() instead of recreating them. get_gf_res_prev() now only selects the arena for the current batch. Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
jinja : support chained comparisons and argument unpacking (#32072 )
2026-11-11 13:27:40 -06:00 · 2026-11-11 19:58:39 +02:00 · 2026-11-11 11:53:04 -06:00 · 2026-11-11 17:34:55 +01:00 · 2026-11-11 15:13:14 +01:00 · 2026-11-11 15:16:16 +02:00
389 changed files with 14272 additions and 7591 deletions
@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2026.1.1-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2026.2.0-devel-ubuntu24.04
 ARG BUILD_DATE=N/A
 ARG APP_VERSION=N/A
 ARG APP_REVISION=N/A
@@ -22,20 +22,20 @@ RUN LLAMA_BUILD_NUMBER="$APP_VERSION" npm run build
 FROM docker.io/intel/deep-learning-essentials:$ONEAPI_VERSION AS build

 RUN echo "Download oneDNN installation package" && \
-    wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/94c3dbac-0852-45be-a57d-21c204cada3e/intel-onednn-2026.0.2.46_offline.sh -O intel-onednn_offline.sh && \
+    wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/46452a4a-7b7f-49f2-be4c-602a900180bb/intel-onednn-2026.1.0.625_offline.sh -O intel-onednn_offline.sh && \
    echo "Install oneDNN components" && \
    bash intel-onednn_offline.sh -s -a --silent --eula accept

 ARG GGML_SYCL_F16=ON
-ARG LEVEL_ZERO_VERSION=1.28.2
+ARG LEVEL_ZERO_VERSION=1.34.0
 ARG LEVEL_ZERO_UBUNTU_VERSION=u24.04
 RUN apt-get update && \
    apt-get install -y git libssl-dev wget ca-certificates && \
    cd /tmp && \
-    wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero.deb && \
-    wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb && \
-    apt-get -o Dpkg::Options::="--force-overwrite" install -y ./level-zero.deb ./level-zero-devel.deb && \
-    rm -f /tmp/level-zero.deb /tmp/level-zero-devel.deb
+    wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze1_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze1.deb && \
+    wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/libze-dev_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O libze-dev.deb && \
+    apt-get -o Dpkg::Options::="--force-overwrite" install -y ./libze1.deb ./libze-dev.deb && \
+    rm -f /tmp/libze1.deb /tmp/libze-dev.deb

 WORKDIR /app

@@ -91,11 +91,11 @@ COPY --from=build /opt/intel/oneapi/dnnl /opt/intel/oneapi/dnnl
 #ARG IGDGMM_VERSION=22.8.2


-ARG IGC_VERSION=v2.34.4
-ARG IGC_VERSION_FULL=2_2.34.4+21428
-ARG COMPUTE_RUNTIME_VERSION=26.18.38308.1
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.18.38308.1-0
-ARG IGDGMM_VERSION=22.10.0
+ARG IGC_VERSION=v2.42.6
+ARG IGC_VERSION_FULL=2_2.42.6+23070
+ARG COMPUTE_RUNTIME_VERSION=26.40.40196.8
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.40.40196.8-0
+ARG IGDGMM_VERSION=22.10.2
 RUN mkdir /tmp/neo/ && cd /tmp/neo/ \
  && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-core-${IGC_VERSION_FULL}_amd64.deb \
  && wget https://github.com/intel/intel-graphics-compiler/releases/download/$IGC_VERSION/intel-igc-opencl-${IGC_VERSION_FULL}_amd64.deb \
@@ -37,7 +37,7 @@ env:
  LLAMA_ARG_LOG_PREFIX: 1
  LLAMA_ARG_LOG_TIMESTAMPS: 1
  # TODO: fix failing tests on OpenVINO backend
-  CTEST_EXCLUDE: "test-llama-archs|test-llama-context"
+  CTEST_EXCLUDE: "test-llama-archs|test-llama-context|test-fp8-kv"

 jobs:
  ubuntu-24-openvino:
@@ -52,8 +52,8 @@ jobs:

    env:
      ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2026.1"
-      LEVEL_ZERO_VERSION: "1.33.1"
+      ONEAPI_INSTALLER_VERSION: "2026.2"
+      LEVEL_ZERO_VERSION: "1.34.0"
      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"

    continue-on-error: true
@@ -68,12 +68,12 @@ jobs:
        run: |
          cd /tmp
          echo "Download oneAPI Deep Learning Essentials installation package"
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b5c100a7-8052-4e88-90e2-63d9916a998b/intel-deep-learning-essentials-2026.1.4.17_offline.sh -O intel-deep-learning-essentials_offline.sh
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/aef232d2-aac5-4cd6-bceb-911019bbb6a8/intel-deep-learning-essentials-2026.2.0.640_offline.sh -O intel-deep-learning-essentials_offline.sh
          echo "Install oneAPI Deep Learning Essentials components"
          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept

          echo "Download oneDNN installation package"
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/94c3dbac-0852-45be-a57d-21c204cada3e/intel-onednn-2026.0.2.46_offline.sh -O intel-onednn_offline.sh
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/46452a4a-7b7f-49f2-be4c-602a900180bb/intel-onednn-2026.1.0.625_offline.sh -O intel-onednn_offline.sh
          echo "Install oneDNN components"
          sudo bash intel-onednn_offline.sh -s -a --silent --eula accept

@@ -150,13 +150,13 @@ jobs:
        shell: bash

    env:
-      WINDOWS_DLE_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/a51104eb-6f4a-4a0e-9dae-1b81b0c81ae1/intel-deep-learning-essentials-2026.1.4.16_offline.exe
-      WINDOWS_ONEDNN_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/34e28083-d04b-46a3-826c-8c1d00bf1e1c/intel-onednn-2026.0.2.47_offline.exe
+      WINDOWS_DLE_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b177a006-5cd9-4cb2-807f-8199cc73d53c/intel-deep-learning-essentials-2026.2.0.638_offline.exe
+      WINDOWS_ONEDNN_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/8f905666-c85c-43fb-9bd0-2618c7ecd33b/intel-onednn-2026.1.0.622_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.tbb.devel
      WINDOWS_DNNL: intel.oneapi.win.dnnl
-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.34.0/level-zero-win-sdk-1.34.0.zip
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2026.1"
+      ONEAPI_INSTALLER_VERSION: "2026.2"
    steps:
      - name: Clone
        id: checkout
@@ -58,12 +58,7 @@ jobs:
      - name: Install dependencies
        run: |
          sudo apt update
-          sudo apt install -y software-properties-common
-          sudo add-apt-repository ppa:deadsnakes/ppa
-          sudo apt update
-          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3.12 python3.12-venv
-          sudo update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 1
-          sudo update-alternatives --config python3
+          sudo apt install -y build-essential cmake libxcb-xinput0 libxcb-xinerama0 libxcb-cursor-dev libvulkan1 mesa-vulkan-drivers libglvnd0 libgl1 libglx0 libegl1 libgles2 libssl-dev time unzip wget python3 python3-venv python3-pip

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.24
@@ -35,7 +35,7 @@ jobs:
        uses: actions/setup-python@v6
        with:
          python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.88
+          pip-install: -r requirements/requirements-all.txt ty==0.0.100
      # - name: Type-check with Pyright
      #   uses: jakebailey/pyright-action@v2
      #   with:
@@ -1307,14 +1307,14 @@ jobs:
        shell: bash

    env:
-      WINDOWS_DLE_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/a51104eb-6f4a-4a0e-9dae-1b81b0c81ae1/intel-deep-learning-essentials-2026.1.4.16_offline.exe
-      WINDOWS_ONEDNN_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/34e28083-d04b-46a3-826c-8c1d00bf1e1c/intel-onednn-2026.0.2.47_offline.exe
+      WINDOWS_DLE_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b177a006-5cd9-4cb2-807f-8199cc73d53c/intel-deep-learning-essentials-2026.2.0.638_offline.exe
+      WINDOWS_ONEDNN_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/8f905666-c85c-43fb-9bd0-2618c7ecd33b/intel-onednn-2026.1.0.622_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.tbb.devel
      WINDOWS_DNNL: intel.oneapi.win.dnnl

-      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.33.1/level-zero-win-sdk-1.33.1.zip
+      LEVEL_ZERO_SDK_URL: https://github.com/oneapi-src/level-zero/releases/download/v1.34.0/level-zero-win-sdk-1.34.0.zip
      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-      ONEAPI_INSTALLER_VERSION: "2026.1"
+      ONEAPI_INSTALLER_VERSION: "2026.2"

    steps:
      - name: Clone
@@ -1363,11 +1363,12 @@ jobs:
          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_def.3.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_avx2.3.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_intel_thread.3.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/libimalloc.dll" ./build/bin

          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_level_zero_v2.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_adapter_opencl.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_loader_shared.dll" ./build/bin
          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/ur_win_proxy_loader.dll" ./build/bin
          ZE_LOADER_DLL=$(find "${{ env.ONEAPI_ROOT }}" "$LEVEL_ZERO_V1_SDK_PATH" -iname ze_loader.dll -print -quit 2>/dev/null || true)
          if [ -n "$ZE_LOADER_DLL" ]; then
@@ -1421,8 +1422,8 @@ jobs:

    env:
      ONEAPI_ROOT: /opt/intel/oneapi/
-      ONEAPI_INSTALLER_VERSION: "2026.1"
-      LEVEL_ZERO_VERSION: "1.33.1"
+      ONEAPI_INSTALLER_VERSION: "2026.2"
+      LEVEL_ZERO_VERSION: "1.34.0"
      LEVEL_ZERO_UBUNTU_VERSION: "u24.04"

    steps:
@@ -1437,12 +1438,12 @@ jobs:
        run: |
          cd /tmp
          echo "Download oneAPI Deep Learning Essentials installation package"
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b5c100a7-8052-4e88-90e2-63d9916a998b/intel-deep-learning-essentials-2026.1.4.17_offline.sh -O intel-deep-learning-essentials_offline.sh
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/aef232d2-aac5-4cd6-bceb-911019bbb6a8/intel-deep-learning-essentials-2026.2.0.640_offline.sh -O intel-deep-learning-essentials_offline.sh
          echo "Install oneAPI Deep Learning Essentials components"
          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept

          echo "Download oneDNN installation package"
-          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/94c3dbac-0852-45be-a57d-21c204cada3e/intel-onednn-2026.0.2.46_offline.sh -O intel-onednn_offline.sh
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/46452a4a-7b7f-49f2-be4c-602a900180bb/intel-onednn-2026.1.0.625_offline.sh -O intel-onednn_offline.sh
          echo "Install oneDNN components"
          sudo bash intel-onednn_offline.sh -s -a --silent --eula accept

@@ -1,6 +1,6 @@
 # llama.cpp

-![llama](https://raw.githubusercontent.com/ggml-org/llama.brand/refs/heads/master/cover/llama-cpp/cover-llama-cpp-dark.svg)
+![llama.cpp](https://raw.githubusercontent.com/ggml-org/llama.brand/refs/heads/master/logo/svg/cover/llama-cpp/cover-llama-cpp-dark.svg)

 <div align="center">

@@ -180,7 +180,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

    # TODO: fix failing tests on OpenVINO backend
-    CTEST_EXTRA="-E test-llama-archs|test-llama-context"
+    CTEST_EXTRA="-E test-llama-archs|test-llama-context|test-fp8-kv"
 fi

 ## helpers
@@ -66,6 +66,8 @@ add_library(${TARGET}
    chat-diff-analyzer.cpp
    chat-peg-parser.cpp
    chat-peg-parser.h
+    chat-tts-norm.cpp
+    chat-tts-norm.h
    chat.cpp
    chat.h
    common.cpp
@@ -282,6 +282,7 @@ const std::vector<ggml_type> kv_cache_types = {
    GGML_TYPE_F32,
    GGML_TYPE_F16,
    GGML_TYPE_BF16,
+    GGML_TYPE_F8_E4M3,
    GGML_TYPE_Q8_0,
    GGML_TYPE_Q4_0,
    GGML_TYPE_Q4_1,
@@ -3828,15 +3829,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.prefill_assistant = value;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_PREFILL_ASSISTANT"));
-    add_opt(common_arg(
-        {"--chat-input-special"},
-        {"--no-chat-input-special"},
-        string_format("experimental: whether to parse special tokens from user input messages in chat completions, does not affect the raw /completions endpoint (default: %s)\n"
-                      "note: this is not a security feature, do not rely on it to prevent prompt injection", params.chat_input_special ? "enabled" : "disabled"),
-        [](common_params & params, bool value) {
-            params.chat_input_special = value;
-        }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_INPUT_SPECIAL"));
    add_opt(common_arg(
        {"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
        string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),
@@ -4500,7 +4492,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.tts_lang = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_TTS}));
+    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
    add_opt(common_arg(
        {"--tts-voice"}, "VOICE",
        "voice for audio generation (model-dependent)\n"
@@ -4508,7 +4500,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        [](common_params & params, const std::string & value) {
            params.tts_voice = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_TTS}));
+    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
    add_opt(common_arg(
        {"--tts-speaker-file"}, "FNAME",
        "speaker file path for audio generation",
@@ -131,6 +131,12 @@ static std::string normalize_quotes_to_json(const std::string & input) {
            continue;
        }

+        // Models can emit raw control characters inside string literals; escape them to keep the JSON valid.
+        if ((in_single_quoted || in_double_quoted) && static_cast<unsigned char>(c) < 0x20) {
+            result += escape_json_string_inner(std::string(1, c));
+            continue;
+        }
+
        // Handle quote characters
        if (c == '"') {
            if (in_single_quoted) {
@@ -0,0 +1,864 @@
+#include "chat-tts-norm.h"
+
+#include "jinja/runtime.h"
+#include "jinja/value.h"
+
+#include <algorithm>
+#include <climits>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+// TTS normalization patterns, see docs/tts-norm-patterns.md for the specification
+// keep this file in sync with the python reference: scripts/chat_tts_norm.py
+
+using namespace jinja;
+
+//
+// character classes
+// they are defined with explicit codepoint ranges, independent of any unicode database version,
+// so that the C++ and python implementations agree on every codepoint
+//
+
+struct tts_norm_range {
+    uint32_t lo;
+    uint32_t hi;
+};
+
+static const tts_norm_range TTS_NORM_SPACE[] = {
+    {0x0009, 0x000D}, {0x0020, 0x0020}, {0x0085, 0x0085}, {0x00A0, 0x00A0}, {0x1680, 0x1680},
+    {0x2000, 0x200A}, {0x2028, 0x2029}, {0x202F, 0x202F}, {0x205F, 0x205F}, {0x3000, 0x3000},
+};
+
+static const tts_norm_range TTS_NORM_PUNCT[] = {
+    {0x0021, 0x002F}, {0x003A, 0x0040}, {0x005B, 0x0060}, {0x007B, 0x007E},
+    {0x00A1, 0x00A9}, {0x00AB, 0x00B4}, {0x00B6, 0x00B9}, {0x00BB, 0x00BF}, {0x00D7, 0x00D7}, {0x00F7, 0x00F7},
+    {0x2010, 0x2027}, {0x2030, 0x205E}, {0x20A0, 0x20CF}, {0x2100, 0x218F}, {0x2190, 0x2BFF},
+    {0x3001, 0x3003}, {0x3008, 0x3020}, {0x3030, 0x3030}, {0x303D, 0x303D}, {0x30FB, 0x30FB},
+    {0xFE10, 0xFE19}, {0xFE30, 0xFE4F}, {0xFE50, 0xFE6B},
+    {0xFF01, 0xFF0F}, {0xFF1A, 0xFF20}, {0xFF3B, 0xFF40}, {0xFF5B, 0xFF65},
+    {0x1F000, 0x1FAFF},
+};
+
+static const tts_norm_range TTS_NORM_CONTROL[] = {
+    {0x0000, 0x001F}, {0x007F, 0x009F}, {0x200B, 0x200F}, {0x2060, 0x206F}, {0xFEFF, 0xFEFF},
+};
+
+static const tts_norm_range TTS_NORM_HAN[] = {
+    {0x2E80, 0x2FDF}, {0x3005, 0x3005}, {0x3007, 0x3007}, {0x3021, 0x3029}, {0x3038, 0x303B},
+    {0x3400, 0x4DBF}, {0x4E00, 0x9FFF}, {0xF900, 0xFAFF}, {0x20000, 0x323AF},
+};
+
+template<size_t N>
+static bool tts_norm_in(const tts_norm_range (&ranges)[N], uint32_t c) {
+    for (const auto & r : ranges) {
+        if (c >= r.lo && c <= r.hi) {
+            return true;
+        }
+    }
+    return false;
+}
+
+static bool tts_norm_is_alpha(uint32_t c) {
+    if ((c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z')) {
+        return true;
+    }
+    if (c == 0xAA || c == 0xB5 || c == 0xBA) {
+        return true;
+    }
+    return c >= 0xC0 && !tts_norm_in(TTS_NORM_SPACE, c) && !tts_norm_in(TTS_NORM_PUNCT, c) && !tts_norm_in(TTS_NORM_CONTROL, c);
+}
+
+// kind is one of "dspawulhc"
+static bool tts_norm_class_match(char kind, uint32_t c) {
+    switch (kind) {
+        case 'd': return c >= '0' && c <= '9';
+        case 's': return tts_norm_in(TTS_NORM_SPACE, c);
+        case 'p': return tts_norm_in(TTS_NORM_PUNCT, c);
+        case 'a': return tts_norm_is_alpha(c);
+        case 'w': return tts_norm_is_alpha(c) || (c >= '0' && c <= '9');
+        case 'u': return c >= 'A' && c <= 'Z';
+        case 'l': return c >= 'a' && c <= 'z';
+        case 'h': return tts_norm_in(TTS_NORM_HAN, c);
+        case 'c': return tts_norm_in(TTS_NORM_CONTROL, c);
+    }
+    return false;
+}
+
+static bool tts_norm_is_class_letter(uint32_t c) {
+    if (c >= 'A' && c <= 'Z') {
+        c += 'a' - 'A';
+    }
+    return c < 0x80 && std::string("dspawulhc").find(static_cast<char>(c)) != std::string::npos;
+}
+
+static uint32_t tts_norm_ascii_lower(uint32_t c) {
+    return (c >= 'A' && c <= 'Z') ? c + ('a' - 'A') : c;
+}
+
+static uint32_t tts_norm_ascii_swapcase(uint32_t c) {
+    if (c >= 'A' && c <= 'Z') return c + ('a' - 'A');
+    if (c >= 'a' && c <= 'z') return c - ('a' - 'A');
+    return c;
+}
+
+static bool tts_norm_is_ascii_alnum(uint32_t c) {
+    return (c >= '0' && c <= '9') || (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+// invalid bytes are decoded as U+FFFD, one per byte; offs[i] is the byte offset of codepoint i, offs[n] = text size
+static void tts_norm_decode_utf8(const std::string & s, std::vector<uint32_t> & cpts, std::vector<size_t> & offs) {
+    cpts.clear();
+    offs.clear();
+    size_t i = 0;
+    const size_t n = s.size();
+    while (i < n) {
+        offs.push_back(i);
+        const unsigned char c = static_cast<unsigned char>(s[i]);
+        uint32_t cp = 0;
+        size_t len = 0;
+        if (c < 0x80) {
+            cp = c; len = 1;
+        } else if ((c & 0xE0) == 0xC0) {
+            cp = c & 0x1F; len = 2;
+        } else if ((c & 0xF0) == 0xE0) {
+            cp = c & 0x0F; len = 3;
+        } else if ((c & 0xF8) == 0xF0) {
+            cp = c & 0x07; len = 4;
+        }
+        bool ok = len > 0 && i + len <= n;
+        for (size_t k = 1; ok && k < len; k++) {
+            const unsigned char cc = static_cast<unsigned char>(s[i + k]);
+            if ((cc & 0xC0) != 0x80) {
+                ok = false;
+            } else {
+                cp = (cp << 6) | (cc & 0x3F);
+            }
+        }
+        if (!ok) {
+            cpts.push_back(0xFFFD);
+            i += 1;
+        } else {
+            cpts.push_back(cp);
+            i += len;
+        }
+    }
+    offs.push_back(n);
+}
+
+//
+// AST
+//
+
+enum class tts_norm_op {
+    LITERAL,    // one codepoint
+    SET,        // one codepoint from a set; class escapes like %d are sets with a single class
+    ANY,        // any codepoint
+    START,      // ^
+    END,        // $
+    BOUNDARY,   // %b
+    GROUP,      // (...), (?:...), (?<name>...)
+    LOOKAHEAD,  // (?=...), (?!...)
+    LOOKBEHIND, // (?<=c), (?<!c) - exactly one codepoint
+};
+
+struct tts_norm_class {
+    char kind;
+    bool neg;
+};
+
+struct tts_norm_set {
+    bool neg = false;
+    std::vector<std::pair<uint32_t, uint32_t>> ranges;
+    std::vector<tts_norm_class> classes;
+};
+
+struct tts_norm_item;
+using tts_norm_seq = std::vector<tts_norm_item>;
+
+struct tts_norm_pattern_node {
+    tts_norm_op op = tts_norm_op::GROUP;
+    uint32_t cp = 0;           // LITERAL
+    tts_norm_set set;               // SET
+    std::vector<tts_norm_seq> alts; // GROUP, LOOKAHEAD; LOOKBEHIND holds a single item
+    int capture = -1;          // GROUP: capture index, -1 for non-capturing
+    bool neg = false;          // LOOKAHEAD, LOOKBEHIND
+};
+
+static constexpr int TTS_NORM_INF = INT_MAX;
+static constexpr int TTS_NORM_MAX_REPEAT = 1000;
+static constexpr size_t TTS_NORM_MAX_PATTERN_LEN = 4096;
+
+struct tts_norm_item {
+    tts_norm_pattern_node atom;
+    int min = 1;
+    int max = 1;
+};
+
+static bool tts_norm_is_zero_width(tts_norm_op op) {
+    return op == tts_norm_op::START || op == tts_norm_op::END || op == tts_norm_op::BOUNDARY
+        || op == tts_norm_op::LOOKAHEAD || op == tts_norm_op::LOOKBEHIND;
+}
+
+//
+// parser
+//
+
+struct tts_norm_parser {
+    std::vector<uint32_t> s;
+    size_t p = 0;
+    size_t & n_caps;
+
+    tts_norm_parser(std::vector<uint32_t> src, size_t & n_caps) : s(std::move(src)), n_caps(n_caps) {}
+
+    [[noreturn]] void fail(const std::string & msg) const {
+        throw std::runtime_error("invalid TTS norm pattern: " + msg + " at position " + std::to_string(p));
+    }
+
+    bool eof() const { return p >= s.size(); }
+    uint32_t peek() const { return s[p]; }
+    bool peek_is(uint32_t c) const { return !eof() && s[p] == c; }
+
+    void expect(uint32_t c, const char * what) {
+        if (!peek_is(c)) {
+            fail(std::string("expected ") + what);
+        }
+        p++;
+    }
+
+    std::vector<tts_norm_seq> parse_alts() {
+        std::vector<tts_norm_seq> alts;
+        alts.push_back(parse_seq());
+        while (peek_is('|')) {
+            p++;
+            alts.push_back(parse_seq());
+        }
+        return alts;
+    }
+
+    tts_norm_seq parse_seq() {
+        tts_norm_seq seq;
+        while (!eof() && peek() != '|' && peek() != ')') {
+            tts_norm_item item;
+            item.atom = parse_atom();
+            parse_quantifier(item);
+            seq.push_back(std::move(item));
+        }
+        return seq;
+    }
+
+    int parse_int() {
+        if (eof() || peek() < '0' || peek() > '9') {
+            fail("expected number in quantifier");
+        }
+        int v = 0;
+        while (!eof() && peek() >= '0' && peek() <= '9') {
+            v = v * 10 + static_cast<int>(peek() - '0');
+            if (v > TTS_NORM_MAX_REPEAT) {
+                fail("repetition count exceeds " + std::to_string(TTS_NORM_MAX_REPEAT));
+            }
+            p++;
+        }
+        return v;
+    }
+
+    void parse_quantifier(tts_norm_item & item) {
+        if (eof()) {
+            return;
+        }
+        const uint32_t c = peek();
+        if (c != '?' && c != '*' && c != '+' && c != '{') {
+            return;
+        }
+        if (tts_norm_is_zero_width(item.atom.op)) {
+            fail("zero-width assertion cannot be repeated");
+        }
+        p++;
+        if (c == '?') {
+            item.min = 0; item.max = 1;
+        } else if (c == '*') {
+            item.min = 0; item.max = TTS_NORM_INF;
+        } else if (c == '+') {
+            item.min = 1; item.max = TTS_NORM_INF;
+        } else {
+            item.min = parse_int();
+            item.max = item.min;
+            if (peek_is(',')) {
+                p++;
+                item.max = peek_is('}') ? TTS_NORM_INF : parse_int();
+            }
+            expect('}', "'}' to close quantifier");
+            if (item.max < item.min) {
+                fail("quantifier max is smaller than min");
+            }
+        }
+        if (!eof() && (peek() == '?' || peek() == '*' || peek() == '+' || peek() == '{')) {
+            fail("multiple quantifiers");
+        }
+    }
+
+    tts_norm_pattern_node parse_group() {
+        // '(' already consumed
+        tts_norm_pattern_node node;
+        node.op = tts_norm_op::GROUP;
+        if (peek_is('?')) {
+            p++;
+            if (eof()) {
+                fail("unterminated group");
+            }
+            const uint32_t c = peek();
+            if (c == ':') {
+                p++;
+            } else if (c == '=' || c == '!') {
+                p++;
+                node.op  = tts_norm_op::LOOKAHEAD;
+                node.neg = c == '!';
+            } else if (c == '<' && p + 1 < s.size() && (s[p + 1] == '=' || s[p + 1] == '!')) {
+                node.op  = tts_norm_op::LOOKBEHIND;
+                node.neg = s[p + 1] == '!';
+                p += 2;
+                tts_norm_item item;
+                if (eof() || peek() == ')') {
+                    fail("lookbehind must contain exactly one codepoint matcher");
+                }
+                item.atom = parse_atom();
+                if (item.atom.op != tts_norm_op::LITERAL && item.atom.op != tts_norm_op::SET && item.atom.op != tts_norm_op::ANY) {
+                    fail("lookbehind must contain exactly one codepoint matcher");
+                }
+                if (!peek_is(')')) {
+                    fail("lookbehind must contain exactly one codepoint matcher");
+                }
+                p++;
+                node.alts.push_back(tts_norm_seq{});
+                node.alts[0].push_back(std::move(item));
+                return node;
+            } else if (c == 'i') {
+                fail("(?i) is only allowed at the start of the pattern");
+            } else {
+                fail("unknown group type");
+            }
+        } else {
+            node.capture = static_cast<int>(n_caps++);
+        }
+        node.alts = parse_alts();
+        expect(')', "')' to close group");
+        return node;
+    }
+
+    // after '%' inside or outside of a set
+    // returns true if it is a class (stored in cls), false if it is a literal (stored in lit)
+    bool parse_escape(bool in_set, tts_norm_class & cls, uint32_t & lit, bool & boundary) {
+        boundary = false;
+        if (eof()) {
+            fail("dangling '%'");
+        }
+        const uint32_t e = s[p++];
+        if (tts_norm_is_class_letter(e)) {
+            cls.kind = static_cast<char>(tts_norm_ascii_lower(e));
+            cls.neg  = e >= 'A' && e <= 'Z';
+            return true;
+        }
+        if (e == 'b' && !in_set) {
+            boundary = true;
+            return false;
+        }
+        if (tts_norm_is_ascii_alnum(e)) {
+            p--;
+            fail(std::string("unknown escape '%") + static_cast<char>(e) + "'");
+        }
+        lit = e;
+        return false;
+    }
+
+    tts_norm_pattern_node parse_set() {
+        // '[' already consumed
+        tts_norm_pattern_node node;
+        node.op = tts_norm_op::SET;
+        if (peek_is('^')) {
+            node.set.neg = true;
+            p++;
+        }
+        bool empty = true;
+        while (true) {
+            if (eof()) {
+                fail("unterminated set");
+            }
+            if (peek() == ']') {
+                p++;
+                break;
+            }
+            empty = false;
+            uint32_t lo = 0;
+            if (peek() == '%') {
+                p++;
+                tts_norm_class cls{};
+                bool boundary = false;
+                if (parse_escape(true, cls, lo, boundary)) {
+                    node.set.classes.push_back(cls);
+                    continue;
+                }
+            } else {
+                lo = s[p++];
+            }
+            uint32_t hi = lo;
+            if (peek_is('-') && p + 1 < s.size() && s[p + 1] != ']') {
+                p++;
+                if (peek() == '%') {
+                    p++;
+                    tts_norm_class cls{};
+                    bool boundary = false;
+                    if (parse_escape(true, cls, hi, boundary)) {
+                        fail("character class cannot be used in a range");
+                    }
+                } else {
+                    hi = s[p++];
+                }
+                if (hi < lo) {
+                    fail("invalid range in set");
+                }
+            }
+            node.set.ranges.emplace_back(lo, hi);
+        }
+        if (empty) {
+            fail("empty set");
+        }
+        return node;
+    }
+
+    tts_norm_pattern_node parse_atom() {
+        const uint32_t c = s[p++];
+        tts_norm_pattern_node node;
+        switch (c) {
+            case '(':
+                return parse_group();
+            case '[':
+                return parse_set();
+            case '.':
+                node.op = tts_norm_op::ANY;
+                return node;
+            case '^':
+                node.op = tts_norm_op::START;
+                return node;
+            case '$':
+                node.op = tts_norm_op::END;
+                return node;
+            case '%': {
+                tts_norm_class cls{};
+                uint32_t lit = 0;
+                bool boundary = false;
+                if (parse_escape(false, cls, lit, boundary)) {
+                    node.op = tts_norm_op::SET;
+                    node.set.classes.push_back(cls);
+                } else if (boundary) {
+                    node.op = tts_norm_op::BOUNDARY;
+                } else {
+                    node.op = tts_norm_op::LITERAL;
+                    node.cp = lit;
+                }
+                return node;
+            }
+            case '?':
+            case '*':
+            case '+':
+            case '{':
+                p--;
+                fail("nothing to repeat");
+            case ']':
+            case '}':
+                p--;
+                fail(std::string("unescaped '") + static_cast<char>(c) + "', use '%" + static_cast<char>(c) + "'");
+            default:
+                node.op = tts_norm_op::LITERAL;
+                node.cp = c;
+                return node;
+        }
+    }
+};
+
+//
+// matcher
+// PEG-style: alternatives are ordered (the first one that matches wins) and quantifiers are greedy and possessive
+// (they never give back what they consumed), so there is no backtracking
+//
+
+struct tts_norm_matcher {
+    const std::vector<uint32_t> & t;
+    bool icase;
+    std::vector<std::pair<int64_t, int64_t>> caps;
+
+    bool is_w(size_t i) const {
+        return i < t.size() && tts_norm_class_match('w', t[i]);
+    }
+
+    bool set_match(const tts_norm_set & st, uint32_t c) const {
+        auto test = [&](uint32_t x) {
+            for (const auto & r : st.ranges) {
+                if (x >= r.first && x <= r.second) {
+                    return true;
+                }
+            }
+            for (const auto & k : st.classes) {
+                if (tts_norm_class_match(k.kind, x) != k.neg) {
+                    return true;
+                }
+            }
+            return false;
+        };
+        bool m = test(c);
+        if (!m && icase) {
+            const uint32_t o = tts_norm_ascii_swapcase(c);
+            if (o != c) {
+                m = test(o);
+            }
+        }
+        return m != st.neg;
+    }
+
+    bool match_alts(const std::vector<tts_norm_seq> & alts, size_t i, size_t & end) {
+        for (const auto & seq : alts) {
+            auto saved = caps;
+            if (match_seq(seq, i, end)) {
+                return true;
+            }
+            caps = std::move(saved);
+        }
+        return false;
+    }
+
+    bool match_seq(const tts_norm_seq & seq, size_t i, size_t & end) {
+        size_t pos = i;
+        for (const auto & it : seq) {
+            int count = 0;
+            while (count < it.max) {
+                const bool is_group = it.atom.op == tts_norm_op::GROUP;
+                std::vector<std::pair<int64_t, int64_t>> saved;
+                if (is_group) {
+                    saved = caps;
+                }
+                size_t e = pos;
+                if (!match_once(it.atom, pos, e)) {
+                    if (is_group) {
+                        caps = std::move(saved);
+                    }
+                    break;
+                }
+                if (e == pos) {
+                    // an empty iteration would repeat forever, it satisfies any remaining minimum
+                    count = std::max(count + 1, it.min);
+                    break;
+                }
+                pos = e;
+                count++;
+            }
+            if (count < it.min) {
+                return false;
+            }
+        }
+        end = pos;
+        return true;
+    }
+
+    bool match_once(const tts_norm_pattern_node & node, size_t i, size_t & end) {
+        const size_t n = t.size();
+        switch (node.op) {
+            case tts_norm_op::LITERAL:
+                if (i < n && (t[i] == node.cp || (icase && tts_norm_ascii_lower(t[i]) == tts_norm_ascii_lower(node.cp)))) {
+                    end = i + 1;
+                    return true;
+                }
+                return false;
+            case tts_norm_op::SET:
+                if (i < n && set_match(node.set, t[i])) {
+                    end = i + 1;
+                    return true;
+                }
+                return false;
+            case tts_norm_op::ANY:
+                if (i < n) {
+                    end = i + 1;
+                    return true;
+                }
+                return false;
+            case tts_norm_op::START:
+                end = i;
+                return i == 0;
+            case tts_norm_op::END:
+                end = i;
+                return i == n;
+            case tts_norm_op::BOUNDARY:
+                end = i;
+                return (i > 0 && is_w(i - 1)) != is_w(i);
+            case tts_norm_op::GROUP: {
+                size_t e = i;
+                if (!match_alts(node.alts, i, e)) {
+                    return false;
+                }
+                if (node.capture >= 0) {
+                    caps[node.capture] = {static_cast<int64_t>(i), static_cast<int64_t>(e)};
+                }
+                end = e;
+                return true;
+            }
+            case tts_norm_op::LOOKAHEAD: {
+                // captures made inside a lookahead are discarded
+                auto saved = caps;
+                size_t e = i;
+                const bool ok = match_alts(node.alts, i, e);
+                caps = std::move(saved);
+                end = i;
+                return ok != node.neg;
+            }
+            case tts_norm_op::LOOKBEHIND: {
+                size_t e = 0;
+                const bool ok = i > 0 && match_once(node.alts[0][0].atom, i - 1, e) && e == i;
+                end = i;
+                return ok != node.neg;
+            }
+        }
+        return false;
+    }
+};
+
+//
+// common_tts_norm_pattern
+//
+
+common_tts_norm_pattern::common_tts_norm_pattern(const std::string & src) {
+    if (src.size() > TTS_NORM_MAX_PATTERN_LEN) {
+        throw std::runtime_error("invalid TTS norm pattern: pattern is longer than " + std::to_string(TTS_NORM_MAX_PATTERN_LEN) + " bytes");
+    }
+    std::vector<uint32_t> cpts;
+    std::vector<size_t> offs;
+    tts_norm_decode_utf8(src, cpts, offs);
+
+    tts_norm_parser parser(std::move(cpts), n_caps);
+    static const std::vector<uint32_t> icase_prefix = {'(', '?', 'i', ')'};
+    if (parser.s.size() >= 4 && std::equal(icase_prefix.begin(), icase_prefix.end(), parser.s.begin())) {
+        icase = true;
+        parser.p = 4;
+    }
+
+    root = std::make_unique<tts_norm_pattern_node>();
+    root->op   = tts_norm_op::GROUP;
+    root->alts = parser.parse_alts();
+    if (!parser.eof()) {
+        parser.fail("unbalanced ')'");
+    }
+}
+
+common_tts_norm_pattern::~common_tts_norm_pattern() = default;
+common_tts_norm_pattern::common_tts_norm_pattern(common_tts_norm_pattern &&) noexcept = default;
+common_tts_norm_pattern & common_tts_norm_pattern::operator=(common_tts_norm_pattern &&) noexcept = default;
+
+bool common_tts_norm_pattern::match_at(const std::vector<uint32_t> & cpts, size_t pos, size_t & end, std::vector<std::pair<int64_t, int64_t>> & groups) const {
+    tts_norm_matcher m{cpts, icase, std::vector<std::pair<int64_t, int64_t>>(n_caps, {-1, -1})};
+    size_t e = pos;
+    if (!m.match_alts(root->alts, pos, e)) {
+        return false;
+    }
+    end = e;
+    groups = std::move(m.caps);
+    return true;
+}
+
+size_t common_tts_norm_pattern::n_groups() const {
+    return n_caps;
+}
+
+//
+// scan / sub
+//
+
+std::vector<common_tts_norm_segment> common_tts_norm_scan(const std::string & text, const std::vector<const common_tts_norm_pattern *> & rules) {
+    std::vector<uint32_t> cpts;
+    std::vector<size_t> offs;
+    tts_norm_decode_utf8(text, cpts, offs);
+
+    std::vector<common_tts_norm_segment> out;
+    const size_t n = cpts.size();
+    size_t i = 0;
+    size_t pending = 0; // start of the current unmatched run
+    std::vector<std::pair<int64_t, int64_t>> groups;
+
+    while (i < n) {
+        bool matched = false;
+        for (size_t r = 0; r < rules.size(); r++) {
+            size_t end = i;
+            if (rules[r]->match_at(cpts, i, end, groups) && end > i) {
+                if (pending < i) {
+                    out.push_back({-1, offs[pending], offs[i], {}});
+                }
+                common_tts_norm_segment seg;
+                seg.rule  = static_cast<int>(r);
+                seg.start = offs[i];
+                seg.end   = offs[end];
+                for (const auto & g : groups) {
+                    if (g.first < 0) {
+                        seg.groups.emplace_back(std::string::npos, std::string::npos);
+                    } else {
+                        seg.groups.emplace_back(offs[g.first], offs[g.second]);
+                    }
+                }
+                out.push_back(std::move(seg));
+                i = end;
+                pending = i;
+                matched = true;
+                break;
+            }
+        }
+        if (!matched) {
+            i++;
+        }
+    }
+    if (pending < n) {
+        out.push_back({-1, offs[pending], offs[n], {}});
+    }
+    return out;
+}
+
+struct tts_norm_repl_piece {
+    std::string literal;
+    int group = -1; // 0 = whole match, 1.. = groups
+};
+
+static std::vector<tts_norm_repl_piece> tts_norm_parse_repl(const std::string & repl, const common_tts_norm_pattern & pattern) {
+    std::vector<tts_norm_repl_piece> pieces;
+    std::string lit;
+    auto flush = [&]() {
+        if (!lit.empty()) {
+            pieces.push_back({lit, -1});
+            lit.clear();
+        }
+    };
+    for (size_t i = 0; i < repl.size(); i++) {
+        if (repl[i] != '%') {
+            lit += repl[i];
+            continue;
+        }
+        if (i + 1 >= repl.size()) {
+            throw std::runtime_error("invalid TTS norm replacement: dangling '%'");
+        }
+        const char c = repl[++i];
+        if (c == '%') {
+            lit += '%';
+        } else if (c >= '0' && c <= '9') {
+            const int g = c - '0';
+            if (static_cast<size_t>(g) > pattern.n_groups()) {
+                throw std::runtime_error("invalid TTS norm replacement: group %" + std::string(1, c) + " does not exist");
+            }
+            flush();
+            pieces.push_back({"", g});
+        } else {
+            throw std::runtime_error("invalid TTS norm replacement: unknown escape '%" + std::string(1, c) + "'");
+        }
+    }
+    flush();
+    return pieces;
+}
+
+std::string common_tts_norm_sub(const std::string & text, const common_tts_norm_pattern & pattern, const std::string & repl) {
+    const auto pieces = tts_norm_parse_repl(repl, pattern);
+    std::string out;
+    for (const auto & seg : common_tts_norm_scan(text, {&pattern})) {
+        if (seg.rule < 0) {
+            out.append(text, seg.start, seg.end - seg.start);
+            continue;
+        }
+        for (const auto & piece : pieces) {
+            if (piece.group < 0) {
+                out += piece.literal;
+            } else if (piece.group == 0) {
+                out.append(text, seg.start, seg.end - seg.start);
+            } else {
+                const auto & g = seg.groups[piece.group - 1];
+                if (g.first != std::string::npos) {
+                    out.append(text, g.first, g.second - g.first);
+                }
+            }
+        }
+    }
+    return out;
+}
+
+//
+// template builtins
+//
+
+// rules can be: a pattern string (rule name "match"), a dict {name: pattern}, or a list of [name, pattern] pairs
+static std::vector<std::pair<std::string, std::string>> tts_norm_rules_from_value(const value & v) {
+    std::vector<std::pair<std::string, std::string>> rules;
+    if (is_val<value_string>(v)) {
+        rules.emplace_back("match", v->as_string().str());
+    } else if (is_val<value_object>(v)) {
+        for (const auto & kv : v->as_ordered_object()) {
+            if (!is_val<value_string>(kv.second)) {
+                throw std::runtime_error("tts_norm_scan: rule pattern must be a string");
+            }
+            rules.emplace_back(kv.first->as_string().str(), kv.second->as_string().str());
+        }
+    } else if (is_val<value_array>(v)) {
+        for (const auto & item : v->as_array()) {
+            if (!is_val<value_array>(item) || item->as_array().size() != 2
+                    || !is_val<value_string>(item->as_array()[0]) || !is_val<value_string>(item->as_array()[1])) {
+                throw std::runtime_error("tts_norm_scan: each rule must be a [name, pattern] pair of strings");
+            }
+            rules.emplace_back(item->as_array()[0]->as_string().str(), item->as_array()[1]->as_string().str());
+        }
+    } else {
+        throw std::runtime_error("tts_norm_scan: rules must be a string, a dict or a list of [name, pattern] pairs, got " + v->type());
+    }
+    return rules;
+}
+
+static value tts_norm_builtin_scan(const func_args & args) {
+    args.ensure_count(2, 2);
+    args.ensure_val<value_string>(args.get_pos(0));
+    const string text_val = args.get_pos(0)->as_string();
+    const std::string text = text_val.str();
+    const bool is_input = text_val.all_parts_are_input();
+
+    const auto rule_srcs = tts_norm_rules_from_value(args.get_pos(1));
+    std::vector<common_tts_norm_pattern> patterns;
+    patterns.reserve(rule_srcs.size());
+    std::vector<const common_tts_norm_pattern *> ptrs;
+    for (const auto & r : rule_srcs) {
+        patterns.emplace_back(r.second);
+    }
+    for (const auto & p : patterns) {
+        ptrs.push_back(&p);
+    }
+
+    auto mk_str = [&](size_t start, size_t end) -> value {
+        return mk_val<value_string>(string(text.substr(start, end - start), is_input));
+    };
+
+    auto out = mk_val<value_array>();
+    for (const auto & seg : common_tts_norm_scan(text, ptrs)) {
+        auto groups = mk_val<value_array>();
+        for (const auto & g : seg.groups) {
+            // groups that did not participate are empty strings, so they render as nothing and are falsy
+            groups->push_back(g.first == std::string::npos ? mk_str(0, 0) : mk_str(g.first, g.second));
+        }
+        auto obj = mk_val<value_object>();
+        obj->insert("rule", seg.rule < 0 ? value(mk_val<value_none>()) : value(mk_val<value_string>(rule_srcs[seg.rule].first)));
+        obj->insert("text", mk_str(seg.start, seg.end));
+        obj->insert("groups", groups);
+        out->push_back(obj);
+    }
+    return out;
+}
+
+static value tts_norm_builtin_sub(const func_args & args) {
+    args.ensure_count(3, 3);
+    args.ensure_vals<value_string, value_string, value_string>();
+    const string text_val = args.get_pos(0)->as_string();
+    const common_tts_norm_pattern pattern(args.get_pos(1)->as_string().str());
+    const std::string res = common_tts_norm_sub(text_val.str(), pattern, args.get_pos(2)->as_string().str());
+    return mk_val<value_string>(string(res, text_val.all_parts_are_input()));
+}
+
+void common_chat_tts_norm_add_builtins(context & ctx) {
+    ctx.set_val("tts_norm_scan", mk_val<value_func>("tts_norm_scan", tts_norm_builtin_scan));
+    ctx.set_val("tts_norm_sub",  mk_val<value_func>("tts_norm_sub",  tts_norm_builtin_sub));
+}
+
@@ -0,0 +1,62 @@
+#pragma once
+
+#include <cstdint>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+// TTS normalization patterns: a small, fully specified matching language for TTS text normalization templates
+//
+// it is NOT a regex engine: the semantics are defined by llama.cpp (see docs/tts-norm-patterns.md), so that this
+// implementation and the python reference (scripts/chat_tts_norm.py) give the same result for every input
+//
+// two functions are exposed to task templates (e.g. "tts") via common_chat_tts_norm_add_builtins():
+//   tts_norm_scan(text, rules)        -> list of segments {rule, text, groups}
+//   tts_norm_sub(text, pattern, repl) -> string
+
+namespace jinja {
+struct context;
+}
+
+struct tts_norm_pattern_node;
+
+class common_tts_norm_pattern {
+public:
+    // throws std::runtime_error on invalid pattern
+    explicit common_tts_norm_pattern(const std::string & src);
+    ~common_tts_norm_pattern();
+    common_tts_norm_pattern(common_tts_norm_pattern &&) noexcept;
+    common_tts_norm_pattern & operator=(common_tts_norm_pattern &&) noexcept;
+
+    // match anchored at codepoint index pos; on success, end is the codepoint index after the match
+    // groups[i] holds [start, end) codepoint indices of capture i+1, or {-1, -1} if the group did not participate
+    bool match_at(const std::vector<uint32_t> & cpts, size_t pos, size_t & end, std::vector<std::pair<int64_t, int64_t>> & groups) const;
+
+    size_t n_groups() const;
+
+private:
+    std::unique_ptr<tts_norm_pattern_node> root;
+    size_t n_caps = 0;
+    bool icase = false;
+};
+
+// segment produced by common_tts_norm_scan(); byte offsets into the input text
+struct common_tts_norm_segment {
+    int    rule  = -1; // index of the matched rule, -1 for unmatched text
+    size_t start = 0;
+    size_t end   = 0;
+    std::vector<std::pair<size_t, size_t>> groups; // byte offsets; {npos, npos} for groups that did not participate
+};
+
+// split text into matched and unmatched segments
+// at each position, rules are tried in order and the first non-empty match wins; otherwise one codepoint is skipped
+std::vector<common_tts_norm_segment> common_tts_norm_scan(const std::string & text, const std::vector<const common_tts_norm_pattern *> & rules);
+
+// replace every non-overlapping match of pattern with repl
+// repl references: %0 (whole match), %1..%9 (groups), %% (literal percent)
+std::string common_tts_norm_sub(const std::string & text, const common_tts_norm_pattern & pattern, const std::string & repl);
+
+// add tts_norm_scan() and tts_norm_sub() to the context of a task template
+// they are not jinja global builtins on purpose: chat templates are not affected
+void common_chat_tts_norm_add_builtins(jinja::context & ctx);
@@ -3,6 +3,7 @@
 #include "chat-auto-parser-helpers.h"
 #include "chat-auto-parser.h"
 #include "chat-peg-parser.h"
+#include "chat-tts-norm.h"
 #include "common.h"
 #include "ggml.h"
 #include "json-schema-to-grammar.h"
@@ -807,6 +808,8 @@ bool common_chat_templates_was_explicit(const struct common_chat_templates * tmp
 // renders a task template (e.g. "tts") with the given variables
 static std::string common_chat_template_render(const common_chat_template & tmpl, const json & vars) {
    jinja::context ctx(tmpl.source());
+    // text normalization helpers: tts_norm_scan() and tts_norm_sub(), see docs/tts-norm-patterns.md
+    common_chat_tts_norm_add_builtins(ctx);
    jinja::global_from_json(ctx, vars, false);
    ctx.set_val("bos_token", jinja::mk_val<jinja::value_string>(tmpl.bos_token()));
    ctx.set_val("eos_token", jinja::mk_val<jinja::value_string>(tmpl.eos_token()));
@@ -819,21 +822,29 @@ static bool is_lfm3_template(const std::string & src) {
    return src.find("<additional_tools>") != std::string::npos;
 }

-// LFM3-Audio: a known language code maps to its name, no language lets the model detect it
-static std::string lfm3_asr_system_prompt(const std::string & language) {
+static std::string string_lower(std::string s) {
+    std::transform(s.begin(), s.end(), s.begin(), [](unsigned char c) { return (char) std::tolower(c); });
+    return s;
+}
+
+// LFM3-Audio: language code -> name
+static std::string lfm3_language_name(const std::string & language) {
    static const std::map<std::string, std::string> names = {
        { "ar", "Arabic" }, { "da", "Danish" }, { "de", "German" }, { "el", "Greek" }, { "en", "English" },
        { "es", "Spanish" }, { "fi", "Finnish" }, { "fr", "French" }, { "he", "Hebrew" }, { "it", "Italian" },
        { "ja", "Japanese" }, { "ko", "Korean" }, { "nl", "Dutch" }, { "pl", "Polish" }, { "pt", "Portuguese" },
        { "sv", "Swedish" }, { "tr", "Turkish" }, { "zh", "Chinese" },
    };
+    const auto it = names.find(string_lower(language));
+    return it != names.end() ? it->second : language;
+}
+
+// LFM3-Audio: no language = auto detect
+static std::string lfm3_asr_system_prompt(const std::string & language) {
    if (language.empty()) {
        return "Perform ASR.";
    }
-    std::string code = language;
-    std::transform(code.begin(), code.end(), code.begin(), [](unsigned char c) { return (char) std::tolower(c); });
-    const auto it = names.find(code);
-    return "Perform ASR in " + (it != names.end() ? it->second : language) + ".";
+    return "Perform ASR in " + lfm3_language_name(language) + ".";
 }

 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates, const std::string & language) {
@@ -868,6 +879,27 @@ std::string common_chat_get_tts_prompt(const common_chat_templates * chat_templa
    });
 }

+std::string common_chat_get_speech_system_prompt(const common_chat_templates * chat_templates,
+                                                 const std::string & language, const std::string & voice) {
+    if (!chat_templates || !chat_templates->template_default || !is_lfm3_template(chat_templates->template_default->source())) {
+        return "";
+    }
+    // same voices as the "tts" template
+    const std::string v = string_lower(voice);
+    const bool has_voice = v == "female" || v == "male";
+    if (language.empty() && !has_voice) {
+        return "";
+    }
+    std::string out = "Speak";
+    if (!language.empty()) {
+        out += " in " + lfm3_language_name(language);
+    }
+    if (has_voice) {
+        out += " using a " + v + " voice";
+    }
+    return out + ".";
+}
+
 std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant) {
    if (!variant.empty()) {
        if (variant == "tool_use") {
@@ -500,3 +500,7 @@ common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates
 // language and voice are optional, the template defines their defaults
 std::string common_chat_get_tts_prompt(const common_chat_templates * chat_templates, const std::string & text,
                                       const std::string & language = "", const std::string & voice = "");
+
+// system prompt line for the speech language and voice (e.g. LFM3-Audio), empty if not used
+std::string common_chat_get_speech_system_prompt(const common_chat_templates * chat_templates,
+                                                 const std::string & language, const std::string & voice);
@@ -679,7 +679,6 @@ struct common_params {
    common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
    int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
    bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
-    bool chat_input_special = true; // if false, special tokens in user input are tokenized as plain text
    int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time

    std::vector<std::string> api_keys;
@@ -37,6 +37,7 @@ struct token {
        call_operator, // ()
        additive_binary_operator, // + - ~
        multiplicative_binary_operator, // * / %
+        power_operator, // **
        comparison_binary_operator, // < > <= >= == !=
        unary_operator, // ! - +
        comment, // {# ... #}
@@ -71,6 +72,7 @@ static std::string type_to_string(token::type t) {
        case token::call_operator: return "call_operator";
        case token::additive_binary_operator: return "additive_binary_operator";
        case token::multiplicative_binary_operator: return "multiplicative_binary_operator";
+        case token::power_operator: return "power_operator";
        case token::comparison_binary_operator: return "comparison_binary_operator";
        case token::unary_operator: return "unary_operator";
        case token::comment: return "comment";
@@ -137,6 +139,7 @@ struct lexer {
        {"+", token::additive_binary_operator},
        {"-", token::additive_binary_operator},
        {"~", token::additive_binary_operator},
+        {"**", token::power_operator},
        {"*", token::multiplicative_binary_operator},
        {"/", token::multiplicative_binary_operator},
        {"%", token::multiplicative_binary_operator},
@@ -178,7 +178,7 @@ private:
            // bool has_caller_args = false;
            if (is(token::open_paren)) {
                // Optional caller arguments, e.g. {% call(user) dump_users(...) %}
-                caller_args = parse_args();
+                caller_args = parse_args(false);
                // has_caller_args = true;
            }
            auto callee = parse_primary_expression();
@@ -283,7 +283,7 @@ private:

    statement_ptr parse_macro_statement(size_t start_pos) {
        auto name = parse_primary_expression();
-        auto args = parse_args();
+        auto args = parse_args(false);
        expect(token::close_statement, "Expected %}");
        statements body;
        // Keep going until we hit {% endmacro
@@ -397,11 +397,13 @@ private:

    statement_ptr parse_comparison_expression() {
        // NOTE: membership has same precedence as comparison
-        // e.g., ('a' in 'apple' == 'b' in 'banana') evaluates as ('a' in ('apple' == ('b' in 'banana')))
-        auto left = parse_additive_expression();
+        // e.g., ('a' in 'apple' == 'b' in 'banana') evaluates as ('a' in 'apple' and 'apple' == 'b' and 'b' in 'banana')
+        statements operands;
+        std::vector<token> ops;
+        operands.push_back(parse_additive_expression());
+        size_t pos = current;
        while (true) {
            token op;
-            size_t start_pos = current;
            if (is_identifier("not") && peek(1).t == token::identifier && peek(1).value == "in") {
                op = {token::identifier, "not in", tokens[current].pos};
                ++current; // consume 'not'
@@ -411,9 +413,17 @@ private:
            } else if (is(token::comparison_binary_operator)) {
                op = next();
            } else break;
-            left = mk_stmt<binary_expression>(start_pos, op, std::move(left), parse_additive_expression());
+            ops.push_back(std::move(op));
+            operands.push_back(parse_additive_expression());
        }
-        return left;
+        if (ops.empty()) {
+            return std::move(operands[0]);
+        }
+        if (ops.size() == 1) {
+            return mk_stmt<binary_expression>(pos, ops[0], std::move(operands[0]), std::move(operands[1]));
+        }
+        // chained comparison, e.g. 1 <= x <= 100
+        return mk_stmt<comparison_chain_expression>(pos, std::move(operands), std::move(ops));
    }

    statement_ptr parse_additive_expression() {
@@ -500,17 +510,22 @@ private:
        return parse_call_expression(std::move(member)); // foo.x()()
    }

-    statements parse_args() {
+    statements parse_args(bool allow_spread = true) {
        // comma-separated arguments list
        expect(token::open_paren, "Expected (");
        statements args;
        while (!is(token::close_paren)) {
            statement_ptr arg;
-            // unpacking: *expr
-            if (peek().t == token::multiplicative_binary_operator && peek().value == "*") {
+            // unpacking: *expr or **expr
+            const auto & t = peek();
+            if (t.t == token::power_operator || (t.t == token::multiplicative_binary_operator && t.value == "*")) {
+                if (!allow_spread) {
+                    throw std::runtime_error("Unexpected token: " + t.value + " of type " + std::to_string(t.t));
+                }
                size_t start_pos = current;
-                ++current; // consume *
-                arg = mk_stmt<spread_expression>(start_pos, parse_expression());
+                ++current; // consume * or **
+                bool is_kwargs = t.t == token::power_operator;
+                arg = mk_stmt<spread_expression>(start_pos, parse_expression(), is_kwargs);
            } else {
                arg = parse_expression();
                if (is(token::equals)) {
@@ -117,20 +117,8 @@ value object_literal::execute_impl(context & ctx) const {
    return obj;
 }

-value binary_expression::execute_impl(context & ctx) const {
-    value left_val = left->execute(ctx);
-
-    // Logical operators
-    if (op.value == "and") {
-        JJ_DEBUG("Executing logical test: %s AND %s", left->type().c_str(), right->type().c_str());
-        return left_val->as_bool() ? right->execute(ctx) : std::move(left_val);
-    } else if (op.value == "or") {
-        JJ_DEBUG("Executing logical test: %s OR %s", left->type().c_str(), right->type().c_str());
-        return left_val->as_bool() ? std::move(left_val) : right->execute(ctx);
-    }
-
+static value exec_binary_op(context & ctx, const token & op, value left_val, value right_val) {
    // Equality operators
-    value right_val = right->execute(ctx);
    JJ_DEBUG("Executing binary expression %s '%s' %s", left_val->type().c_str(), op.value.c_str(), right_val->type().c_str());
    if (op.value == "==") {
        return mk_val<value_bool>(*left_val == *right_val);
@@ -307,6 +295,67 @@ value binary_expression::execute_impl(context & ctx) const {
    throw std::runtime_error("Unknown operator \"" + op.value + "\" between " + left_val->type() + " and " + right_val->type());
 }

+value binary_expression::execute_impl(context & ctx) const {
+    value left_val = left->execute(ctx);
+
+    // Logical operators
+    if (op.value == "and") {
+        JJ_DEBUG("Executing logical test: %s AND %s", left->type().c_str(), right->type().c_str());
+        return left_val->as_bool() ? right->execute(ctx) : std::move(left_val);
+    } else if (op.value == "or") {
+        JJ_DEBUG("Executing logical test: %s OR %s", left->type().c_str(), right->type().c_str());
+        return left_val->as_bool() ? std::move(left_val) : right->execute(ctx);
+    }
+
+    return exec_binary_op(ctx, op, std::move(left_val), right->execute(ctx));
+}
+
+value comparison_chain_expression::execute_impl(context & ctx) const {
+    value left_val = operands[0]->execute(ctx);
+    value res;
+    for (size_t i = 0; i < ops.size(); ++i) {
+        value right_val = operands[i + 1]->execute(ctx);
+        res = exec_binary_op(ctx, ops[i], left_val, right_val);
+        if (!res->as_bool()) {
+            break; // same as "and", the next operands are not executed
+        }
+        left_val = std::move(right_val);
+    }
+    return res;
+}
+
+static void exec_call_args(context & ctx, const statements & arg_stmts, func_args & args) {
+    for (const auto & arg_stmt : arg_stmts) {
+        if (!is_stmt<spread_expression>(arg_stmt)) {
+            auto arg_val = arg_stmt->execute(ctx);
+            JJ_DEBUG("  Argument type: %s", arg_val->type().c_str());
+            args.push_back(arg_val);
+            continue;
+        }
+        auto spread = cast_stmt<spread_expression>(arg_stmt);
+        value val = spread->argument->execute(ctx);
+        JJ_DEBUG("  Unpacked argument type: %s", val->type().c_str());
+        if (spread->is_kwargs) {
+            if (!is_val<value_object>(val)) {
+                throw std::runtime_error("Argument after ** must be an object, got " + val->type());
+            }
+            for (const auto & pair : val->as_ordered_object()) {
+                if (!is_val<value_string>(pair.first)) {
+                    throw std::runtime_error("Keyword argument key must be a string, got " + pair.first->type());
+                }
+                args.push_back(mk_val<value_kwarg>(pair.first->as_string().str(), pair.second));
+            }
+        } else {
+            if (!is_val<value_array>(val)) {
+                throw std::runtime_error("Argument after * must be an array, got " + val->type());
+            }
+            for (const auto & item : val->as_array()) {
+                args.push_back(item);
+            }
+        }
+    }
+}
+
 static value apply_filter(context & ctx, const statement_ptr & filter, value input) {
    JJ_DEBUG("Applying filter to %s", input->type().c_str());

@@ -324,9 +373,7 @@ static value apply_filter(context & ctx, const statement_ptr & filter, value inp

        JJ_DEBUG("Applying filter '%s' with arguments to %s", filter_id.c_str(), input->type().c_str());
        func_args args(ctx);
-        for (const auto & arg_expr : call->args) {
-            args.push_back(arg_expr->execute(ctx));
-        }
+        exec_call_args(ctx, call->args, args);

        return apply_filter_by_name(ctx, filter_id, input, args);

@@ -369,9 +416,7 @@ value test_expression::execute_impl(context & ctx) const {
        test_id = cast_stmt<identifier>(call->callee)->val;

        JJ_DEBUG("Applying test '%s' with arguments to %s", test_id.c_str(), input->type().c_str());
-        for (const auto & arg_expr : call->args) {
-            args.push_back(arg_expr->execute(ctx));
-        }
+        exec_call_args(ctx, call->args, args);

    } else {
        throw std::runtime_error("Invalid test expression");
@@ -762,12 +807,7 @@ value call_statement::execute_impl(context & ctx) const {
    call_ctx.set_val("caller", mk_val<value_func>("caller", func));

    func_args args(call_ctx);
-
-    for (const auto & arg_expr : call_expr->args) {
-        auto arg_val = arg_expr->execute(ctx);
-        JJ_DEBUG("  Argument type: %s", arg_val->type().c_str());
-        args.push_back(arg_val);
-    }
+    exec_call_args(ctx, call_expr->args, args);

    JJ_DEBUG("Calling macro '%s' with %zu arguments", callee_func->name.c_str(), args.count());
    return callee_func->invoke(args);
@@ -907,11 +947,7 @@ value member_expression::execute_impl(context & ctx) const {
 value call_expression::execute_impl(context & ctx) const {
    // gather arguments
    func_args args(ctx);
-    for (auto & arg_stmt : this->args) {
-        auto arg_val = arg_stmt->execute(ctx);
-        JJ_DEBUG("  Argument type: %s", arg_val->type().c_str());
-        args.push_back(arg_val);
-    }
+    exec_call_args(ctx, this->args, args);
    // execute callee
    value callee_val = callee->execute(ctx);
    if (!is_val<value_func>(callee_val)) {
@@ -471,6 +471,28 @@ struct binary_expression : public expression {
    }
 };

+/**
+ * Two or more comparisons in a row, e.g. `a < b <= c`
+ * It is the same as `a < b and b <= c`, but each operand is executed only one time.
+ */
+struct comparison_chain_expression : public expression {
+    statements operands;
+    std::vector<token> ops;
+
+    comparison_chain_expression(statements && operands, std::vector<token> && ops)
+        : operands(std::move(operands)), ops(std::move(ops)) {
+        assert(this->operands.size() == this->ops.size() + 1);
+        for (const auto & operand : this->operands) chk_type<expression>(operand);
+    }
+    std::string type() const override { return "ComparisonChainExpression"; }
+    value execute_impl(context & ctx) const override;
+    void visit(context & ctx) const override {
+        ctx.visitor(false, this, {
+            {"operands", stmts_to_ptr(operands)}
+        });
+    }
+};
+
 /**
 * An operation with two sides, separated by the | operator.
 * Operator precedence: https://github.com/pallets/jinja/issues/379#issuecomment-168076202
@@ -633,7 +655,9 @@ struct keyword_argument_expression : public expression {

 struct spread_expression : public expression {
    statement_ptr argument;
-    explicit spread_expression(statement_ptr && argument) : argument(std::move(argument)) {
+    bool is_kwargs;
+    spread_expression(statement_ptr && argument, bool is_kwargs)
+        : argument(std::move(argument)), is_kwargs(is_kwargs) {
        chk_type<expression>(this->argument);
    }
    std::string type() const override { return "SpreadExpression"; }
@@ -860,6 +860,7 @@ public:
        std::unordered_set<std::string> required;
        std::vector<std::pair<std::string, const common_chat_schema *>> properties;
        std::map<std::string, size_t> enum_values;
+        std::vector<const common_chat_schema *> leaves;
        std::function<void(const common_chat_schema &, bool, int)> add_component = [&](const common_chat_schema & comp, bool is_required, int depth) {
            if (depth >= COMMON_RECURSIVE_LIMIT) {
                throw std::invalid_argument("JSON schema is nested more than " + std::to_string(COMMON_RECURSIVE_LIMIT) + " levels deep");
@@ -879,6 +880,10 @@ public:
                for (const auto & v : as<common_chat_schema_enum>(comp).values) {
                    enum_values[_generate_constant_rule(v)] += 1;
                }
+            } else if (is_required && comp.kind() != common_chat_schema::KIND_ANY) {
+                // Direct non-structural children are used for merging. anyOf alternatives are not merged.
+                // Empty schemas are skipped.
+                leaves.push_back(&comp);
            }
        };
        for (const auto & child : schema.children) {
@@ -901,6 +906,38 @@ public:
                return _add_rule(rule_name, "(" + string_join(enum_intersection, " | ") + ")");
            }
        }
+
+        if (properties.empty() && !leaves.empty()) {
+            if (leaves.size() == 1) {
+                return visit(*leaves[0], name);
+            }
+            bool all_numeric = true;
+            bool any_integer = false;
+            int64_t lo = std::numeric_limits<int64_t>::min();
+            int64_t hi = std::numeric_limits<int64_t>::max();
+            for (const auto * leaf : leaves) {
+                if (leaf->kind() == common_chat_schema::KIND_INTEGER) {
+                    const auto & i = as<common_chat_schema_integer>(*leaf);
+                    any_integer = true;
+                    lo = std::max(lo, i.minimum);
+                    hi = std::min(hi, i.maximum);
+                } else if (leaf->kind() != common_chat_schema::KIND_NUMBER) {
+                    all_numeric = false;
+                }
+            }
+            if (all_numeric) {
+                if (lo > hi) {
+                    throw std::invalid_argument("allOf has an empty numeric range");
+                }
+                if (any_integer) {
+                    common_chat_schema_integer merged;
+                    merged.minimum = lo;
+                    merged.maximum = hi;
+                    return visit(merged, name);
+                }
+                return visit(*leaves[0], name);
+            }
+        }
        return _add_rule(rule_name, _build_object_rule(properties, required, name, nullptr));
    }

@@ -300,7 +300,7 @@ common_chat_params common_chat_params_init_deepseek_v3_2(const common_chat_templ
    });

    if (include_grammar) {
-        data.grammar_lazy = has_tools && !require_tools;
+        data.grammar_lazy = !has_response_format && has_tools && !require_tools;
        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
            data.parser.build_grammar(builder, data.grammar_lazy);
        });
@@ -109,7 +109,7 @@ common_chat_params common_chat_params_init_ministral_3(const common_chat_templat
    });

    if (include_grammar) {
-        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar_lazy = !has_response_format && has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;

        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
            data.parser.build_grammar(builder, data.grammar_lazy);
@@ -46,9 +46,10 @@ common_chat_params common_chat_params_init_muse_spark(const common_chat_template

    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;

-    auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
-    // Constrained grammar whenever tools are offered.
-    auto include_grammar = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+    auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
+    auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
+    // Constrained grammar whenever tools are offered or a response format is requested.
+    auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);

    data.parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto start = p.rule("start", p.literal("<|start|>assistant"));
@@ -75,6 +76,15 @@ common_chat_params common_chat_params_init_muse_spark(const common_chat_template
        auto final_msg  = p.rule("final", recipient + p.literal("<|message|>") +
                                              p.content(p.until_one_of({ "<|eot|>", "<|eom|>" })));

+        if (has_response_format) {
+            auto response_json   = p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema));
+            auto response_format = p.rule("response-format",
+                recipient + p.literal("<|message|>") +
+                ((p.literal("```json") + p.space() + response_json + p.space() + p.literal("```")) | response_json));
+
+            return p.zero_or_more(start + analysis) + start + response_format;
+        }
+
        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
            auto string_value = p.ac(
                p.tool_arg_string_value(p.until("</atem:parameter>")) + p.tool_arg_close(p.literal("</atem:parameter>")),
@@ -132,7 +142,7 @@ common_chat_params common_chat_params_init_muse_spark(const common_chat_template
    });

    if (include_grammar) {
-        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar_lazy = !(has_response_format || (has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED));
        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
            data.parser.build_grammar(builder, data.grammar_lazy);
        });
@@ -191,7 +191,7 @@ common_chat_params common_chat_params_init_qwen3_coder(const common_chat_templat
    });

    if (include_grammar) {
-        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar_lazy = !has_response_format && has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;

        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
            data.parser.build_grammar(builder, data.grammar_lazy);
@@ -371,6 +371,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "InternVisionModel": "internvl",
    "JanusForConditionalGeneration": "januspro",
    "KimiK25ForConditionalGeneration": "kimivl",
+    "KimiK31ForConditionalGeneration": "kimivl",
    "KimiVLForConditionalGeneration": "kimivl",
    "Lfm2AudioForConditionalGeneration": "lfm2",
    "D1OmniModel": "lfm2",
@@ -1,5 +1,6 @@
 from __future__ import annotations

+import re
 from typing import Callable, Iterable, TYPE_CHECKING

 import torch
@@ -198,6 +199,18 @@ class SarvamMoEModel(BailingMoeV2Model):
    #  - full rotary (no partial_rotary_factor)
    #  - expert bias is zero-mean normalized at load time

+    def set_vocab(self):
+        # BPE with byte fallback: mark <0xXX> tokens as BYTE so they decode to raw bytes
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        toktypes = [gguf.TokenType.BYTE if re.fullmatch(r"<0x[0-9A-F]{2}>", tok) else typ for tok, typ in zip(tokens, toktypes)]
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
    def set_gguf_parameters(self):
        super().set_gguf_parameters()
        hparams = self.hparams
@@ -7,6 +7,7 @@ import ast
 import logging
 import contextlib
 import json
+import math
 import os
 import re
 import sys
@@ -31,10 +32,10 @@ import gguf
 from gguf.vocab import MistralTokenizerType, MistralVocab

 try:
-    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: ignore[import-not-found, ty:unresolved-import]
-    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # type: ignore[import-not-found, ty:unresolved-import]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found, ty:unresolved-import]
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.base import TokenizerVersion # ty: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.multimodal import DATASET_MEAN as _MISTRAL_COMMON_DATASET_MEAN, DATASET_STD as _MISTRAL_COMMON_DATASET_STD # ty: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # ty: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.sentencepiece import ( # ty: ignore[import-not-found]
        SentencePieceTokenizer,
    )

@@ -272,7 +273,7 @@ class ModelBase:
                    if weight_map is None or not isinstance(weight_map, dict):
                        raise ValueError(f"Can't load 'weight_map' from {index_name!r}")
                    tensor_names_from_index.update(weight_map.keys())
-                    part_dict: dict[str, None] = dict.fromkeys(weight_map.values(), None) # ty: ignore[invalid-assignment]
+                    part_dict: dict[str, None] = dict.fromkeys(weight_map.values(), None)
                    part_names = sorted(part_dict.keys())
            else:
                weight_map = {}
@@ -343,6 +344,36 @@ class ModelBase:
            logger.info(f"  + {scale_name} (per-expert scale, shape [{len(scales)}])")
            self.gguf_writer.add_tensor(scale_name, scale_vals)

+    def _prepare_kv_cache_scales(self):
+        for name in list(self.model_tensors.keys()):
+            if not name.endswith((".k_scale", ".v_scale")):
+                continue
+
+            mapped = self.tensor_map.get_type_and_name(key=name, try_suffixes=(".k_scale", ".v_scale"))
+            if mapped is None:
+                continue
+
+            tensor_type, new_name = mapped
+            if tensor_type not in (gguf.MODEL_TENSOR.ATTN_K, gguf.MODEL_TENSOR.ATTN_V, gguf.MODEL_TENSOR.ATTN_QKV):
+                continue
+
+            bid = int(new_name.split(".")[1])
+            scale_type = gguf.MODEL_TENSOR.ATTN_K if name.endswith(".k_scale") else gguf.MODEL_TENSOR.ATTN_V
+            suffix = ".k_scale" if name.endswith(".k_scale") else ".v_scale"
+            new_name = gguf.TENSOR_NAMES[scale_type].format(bid=bid) + suffix
+
+            scale = LazyTorchTensor.to_eager(self.model_tensors.pop(name)())
+            if scale.dtype not in (torch.bfloat16, torch.float32) or scale.numel() != 1:
+                raise ValueError(f"KV cache scale {name!r} must be a scalar BF16 or FP32 tensor")
+
+            scale = scale.float()
+            value = float(scale.item())
+            if not math.isfinite(value) or value == 0.0:
+                raise ValueError(f"KV cache scale {name!r} must be finite and nonzero")
+
+            logger.info(f"  + {new_name} (KV cache scale, shape [1])")
+            self.gguf_writer.add_tensor(new_name, scale.flatten().numpy())
+
    def dequant_model(self):
        # If all quantized tensors were already handled (e.g. pure NVFP4), skip
        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv", ".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")) for k in self.model_tensors):
@@ -1455,6 +1486,8 @@ class ModelBase:
            unmatched = layer_algo.keys() - matched
            assert not unmatched, f"quantized_layers entries not in the model tensors: {sorted(unmatched)}"

+        self._prepare_kv_cache_scales()
+
        # NVFP4 weights are repacked and written directly to gguf_writer.
        # This must run before dequant_model so NVFP4 tensors are removed
        # from model_tensors, leaving only non-NVFP4 (e.g. FP8) for dequant.
@@ -2022,15 +2055,16 @@ class TextModel(ModelBase):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
-        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
-        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
+        assert max(tokenizer.vocab.values()) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()

-        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = tokenizer.added_tokens_decoder

        for i in range(vocab_size):
            if i not in reverse_vocab:
@@ -2043,7 +2077,7 @@ class TextModel(ModelBase):
                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
                    if not added_tokens_decoder[i].normalized:
                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[invalid-assignment]
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

@@ -2453,6 +2487,7 @@ class TextModel(ModelBase):
    def _set_vocab_hybriddna(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        vocab = tokenizer.get_vocab()
        vocab_size = self.hparams.get("vocab_size", len(vocab))
        assert max(vocab.values()) < vocab_size
@@ -2466,11 +2501,11 @@ class TextModel(ModelBase):
        if tokenizer.k != 6 or tail not in ("pad_a", "pad_n"):
            raise NotImplementedError(f"HybridDNATokenizer with k={tokenizer.k}, tail={tail!r}")
        for kmer in [*tokenizer.kmers, *getattr(tokenizer, "npad_tokens", [])]:
-            reverse_vocab[tokenizer.dna_token_to_id[kmer]] = kmer + "\ue000"  # ty: ignore[unresolved-attribute]
+            reverse_vocab[tokenizer.dna_token_to_id[kmer]] = kmer + "\ue000"
        # older HybridDNATokenizer versions (Carbon-500M/3B/8B) do not register the base tokenizer's added tokens (<|im_start|>, <think>, ...) as their own
        # but they are still encoded as special tokens: take their types from the base tokenizer
        base_tokenizer = getattr(tokenizer, "_base_tokenizer", tokenizer)
-        added_tokens_decoder = {**base_tokenizer.added_tokens_decoder, **tokenizer.added_tokens_decoder}  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = {**base_tokenizer.added_tokens_decoder, **tokenizer.added_tokens_decoder}

        tokens: list[str] = []
        toktypes: list[int] = []
@@ -2517,24 +2552,25 @@ class TextModel(ModelBase):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        vocab_size = hparams["vocab_size"]
-        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.get_vocab().values()) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

        merges = []
        vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
+        mergeable_ranks = tokenizer.mergeable_ranks
        for token, rank in mergeable_ranks.items():
            vocab[QwenModel.token_bytes_to_string(token)] = rank
            if len(token) == 1:
                continue
            merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)
            assert len(merged) == 2
-            merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
+            merges.append(str(' '.join(map(QwenModel.token_bytes_to_string, merged))))

        # for this kind of tokenizer, added_vocab is not a subset of vocab, so they need to be combined
-        added_vocab = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.special_tokens
        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **added_vocab}.items()}

        for i in range(vocab_size):
@@ -2557,10 +2593,10 @@ class TextModel(ModelBase):
        special_vocab.merges = merges
        # only add special tokens when they were not already loaded from config.json
        if len(special_vocab.special_token_ids) == 0:
-            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+            special_vocab._set_special_token("bos", tokenizer.special_tokens["<|endoftext|>"])
+            special_vocab._set_special_token("eos", tokenizer.special_tokens["<|endoftext|>"])
        # this one is usually not in config.json anyway
-        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("unk", tokenizer.special_tokens["<|endoftext|>"])
        special_vocab.add_to_gguf(self.gguf_writer)

    def _set_vocab_sentencepiece(self, add_to_gguf=True):
@@ -2834,21 +2870,23 @@ class TextModel(ModelBase):
    def _set_vocab_glmedge(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert tokenizer is not None
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
        self.gguf_writer.add_tokenizer_pre(tokpre)
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
        special_vocab.add_to_gguf(self.gguf_writer)

    def _set_vocab_glm(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert tokenizer is not None
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
@@ -2857,10 +2895,10 @@ class TextModel(ModelBase):
        self.gguf_writer.add_token_types(toktypes)
        # Special tokens
        # Note: Using <|endoftext|> (151329) for eot causes endless generation
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # ty: ignore[unresolved-attribute]  # 151331
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]  # 151336
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]  # 151329
-        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # ty: ignore[unresolved-attribute]  # 151338
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["[gMASK]"])  # 151331
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # 151336
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # 151329
+        special_vocab._set_special_token("eom", tokenizer.get_added_vocab()["<|observation|>"])  # 151338
        special_vocab.add_to_gguf(self.gguf_writer)

    def _set_vocab_interns1(self):
@@ -2869,16 +2907,17 @@ class TextModel(ModelBase):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        vocab = getattr(tokenizer, 'vocab', tokenizer.get_vocab())
        vocab_size = self.hparams.get("vocab_size", len(vocab))
        assert max(vocab.values()) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab.items()}
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()

-        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = tokenizer.added_tokens_decoder

        for i in range(vocab_size):
            if i not in reverse_vocab:
@@ -2891,7 +2930,7 @@ class TextModel(ModelBase):
                    # To avoid unexpected issues - we make sure to normalize non-normalized tokens
                    if not added_tokens_decoder[i].normalized:
                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[invalid-assignment]
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

@@ -60,7 +60,8 @@ class BeamModel(TextModel):
        # otherwise EOT is guessed from a hash-ordered scan
        from transformers import AutoTokenizer
        tok = AutoTokenizer.from_pretrained(self.dir_model)
-        ret_id = tok.convert_tokens_to_ids("<|return|>")  # ty: ignore[unresolved-attribute]
+        assert tok is not None
+        ret_id = tok.convert_tokens_to_ids("<|return|>")
        if isinstance(ret_id, int) and ret_id >= 0:
            self.gguf_writer.add_eot_token_id(ret_id)

@@ -132,6 +132,7 @@ class BertModel(TextModel):
            from base64 import b64decode
            from transformers import AutoTokenizer
            tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+            assert tokenizer is not None

            with open(tokenizer_path, "r", encoding="utf-8") as fp:
                tokenizer_json = json.load(fp)
@@ -140,13 +141,13 @@ class BertModel(TextModel):
                with open(tokenizer_config_path, "r", encoding="utf-8") as fp:
                    tokenizer_config_json = json.load(fp)

-            add_prefix = tokenizer.add_prefix_space  # ty: ignore[unresolved-attribute]
-            remove_whitespaces = tokenizer.clean_up_tokenization_spaces  # ty: ignore[unresolved-attribute]
+            add_prefix = tokenizer.add_prefix_space
+            remove_whitespaces = tokenizer.clean_up_tokenization_spaces
            precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])

-            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)  # ty: ignore[unresolved-attribute]
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)
        else:
-            sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
            sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM

@@ -183,20 +184,20 @@ class BertModel(TextModel):
                scores[token_id] = score
                toktypes[token_id] = toktype
        else:
-            added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+            added_vocab = tokenizer.get_added_vocab()
            unk_token = tokenizer_config_json.get("unk_token")
-            unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))  # ty: ignore[no-matching-overload]
+            unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))

-            for token_id in range(tokenizer.vocab_size):  # ty: ignore[unresolved-attribute]
-                piece = tokenizer._convert_id_to_token(token_id)  # ty: ignore[unresolved-attribute]
-                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:  # ty: ignore[unresolved-attribute]
+            for token_id in range(tokenizer.vocab_size):
+                piece = tokenizer._convert_id_to_token(token_id)
+                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:
                    text = piece.encode("utf-8")
                    score = tokenizer_json["model"]["vocab"][token_id][1]

                    toktype = SentencePieceTokenTypes.NORMAL
                    if token_id == unk_token_id:
                        toktype = SentencePieceTokenTypes.UNKNOWN
-                    elif token_id in tokenizer.all_special_ids:  # ty: ignore[unresolved-attribute]
+                    elif token_id in tokenizer.all_special_ids:
                        toktype = SentencePieceTokenTypes.CONTROL
                    elif token_id in added_vocab.values():
                        toktype = SentencePieceTokenTypes.USER_DEFINED
@@ -22,12 +22,13 @@ class ChatGLMModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
-        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))  # ty: ignore[unresolved-attribute]
-        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))
+        assert max(tokenizer.get_vocab().values()) < vocab_size
        role_special_tokens = ["<|system|>", "<|user|>", "<|assistant|>", "<|observation|>"]
        special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"] + role_special_tokens
        for token_id in range(vocab_size):
-            piece = tokenizer._convert_id_to_token(token_id)  # ty: ignore[unresolved-attribute]
+            piece = tokenizer._convert_id_to_token(token_id)
            if token_id == 0:
                piece = "<unk>"
            elif token_id == 1:
@@ -35,17 +36,18 @@ class ChatGLMModel(TextModel):
            elif token_id == 2:
                piece = "<eos>"

-            text = piece.encode("utf-8")  # ty: ignore[unresolved-attribute]
+            assert isinstance(piece, str)
+            text = piece.encode("utf-8")
            score = 0.0
            # Referencing the tokenizer Python implementation(https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py),
            # it is only valid if it is less than tokenizer.tokenizer.sp_model.vocab_size()
-            if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():  # ty: ignore[unresolved-attribute, invalid-argument-type]
-                score = tokenizer.tokenizer.sp_model.get_score(token_id)  # ty: ignore[unresolved-attribute]
+            if len(piece) != 0 and token_id < tokenizer.tokenizer.sp_model.vocab_size():
+                score = tokenizer.tokenizer.sp_model.get_score(token_id)

-            if token_id >= tokenizer.tokenizer.sp_model.vocab_size():  # ty: ignore[unresolved-attribute]
+            if token_id >= tokenizer.tokenizer.sp_model.vocab_size():
                if piece in special_tokens:
                    toktype = SentencePieceTokenTypes.CONTROL
-                elif len(piece) == 0:  # ty: ignore[invalid-argument-type]
+                elif len(piece) == 0:
                    text = f"[PAD{token_id}]".encode("utf-8")
                    toktype = SentencePieceTokenTypes.UNUSED
                else:
@@ -56,13 +58,13 @@ class ChatGLMModel(TextModel):
                continue

            toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.tokenizer.sp_model.is_unknown(token_id):  # ty: ignore[unresolved-attribute]
+            if tokenizer.tokenizer.sp_model.is_unknown(token_id):
                toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.tokenizer.sp_model.is_control(token_id):  # ty: ignore[unresolved-attribute]
+            elif tokenizer.tokenizer.sp_model.is_control(token_id):
                toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.tokenizer.sp_model.is_unused(token_id):  # ty: ignore[unresolved-attribute]
+            elif tokenizer.tokenizer.sp_model.is_unused(token_id):
                toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.tokenizer.sp_model.is_byte(token_id):  # ty: ignore[unresolved-attribute]
+            elif tokenizer.tokenizer.sp_model.is_byte(token_id):
                toktype = SentencePieceTokenTypes.BYTE

            tokens.append(text)
@@ -115,8 +117,9 @@ class ChatGLMModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        vocab_size = hparams.get("padded_vocab_size",hparams["vocab_size"])
-        assert max(tokenizer.get_vocab().values()) < vocab_size  # ty: ignore[unresolved-attribute]
+        assert max(tokenizer.get_vocab().values()) < vocab_size

        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
@@ -125,10 +128,10 @@ class ChatGLMModel(TextModel):
        self.gguf_writer.add_token_types(toktypes)
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        # only add special tokens when they were not already loaded from config.json
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
        # this one is usually not in config.json anyway
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
        special_vocab.add_to_gguf(self.gguf_writer)

    def set_gguf_parameters(self):
@@ -277,13 +277,14 @@ class DeepseekV2Model(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        tokpre = self.get_vocab_base_pre(tokenizer)

        if tokpre == "kimi-k2":
            # Build merges list using the approach similar to HunYuanMoE
            merges = []
            vocab = {}
-            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]
+            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[missing-union-attribute]
            for token, rank in mergeable_ranks.items():
                vocab[QwenModel.token_bytes_to_string(token)] = rank
                if len(token) == 1:
@@ -294,7 +295,7 @@ class DeepseekV2Model(TextModel):

            # Build token list
            vocab_size = self.hparams["vocab_size"]
-            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
+            special_tokens = tokenizer.special_tokens
            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
            tokens: list[str] = []
            toktypes: list[int] = []
@@ -62,6 +62,7 @@ def engram_compressed_token_map(tokenizer_dir: Path) -> tuple[list[int], int]:
    from transformers import AutoTokenizer

    tokenizer = AutoTokenizer.from_pretrained(tokenizer_dir)
+    assert tokenizer is not None
    sentinel = "\ue000"
    normalizer = normalizers.Sequence([
        normalizers.NFKC(),
@@ -80,6 +81,7 @@ def engram_compressed_token_map(tokenizer_dir: Path) -> tuple[list[int], int]:
        text = backend.decode([token_id], skip_special_tokens=False)
        if "\ufffd" in text:
            key = backend.id_to_token(token_id)
+            assert key is not None
        else:
            normalized = normalizer.normalize_str(text)
            key = normalized if normalized else text
@@ -75,13 +75,14 @@ class Dots3NoteModel(DeepseekV2Model):
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert tokenizer is not None
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
        self.gguf_writer.add_tokenizer_pre(tokpre)
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])
        special_vocab.add_to_gguf(self.gguf_writer)

    @classmethod
@@ -19,15 +19,16 @@ class DreamModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None

-        vocab_dict = tokenizer.get_vocab()  # ty: ignore[unresolved-attribute]
+        vocab_dict = tokenizer.get_vocab()
        vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
        assert max(vocab_dict.values()) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()

        for i in range(vocab_size):
            if i not in reverse_vocab:
@@ -31,6 +31,7 @@ class Glm4Model(TextModel):
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
@@ -38,10 +39,10 @@ class Glm4Model(TextModel):
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
        special_vocab.add_to_gguf(self.gguf_writer)

    def set_gguf_parameters(self):
@@ -391,16 +392,17 @@ class SolarOpenModel(Glm4MoeModel):
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert tokenizer is not None
        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
        tokens, toktypes, tokpre = self.get_vocab_base()
        self.gguf_writer.add_tokenizer_model("gpt2")
        self.gguf_writer.add_tokenizer_pre(tokpre)
        self.gguf_writer.add_token_list(tokens)
        self.gguf_writer.add_token_types(toktypes)
-        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endoftext|>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<unk>"])  # ty: ignore[unresolved-attribute]
-        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|startoftext|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<unk>"])
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|startoftext|>"])
        special_vocab.add_to_gguf(self.gguf_writer)


@@ -85,6 +85,16 @@ class GptOssModel(TextModel):
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item

+        # Normalize erroneous GPT-OSS MXFP4 tensor naming.
+        # https://github.com/p-e-w/heretic/pull/211#issuecomment-5094370390
+        if "mlp.experts." in name:
+            for proj in ("down_proj", "gate_up_proj"):
+                name = name.replace(
+                    f"{proj}_blocks_blocks", f"{proj}_blocks"
+                ).replace(
+                    f"{proj}_blocks_scales", f"{proj}_scales"
+                )
+
        if "sinks" in name:
            name += ".weight"

@@ -24,6 +24,7 @@ class HunYuanMoEModel(TextModel):
    def set_vocab(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None

        # 1. Get the pre-tokenizer identifier hash
        tokpre = self.get_vocab_base_pre(tokenizer)
@@ -31,7 +32,7 @@ class HunYuanMoEModel(TextModel):
        # 2. Reverse-engineer the merges list from mergeable_ranks
        merges = []
        vocab = {}
-        mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
+        mergeable_ranks = tokenizer.mergeable_ranks
        for token, rank in mergeable_ranks.items():
            vocab[QwenModel.token_bytes_to_string(token)] = rank
            if len(token) == 1:
@@ -42,8 +43,8 @@ class HunYuanMoEModel(TextModel):

        # 3. Generate the tokens and toktypes lists
        vocab_size = self.hparams["vocab_size"]
-        assert tokenizer.vocab_size == vocab_size  # ty: ignore[unresolved-attribute]
-        special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
+        assert tokenizer.vocab_size == vocab_size
+        special_tokens = tokenizer.special_tokens
        reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
        tokens: list[str] = []
        toktypes: list[int] = []
@@ -184,6 +185,7 @@ class HunYuanModel(TextModel):
        else:
            from transformers import AutoTokenizer
            tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+            assert tokenizer is not None

            # 1. Get the pre-tokenizer identifier hash
            tokpre = self.get_vocab_base_pre(tokenizer)
@@ -191,7 +193,7 @@ class HunYuanModel(TextModel):
            # 2. Reverse-engineer the merges list from mergeable_ranks
            merges = []
            vocab = {}
-            mergeable_ranks = tokenizer.mergeable_ranks  # ty: ignore[unresolved-attribute]
+            mergeable_ranks = tokenizer.mergeable_ranks
            for token, rank in mergeable_ranks.items():
                vocab[QwenModel.token_bytes_to_string(token)] = rank
                if len(token) == 1:
@@ -202,8 +204,8 @@ class HunYuanModel(TextModel):

            # 3. Generate the tokens and toktypes lists
            vocab_size = self.hparams["vocab_size"]
-            assert tokenizer.vocab_size == vocab_size  # ty: ignore[unresolved-attribute]
-            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
+            assert tokenizer.vocab_size == vocab_size
+            special_tokens = tokenizer.special_tokens
            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
            tokens: list[str] = []
            toktypes: list[int] = []
@@ -36,7 +36,7 @@ class InternLM2Model(TextModel):
            logger.error(f'Error: Missing {tokenizer_path}')
            sys.exit(1)

-        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix

@@ -29,13 +29,14 @@ class KimiLinearModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None
        tokpre = self.get_vocab_base_pre(tokenizer)

        if tokpre == "kimi-k2":
            # Build merges list using the approach similar to HunYuanMoE
            merges = []
            vocab = {}
-            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]
+            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[missing-union-attribute]
            for token, rank in mergeable_ranks.items():
                vocab[QwenModel.token_bytes_to_string(token)] = rank
                if len(token) == 1:
@@ -45,7 +46,7 @@ class KimiLinearModel(TextModel):
                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))
            # Build token list
            vocab_size = self.hparams["vocab_size"]
-            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]
+            special_tokens = tokenizer.special_tokens
            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}
            tokens: list[str] = []
            toktypes: list[int] = []
@@ -71,7 +72,7 @@ class KimiLinearModel(TextModel):
            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
            special_vocab.add_to_gguf(self.gguf_writer)
            # override eos id in config.json with tiktoken eos id
-            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)  # ty: ignore[unresolved-attribute]
+            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)
        else:
            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")

@@ -57,6 +57,8 @@ class KimiVLModel(MmprojModel):
 class KimiK25Model(MmprojModel):
    """Kimi-K2.5 with MoonViT3d vision encoder"""

+    projector_type = gguf.VisionProjectorType.KIMIK25
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

@@ -78,7 +80,7 @@ class KimiK25Model(MmprojModel):
        super().set_gguf_parameters()
        assert self.hparams_vision is not None

-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.KIMIK25)
+        self.gguf_writer.add_clip_projector_type(self.projector_type)

        # Position embedding parameters (for interpolation)
        self.gguf_writer.add_uint32("vision.pos_emb_height", self.hparams_vision.get("init_pos_emb_height", 64))
@@ -87,7 +89,11 @@ class KimiK25Model(MmprojModel):

        # Projector parameters
        self.gguf_writer.add_vision_use_gelu(self.hparams_vision.get("projector_hidden_act", "gelu") == "gelu")
-        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("projector_ln_eps", 1e-5))
+        if self.hparams_vision.get("norm_type", "layernorm") == "rmsnorm":
+            # nn.RMSNorm without eps uses finfo(float32).eps, also for bf16 inputs
+            self.gguf_writer.add_vision_attention_layernorm_eps(torch.finfo(torch.float32).eps)
+        else:
+            self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("projector_ln_eps", 1e-5))
        self.gguf_writer.add_vision_projector_scale_factor(self.merge_kernel_size[0])

        # Image size limits
@@ -120,7 +126,7 @@ class KimiK25Model(MmprojModel):

    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
        assert self.hparams_vision is not None
-        n_head = self.hparams_vision.get("num_attention_heads", 16)
+        n_head = self.find_vparam(["num_attention_heads", "vt_num_attention_heads"], optional=True) or 16

        # Permute Q/K weights/biases from interleaved to split RoPE format
        # This allows using build_rope_2d at runtime without post-permutation.
@@ -171,3 +177,19 @@ class Glm5vModel(KimiK25Model):
            name = name.replace("mm_projector.linear_", "mm_projector.proj.linear_", 1)

        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("KimiK31ForConditionalGeneration")
+@ModelBase.example("moonshotai/Kimi-K3.1-Pro")
+class KimiK31VisionModel(KimiK25Model):
+    """Kimi-K2.5 MoonViT3d with RMSNorm, a wider QKV and the patchmergerv2 projector"""
+
+    projector_type = gguf.VisionProjectorType.KIMIK31
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        assert self.hparams_vision["norm_type"] == "rmsnorm"
+        assert self.hparams_vision["mm_projector_type"] == "patchmergerv2"
+        n_head = self.find_vparam(["vt_num_attention_heads"])
+        self.gguf_writer.add_vision_head_dim(self.hparams_vision["qkv_hidden_size"] // n_head)
@@ -77,10 +77,10 @@ class _DecisionLoraMixin:
            delta = self.lora_scale * (lora["B"] @ lora["A"])
            data_torch = data_torch.float() + LazyTorchTensor.from_eager(delta)
            self.lora_merged.add(name[name.index("layers."):])
-        yield from super().modify_tensors(data_torch, name, bid)  # ty: ignore[unresolved-attribute]
+        yield from super().modify_tensors(data_torch, name, bid)  # ty: ignore[missing-attribute]

    def prepare_tensors(self):
-        super().prepare_tensors()  # ty: ignore[unresolved-attribute]
+        super().prepare_tensors()  # ty: ignore[missing-attribute]
        if len(self.lora_merged) != len(self.lora):
            raise ValueError(f"only {len(self.lora_merged)} of {len(self.lora)} LoRA tensors were merged into the base model")

@@ -22,15 +22,16 @@ class LLaDAModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None

-        vocab_dict = tokenizer.get_vocab()  # ty: ignore[unresolved-attribute]
+        vocab_dict = tokenizer.get_vocab()
        vocab_size = self.hparams.get("vocab_size", len(vocab_dict))
        assert max(vocab_dict.values()) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in vocab_dict.items()}
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()

        for i in range(vocab_size):
            if i not in reverse_vocab:
@@ -263,6 +263,7 @@ class MiniCPMV4_7VisionModel(MiniCPMV4_6VisionModel):
            # the tile separators are text tokens; clip appends their embeddings so that one chunk holds the whole image
            from transformers import AutoTokenizer
            tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+            assert tokenizer is not None
            ids = []
            for text in self.tok_embd_sep:
                tok = tokenizer.encode(text, add_special_tokens=False)
@@ -25,7 +25,8 @@ class MiniMaxText01Model(TextModel):
        # they get in the way of the token sampling process and must be suppressed

        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
-        tokenizer_vocab_size = tokenizer.vocab_size  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        tokenizer_vocab_size = tokenizer.vocab_size

        with open(self.dir_model / "model.safetensors.index.json", "r", encoding="utf-8") as f:
            weight_map = json.load(f)["weight_map"]
@@ -12,9 +12,9 @@ from .deepseek import DeepseekV2Model
 from .llama import LlamaModel

 if _mistral_common_installed:
-    from mistral_common.tokens.tokenizers.base import TokenizerVersion  # type: ignore[import-not-found, ty:unresolved-import]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer  # type: ignore[import-not-found, ty:unresolved-import]
-    from mistral_common.tokens.tokenizers.sentencepiece import SentencePieceTokenizer  # type: ignore[import-not-found, ty:unresolved-import]
+    from mistral_common.tokens.tokenizers.base import TokenizerVersion  # ty: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.tekken import Tekkenizer  # ty: ignore[import-not-found]
+    from mistral_common.tokens.tokenizers.sentencepiece import SentencePieceTokenizer  # ty: ignore[import-not-found]
 else:
    TokenizerVersion = None  # type: ignore[assignment]
    Tekkenizer = None  # type: ignore[assignment]
@@ -37,7 +37,8 @@ class MuseGlimmerModel(TextModel):

        from transformers import AutoTokenizer
        tok = AutoTokenizer.from_pretrained(self.dir_model)
-        eot_id = tok.convert_tokens_to_ids("<|eot|>")  # ty: ignore[unresolved-attribute]
+        assert tok is not None
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")
        if isinstance(eot_id, int) and eot_id >= 0:
            self.gguf_writer.add_eot_token_id(eot_id)

@@ -152,7 +153,7 @@ class MuseGlimmerAssistantModel(TextModel):
            target_arch = json.load(f)["architectures"][0]
        target_cls = get_model_class(target_arch)
        if target_cls is not type(self):
-            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
+            target_cls.set_vocab(self)  # ty: ignore[missing-attribute]
        else:
            super().set_vocab()

@@ -44,7 +44,8 @@ class MuseSparkModel(TextModel):

        from transformers import AutoTokenizer
        tok = AutoTokenizer.from_pretrained(self.dir_model)
-        eot_id = tok.convert_tokens_to_ids("<|eot|>")  # ty: ignore[unresolved-attribute]
+        assert tok is not None
+        eot_id = tok.convert_tokens_to_ids("<|eot|>")
        if isinstance(eot_id, int) and eot_id >= 0:
            self.gguf_writer.add_eot_token_id(eot_id)

@@ -73,7 +73,7 @@ class LiLiCorrDraftModel(DFlashModel):
        try:
            if target_arch == "NemotronHForCausalLM":
                self.is_moe = "num_experts_per_tok" in target_text_config
-            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
+            target_cls.set_vocab(self)  # ty: ignore[missing-attribute]
        finally:
            self.dir_model = original_dir
        self.gguf_writer.add_mask_token_id(self.hparams["dflash_config"]["mask_token_id"])
@@ -609,6 +609,7 @@ class NemotronHModel(GraniteHybridModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        assert tokenizer is not None

        # Pad vocab size (from Mamba2Model/GraniteHybridModel)
        self.hparams["pad_vocab_size_multiple"] = 8 # Setting this here since GraniteHybridModel.set_vocab() isn't being invoked now.
@@ -623,14 +624,14 @@ class NemotronHModel(GraniteHybridModel):
        vocab_skip = getattr(self, "_vocab_skip", ())
        user_defined_tokens = getattr(self, "_user_defined_tokens", None)

-        assert max(id_ for tok, id_ in tokenizer.vocab.items() if tok not in vocab_skip) < vocab_size  # ty: ignore[unresolved-attribute]
+        assert max(id_ for tok, id_ in tokenizer.vocab.items() if tok not in vocab_skip) < vocab_size

        tokpre = self.get_vocab_base_pre(tokenizer)

-        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()

-        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+        added_tokens_decoder = tokenizer.added_tokens_decoder

        for i in range(vocab_size):
            if i not in reverse_vocab:
@@ -641,7 +642,7 @@ class NemotronHModel(GraniteHybridModel):
                if token in added_vocab:
                    if not added_tokens_decoder[i].normalized:
                        previous_token = token
-                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[invalid-assignment]
                        if previous_token != token:
                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")

@@ -191,7 +191,7 @@ class PocketTTSModel(TextModel):
        # unigram segmentation, so use the UGM tokenizer instead
        from sentencepiece import sentencepiece_model_pb2 as model

-        proto = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        proto = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        proto.ParseFromString(open(self.dir_model / "tokenizer.model", "rb").read())
        assert proto.trainer_spec.model_type == 1, "expected a unigram tokenizer"

@@ -225,12 +225,13 @@ class Qwen3Model(Qwen2Model):
    def _find_rerank_config(self):
        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        assert tokenizer is not None

        self.is_rerank = True
        self.is_tied_embeddings = self.hparams.get("tie_word_embeddings", False)
-        self.token_false_id = tokenizer.convert_tokens_to_ids("no")  # ty: ignore[unresolved-attribute, invalid-assignment]
-        self.token_true_id = tokenizer.convert_tokens_to_ids("yes")  # ty: ignore[unresolved-attribute, invalid-assignment]
-        self.sep_token_id = tokenizer.convert_tokens_to_ids("|")  # ty: ignore[unresolved-attribute]
+        self.token_false_id = tokenizer.convert_tokens_to_ids("no")  # ty: ignore[invalid-assignment]
+        self.token_true_id = tokenizer.convert_tokens_to_ids("yes")  # ty: ignore[invalid-assignment]
+        self.sep_token_id = tokenizer.convert_tokens_to_ids("|")

        assert self.token_false_id is not None and self.token_true_id is not None

@@ -329,7 +330,7 @@ class _QwenMtpMixin:
        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
        type(self)._original_block_count = hparams.get(key)
        type(self).opt_num_mtp_layers = 0
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[missing-attribute]

    @classmethod
    def filter_tensors(cls, item):
@@ -366,7 +367,7 @@ class _QwenMtpMixin:
        return name, gen

    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        super().set_gguf_parameters()  # ty: ignore[missing-attribute]
        if self.no_mtp:
            return
        if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0:
@@ -374,15 +375,15 @@ class _QwenMtpMixin:

    def prepare_metadata(self, vocab_only: bool):
        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
+        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[missing-attribute]

        if not self.mtp_only or not from_dir:
            return

-        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"


@@ -685,7 +686,7 @@ class _Qwen35MRopeMixin:
    rope_parameters: dict

    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        super().set_gguf_parameters()  # ty: ignore[missing-attribute]
        if "mrope_section" not in self.rope_parameters:
            self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)

@@ -843,7 +844,7 @@ class DFlashModel(Qwen3Model):
        if target_cls is not type(self):
            if target_arch == "NemotronHForCausalLM":
                setattr(self, "is_moe", "num_experts_per_tok" in target_hparams)
-            target_cls.set_vocab(self)  # ty: ignore[unresolved-attribute]
+            target_cls.set_vocab(self)  # ty: ignore[missing-attribute]
        else:
            super().set_vocab()

@@ -40,7 +40,7 @@ class T5Model(TextModel):
        if not tokenizer_path.is_file():
            raise FileNotFoundError(f"File not found: {tokenizer_path}")

-        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())

        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
@@ -178,7 +178,7 @@ class T5EncoderModel(TextModel):
        if not tokenizer_path.is_file():
            raise FileNotFoundError(f"File not found: {tokenizer_path}")

-        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[missing-attribute]
        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())

        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
@@ -25,15 +25,16 @@ class XverseModel(TextModel):

        from transformers import AutoTokenizer
        tokenizer = AutoTokenizer.from_pretrained(dir_model)
-        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
        # Since we are checking the maximum index, we need to ensure it's strictly less than vocab_size,
        # because vocab_size is the count of items, and indexes start at 0.
-        max_vocab_index = max(tokenizer.get_vocab().values())  # ty: ignore[unresolved-attribute]
+        max_vocab_index = max(tokenizer.get_vocab().values())
        if max_vocab_index >= vocab_size:
            raise ValueError("Vocabulary size exceeds expected maximum size.")

-        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
-        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()

        for token_id in range(vocab_size):
            token_text = reverse_vocab[token_id].encode('utf-8')
@@ -44,7 +45,7 @@ class XverseModel(TextModel):
            elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
                toktype = gguf.TokenType.BYTE  # special
            elif reverse_vocab[token_id] in added_vocab:
-                if tokenizer.added_tokens_decoder[token_id].special:  # ty: ignore[unresolved-attribute]
+                if tokenizer.added_tokens_decoder[token_id].special:
                    toktype = gguf.TokenType.CONTROL
                else:
                    toktype = gguf.TokenType.USER_DEFINED
@@ -325,7 +325,8 @@ for model in [*pre_computed_hashes, *all_models]:
        except Exception as e:
            raise OSError(f"Error loading tokenizer for model {name}.") from e

-        chktok = tokenizer.encode(CHK_TXT)  # ty: ignore[unresolved-attribute]
+        assert tokenizer is not None
+        chktok = tokenizer.encode(CHK_TXT)
        chkhsh = sha256(str(chktok).encode()).hexdigest()

        logger.info(f"model: {name}")
@@ -486,6 +487,7 @@ for model in models:
        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
        continue  # Skip this model and continue with the next one in the loop

+    assert tokenizer is not None
    if not os.path.exists(f"models/ggml-vocab-{name}.gguf"):
        logger.info(f"Skip vocab files for model {name}, no GGUF file found")
        continue
@@ -497,7 +499,7 @@ for model in models:

    with open(f"models/ggml-vocab-{name}.gguf.out", "w") as f:
        for text in tests:
-            res = tokenizer.encode(text, add_special_tokens=False)  # ty: ignore[unresolved-attribute]
+            res = tokenizer.encode(text, add_special_tokens=False)
            for r in res:
                f.write(f" {r}")
            f.write("\n")
@@ -443,7 +443,8 @@ if __name__ == '__main__':
                    # the invocation string includes the "<|start_of_turn|>"
                    # token, but the adapters themselves were trained to
                    # activate _after_ that first token, so we drop it here.
-                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]  # ty: ignore[call-non-callable]
+                    assert tokenizer is not None
+                    alora_invocation_tokens = tokenizer(invocation_string)["input_ids"][1:]
                if alora_invocation_tokens:
                    logger.debug("GGUF KV: %s = %s", gguf.Keys.Adapter.ALORA_INVOCATION_TOKENS, alora_invocation_tokens)
                    self.gguf_writer.add_key_value(
@@ -52,6 +52,9 @@ The packages for FP32 and FP16 would have different accuracy and performance on

 ## News

+- 2026.11
+  - Update the CI and release build for oneAPI 2026.2 and Level Zero SDK 1.34.0. oneDNN is updated to the 2026.1 stand-alone package.
+
 - 2026.09
  - Update the CI build environment for oneAPI 2026.1 (unified oneAPI Toolkit). oneDNN is removed from the Deep Learning Essentials package in 2026.0, so the CI now uses the oneAPI Toolkit installer which still includes oneDNN.
  - oneAPI 2026.1 improves the SYCL build performance: measured with the same code on Arc B570, prompt processing 1331 vs 434 t/s (3.1x) vs the 2025.3-based release build.
@@ -285,6 +288,7 @@ Upon a successful installation, SYCL is enabled for the available Intel devices,

 |Verified release|
 |-|
+|2026.2 |
 |2026.1 |
 |2025.3.3 |
 |2025.2.1|
@@ -825,7 +829,8 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_XMX_ESIMD | 0 or 1 (default) | Use the XMX (DPAS) multi-column mat-vec kernels for reordered Q2_K to Q6_K and Q8_0 weights on Xe2 GPUs. Speeds up speculative decoding and short prompt batches. |
 | GGML_SYCL_XMX_MAX_COLS | non-negative integer, 80 (default) | Most activation columns (batch size) the `GGML_SYCL_ENABLE_XMX_ESIMD` kernels take; wider batches use the other mat-mul kernels. A larger value keeps XMX for bigger batches (as repeated launches), `0` never uses it. |
 | GGML_SYCL_XMX_GLU_MAX_COLS | 0 to 16, 16 (default) | Most columns the fused gate and up XMX kernel takes; wider batches run gate and up as separate mat-vecs. 16 is the most the kernel is built for, larger values are clamped to 16 and `0` turns the fused kernel off. |
-| GGML_SYCL_XMX_GATHER_TYPES | decimal bitmask, all bits set (default) | Weight formats that may use the XMX dequant-GEMM paths, which dequantize weights straight into the XMX tiles. This speeds up prompt processing of MoE models on GPUs with XMX units (Arc A- and B-series, Arc Pro, Data Center GPU Max), for example pp512 of Qwen3-30B-A3B UD-IQ3_XXS by about 50% on an Arc Pro B60. Bits:<br>* 1: IQ4_NL, 2: IQ3_S, 4: IQ4_XS, 8: IQ3_XXS, 16: IQ2_XXS, 32: IQ2_XS, 64: IQ2_S, 128: IQ1_S, 256: IQ1_M<br>* 512: Q8_0, 1024: Q4_K, 2048: Q5_K, 4096: Q6_K (MoE `MUL_MAT_ID` only)<br>Add values to combine them, for example `3` for IQ4_NL and IQ3_S; `0` disables the paths. A set bit does not force the path: batches of more than 64 tokens per expert or row lengths that are not a multiple of 256 (32 for IQ4_NL and Q8_0) use the library GEMM. |
+| GGML_SYCL_XMX_GATHER_TYPES | decimal bitmask, all bits set (default) | Weight formats that may use the XMX dequant-GEMM paths, which dequantize weights straight into the XMX tiles. This speeds up prompt processing of MoE models on GPUs with XMX units (Arc A- and B-series, Arc Pro, Data Center GPU Max), for example pp512 of Qwen3-30B-A3B UD-IQ3_XXS by about 50% on an Arc Pro B60. Bits:<br>* 1: IQ4_NL, 2: IQ3_S, 4: IQ4_XS, 8: IQ3_XXS, 16: IQ2_XXS, 32: IQ2_XS, 64: IQ2_S, 128: IQ1_S, 256: IQ1_M<br>* 512: Q8_0, 1024: Q4_K, 2048: Q5_K, 4096: Q6_K, 8192: Q2_K, 16384: Q3_K (MoE `MUL_MAT_ID` only)<br>Add values to combine them, for example `3` for IQ4_NL and IQ3_S; `0` disables the paths. A set bit does not force the path: batches of more than 64 tokens per expert use the library GEMM unless the device schedule width heuristic is overridden with `GGML_SYCL_MMID_SCHED=3`. Row lengths must be a multiple of 256 (32 for IQ4_NL and Q8_0). |
+| GGML_SYCL_MMID_SCHED | decimal bitmask, 1 (default) | Build the grouped MoE routing and tile schedule on the GPU, avoiding the expert-ID host readback and CPU sort.<br>* 1: enable the device schedule.<br>* 2: override the fallback for models with few routed experts by disabling the performance heuristic that rejects batches averaging more than 64 routed rows per expert. This may reduce performance; type, shape, precision and device capability checks still apply.<br>Use `3` to enable the device schedule and override this heuristic, or `0` to use the host schedule. Bit 2 alone does not enable the device schedule. Higher bits are ignored. |
 | GGML_SYCL_XMX_GATHER_SHAPES | decimal bitmask, 255 (default) | XMX `joint_matrix` combinations the paths of `GGML_SYCL_XMX_GATHER_TYPES` may use; the operand type comes from `GGML_SYCL_DYNAMIC_PRECISION` and the best supported combination is picked automatically (logged as `fg_pick_combo`). Bits:<br>* Xe2, Xe3, Xe-HPC: 1: f16 8x16x16, 2: f16 16x16x16, 4: f16 32x64x16, 8: f16 32x64x32, 32: tf32 8x16x8, 64: bf16 8x16x16<br>* Xe-HPG (Arc A770, ARL-H): 16: f16 8x8x16, 128: bf16 8x8x16<br>Clear a bit to exclude a combination, or set a single bit to force one for testing. |
 | GGML_SYCL_DYNAMIC_PRECISION | `F16` (default with `GGML_SYCL_F16=ON`), `BF16`, `TF32` or `F32` (default otherwise) | Operand type of the XMX dequant-GEMM paths (`GGML_SYCL_XMX_GATHER_TYPES`); accumulation is always f32. `F16` is the fastest, but activations above 65504 overflow. `BF16` keeps the f32 range at a 7-bit mantissa, `TF32` keeps the range and the f16 mantissa but is about 30% slower and needs Xe2, Xe3 or Xe-HPC, and `F32` turns the XMX paths off. Ops that request a higher src1 precision ([TAG_GGML_PREC]) get it regardless of this setting. |
 | GGML_SYCL_DYNAMIC_REQUIRED_PRECISION | `F32` (default), `TF32`, `BF16` or `F16` | Lowest type the XMX paths may use for an op that requests an F32 src1, such as Mistral 4 `ffn_down_exps`. The default runs such ops on the library f32 GEMM; `TF32` or `BF16` trade mantissa for speed while keeping the f32 range. `F16` ignores the request and can overflow; it is meant for testing only. |
@@ -86,7 +86,7 @@ Legend:
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -12336,8 +12336,13 @@
 "WebGPU: WebGPU","PAD","type=f32,ne_a=[100,1,1,1],pad_0=100,pad_1=0,circular=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","PAD","type=f32,ne_a=[100,1,1,1],pad_0=0,pad_1=100,circular=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","PAD","type=f32,ne_a=[100,100,1,1],pad_0=50,pad_1=50,circular=0","support","1","yes","WebGPU"
-"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","WebGPU"
-"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","no","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9,view=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[1,3,2,2],pad_0=0,pad_1=0,view=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[5,3,2,2],pad_0=4,pad_1=1,view=1","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[281,128,1,1],pad_0=2,pad_1=2,view=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[257,7,1,1],pad_0=0,pad_1=256,view=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[2,16777217,1,1],pad_0=1,pad_1=1,view=0","support","1","yes","WebGPU"
+"WebGPU: WebGPU","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9,view=0","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","0","no","WebGPU"
 "WebGPU: WebGPU","ARANGE","type=f32,start=0.000000,stop=10.000000,step=1.000000","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARANGE","type=f32,start=3.000000,stop=3.500000,step=1.000000","support","1","yes","WebGPU"
@@ -0,0 +1,125 @@
+# TTS normalization patterns
+
+Text-to-speech models often need their input normalized before tokenization: numbers, currencies, dates and symbols spelled out, punctuation mapped, spacing fixed. The exact rules are model-specific (they must match what the model was trained on), so they live in the model's `tts` task template rather than in C++.
+
+Jinja has no regex support, and exposing `std::regex` / Python `re` would not help: the two differ in many details (`\d`, `\w`, `\b`, lookbehind, case folding, backtracking limits), so a template could give different results in llama.cpp and in Python. Instead, task templates get two functions with a small matching language whose semantics are defined here:
+
+- `tts_norm_scan(text, rules)`: split text into matched and unmatched segments
+- `tts_norm_sub(text, pattern, repl)`: replace every match
+
+Implementations:
+- C++: [common/chat-tts-norm.cpp](../common/chat-tts-norm.cpp), injected into task templates by `common_chat_template_render()` in [common/chat.cpp](../common/chat.cpp). They are not available to chat templates.
+- Python reference: [scripts/chat_tts_norm.py](../scripts/chat_tts_norm.py), for use with `jinja2` (`env.globals["tts_norm_scan"] = chat_tts_norm.tts_norm_scan`)
+- Both are checked against each other by `tests/test-jinja.cpp` (`test-jinja -py "tts norm patterns.*"`)
+
+## Example
+
+```jinja
+{%- for s in tts_norm_scan(text, [
+    ['money',   '%$(%d+)(?:%.(%d%d))?'],
+    ['percent', '(%d+)%%'],
+    ['time',    '(%d%d?):(%d%d)(?!%d)'],
+    ['neg',     '(?<!%w)-(%d+)'],
+    ['number',  '%d+'],
+]) -%}
+  {%- if   s.rule == 'money'   -%}{%- set whole, cents = s.groups -%}{{ num(whole) }} dollars{{ ' and ' ~ num(cents) ~ ' cents' if cents else '' }}
+  {%- elif s.rule == 'percent' -%}{{ num(s.groups[0]) }} percent
+  {%- elif s.rule == 'time'    -%}{%- set h, m = s.groups -%}{{ num(h) }} {{ num(m) }}
+  {%- elif s.rule == 'neg'     -%}minus {{ num(s.groups[0]) }}
+  {%- elif s.rule == 'number'  -%}{{ num(s.text) }}
+  {%- else -%}{{ s.text }}
+  {%- endif -%}
+{%- endfor -%}
+```
+
+`It costs $3.50, 20% off, at 10:45, -5 degrees, 7 cats.` becomes `It costs three dollars and fifty cents, twenty percent off, at ten forty-five, minus five degrees, seven cats.` (with a `num` macro that spells numbers, see the "mini text normalizer" test in `tests/test-jinja.cpp`).
+
+## Functions
+
+### `tts_norm_scan(text, rules)`
+
+`rules` is one of:
+- a list of `[name, pattern]` pairs (recommended, the order is explicit)
+- a dict `{name: pattern}`, in insertion order
+- a single pattern string, the rule is named `match`
+
+Scanning starts at the first codepoint. At each position, the rules are tried in order and the first one that produces a **non-empty** match wins; the scan continues after the match. If no rule matches, one codepoint is skipped. Matches never overlap.
+
+The result is a list of segments covering the whole text, in order. Consecutive unmatched codepoints are merged into one segment. Each segment is a dict:
+
+| key | matched segment | unmatched segment |
+|---|---|---|
+| `rule` | rule name | `none` |
+| `text` | matched text | unmatched text |
+| `groups` | list of capture groups, in order of their `(` | `[]` |
+
+There are no named groups, unpack `groups` instead: `{% set whole, cents = s.groups %}`.
+
+A group that did not participate in the match (e.g. inside a `?` that matched nothing) is an empty string, so it renders as nothing and is falsy.
+
+### `tts_norm_sub(text, pattern, repl)`
+
+Replaces every match of `pattern`, with the same scanning rules as `tts_norm_scan`. `repl` is a literal string with references:
+
+| reference | value |
+|---|---|
+| `%0` | whole match |
+| `%1` .. `%9` | group |
+| `%%` | a literal `%` |
+
+Any other `%` sequence in `repl`, or a reference to a group that does not exist, is an error.
+
+## Pattern syntax
+
+Patterns operate on Unicode codepoints. Invalid UTF-8 bytes (C++ only) are seen as U+FFFD, one per byte, and preserved in the output.
+
+| syntax | meaning |
+|---|---|
+| `x` | the literal codepoint `x` |
+| `%x` | escape: `%` followed by any non-alphanumeric codepoint is that literal codepoint (`%.`, `%$`, `%%`, `%(`, ...) |
+| `.` | any codepoint, including newlines |
+| `%d` `%s` `%p` `%a` `%w` `%u` `%l` `%h` `%c` | character classes, see below; the uppercase form (`%D`, ...) is the negation |
+| `[...]` | set: literal codepoints, ranges `a-z`, classes `%d`; `[^...]` is the negation |
+| `(...)` | capture group |
+| `(?:...)` | group without capture |
+| `a\|b` | ordered alternation |
+| `?` `*` `+` `{m}` `{m,}` `{m,n}` | quantifiers, greedy and possessive; `m`, `n` up to 1000 |
+| `^` `$` | start / end of the whole text |
+| `%b` | boundary: between a `%w` codepoint and a non-`%w` codepoint (the edges of the text count as non-`%w`) |
+| `(?=...)` `(?!...)` | lookahead, positive / negative |
+| `(?<=c)` `(?<!c)` | lookbehind of exactly one codepoint: `c` is a literal, a set, a class or `.` |
+| `(?i)` | at the very start of the pattern only: ignore ASCII case |
+
+These characters must be escaped with `%` to be matched literally: `( ) [ ] { } . | ? * + ^ $ %`. Inside a set, only `]`, `%`, `-` (between two codepoints) and `^` (first position) are special.
+
+### Matching semantics
+
+The matcher is a PEG, not a backtracking regex:
+
+- **Ordered alternation**: alternatives are tried left to right and the first one that matches is committed to. If what follows fails, the other alternatives are not retried. `(?:a|ab)c` never matches `abc`; write `(?:ab|a)c`.
+- **Possessive quantifiers**: quantifiers match as many repetitions as possible and never give any back. `%d+0` never matches; `.*x` never matches. Write patterns that describe exactly what to consume.
+- A repetition that matches the empty string ends the loop.
+- Captures inside lookaheads are discarded. A capture inside a repeated group holds its last repetition.
+
+This gives a simple and linear-time implementation with identical results in C++ and Python.
+
+### Character classes
+
+Classes are defined by explicit codepoint ranges, not by a Unicode database, so the result does not depend on the Unicode version of the platform.
+
+| class | definition |
+|---|---|
+| `%d` | ASCII digits `0-9` only |
+| `%s` | U+0009..U+000D, U+0020, U+0085, U+00A0, U+1680, U+2000..U+200A, U+2028, U+2029, U+202F, U+205F, U+3000 |
+| `%p` | punctuation and symbols: ASCII punctuation, U+00A1..U+00BF (except U+00AA, U+00B5, U+00BA), U+00D7, U+00F7, U+2010..U+2027, U+2030..U+205E, U+20A0..U+20CF (currency), U+2100..U+2BFF (letterlike, number forms, arrows, math, technical, box drawing, shapes, misc symbols, dingbats), CJK punctuation (U+3001..U+3003, U+3008..U+3020, U+3030, U+303D, U+30FB), U+FE10..U+FE19, U+FE30..U+FE6B, fullwidth punctuation (U+FF01..U+FF0F, U+FF1A..U+FF20, U+FF3B..U+FF40, U+FF5B..U+FF65), U+1F000..U+1FAFF (emoji) |
+| `%c` | control and invisible format characters: U+0000..U+001F, U+007F..U+009F, U+200B..U+200F, U+2060..U+206F, U+FEFF |
+| `%a` | ASCII letters, U+00AA, U+00B5, U+00BA, and every codepoint from U+00C0 that is not in `%s`, `%p` or `%c` |
+| `%w` | `%a` or `%d` |
+| `%u` / `%l` | ASCII uppercase / lowercase letters only |
+| `%h` | Han: U+2E80..U+2FDF, U+3005, U+3007, U+3021..U+3029, U+3038..U+303B, U+3400..U+4DBF, U+4E00..U+9FFF, U+F900..U+FAFF, U+20000..U+323AF |
+
+Classes may overlap (e.g. U+0085 is both `%s` and `%c`). `%a` intentionally includes all scripts, combining marks and CJK, so `(?<!%w)` and `%b` work as "not inside a word" for any language.
+
+## Errors
+
+Invalid patterns are rejected when the function is called, with a message and a position, for example: unbalanced parentheses, a quantifier with nothing to repeat, multiple quantifiers (`a+?`), repeating a zero-width assertion (`%b+`), unknown escapes (`%q`), unescaped `]` or `}`, an empty set, a lookbehind wider than one codepoint and `(?i)` not at the start.
@@ -7,7 +7,7 @@ import os

 # Add utils directory to path for direct script execution
 sys.path.insert(0, str(Path(__file__).parent.parent / "utils"))
-from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
+from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # ty: ignore[import-not-found]

 def quick_logits_check(pytorch_file, llamacpp_file):
    """Lightweight sanity check before NMSE"""
@@ -33,6 +33,7 @@ print("EOS token id:     ", config.eos_token_id)

 print("Loading model and tokenizer using AutoTokenizer:", model_path)
 tokenizer = AutoTokenizer.from_pretrained(model_path)
+assert tokenizer is not None

 if unreleased_model_name:
    model_name_lower = unreleased_model_name.lower()
@@ -56,11 +57,11 @@ model_name = os.path.basename(model_path)
 print(f"Model name: {model_name}")

 prompt = "Hello world today"
-input_ids = tokenizer(prompt, return_tensors="pt").input_ids  # ty: ignore[call-non-callable]
+input_ids = tokenizer(prompt, return_tensors="pt").input_ids
 token_ids = input_ids[0].cpu().tolist()
 print(f"Input tokens: {input_ids}")
 print(f"Input text: {repr(prompt)}")
-print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")  # ty: ignore[unresolved-attribute]
+print(f"Tokenized: {tokenizer.convert_ids_to_tokens(input_ids[0])}")

 with torch.no_grad():
    outputs = model(input_ids, output_hidden_states=True)
@@ -83,7 +84,7 @@ with torch.no_grad():

    # Print embeddings per token in the requested format
    print("\nToken embeddings:")
-    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])  # ty: ignore[unresolved-attribute]
+    tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
    for i, embedding in enumerate(token_embeddings):
        # Format: show first few values, ..., then last few values
        if len(embedding) > 10:
@@ -64,7 +64,7 @@ def load_model_and_tokenizer(model_path, use_sentence_transformers=False, device
        print("Using SentenceTransformer to apply all numbered layers")
        model = SentenceTransformer(model_path)
        tokenizer = model.tokenizer
-        config = model[0].auto_model.config  # ty: ignore[unresolved-attribute]
+        config = model[0].auto_model.config  # ty: ignore[missing-union-attribute]
    else:
        tokenizer = AutoTokenizer.from_pretrained(model_path)
        config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
@@ -5,7 +5,7 @@ import sys
 import os
 import argparse
 from pathlib import Path
-from common import get_model_name_from_env_path  # type: ignore[import-not-found, ty:unresolved-import]
+from common import get_model_name_from_env_path  # ty: ignore[import-not-found]

 def calculate_nmse(reference, test):
    mse = np.mean((test - reference) ** 2)
@@ -2,7 +2,7 @@

 import argparse
 import sys
-from common import compare_tokens  # type: ignore[import-not-found, ty:unresolved-import]
+from common import compare_tokens  # ty: ignore[import-not-found]


 def parse_arguments():
@@ -7,7 +7,7 @@ import importlib
 from pathlib import Path

 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
-from common import compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
+from common import compare_tokens, exit_with_warning  # ty: ignore[import-not-found]

 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')

@@ -183,6 +183,7 @@ def main():
    # Load the python model to get configuration information and also to load the tokenizer.
    print("Loading model and tokenizer using AutoTokenizer:", args.model_path)
    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    assert tokenizer is not None
    config = AutoConfig.from_pretrained(args.model_path, trust_remote_code=True)

    if unreleased_model_name:
@@ -207,8 +208,8 @@ def main():
        else:
            model = AutoModel.from_pretrained(args.model_path, trust_remote_code=True)

-    encoded = tokenizer(prompt, return_tensors="pt")  # ty: ignore[call-non-callable]
-    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])  # ty: ignore[unresolved-attribute]
+    encoded = tokenizer(prompt, return_tensors="pt")
+    tokens = tokenizer.convert_ids_to_tokens(encoded['input_ids'][0])
    n_tokens = len(tokens)
    print(f"n_tokens: {n_tokens}");
    print(f"hidden_size: {model.config.hidden_size}")
@@ -1285,7 +1285,7 @@ def convert_dictionary_to_pydantic_model(dictionary: dict[str, Any], model_name:
                    if items != {}:
                        array = {"properties": items}
                        array_type = convert_dictionary_to_pydantic_model(array, f"{model_name}_{field_name}_items")
-                        fields[field_name] = (list[array_type], ...)  # ty: ignore[invalid-type-form]
+                        fields[field_name] = (list[array_type], ...)
                    else:
                        fields[field_name] = (list, ...)
                elif field_type == "object":
@@ -205,7 +205,7 @@ option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copie
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA                         "ggml: compile ggml FlashAttention CUDA kernels"  ON)
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
-set   (GGML_CUDA_FA_QUANTS "q4_0-q4_0;q8_0-q8_0;f16-f16;bf16-bf16" CACHE STRING
+set   (GGML_CUDA_FA_QUANTS "q4_0-q4_0;q8_0-q8_0;f16-f16;bf16-bf16;f8_e4m3-f8_e4m3" CACHE STRING
                                            "ggml: FlashAttention K-V type combinations to compile, \"all\" or a list such as \"q8_0-q8_0;q8_0-q4_0\"")
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 option(GGML_CUDA_NCCL                       "ggml: use NVIDIA Collective Comm. Library"       ON)
@@ -58,7 +58,7 @@ endfunction()

 # Determines which FlashAttention vector kernel template instances to compile, returns them in OUT_SRCS.
 function(ggml_cuda_fattn_vec_instances DIR OUT_SRCS)
-    set(FA_TYPES q4_0 q4_1 q5_0 q5_1 q8_0 bf16 f16)
+    set(FA_TYPES q4_0 q4_1 q5_0 q5_1 q8_0 bf16 f16 f8_e4m3)

    string(TOLOWER "${GGML_CUDA_FA_QUANTS}" FA_QUANTS)
    string(STRIP   "${FA_QUANTS}" FA_QUANTS)
@@ -131,15 +131,9 @@ endif()
 # POSIX conformance
 #

-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+# POSIX.1-2001 with XSI extensions provides clock_gettime, CLOCK_MONOTONIC, posix_memalign and M_PI.
+# OpenBSD needs POSIX.1-2008 for locale_t. FreeBSD needs it for F_DUPFD_CLOEXEC and strnlen.
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD|FreeBSD")
    add_compile_definitions(_XOPEN_SOURCE=700)
 elseif (CMAKE_SYSTEM_NAME MATCHES "AIX")
    # Don't define _XOPEN_SOURCE.  We need _ALL_SOURCE, which is the default,
@@ -1398,34 +1398,35 @@ static void ggml_backend_cann_buffer_get_tensor(ggml_backend_buffer_t buffer,
 static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
                                                const ggml_tensor *   src,
                                                ggml_tensor *         dst) {
-    if (ggml_backend_buft_is_cann(src->buffer->buft)) {
-        ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
-        ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
+    if (!ggml_backend_buft_is_cann(src->buffer->buft)) {
+        return false;
+    }

-        size_t memcpy_size = ggml_nbytes(src);
-        // Same device.
-        if (src_ctx->device == dst_ctx->device) {
-            ACL_CHECK(aclrtMemcpy((char *) dst->data, memcpy_size, (const char *) src->data, memcpy_size,
-                                  ACL_MEMCPY_DEVICE_TO_DEVICE));
-            return true;
-        } else {
+    ggml_backend_cann_buffer_context * src_ctx = (ggml_backend_cann_buffer_context *) src->buffer->context;
+    ggml_backend_cann_buffer_context * dst_ctx = (ggml_backend_cann_buffer_context *) buffer->context;
+
+    size_t memcpy_size = ggml_nbytes(src);
+    // Same device.
+    if (src_ctx->device == dst_ctx->device) {
+        ACL_CHECK(aclrtMemcpy((char *) dst->data, memcpy_size, (const char *) src->data, memcpy_size,
+                                ACL_MEMCPY_DEVICE_TO_DEVICE));
+        return true;
+    } else {
 #ifdef ASCEND_310P
-            // TODO: Support 310p P2P copy
-            return false;
+        // TODO: Support 310p P2P copy
+        return false;
 #endif
-            // Different device but can access by peer.
-            int32_t canAccessPeer = 0;
-            ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device, dst_ctx->device));
-            if (canAccessPeer) {
-                ggml_cann_set_device(src_ctx->device);
-                ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
-                ACL_CHECK(aclrtMemcpy((char *) dst->data, memcpy_size, (const char *) src->data, memcpy_size,
-                                      ACL_MEMCPY_DEVICE_TO_DEVICE));
-                return true;
-            }
+        // Different device but can access by peer.
+        int32_t canAccessPeer = 0;
+        ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device, dst_ctx->device));
+        if (canAccessPeer) {
+            ggml_cann_set_device(src_ctx->device);
+            ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
+            ACL_CHECK(aclrtMemcpy((char *) dst->data, memcpy_size, (const char *) src->data, memcpy_size,
+                                    ACL_MEMCPY_DEVICE_TO_DEVICE));
+            return true;
        }
    }
-    return false;
 }

 /**
@@ -2446,6 +2447,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            break;
        case GGML_OP_MUL_MAT:
            {
+                if (ggml_prec_get_src(op, 1) == GGML_PREC_BF16 || ggml_prec_get_src(op, 1) == GGML_PREC_F32) {
+                    return false;
+                }
                switch (op->src[0]->type) {
 #ifndef ASCEND_310P
                    case GGML_TYPE_BF16:
@@ -2864,12 +2868,13 @@ static ggml_backend_t ggml_backend_cann_device_init(ggml_backend_dev_t dev, cons
 *              otherwise false.
 */
 static bool ggml_backend_cann_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    if (ggml_backend_buft_is_cann(buft)) {
-        ggml_backend_cann_device_context *      dev_ctx  = (ggml_backend_cann_device_context *) dev->context;
-        ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
-        return buft_ctx->device == dev_ctx->device;
+    if (!ggml_backend_buft_is_cann(buft)) {
+        return false;
    }
-    return false;
+
+    ggml_backend_cann_device_context *      dev_ctx  = (ggml_backend_cann_device_context *) dev->context;
+    ggml_backend_cann_buffer_type_context * buft_ctx = (ggml_backend_cann_buffer_type_context *) buft->context;
+    return buft_ctx->device == dev_ctx->device;
 }

 static ggml_backend_buffer_type_t ggml_backend_cann_device_get_buffer_type(ggml_backend_dev_t dev) {
@@ -8911,14 +8911,9 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(

        float       * VKQ32 = (float       *) params->wdata + ith*(1*DK + 2*DV + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator
        float       * V32   =                 (VKQ32 + 1*DV); // (temporary) FP32 V buffer
-        ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*DV); // (temporary) FP16 VKQ accumulator
        ggml_fp16_t * Q_q   = (ggml_fp16_t *) (VKQ32 + 2*DV); // (temporary) buffer for Q converted to quantized/FP16

-        if (v->type == GGML_TYPE_F16) {
-            memset(VKQ16, 0, DV*sizeof(ggml_fp16_t));
-        } else {
-            memset(VKQ32, 0, DV*sizeof(float));
-        }
+        memset(VKQ32, 0, DV*sizeof(float));

        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1] + (iq2%mask->ne[2])*mask->nb[2] + (iq3%mask->ne[3])*mask->nb[3]) : NULL;

@@ -8970,53 +8965,32 @@ static void ggml_compute_forward_flash_attn_ext_f16_one_chunk(

            const char * v_data = ((const char *) v->data + (ik1*nbv1 + iv2*nbv2 + iv3*nbv3));

-            if (v->type == GGML_TYPE_F16) {
-                if (s > M) {
-                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
-                    M = s;
-                    ms = expf(Mold - M);
+            if (s > M) {
+                // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
+                M = s;
+                ms = expf(Mold - M);

-                    // V = V*expf(Mold - M)
-                    ggml_vec_scale_f16(DV, VKQ16, ms);
-                } else {
-                    // no new maximum, ms == 1.0f, vs != 1.0f
-                    vs = expf(s - M);
-                }
-
-                // V += v*expf(s - M)
-                ggml_vec_mad_f16(DV, VKQ16, (const ggml_fp16_t *) v_data, vs);
+                // V = V*expf(Mold - M)
+                ggml_vec_scale_f32(DV, VKQ32, ms);
            } else {
-                if (s > M) {
-                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
-                    M = s;
-                    ms = expf(Mold - M);
+                // no new maximum, ms == 1.0f, vs != 1.0f
+                vs = expf(s - M);
+            }

-                    // V = V*expf(Mold - M)
-                    ggml_vec_scale_f32(DV, VKQ32, ms);
-                } else {
-                    // no new maximum, ms == 1.0f, vs != 1.0f
-                    vs = expf(s - M);
-                }
-
-                // V += v*expf(s - M)
-                if (v_to_float) {
-                    v_to_float(v_data, V32, DV);
-                    ggml_vec_mad_f32(DV, VKQ32, V32, vs);
-                } else {
-                    // V is F32
-                    ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs);
-                }
+            // V += v*expf(s - M)
+            if (v->type == GGML_TYPE_F16) {
+                ggml_vec_mad_f16_f32(DV, VKQ32, (const ggml_fp16_t *) v_data, vs);
+            } else if (v_to_float) {
+                v_to_float(v_data, V32, DV);
+                ggml_vec_mad_f32(DV, VKQ32, V32, vs);
+            } else {
+                // V is F32
+                ggml_vec_mad_f32(DV, VKQ32, (const float *) v_data, vs);
            }

            S = S*ms + vs; // scale and increment sum with partial sum
        }

-        if (v->type == GGML_TYPE_F16) {
-            for (int64_t d = 0; d < DV; ++d) {
-                VKQ32[d] = GGML_CPU_FP16_TO_FP32(VKQ16[d]);
-            }
-        }
-
        // sinks - apply only on the first kv-chunk
        if (sinks && ic_start == 0) {
            const float s = ((float *)((char *) sinks->data))[h];
@@ -9569,8 +9543,8 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 void ggml_compute_forward_flash_attn_ext(
        const ggml_compute_params * params,
        ggml_tensor * dst) {
-    switch (dst->op_params[3]) {
-        case GGML_PREC_DEFAULT:
+    switch (ggml_prec_get_acc(dst)) {
+        case GGML_PREC_UNDEFINED:
        case GGML_PREC_F32:
        case GGML_PREC_BF16:
            {
@@ -1180,7 +1180,7 @@ class tensor_traits_common : public tensor_traits_base {
        const int64_t DK = nek0;
        const int64_t DV = nev0;

-        const bool supported_prec  = (dst->op_params[3] == GGML_PREC_F32 || dst->op_params[3] == GGML_PREC_BF16 || dst->op_params[3] == GGML_PREC_DEFAULT);
+        const bool supported_prec  = (ggml_prec_get_acc(dst) == GGML_PREC_F32 || ggml_prec_get_acc(dst) == GGML_PREC_BF16 || ggml_prec_get_acc(dst) == GGML_PREC_UNDEFINED);
        const bool supported_types = (q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 && v->type == GGML_TYPE_F16);
        const bool supported_shape = (DK > 0 && DK <= 128 && DV > 0 && DV <= 128);
        const bool supported_vlen  = (__riscv_vlenb() == 128);
@@ -436,6 +436,41 @@ inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const
 #endif
 }

+// y (F32) += x (F16)*v
+// on AVX512, AVX with F16C, NEON and RVV, x is widened to F32 in-register and fused into the multiply-add
+inline static void ggml_vec_mad_f16_f32(const int n, float * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
+    int i = 0;
+#if defined(__AVX512F__)
+    const __m512 vx = _mm512_set1_ps(v);
+    for (; i + 16 <= n; i += 16) {
+        const __m512 ax = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x + i)));
+        _mm512_storeu_ps(y + i, _mm512_fmadd_ps(ax, vx, _mm512_loadu_ps(y + i)));
+    }
+#elif defined(__AVX__) && defined(__F16C__)
+    const __m256 vx = _mm256_set1_ps(v);
+    for (; i + 8 <= n; i += 8) {
+        const __m256 ax = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x + i)));
+        _mm256_storeu_ps(y + i, GGML_F32x8_FMA(_mm256_loadu_ps(y + i), ax, vx));
+    }
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) && defined(__ARM_FP16_FORMAT_IEEE)
+    const float32x4_t vx = vdupq_n_f32(v);
+    for (; i + 8 <= n; i += 8) {
+        const float16x8_t ax = vld1q_f16((const __fp16 *)(x + i));
+        vst1q_f32(y + i + 0, vfmaq_f32(vld1q_f32(y + i + 0), vcvt_f32_f16(vget_low_f16 (ax)), vx));
+        vst1q_f32(y + i + 4, vfmaq_f32(vld1q_f32(y + i + 4), vcvt_f32_f16(vget_high_f16(ax)), vx));
+    }
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfhmin)
+    for (int vl; i < n; i += vl) {
+        vl = __riscv_vsetvl_e16m2(n - i);
+        const vfloat32m4_t ax = __riscv_vfwcvt_f_f_v_f32m4(__riscv_vle16_v_f16m2((const _Float16 *)x + i, vl), vl);
+        __riscv_vse32_v_f32m4(y + i, __riscv_vfmacc_vf_f32m4(__riscv_vle32_v_f32m4(y + i, vl), v, ax, vl), vl);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] += GGML_CPU_FP16_TO_FP32(x[i])*v;
+    }
+}
+
 inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
    const int sve_register_length = svcntb() * 8;
@@ -163,7 +163,7 @@ if (CUDAToolkit_FOUND)
    if (GGML_STATIC)
        if (WIN32)
            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas)
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
        else ()
            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "10.1")
                target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -172,7 +172,7 @@ if (CUDAToolkit_FOUND)
            endif()
        endif()
    else()
-        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas)
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
    endif()

    if (GGML_CUDA_NO_VMM)
@@ -7,7 +7,6 @@
 #include <cstdint>
 #include <cstdlib>
 #include <memory>
-#include <mutex>

 #if defined(GGML_USE_HIP)
 #define GGML_COMMON_DECL_HIP
@@ -389,6 +388,10 @@ static bool blackwell_mma_available(const int cc) {
           ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_RUBIN;
 }

+static bool fp8_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_ADA_LOVELACE && cc < GGML_CUDA_CC_RUBIN;
+}
+
 // Checks whether the tensor's base data pointer and higher-dimensional strides are byte-aligned to `alignment` bytes.
 static bool ggml_cuda_is_aligned(const ggml_tensor * tensor, const size_t alignment) {
    GGML_ASSERT(tensor != nullptr);
@@ -884,6 +887,67 @@ static __device__ __forceinline__ float ggml_cuda_ue4m3_to_fp32(uint8_t x) {
 #endif // defined(GGML_USE_HIP) && defined(CDNA3) && defined(FP8_AVAILABLE) && HIP_VERSION >= 60200000
 }

+static __host__ __device__ __forceinline__ float ggml_cuda_f8_e4m3_to_fp32(uint8_t x) {
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    __nv_fp8_e4m3 xf;
+    xf.__x = x;
+    return static_cast<float>(xf);
+#else
+    const uint8_t magnitude = x & 0x7F;
+    if (magnitude == 0x7F) {
+        return NAN;
+    }
+
+    const int exp = (magnitude >> 3) & 0x0F;
+    const int man = magnitude & 0x07;
+    float value;
+    if (exp == 0) {
+        value = ldexpf((float) man, -9);
+    } else {
+        value = ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    }
+    return x & 0x80 ? -value : value;
+#endif // defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
+static __host__ __device__ __forceinline__ uint8_t ggml_cuda_fp32_to_f8_e4m3(float x) {
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    return __nv_cvt_float_to_fp8(x, __NV_SATFINITE, __NV_E4M3);
+#elif !defined(__CUDA_ARCH__) && !defined(__HIP_DEVICE_COMPILE__) && !defined(__MUSA_ARCH__)
+    // ggml_cuda_cast also runs on the host, where __float2int_rn is not available.
+    return ggml_fp32_to_f8_e4m3(x);
+#else
+    const uint8_t sign = signbit(x) ? 0x80 : 0;
+    x = fabsf(x);
+
+    if (isnan(x)) {
+        return sign | 0x7F;
+    }
+    if (x == 0.0f) {
+        return sign;
+    }
+    if (isinf(x) || x >= 448.0f) {
+        return sign | 0x7E;
+    }
+    if (x < 0.015625f) {
+        return sign | (uint8_t) __float2int_rn(x * 512.0f);
+    }
+
+    int exp;
+    const float mantissa = frexpf(x, &exp) * 2.0f;
+    int encoded_exp = exp + 6;
+    int encoded_man = __float2int_rn((mantissa - 1.0f) * 8.0f);
+    if (encoded_man == 8) {
+        encoded_man = 0;
+        encoded_exp++;
+    }
+    if (encoded_exp > 15 || (encoded_exp == 15 && encoded_man > 6)) {
+        return sign | 0x7E;
+    }
+    return sign | (uint8_t) (encoded_exp << 3) | (uint8_t) encoded_man;
+#endif // defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
 static __device__ __forceinline__ uint8_t ggml_cuda_fp32_to_ue4m3(float x) {
 #if defined(BLACKWELL_MMA_AVAILABLE) // This is used for NVFP4 subblock scale quantizations only
    if (!(x > 0.0f)) {
@@ -1271,186 +1335,9 @@ struct ggml_tensor_extra_gpu {
 #define USE_CUDA_GRAPH
 #endif

-struct ggml_cuda_graph {
-#ifdef USE_CUDA_GRAPH
-    ~ggml_cuda_graph() {
-        for (cudaGraphExec_t instance : instances) {
-            CUDA_CHECK(cudaGraphExecDestroy(instance));
-        }
-    }
-    // the ggml graph is captured as a series of CUDA graphs, so the GPU can run one chunk while the next is enqueued
-    std::vector<cudaGraphExec_t> instances;
-    size_t num_nodes = 0;
-    std::vector<cudaGraphNode_t> nodes;
-    bool disable_due_to_gpu_arch = false;
-    bool warmup_complete = false;
-    uint64_t uid = 0;
-    int64_t last_used_time = 0;
-    struct node_properties {
-        ggml_tensor node;
-        void *   node_src_data_ptrs[GGML_MAX_SRC];
-        int64_t  node_src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
-        size_t   node_src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
-    };
-    std::vector<node_properties> node_props;
-
-    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);
-    }
-#endif
-};
-
-struct ggml_cuda_concurrent_event {
-    std::vector<cudaEvent_t> join_events;
-    cudaEvent_t              fork_event = nullptr;
-
-    int                                          n_streams = 0;
-    std::unordered_map<const ggml_tensor *, int> stream_mapping;
-
-    // Original order of nodes in this concurrent region (before interleaving)
-    // Used to restore grouping for fusion within streams
-    std::vector<const ggml_tensor *> original_order;
-
-    const ggml_tensor * join_node;
-
-    ggml_cuda_concurrent_event() = default;
-
-    ggml_cuda_concurrent_event(const ggml_cuda_concurrent_event &) = delete;
-    ggml_cuda_concurrent_event & operator=(const ggml_cuda_concurrent_event &) = delete;
-
-    explicit ggml_cuda_concurrent_event(int n_streams) : n_streams(n_streams) {
-        join_events.resize(n_streams);
-
-        for (size_t i = 0; i < join_events.size(); ++i) {
-            CUDA_CHECK(cudaEventCreateWithFlags(&join_events[i], cudaEventDisableTiming));
-        }
-
-        CUDA_CHECK(cudaEventCreateWithFlags(&fork_event, cudaEventDisableTiming));
-    }
-
-    ggml_cuda_concurrent_event(ggml_cuda_concurrent_event && other) noexcept
-    : join_events(std::move(other.join_events))
-    , fork_event(other.fork_event)
-    , n_streams(other.n_streams)
-    , stream_mapping(std::move(other.stream_mapping))
-    , original_order(std::move(other.original_order))
-    , join_node(other.join_node) {
-        other.fork_event = nullptr;
-    }
-
-    // 1. check if any branches write to overlapping memory ranges (except the join node)
-    // 2. check whether all srcs are either within the branch or outside the nodes covered by ggml_cuda_concurrent_event
-    // we assume all nodes have the same buffer
-    bool is_valid() const {
-        std::vector<std::vector<std::pair<int64_t, int64_t>>> write_ranges;
-        write_ranges.resize(n_streams);
-
-        // get join_node's memory range to exclude from overlap checking.
-        // multiple nodes can use join_node's buffer; we synchronize on the join node.
-        const ggml_tensor * join_t     = join_node->view_src ? join_node->view_src : join_node;
-        const int64_t       join_start = (int64_t) join_t->data;
-        const int64_t       join_end   = join_start + ggml_nbytes(join_t);
-
-        for (const auto & [tensor, stream] : stream_mapping) {
-            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
-            const int64_t       t_start = (int64_t) t->data;
-            const int64_t       t_end   = t_start + ggml_nbytes(t);
-
-            // skip tensors that overlap with join_node's buffer.
-            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
-                continue;
-            }
-
-            // concurrent streams begin from 1
-            write_ranges[stream - 1].emplace_back(t_start, t_end);
-        }
-
-        for (int i = 0; i < n_streams; ++i) {
-            // sorts first by start then by end of write range
-            std::sort(write_ranges[i].begin(), write_ranges[i].end());
-        }
-
-        bool writes_overlap = false;
-        bool dependent_srcs = false;
-        for (const auto & [tensor, stream] : stream_mapping) {
-            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
-            const int64_t       t_start = (int64_t) t->data;
-            const int64_t       t_end   = t_start + ggml_nbytes(t);
-
-            // skip tensors that overlap with join_node's buffer
-            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
-                continue;
-            }
-
-            // check if this buffer's write data overlaps with another stream's
-            std::pair<int64_t, int64_t> data_range = std::make_pair(t_start, t_end);
-            for (int i = 0; i < n_streams; ++i) {
-                if (i == stream - 1) {
-                    continue;
-                }
-                auto it = std::lower_bound(write_ranges[i].begin(), write_ranges[i].end(), data_range);
-
-                if (it != write_ranges[i].end()) {
-                    const std::pair<int64_t, int64_t> & other = *it;
-
-                    // std::lower_bound returns the first element where other >= data_range (lexicographically).
-                    // This guarantees other.first >= data_range.first.
-                    // Therefore, overlap occurs iff other.first < data_range.second
-                    // (i.e., the other range starts before this range ends).
-                    if (other.first < data_range.second) {
-                        GGML_LOG_DEBUG("Writes overlap for %s", tensor->name);
-                        writes_overlap = true;
-                        break;
-                    }
-                }
-            }
-
-            //check if all srcs are either in branch or don't have a branch
-            for (int i = 0; i < GGML_MAX_SRC; ++i) {
-                if (!tensor->src[i]) {
-                    continue;
-                }
-
-                auto it = stream_mapping.find(tensor->src[i]);
-
-                if (it == stream_mapping.end()) {
-                    continue;
-                }
-
-                if (it->second != stream) {
-                    dependent_srcs = true;
-                    break;
-                }
-            }
-
-            if (dependent_srcs || writes_overlap) {
-                break;
-            }
-        }
-
-        return !writes_overlap && !dependent_srcs;
-    }
-
-    ~ggml_cuda_concurrent_event() {
-        if (fork_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(fork_event));
-        }
-        for (cudaEvent_t e : join_events) {
-            if (e != nullptr) {
-                CUDA_CHECK(cudaEventDestroy(e));
-            }
-        }
-    }
-};
-
-struct ggml_cuda_stream_context {
-    std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
-
-    void reset() {
-        concurrent_events.clear();
-    }
-};
+// defined in graph.cuh
+struct ggml_cuda_graph;
+struct ggml_cuda_stream_context;

 struct ggml_backend_cuda_context {
    int device;
@@ -1461,6 +1348,9 @@ struct ggml_backend_cuda_context {
    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {};
    void * cublas_workspaces[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {};
    size_t cublas_workspace_sizes[GGML_CUDA_MAX_DEVICES] = {};
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+    cublasLtHandle_t cublaslt_handles[GGML_CUDA_MAX_DEVICES] = {};
+#endif

    int curr_stream_no = 0;

@@ -1472,61 +1362,20 @@ struct ggml_backend_cuda_context {
    int64_t next_graph_eviction_time = 0;
    int64_t last_graph_use_time = 0;

-    ggml_cuda_graph * cuda_graph(const void * first_node_ptr, int64_t time_now = ggml_time_us()) {
-        // Give all cached graphs 10s to be reused after an idle period, without updating each graph.
-        if (!cuda_graphs.empty() && time_now - last_graph_use_time >= 10'000'000) {
-            next_graph_eviction_time = time_now + 10'000'000;
-        }
-        last_graph_use_time = time_now;
-
-        // Sweep every 5s during active use, evicting graphs unused for at least 10s.
-        if (time_now >= next_graph_eviction_time) {
-            next_graph_eviction_time = time_now + 5'000'000;
-            for (auto it = cuda_graphs.begin(); it != cuda_graphs.end(); ) {
-                if (it->first != first_node_ptr && time_now - it->second->last_used_time >= 10'000'000) {
-                    it = cuda_graphs.erase(it);
-                } else {
-                    ++it;
-                }
-            }
-        }
-
-        auto it = cuda_graphs.find(first_node_ptr);
-        if (it == cuda_graphs.end()) {
-            it = cuda_graphs.emplace(first_node_ptr, std::make_unique<ggml_cuda_graph>()).first;
-        }
-        it->second->last_used_time = time_now;
-        return it->second.get();
-    }
+    // defined in ggml-cuda.cu, which includes graph.cuh
+    ggml_cuda_graph * cuda_graph(const void * first_node_ptr, int64_t time_now = ggml_time_us());

    // Check if any CUDA graph is enabled for this context (used by kernels that need to know
    // if graphs are in use without having access to the specific graph key)
-    bool any_cuda_graph_enabled() const {
-        for (const auto & [key, graph] : cuda_graphs) {
-            if (graph && graph->is_enabled()) {
-                return true;
-            }
-        }
-        return false;
-    }
+    bool any_cuda_graph_enabled() const;

    // Check if any CUDA graph has an instance for this context
-    bool any_cuda_graph_has_instance() const {
-        for (const auto & [key, graph] : cuda_graphs) {
-            if (graph && !graph->instances.empty()) {
-                return true;
-            }
-        }
-        return false;
-    }
+    bool any_cuda_graph_has_instance() const;
 #endif // USE_CUDA_GRAPH

-    explicit ggml_backend_cuda_context(int device) :
-        device(device),
-        name(GGML_CUDA_NAME + std::to_string(device)) {
-    }
+    explicit ggml_backend_cuda_context(int device);

-    ggml_cuda_stream_context concurrent_stream_context;
+    std::unique_ptr<ggml_cuda_stream_context> concurrent_stream_context;

    ~ggml_backend_cuda_context();

@@ -1540,7 +1389,7 @@ struct ggml_backend_cuda_context {

    cudaStream_t stream() { return stream(device, curr_stream_no); }

-    ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }
+    ggml_cuda_stream_context & stream_context() { return *concurrent_stream_context; }

    cublasHandle_t cublas_handle() {
        if (cublas_handles[device][curr_stream_no] == nullptr) {
@@ -1548,10 +1397,14 @@ struct ggml_backend_cuda_context {
            CUBLAS_CHECK(cublasCreate(&cublas_handles[device][curr_stream_no]));
            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device][curr_stream_no], CUBLAS_TF32_TENSOR_OP_MATH));
            CUBLAS_CHECK(cublasSetStream(cublas_handles[device][curr_stream_no], stream()));
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUBLAS_VER_MAJOR > 11 || (CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR >= 2))
+#if (defined(GGML_USE_HIP) && HIP_VERSION >= 70000000) || (!defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUBLAS_VER_MAJOR > 11 || (CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR >= 2)))
            if (cublas_workspace_sizes[device] == 0) {
+#if defined(GGML_USE_HIP)
+                cublas_workspace_sizes[device] = 32 * 1024 * 1024;
+#else
                const int cc = ggml_cuda_info().devices[device].cc;
                cublas_workspace_sizes[device] = (cc >= GGML_CUDA_CC_HOPPER) ? 32 * 1024 * 1024 : 4 * 1024 * 1024;
+#endif
            }
            CUDA_CHECK(cudaMalloc(&cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
            CUBLAS_CHECK(cublasSetWorkspace(cublas_handles[device][curr_stream_no], cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
@@ -1560,6 +1413,16 @@ struct ggml_backend_cuda_context {
        return cublas_handles[device][curr_stream_no];
    }

+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+    cublasLtHandle_t cublaslt_handle() {
+        if (cublaslt_handles[device] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasLtCreate(&cublaslt_handles[device]));
+        }
+        return cublaslt_handles[device];
+    }
+#endif
+
    // pool
    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];

@@ -1644,60 +1507,8 @@ struct ggml_cuda_pdl_config {

 };

-static bool ggml_cuda_kernel_can_use_pdl(const void * kernel) {
-    const int device = ggml_cuda_get_device();
-
-    struct cache_key {
-        int          device;
-        const void * kernel;
-
-        bool operator==(const cache_key & other) const { return device == other.device && kernel == other.kernel; }
-    };
-
-    struct cache_key_hash {
-        // MurmurHash3 mixing function for better hash distribution (vs. just std::hash which in some implementations simply returns the identity)
-        static size_t hash_mix(size_t x) {
-            std::uint64_t       y = x;
-            const std::uint64_t m = 0xe9846af9b1a615d;
-
-            y ^= y >> 32;
-            y *= m;
-            y ^= y >> 32;
-            y *= m;
-            y ^= y >> 28;
-
-            return static_cast<size_t>(y);
-        }
-
-        size_t operator()(const cache_key & key) const {
-            // Use a nonzero seed to avoid mapping all-zero keys to zero
-            size_t h = 42;
-            h        = hash_mix(h + key.device);
-            h        = hash_mix(h + reinterpret_cast<size_t>(key.kernel));
-            return h;
-        }
-    };
-
-    static std::mutex                                          cache_mutex;
-    static std::unordered_map<cache_key, bool, cache_key_hash> cache;
-
-    const cache_key             key = { device, kernel };
-    std::lock_guard<std::mutex> lock(cache_mutex);
-    const auto                  it = cache.find(key);
-    if (it != cache.end()) {
-        return it->second;
-    }
-
-    cudaFuncAttributes attr = {};
-    CUDA_CHECK(cudaFuncGetAttributes(&attr, kernel));
-
-    // PDL device-side primitives are emitted only for PTX versions >= 90.
-    // We have to guard on a loaded kernel's PTX version so a kernel forward-JIT'ed
-    // from pre-Hopper PTX to a Hopper-or-newer GPU does not opt into PDL.
-    const bool can_use_pdl = attr.ptxVersion >= 90;
-    cache.emplace(key, can_use_pdl);
-    return can_use_pdl;
-}
+// whether kernel can be launched with PDL on the current device (also honors GGML_CUDA_PDL=0), defined in ggml-cuda.cu
+bool ggml_cuda_kernel_can_use_pdl(const void * kernel);

 #endif //defined(GGML_CUDA_USE_PDL)

@@ -1711,13 +1522,7 @@ static bool ggml_cuda_kernel_can_use_pdl(const void * kernel) {
 template<typename Kernel, typename... Args>
 static __inline__ void ggml_cuda_kernel_launch(Kernel kernel, const ggml_cuda_kernel_launch_params & launch_params, Args&&... args) {
 #if defined(GGML_CUDA_USE_PDL)
-
-    static const bool env_pdl_enabled = []() {
-        const char * env = getenv("GGML_CUDA_PDL");
-        return env == nullptr || std::atoi(env) != 0;
-    }();
-
-    if (env_pdl_enabled && ggml_cuda_kernel_can_use_pdl(reinterpret_cast<const void *>(kernel))) {
+    if (ggml_cuda_kernel_can_use_pdl(reinterpret_cast<const void *>(kernel))) {
        auto pdl_cfg = ggml_cuda_pdl_config(launch_params);

        CUDA_CHECK(cudaLaunchKernelEx(&pdl_cfg.cfg, kernel, std::forward<Args>(args)... ));
@@ -541,6 +541,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
            return dequantize_row_mxfp4_cuda;
        case GGML_TYPE_NVFP4:
            return dequantize_row_nvfp4_cuda;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cont_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_F32:
            return convert_unary_cont_cuda<float>;
        case GGML_TYPE_F16:
@@ -601,6 +603,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
            return dequantize_row_mxfp4_cuda;
        case GGML_TYPE_NVFP4:
            return dequantize_row_nvfp4_cuda;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cont_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_F32:
            return convert_unary_cont_cuda<float>;
        case GGML_TYPE_BF16:
@@ -658,6 +662,8 @@ to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
            return dequantize_row_mxfp4_cuda;
        case GGML_TYPE_NVFP4:
            return dequantize_row_nvfp4_cuda;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cont_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_F16:
            return convert_unary_cont_cuda<half>;
        case GGML_TYPE_BF16:
@@ -685,6 +691,8 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
        case GGML_TYPE_Q8_0:
            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_BF16:
            return convert_unary_cuda<nv_bfloat16>;
        default:
@@ -710,6 +718,8 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
        case GGML_TYPE_Q8_0:
            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_F16:
            return convert_unary_cuda<half, nv_bfloat16>;
        default:
@@ -735,6 +745,8 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
        case GGML_TYPE_Q8_0:
            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_F8_E4M3:
+            return convert_unary_cuda<ggml_fp8_e4m3_t>;
        case GGML_TYPE_BF16:
            return convert_unary_cuda<nv_bfloat16, float>;
        default:
@@ -35,6 +35,36 @@ template<typename dst_t, typename src_t>
 __host__ __device__ inline dst_t ggml_cuda_cast(src_t x) {
    if constexpr (std::is_same_v<dst_t, src_t>) {
        return x;
+    } else if constexpr (std::is_same_v<dst_t, ggml_fp8_e4m3_t>) {
+        return { ggml_cuda_fp32_to_f8_e4m3(ggml_cuda_cast<float>(x)) };
+    } else if constexpr (std::is_same_v<src_t, ggml_fp8_e4m3_t>) {
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if CUDART_VERSION >= 13040
+        if constexpr (std::is_same_v<dst_t, nv_bfloat16>) {
+            return static_cast<nv_bfloat16>(__nv_cvt_fp8_to_bf16raw(x.bits, __NV_E4M3));
+        }
+#endif
+        if constexpr (std::is_same_v<dst_t, half>) {
+            return static_cast<half>(__nv_cvt_fp8_to_halfraw(x.bits, __NV_E4M3));
+        }
+#endif
+        return ggml_cuda_cast<dst_t>(ggml_cuda_f8_e4m3_to_fp32(x.bits));
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    } else if constexpr (std::is_same_v<src_t, __nv_fp8x2_e4m3> && std::is_same_v<dst_t, nv_bfloat162>) {
+#if CUDART_VERSION >= 13040
+        return static_cast<nv_bfloat162>(__nv_cvt_fp8x2_to_bf162raw(x.__x, __NV_E4M3));
+#else
+        return make_bfloat162(
+            ggml_cuda_cast<nv_bfloat16>(ggml_cuda_f8_e4m3_to_fp32(x.__x & 0xFF)),
+            ggml_cuda_cast<nv_bfloat16>(ggml_cuda_f8_e4m3_to_fp32(x.__x >> 8)));
+#endif
+    } else if constexpr (std::is_same_v<src_t, __nv_fp8x2_e4m3> && std::is_same_v<dst_t, float2>) {
+#if CUDART_VERSION >= 13040
+        return ggml_cuda_cast<float2>(ggml_cuda_cast<nv_bfloat162>(x));
+#else
+        return __half22float2(static_cast<half2>(__nv_cvt_fp8x2_to_halfraw2(x.__x, __NV_E4M3)));
+#endif
+#endif
    } else if constexpr(std::is_same_v<dst_t, nv_bfloat16>) {
        return __float2bfloat16(float(x));
    } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
@@ -477,6 +477,41 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
    } else if (ggml_cuda_cpy_as_memcpy_2d(src0, src1, mc_width, mc_height, mc_spitch, mc_dpitch)) {
        CUDA_CHECK(cudaMemcpy2DAsync(src1_ddc, mc_dpitch, src0_ddc, mc_spitch,
                                     mc_width, mc_height, cudaMemcpyDeviceToDevice, main_stream));
+    } else if (src0->type == GGML_TYPE_F8_E4M3 && src1->type == GGML_TYPE_F8_E4M3) {
+        ggml_cpy_scalar_cuda<uint8_t, uint8_t>
+            (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F8_E4M3) {
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<float, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<float, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F8_E4M3) {
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<half, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<half, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
+    } else if (src0->type == GGML_TYPE_BF16 && src1->type == GGML_TYPE_F8_E4M3) {
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<nv_bfloat16, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<nv_bfloat16, ggml_fp8_e4m3_t>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
+    } else if (src0->type == GGML_TYPE_F8_E4M3 && src1->type == GGML_TYPE_F32) {
+        if (contiguous_srcs) {
+            ggml_cpy_scalar_contiguous_cuda<ggml_fp8_e4m3_t, float>
+                (src0_ddc, src1_ddc, ne, main_stream);
+        } else {
+            ggml_cpy_scalar_cuda<ggml_fp8_e4m3_t, float>
+                (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+        }
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
        if (can_be_transposed) {
            ggml_cpy_scalar_cuda<float, float, true>
@@ -146,6 +146,34 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_bf16(
    return sum;
 }

+template <int D, int nthreads>
+static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f8_e4m3(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const uint8_t * K_f8 = (const uint8_t *) K_c;
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += nthreads*4) {
+        uint8_t tmp[4];
+        ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_f8 + k_KQ_0 + (threadIdx.x % nthreads)*4);
+#pragma unroll
+        for (int k_KQ_1 = 0; k_KQ_1 < 4; ++k_KQ_1) {
+            const float k = ggml_cuda_f8_e4m3_to_fp32(tmp[k_KQ_1]);
+#ifdef V_DOT2_F32_F16_AVAILABLE
+            sum += k * __half2float(((const half *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
+#else
+            sum += k * ((const float *) Q_v)[k_KQ_0/nthreads + k_KQ_1];
+#endif // V_DOT2_F32_F16_AVAILABLE
+        }
+    }
+
+    return sum;
+}
+
 template<int D, int nthreads>
 static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_q4_0(
    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
@@ -406,6 +434,19 @@ static __device__ __forceinline__ void dequantize_V_bf16(const void * __restrict
    }
 }

+template <typename T, int ne>
+static __device__ __forceinline__ void dequantize_V_f8_e4m3(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
+    static_assert(ne == 2 || ne == 4, "bad ne");
+    static_assert(std::is_same_v<T, half> || std::is_same_v<T, float>, "unsupported type");
+    ggml_fp8_e4m3_t tmp[ne];
+    ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, (const ggml_fp8_e4m3_t *) vx + i0);
+
+#pragma unroll
+    for (int l = 0; l < ne; ++l) {
+        ((T *) dst)[l] = ggml_cuda_cast<T>(tmp[l]);
+    }
+}
+
 template <typename T, int ne>
 static __device__ __forceinline__ void dequantize_V_q4_0(const void * __restrict__ vx, void * __restrict__ dst, const int64_t i0) {
    const block_q4_0 * x = (const block_q4_0 *) vx;
@@ -644,6 +685,8 @@ constexpr __device__ vec_dot_KQ_t get_vec_dot_KQ() {
        return vec_dot_fattn_vec_KQ_q8_0<D, nthreads>;
    } else if constexpr (type_K == GGML_TYPE_BF16) {
        return vec_dot_fattn_vec_KQ_bf16<D, nthreads>;
+    } else if constexpr (type_K == GGML_TYPE_F8_E4M3) {
+        return vec_dot_fattn_vec_KQ_f8_e4m3<D, nthreads>;
    } else {
        static_assert(type_K == -1, "bad type");
        return nullptr;
@@ -666,6 +709,8 @@ constexpr __device__ dequantize_V_t get_dequantize_V() {
        return dequantize_V_q8_0<T, ne>;
    } else if constexpr (type_V == GGML_TYPE_BF16) {
        return dequantize_V_bf16<float, ne>;
+    } else if constexpr (type_V == GGML_TYPE_F8_E4M3) {
+        return dequantize_V_f8_e4m3<T, ne>;
    } else {
        static_assert(type_V == -1, "bad type");
        return nullptr;
@@ -36,11 +36,21 @@ struct fattn_mma_config {
    }                                                                                                                                                              \

 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_ampere(const int DKQ, const int DV, const int ncols) {
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40,  8, 128, 2, 128,  24,  24,  24, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 16, 128, 2,  64,  24,  24,  24, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 32, 128, 2,  64,  24,  24,  24, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 64, 128, 2,  64,  24,  24,  24, 2, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64,  8, 128, 2, 128,  32,  32,  32, 2, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 16, 128, 2,  64,  32,  32,  32, 2, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 32, 128, 2,  64,  32,  32,  32, 2, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 64, 128, 2,  64,  32,  32,  32, 2, true);

+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72,  8, 128, 2, 128,  40,  40,  40, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 16, 128, 2,  64,  40,  40,  40, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 32, 128, 2,  64,  40,  40,  40, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 64, 128, 2,  64,  40,  40,  40, 2, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80,  8, 128, 2, 128,  40,  40,  40, 2, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 16, 128, 2,  64,  40,  40,  40, 2, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 32, 128, 2,  64,  40,  40,  40, 2, true);
@@ -123,16 +133,29 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128,  64, 1, false);

+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2, 32, 128, 128, 64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2, 32, 128, 64, 128, 2, false);
+
    // TODO tune specifically for Volta
    return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
 }

 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_rdna(const int DKQ, const int DV, const int ncols) {
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40,  8, 128, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 16, 128, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 32, 128, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 64, 128, 2,  64,  24,  24,  24, 1, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64,  8, 128, 2,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 16, 128, 2,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 32, 128, 2,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 64, 128, 2,  64,  32,  32,  32, 1, true);

+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72,  8,  64, 2,  32,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 16,  64, 2,  32,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 32, 128, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 64, 128, 2,  64,  40,  40,  40, 1, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80,  8,  64, 2,  32,  40,  40,  40, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 16,  64, 2,  32,  40,  40,  40, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 32, 128, 2,  64,  40,  40,  40, 1, true);
@@ -180,11 +203,21 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
 }

 static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_cdna(const int DKQ, const int DV, const int ncols) {
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40,  8, 256, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 16, 256, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 32, 256, 2,  64,  24,  24,  24, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 40,  40, 64, 256, 2,  64,  24,  24,  24, 1, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64,  8, 128, 1,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 16, 256, 2,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 32, 256, 2,  64,  32,  32,  32, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64,  64, 64, 256, 3,  64,  32,  32,  32, 1, true);

+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72,  8, 256, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 16, 256, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 32, 256, 2,  64,  40,  40,  40, 1, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 72,  72, 64, 256, 2,  64,  40,  40,  40, 1, true);
+
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80,  8, 256, 2,  64,  40,  40,  40, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 16, 256, 2,  64,  40,  40,  40, 1, true);
    GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80,  80, 32, 256, 2,  64,  40,  40,  40, 1, true);
@@ -619,6 +652,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    constexpr bool Q_in_reg        = ggml_cuda_fattn_mma_get_Q_in_reg (DKQ, DV, ncols);
    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages  (DKQ, DV, ncols1, ncols2, use_sparse);

+    constexpr int DKQp = GGML_PAD(DKQ, 16);
+    constexpr int DVp  = GGML_PAD(DV, 16);
+
    constexpr bool swz = ggml_cuda_fattn_mma_get_swizzled(DKQ, DV, ncols);
    constexpr int stride_tile_K = swz ? nbatch_K2 : nbatch_K2 + 4;
    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : (swz ? nbatch_V2 : nbatch_V2 + 4);
@@ -635,12 +671,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    if constexpr (nstages > 1) {
        static_assert(!oob_check, "OOB check incompatible with multi-stage pipeline");
        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
-        static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi stage loading");
+        static_assert(nbatch_K2 == DKQp/2, "batching not implemented for multi stage loading");
        constexpr bool use_cp_async = true;
        cp_async_wait_all();
        __syncthreads();
        flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-            (V_h2, tile_V, nbatch_V2, stride_V, k_VKQ_0, k_VKQ_sup, indices);
+            (V_h2, tile_V, min(nbatch_V2, DV/2), stride_V, k_VKQ_0, k_VKQ_sup, indices);
    } else {
        // the sparse mask values are gathered per element, always load them synchronously
        constexpr bool use_cp_async = nstages == 1 && !use_sparse;
@@ -653,11 +689,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    // For MLA K and V have the same data.
    // Therefore, iterate over K in reverse and later re-use the data if possible.
 #pragma unroll
-    for (int k0_start = (DKQ/2-1) - (DKQ/2-1) % nbatch_K2; k0_start >= 0; k0_start -= nbatch_K2) {
-        const int k0_stop = k0_start + nbatch_K2 < DKQ/2 ? k0_start + nbatch_K2 : DKQ/2;
+    for (int k0_start = (DKQp/2-1) - (DKQp/2-1) % nbatch_K2; k0_start >= 0; k0_start -= nbatch_K2) {
+        const int k0_stop = k0_start + nbatch_K2 < DKQp/2 ? k0_start + nbatch_K2 : DKQp/2;

        if constexpr (nstages <= 1) {
-            const int k0_diff = k0_stop - k0_start;
+            const int k0_diff = min(k0_stop, DKQ/2) - k0_start; // Only load the real head size from global memory.
            constexpr bool use_cp_async = nstages == 1;
            flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
                (K_h2 + k0_start, tile_K, k0_diff, stride_K, k_VKQ_0, k_VKQ_sup, indices);
@@ -691,7 +727,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                }
            }
        } else {
-            constexpr int stride_tile_Q = DKQ/2 + 4;
+            constexpr int stride_tile_Q = DKQp/2 + 4;
 #pragma unroll
            for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -918,7 +954,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        if constexpr (cols_per_warp == 8) {
            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
 #pragma unroll
-            for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
+            for (int i = 0; i < DVp/T_C_VKQ::I; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -929,7 +965,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
            for (int col = 0; col < cols_per_thread; ++col) {
                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
 #pragma unroll
-                for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+                for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
                    for (int l0 = 0; l0 < T_C_VKQ::ne; l0 += 2) {
                        VKQ_C[i].x[l0 + col] *= KQ_max_scale_h2;
@@ -941,7 +977,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        if constexpr (std::is_same_v<decltype(T_C_VKQ::x), half2[T_C_VKQ::ne]>) {
            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[0]);
 #pragma unroll
-            for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+            for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -950,7 +986,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        } else {
            static_assert(std::is_same_v<decltype(T_C_VKQ::x), float[T_C_VKQ::ne]>, "bad VKQ type");
 #pragma unroll
-            for (int i = 0; i < DV/T_C_VKQ::J; ++i) {
+            for (int i = 0; i < DVp/T_C_VKQ::J; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale[0];
@@ -961,7 +997,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        const half2 KQ_max_scale_h2 = make_half2(
            KQ_max_scale[(threadIdx.x / 2) % 2], KQ_max_scale[(threadIdx.x / 2) % 2]);
 #pragma unroll
-        for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+        for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
            for (int l = 0; l < T_C_VKQ::ne; ++l) {
                VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -997,19 +1033,19 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                    (mask_h, tile_mask, stride_mask, k_VKQ_0 + nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
            }
            flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-                (K_h2, tile_K, nbatch_K2, stride_K, k_VKQ_0 + nbatch_fa, k_VKQ_sup, indices);
+                (K_h2, tile_K, min(nbatch_K2, DKQ/2), stride_K, k_VKQ_0 + nbatch_fa, k_VKQ_sup, indices);
        }
    }


    // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
-    for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
-        static_assert(DV % (2*nbatch_V2) == 0, "bad loop size");
+    for (int i0_start = 0; i0_start < DVp; i0_start += 2*nbatch_V2) {
+        static_assert(DVp % (2*nbatch_V2) == 0, "bad loop size");
        const int i0_stop = i0_start + 2*nbatch_V2;

        if constexpr (nstages <= 1) {
-            const int i0_diff = i0_stop - i0_start;
+            const int i0_diff = min(i0_stop, DV) - i0_start; // Only load the real head size from global memory.
            if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
                constexpr bool use_cp_async = nstages == 1;
                flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
@@ -1146,6 +1182,22 @@ template<int ncols> struct mma_tile_sizes<112, ncols> {
    using T_B_VKQ = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // column-major
    using T_C_VKQ = tile<16, 16, float, DATA_LAYOUT_I_MAJOR>;          // column-major
 };
+template<int ncols> struct mma_tile_sizes<40, ncols> {
+    using T_A_KQ  = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
+    using T_B_KQ  = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // column-major
+    using T_C_KQ  = tile<16, 16, float, DATA_LAYOUT_I_MAJOR>;          // column-major
+    using T_A_VKQ = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
+    using T_B_VKQ = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // column-major
+    using T_C_VKQ = tile<16, 16, float, DATA_LAYOUT_I_MAJOR>;          // column-major
+};
+template<int ncols> struct mma_tile_sizes<72, ncols> {
+    using T_A_KQ  = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
+    using T_B_KQ  = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // column-major
+    using T_C_KQ  = tile<16, 16, float, DATA_LAYOUT_I_MAJOR>;          // column-major
+    using T_A_VKQ = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
+    using T_B_VKQ = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // column-major
+    using T_C_VKQ = tile<16, 16, float, DATA_LAYOUT_I_MAJOR>;          // column-major
+};
 #else
 template<int DKQ, int ncols> struct mma_tile_sizes {
    using T_A_KQ  = tile<16,  8, half2, DATA_LAYOUT_I_MAJOR>;           // row-major
@@ -1171,6 +1223,22 @@ template<int ncols> struct mma_tile_sizes<112, ncols> {
    using T_B_VKQ = tile<16,  8, half2>; // column-major
    using T_C_VKQ = tile<16,  8, half2>; // column-major
 };
+template<int ncols> struct mma_tile_sizes<40, ncols> {
+    using T_A_KQ  = tile<16,  8, half2>; // row-major
+    using T_B_KQ  = tile<16,  8, half2>; // column-major
+    using T_C_KQ  = tile<16, 16, float>; // column-major
+    using T_A_VKQ = tile<16,  8, half2>; // row-major
+    using T_B_VKQ = tile<16,  8, half2>; // column-major
+    using T_C_VKQ = tile<16,  8, half2>; // column-major
+};
+template<int ncols> struct mma_tile_sizes<72, ncols> {
+    using T_A_KQ  = tile<16,  8, half2>; // row-major
+    using T_B_KQ  = tile<16,  8, half2>; // column-major
+    using T_C_KQ  = tile<16, 16, float>; // column-major
+    using T_A_VKQ = tile<16,  8, half2>; // row-major
+    using T_B_VKQ = tile<16,  8, half2>; // column-major
+    using T_C_VKQ = tile<16,  8, half2>; // column-major
+};
 #endif // RDNA3
 #elif defined(AMD_MFMA_AVAILABLE)
 template<int DKQ, int ncols> struct mma_tile_sizes {
@@ -1240,6 +1308,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    constexpr bool Q_in_reg        = ggml_cuda_fattn_mma_get_Q_in_reg      (DKQ, DV, ncols);
    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages       (DKQ, DV, ncols1, ncols2, use_sparse);

+    constexpr int DKQp = GGML_PAD(DKQ, 16);
+    constexpr int DVp  = GGML_PAD(DV, 16);
+
    if (cols_per_warp > ncols) {
        NO_DEVICE_CODE;
        return;
@@ -1248,7 +1319,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    static_assert(nwarps * (cols_per_warp/ncols2) % ncols1 == 0, "bad nwarps");

    constexpr bool swz = ggml_cuda_fattn_mma_get_swizzled(DKQ, DV, ncols);
-    constexpr int stride_tile_Q = DKQ/2 + 4;
+    constexpr int stride_tile_Q = DKQp/2 + 4;
    constexpr int stride_tile_K = swz ? nbatch_K2 : nbatch_K2 + 4;
    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : (swz ? nbatch_V2 : nbatch_V2 + 4);
    constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V;
@@ -1258,17 +1329,17 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    half2 * tile_V    =           nstages > 1 ? tile_K + nbatch_fa * stride_tile_K : tile_K;
    half  * tile_mask = (half *) (nstages > 1 ? tile_V + nbatch_fa * stride_tile_V : tile_V + nbatch_fa * stride_tile_KV_max);

-    T_B_KQ    Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
+    T_B_KQ    Q_B[(Q_in_reg ? DKQp/(2*T_B_KQ::J) : 1)];
 #if defined(TURING_MMA_AVAILABLE)
-    T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
+    T_C_VKQ VKQ_C[cols_per_warp == 8 ? DVp/T_C_VKQ::I : DVp/(2*T_C_VKQ::J)];
 #elif defined(AMD_WMMA_256B_AVAILABLE)
-    T_C_VKQ VKQ_C[DV % 32 != 0       ? DV/T_C_VKQ::J : DV/(2*T_C_VKQ::J)];
+    T_C_VKQ VKQ_C[DVp % 32 != 0      ? DVp/T_C_VKQ::J : DVp/(2*T_C_VKQ::J)];
 #elif defined(AMD_MFMA_AVAILABLE)
-    T_C_VKQ VKQ_C[                                     DV/T_C_VKQ::J];
+    T_C_VKQ VKQ_C[                                      DVp/T_C_VKQ::J];
 #elif defined(AMD_WMMA_AVAILABLE)
-    T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
+    T_C_VKQ VKQ_C[                                      DVp/(2*T_C_VKQ::J)];
 #else // Volta
-    T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
+    T_C_VKQ VKQ_C[                                      DVp/(2*T_C_VKQ::J)];
 #endif // defined(TURING_MMA_AVAILABLE)

    float KQ_rowsum[cols_per_thread] = {0.0f};
@@ -1284,8 +1355,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    const half2 scale_h2 = make_half2(scale, scale);
 #pragma unroll
    for (int stride_k : {warp_size, warp_size/2, warp_size/4, warp_size/8}) {
-        const int k0_start  = stride_k == warp_size ? 0 : DKQ/2 - (DKQ/2) % (2*stride_k);
-        const int k0_stop   =                             DKQ/2 - (DKQ/2) % (1*stride_k);
+        const int k0_start  = stride_k == warp_size ? 0 : DKQp/2 - (DKQp/2) % (2*stride_k);
+        const int k0_stop   =                             DKQp/2 - (DKQp/2) % (1*stride_k);
        const int stride_jc = warp_size / stride_k;

        if (k0_start == k0_stop) {
@@ -1308,7 +1379,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);

-                    const float2 tmp = Q_f2[(jt*ncols1 + j)*stride_Q1 + c*stride_Q2 + k];
+                    const float2 tmp = DKQp == DKQ || k < DKQ/2 ?
+                        Q_f2[(jt*ncols1 + j)*stride_Q1 + c*stride_Q2 + k] : make_float2(0.0f, 0.0f);
                    tile_Q[jc*stride_tile_Q + k] = scale_h2 * make_half2(tmp.x, tmp.y);
                }
            } else {
@@ -1328,19 +1400,50 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        const int j0 = (threadIdx.y / np) * cols_per_warp;

 #pragma unroll
-        for (int k0 = 0; k0 < DKQ/2; k0 += T_B_KQ::J) {
+        for (int k0 = 0; k0 < DKQp/2; k0 += T_B_KQ::J) {
            load_ldmatrix(Q_B[k0/T_B_KQ::J], tile_Q + j0*stride_tile_Q + k0, stride_tile_Q);
        }
    }

    __syncthreads();

+    // Zero the pad columns of the K/V tiles once, the tile loads only write the first DKQ/2 (DV/2) half2 of a row.
+    // Done after Q_B is loaded because tile_K may alias tile_Q.
+    if constexpr (DKQp != DKQ || DVp != DV) {
+        static_assert(!swz, "pad zero-fill assumes unswizzled tiles");
+        static_assert(nbatch_K2 == DKQp/2 && nbatch_V2 == DVp/2, "K/V tile batching not implemented for padded head sizes");
+        static_assert(DKQ == DV, "pad zero-fill for DKQ != DV requires separate handling of the aliased V tile");
+
+        const half2 zero[4] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
+        const int tid = threadIdx.y*warp_size + threadIdx.x;
+#pragma unroll
+        for (int i0 = 0; i0 < nbatch_fa; i0 += nwarps*warp_size) {
+            const int i = i0 + tid;
+
+            if (i0 + nwarps*warp_size > nbatch_fa && i >= nbatch_fa) {
+                break;
+            }
+
+#pragma unroll
+            for (int k = DKQ/2; k < DKQp/2; k += 4) {
+                ggml_cuda_memcpy_1<16>(tile_K + i*stride_tile_K + k, zero);
+            }
+            if constexpr (nstages > 1) { // tile_V is only a separate buffer with multi-stage loading.
+#pragma unroll
+                for (int k = DV/2; k < DVp/2; k += 4) {
+                    ggml_cuda_memcpy_1<16>(tile_V + i*stride_tile_V + k, zero);
+                }
+            }
+        }
+        __syncthreads();
+    }
+
    int kb0 = kb0_start;

    // Preload mask and K data for first iteration when using cp_async with multiple stages:
    if constexpr (nstages > 1) {
        static_assert(!use_sparse, "sparse gather not implemented for multi-stage loading");
-        static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi-stage pipeline");
+        static_assert(nbatch_K2 == DKQp/2, "batching not implemented for multi-stage pipeline");
        constexpr bool use_cp_async = true;
        constexpr bool oob_check    = false;
        constexpr int  k_VKQ_sup    = nbatch_fa;
@@ -1349,7 +1452,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                (mask_h, tile_mask, stride_mask, kb0*nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
        }
        flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
-            (K_h2, tile_K, nbatch_K2, stride_K, kb0*nbatch_fa, k_VKQ_sup, indices);
+            (K_h2, tile_K, min(nbatch_K2, DKQ/2), stride_K, kb0*nbatch_fa, k_VKQ_sup, indices);
    }

    // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
@@ -1454,7 +1557,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        if constexpr (cols_per_warp == 8) {
            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
 #pragma unroll
-            for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
+            for (int i = 0; i < DVp/T_C_VKQ::I; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -1465,7 +1568,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
            for (int col = 0; col < cols_per_thread; ++col) {
                const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
 #pragma unroll
-                for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+                for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
                    for (int l0 = 0; l0 < T_C_VKQ::ne; l0 += 2) {
                        VKQ_C[i].x[l0 + col] *= KQ_max_scale_h2;
@@ -1477,7 +1580,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        if constexpr (std::is_same_v<decltype(T_C_VKQ::x), half2[T_C_VKQ::ne]>) {
            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[0]);
 #pragma unroll
-            for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+            for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -1486,7 +1589,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        } else {
            static_assert(std::is_same_v<decltype(T_C_VKQ::x), float[T_C_VKQ::ne]>, "bad VKQ type");
 #pragma unroll
-            for (int i = 0; i < DV/T_C_VKQ::J; ++i) {
+            for (int i = 0; i < DVp/T_C_VKQ::J; ++i) {
 #pragma unroll
                for (int l = 0; l < T_C_VKQ::ne; ++l) {
                    VKQ_C[i].x[l] *= KQ_max_scale[0];
@@ -1497,7 +1600,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        const int col = (threadIdx.x / 2) % 2;
        const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
 #pragma unroll
-        for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+        for (int i = 0; i < (DVp/2)/T_C_VKQ::J; ++i) {
 #pragma unroll
            for (int l = 0; l < T_C_VKQ::ne; ++l) {
                VKQ_C[i].x[l] *= KQ_max_scale_h2;
@@ -1511,7 +1614,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    // So also write VKQ accumulators to shared memory in column-major format if np == 1.

    constexpr int tile_stride = nbatch_combine + 4;
-    static_assert((DV/2) % nbatch_combine == 0, "bad nbatch_combine");
+    static_assert((DVp/2) % nbatch_combine == 0, "bad nbatch_combine");

    if constexpr (cols_per_warp == 8) {
        const int jc_cwmo = (threadIdx.x % (2*T_C_VKQ::J)) / T_C_VKQ::J; // jc combine write meta offset
@@ -1655,7 +1758,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    }

 #pragma unroll
-    for (int k00 = 0; k00 < DV/2; k00 += nbatch_combine) {
+    for (int k00 = 0; k00 < DVp/2; k00 += nbatch_combine) {
        if constexpr (cols_per_warp == 8) {
            static_assert(std::is_same_v<decltype(T_C_VKQ::x), half2[T_C_VKQ::ne]>, "bad VKQ type");
            const int jc_cwd = threadIdx.y*T_B_KQ::I + T_B_KQ::get_i(-1); // jc combine write data
@@ -1754,6 +1857,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                    for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                        const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);

+                        if (DVp != DV && k00 + k >= DV/2) {
+                            continue;
+                        }
+
                        float2 dstk_val = make_float2(0.0f, 0.0f);
 #pragma unroll
                        for (int ip = 0; ip < np; ++ip) {
@@ -2034,6 +2141,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    const int nwarps         = nthreads / warp_size_host;

    constexpr bool V_is_K_view = DKQ == 576; // Guaranteed by the kernel selection logic in fattn.cu
+    constexpr int  DKQp        = GGML_PAD(DKQ, 16);

    // KV tile strides must match flash_attn_ext_f16_iter / _process_tile.
    const bool swz = ggml_cuda_fattn_mma_get_swizzled(DKQ, DV, ncols, cc);
@@ -2041,7 +2149,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    const int stride_tile_V = V_is_K_view ? stride_tile_K : (swz ? nbatch_V2 : nbatch_V2 + 4);
    const size_t nbytes_shared_KV_1stage = nbatch_fa            * std::max(stride_tile_K,  stride_tile_V) * sizeof(half2);
    const size_t nbytes_shared_KV_2stage = nbatch_fa            *         (stride_tile_K + stride_tile_V) * sizeof(half2);
-    const size_t nbytes_shared_Q         = ncols                * (DKQ/2 + 4)                             * sizeof(half2);
+    const size_t nbytes_shared_Q         = ncols                * (DKQp/2 + 4)                            * sizeof(half2);
    const size_t nbytes_shared_mask      = ncols1               * (nbatch_fa/2 + 4)                       * sizeof(half2);
    const size_t nbytes_shared_combine   = nwarps*cols_per_warp * (nbatch_combine + 4)                    * sizeof(half2);

@@ -2126,28 +2234,36 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/ 8,  8); \
    extern DECL_FATTN_MMA_F16_CASE(DKQ, DV, (ncols)/16, 16); \

+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 40,  40,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64,   8)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 72,  72,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  80,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  96,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128,   8)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,   8)

+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 40,  40,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64,  16)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 72,  72,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  80,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  96,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128,  16)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  16)

+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 40,  40,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64,  32)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 72,  72,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  80,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  96,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128,  32)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  32)

+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 40,  40,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 64,  64,  64)
+DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 72,  72,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 80,  80,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2( 96,  96,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112,  64)
@@ -96,7 +96,7 @@ static __global__ void flash_attn_ext_vec(
    constexpr int V_cols_per_iter   = WARP_SIZE / nthreads_V;

    constexpr vec_dot_KQ_t vec_dot_KQ = get_vec_dot_KQ<type_K, D, nthreads_KQ>();
-    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16 && type_K != GGML_TYPE_BF16;
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16 && type_K != GGML_TYPE_BF16 && type_K != GGML_TYPE_F8_E4M3;
 #ifdef V_DOT2_F32_F16_AVAILABLE
    constexpr dequantize_V_t dequantize_V = get_dequantize_V<type_V, half,  V_rows_per_thread>();
 #else
@@ -203,6 +203,27 @@ static __global__ void flash_attn_ext_vec(
        }

        __syncthreads();
+    } else if constexpr (type_K == GGML_TYPE_F8_E4M3) {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float * Q_j = (const float *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D; i0 += nthreads_KQ*4) {
+                const int i = i0 + (threadIdx.x % nthreads_KQ)*4;
+                float4 tmp = make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+                if (ncols == 1 || ic0 + j < int(ne01.z)) {
+                    ggml_cuda_memcpy_1<sizeof(tmp)>(&tmp, Q_j + i);
+                }
+#ifdef V_DOT2_F32_F16_AVAILABLE
+                const half2 scale_h2 = make_half2(scale, scale);
+                Q_reg[j][i0/(2*nthreads_KQ) + 0] = __hmul2(make_half2(tmp.x, tmp.y), scale_h2);
+                Q_reg[j][i0/(2*nthreads_KQ) + 1] = __hmul2(make_half2(tmp.z, tmp.w), scale_h2);
+#else
+                Q_reg[j][i0/(2*nthreads_KQ) + 0] = make_float2(tmp.x*scale, tmp.y*scale);
+                Q_reg[j][i0/(2*nthreads_KQ) + 1] = make_float2(tmp.z*scale, tmp.w*scale);
+#endif
+            }
+        }
    } else {
 #ifdef V_DOT2_F32_F16_AVAILABLE
        const half2 scale_h2 = make_half2(scale, scale);
@@ -587,6 +608,7 @@ void ggml_cuda_flash_attn_ext_vec_case(ggml_backend_cuda_context & ctx, ggml_ten
    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q5_1); \
    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_Q8_0); \
    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_BF16); \
+    extern DECL_FATTN_VEC_CASE(D, type_K, GGML_TYPE_F8_E4M3); \

 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q4_0)
@@ -595,6 +617,7 @@ EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_Q8_0)
 EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_BF16)
+EXTERN_DECL_FATTN_VEC_CASES( 64, GGML_TYPE_F8_E4M3)

 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q4_0)
@@ -603,6 +626,7 @@ EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_Q8_0)
 EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_BF16)
+EXTERN_DECL_FATTN_VEC_CASES(128, GGML_TYPE_F8_E4M3)

 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_F16)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q4_0)
@@ -611,3 +635,4 @@ EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_0)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q5_1)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_Q8_0)
 EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_BF16)
+EXTERN_DECL_FATTN_VEC_CASES(256, GGML_TYPE_F8_E4M3)
@@ -314,10 +314,18 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
    const ggml_tensor * mask = dst->src[3];

    switch (Q->ne[0]) {
+        case 40:
+            GGML_ASSERT(V->ne[0] == 40);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 40,  40>(ctx, dst);
+            break;
        case 64:
            GGML_ASSERT(V->ne[0] == 64);
            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 64,  64>(ctx, dst);
            break;
+        case 72:
+            GGML_ASSERT(V->ne[0] == 72);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 72,  72>(ctx, dst);
+            break;
        case 80:
            GGML_ASSERT(V->ne[0] == 80);
            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2< 80,  80>(ctx, dst);
@@ -462,6 +470,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, F16)
    FATTN_VEC_CASES_ALL_D(Q8_0, F16)
    FATTN_VEC_CASES_ALL_D(BF16, F16)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, F16)

    FATTN_VEC_CASES_ALL_D(F16,  Q4_0)
    FATTN_VEC_CASES_ALL_D(Q4_0, Q4_0)
@@ -470,6 +479,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, Q4_0)
    FATTN_VEC_CASES_ALL_D(Q8_0, Q4_0)
    FATTN_VEC_CASES_ALL_D(BF16, Q4_0)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, Q4_0)

    FATTN_VEC_CASES_ALL_D(F16,  Q4_1)
    FATTN_VEC_CASES_ALL_D(Q4_0, Q4_1)
@@ -478,6 +488,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, Q4_1)
    FATTN_VEC_CASES_ALL_D(Q8_0, Q4_1)
    FATTN_VEC_CASES_ALL_D(BF16, Q4_1)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, Q4_1)

    FATTN_VEC_CASES_ALL_D(F16,  Q5_0)
    FATTN_VEC_CASES_ALL_D(Q4_0, Q5_0)
@@ -486,6 +497,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, Q5_0)
    FATTN_VEC_CASES_ALL_D(Q8_0, Q5_0)
    FATTN_VEC_CASES_ALL_D(BF16, Q5_0)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, Q5_0)

    FATTN_VEC_CASES_ALL_D(F16,  Q5_1)
    FATTN_VEC_CASES_ALL_D(Q4_0, Q5_1)
@@ -494,6 +506,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, Q5_1)
    FATTN_VEC_CASES_ALL_D(Q8_0, Q5_1)
    FATTN_VEC_CASES_ALL_D(BF16, Q5_1)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, Q5_1)

    FATTN_VEC_CASES_ALL_D(F16,  Q8_0)
    FATTN_VEC_CASES_ALL_D(Q4_0, Q8_0)
@@ -502,6 +515,7 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, Q8_0)
    FATTN_VEC_CASES_ALL_D(Q8_0, Q8_0)
    FATTN_VEC_CASES_ALL_D(BF16, Q8_0)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, Q8_0)

    FATTN_VEC_CASES_ALL_D(F16,  BF16)
    FATTN_VEC_CASES_ALL_D(Q4_0, BF16)
@@ -510,6 +524,16 @@ static fattn_vec_case_t ggml_cuda_get_fattn_vec_case(const int64_t head_size, co
    FATTN_VEC_CASES_ALL_D(Q5_1, BF16)
    FATTN_VEC_CASES_ALL_D(Q8_0, BF16)
    FATTN_VEC_CASES_ALL_D(BF16, BF16)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, BF16)
+
+    FATTN_VEC_CASES_ALL_D(F16, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(Q4_0, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(Q4_1, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(Q5_0, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(Q5_1, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(Q8_0, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(BF16, F8_E4M3)
+    FATTN_VEC_CASES_ALL_D(F8_E4M3, F8_E4M3)

    return nullptr;
 }
@@ -553,6 +577,7 @@ static bool ggml_cuda_fattn_kv_type_supported(const ggml_type type) {
        case GGML_TYPE_Q5_0:
        case GGML_TYPE_Q5_1:
        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_F8_E4M3:
            return true;
        default:
            return false;
@@ -666,7 +691,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
    const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && Q->ne[0] != 192 && K->ne[1] % FATTN_KQ_STRIDE == 0;

    // If Turing tensor cores are available, use them:
-    if (turing_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
+    if (turing_mma_available(cc)) {
        if (can_use_vector_kernel) {
            if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                // the sparse gather exists only in the MMA kernel: (DKQ, DV, 1, 8) with GQA > 4
@@ -0,0 +1,206 @@
+#include "fp8.cuh"
+
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+
+struct fp8_abs_src {
+    const float * x;
+    int64_t ne0;
+    int64_t ne1;
+    int64_t ne2;
+    int64_t s1;
+    int64_t s2;
+    int64_t s3;
+
+    __device__ float operator()(int64_t i) const {
+        const int64_t i0 = i % ne0;
+        i /= ne0;
+        const int64_t i1 = i % ne1;
+        i /= ne1;
+        const int64_t i2 = i % ne2;
+        const int64_t i3 = i / ne2;
+        const float value = fabsf(x[i0 + i1*s1 + i2*s2 + i3*s3]);
+        return isfinite(value) ? value : 448.0f;
+    }
+};
+
+static __global__ void fp8_amax_partials(fp8_abs_src src, int64_t ne, float * partials) {
+    float amax = 0.0f;
+    for (int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x; i < ne; i += (int64_t) blockDim.x*gridDim.x) {
+        amax = fmaxf(amax, src(i));
+    }
+
+    __shared__ float shared[WARP_SIZE];
+    amax = block_reduce<block_reduce_method::MAX, 256>(amax, shared);
+    if (threadIdx.x == 0) {
+        partials[blockIdx.x] = amax;
+    }
+}
+
+static __global__ void fp8_amax_final(const float * partials, int n, float * amax) {
+    float value = 0.0f;
+    for (int i = threadIdx.x; i < n; i += blockDim.x) {
+        value = fmaxf(value, partials[i]);
+    }
+
+    __shared__ float shared[WARP_SIZE];
+    value = block_reduce<block_reduce_method::MAX, 256>(value, shared);
+    if (threadIdx.x == 0) {
+        *amax = value;
+    }
+}
+
+static __global__ void quantize_fp8_e4m3(
+        const float * __restrict__ x, uint8_t * __restrict__ y, const float * __restrict__ amax,
+        float * __restrict__ scale, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne, int64_t s1, int64_t s2, int64_t s3) {
+    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+
+    int64_t tmp = i / ne0;
+    const int64_t i0 = i - tmp*ne0;
+    const int64_t i1 = tmp % ne1;
+    tmp /= ne1;
+    const int64_t i2 = tmp % ne2;
+    const int64_t i3 = tmp / ne2;
+
+    const float d = *amax > 0.0f ? *amax / 448.0f : 1.0f;
+    y[i] = ggml_cuda_fp32_to_f8_e4m3(x[i0 + i1*s1 + i2*s2 + i3*s3] / d);
+    if (i == 0) {
+        *scale = d;
+    }
+}
+
+static void fp8_set_batch(cublasLtMatrixLayout_t desc, int32_t count, int64_t stride) {
+    CUBLAS_CHECK(cublasLtMatrixLayoutSetAttribute(desc, CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT, &count, sizeof(count)));
+    CUBLAS_CHECK(cublasLtMatrixLayoutSetAttribute(desc, CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET, &stride, sizeof(stride)));
+}
+
+static void fp8_destroy_matmul(
+        cublasLtMatmulDesc_t op_desc, cublasLtMatrixLayout_t a_desc, cublasLtMatrixLayout_t b_desc,
+        cublasLtMatrixLayout_t d_desc) {
+    CUBLAS_CHECK(cublasLtMatrixLayoutDestroy(d_desc));
+    CUBLAS_CHECK(cublasLtMatrixLayoutDestroy(b_desc));
+    CUBLAS_CHECK(cublasLtMatrixLayoutDestroy(a_desc));
+    CUBLAS_CHECK(cublasLtMatmulDescDestroy(op_desc));
+}
+
+bool ggml_cuda_mul_mat_fp8(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+        const ggml_tensor * weight_scale) {
+    const int cc = ggml_cuda_info().devices[ctx.device].cc;
+    if (!fp8_mma_hardware_available(cc) || src0->type != GGML_TYPE_F8_E4M3 || src1->type != GGML_TYPE_F32 ||
+            dst->type != GGML_TYPE_F32 || !ggml_is_contiguous(dst) || src0->ne[0] % 16 != 0 || src0->ne[1] % 16 != 0 ||
+            src0->nb[0] != sizeof(uint8_t) || src0->nb[1] != (size_t) src0->ne[0] || src1->nb[0] != sizeof(float) ||
+            !ggml_cuda_is_aligned(src0, 16) || !ggml_cuda_is_aligned(dst, 16)) {
+        return false;
+    }
+
+    GGML_ASSERT(!weight_scale || (weight_scale->type == GGML_TYPE_F32 && ggml_is_contiguous(weight_scale) && ggml_nelements(weight_scale) == 1));
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+    GGML_ASSERT(ne10 == ne00);
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    cudaStream_t stream = ctx.stream();
+    const int64_t ne_src1 = ggml_nelements(src1);
+    ggml_cuda_pool_alloc<uint8_t> src1_fp8(ctx.pool(), ne_src1);
+    ggml_cuda_pool_alloc<float> src1_scale(ctx.pool(), 1);
+    ggml_cuda_pool_alloc<float> src1_amax(ctx.pool(), 1);
+
+    const fp8_abs_src abs_src = {
+        (const float *) src1->data, ne10, ne11, ne12,
+        (int64_t) (nb11 / sizeof(float)), (int64_t) (nb12 / sizeof(float)), (int64_t) (nb13 / sizeof(float))
+    };
+    const int reduce_blocks = std::min<int64_t>((ne_src1 + 255)/256, 1024);
+    ggml_cuda_pool_alloc<float> reduce_tmp(ctx.pool(), reduce_blocks);
+    fp8_amax_partials<<<reduce_blocks, 256, 0, stream>>>(abs_src, ne_src1, reduce_tmp.ptr);
+    fp8_amax_final<<<1, 256, 0, stream>>>(reduce_tmp.ptr, reduce_blocks, src1_amax.ptr);
+
+    quantize_fp8_e4m3<<<(ne_src1 + 255)/256, 256, 0, stream>>>(
+        (const float *) src1->data, src1_fp8.ptr, src1_amax.ptr, src1_scale.ptr, ne10, ne11, ne12, ne_src1,
+        nb11 / sizeof(float), nb12 / sizeof(float), nb13 / sizeof(float));
+    CUDA_CHECK(cudaGetLastError());
+
+    cublasLtMatmulDesc_t op_desc;
+    cublasLtMatrixLayout_t a_desc;
+    cublasLtMatrixLayout_t b_desc;
+    cublasLtMatrixLayout_t d_desc;
+    CUBLAS_CHECK(cublasLtMatmulDescCreate(&op_desc, CUBLAS_COMPUTE_32F, CUDA_R_32F));
+    const cublasOperation_t trans_a = CUBLAS_OP_T;
+    CUBLAS_CHECK(cublasLtMatmulDescSetAttribute(op_desc, CUBLASLT_MATMUL_DESC_TRANSA, &trans_a, sizeof(trans_a)));
+    CUBLAS_CHECK(cublasLtMatmulDescSetAttribute(
+        op_desc, CUBLASLT_MATMUL_DESC_B_SCALE_POINTER, &src1_scale.ptr, sizeof(src1_scale.ptr)));
+    if (weight_scale) {
+        const float * scale_ptr = (const float *) weight_scale->data;
+        CUBLAS_CHECK(cublasLtMatmulDescSetAttribute(
+            op_desc, CUBLASLT_MATMUL_DESC_A_SCALE_POINTER, &scale_ptr, sizeof(scale_ptr)));
+    }
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
+    // src1 batches that share a src0 matrix are adjacent in the compact src1 copy and in dst, fold them into N
+    const int64_t n = ne11*(ne02 == 1 ? r2*r3 : r2);
+
+    // batch the remaining matmuls with a fixed src0 stride, or issue one batched call per i3 if there is none
+    int64_t batch_count = ne02;
+    int64_t stride_a    = nb02;
+    int64_t n_calls     = ne13;
+    if (ne02 == 1) {
+        batch_count = ne03;
+        stride_a    = nb03;
+        n_calls     = 1;
+    } else if (r3 == 1 && (ne03 == 1 || nb03 == ne02*nb02)) {
+        batch_count = ne02*ne03;
+        n_calls     = 1;
+    }
+
+    CUBLAS_CHECK(cublasLtMatrixLayoutCreate(&a_desc, CUDA_R_8F_E4M3, ne00, ne01, ne00));
+    CUBLAS_CHECK(cublasLtMatrixLayoutCreate(&b_desc, CUDA_R_8F_E4M3, ne10, n, ne10));
+    CUBLAS_CHECK(cublasLtMatrixLayoutCreate(&d_desc, CUDA_R_32F, ne0, n, ne0));
+    if (batch_count > 1) {
+        fp8_set_batch(a_desc, batch_count, stride_a);
+        fp8_set_batch(b_desc, batch_count, ne10*n);
+        fp8_set_batch(d_desc, batch_count, ne0*n);
+    }
+
+    // TODO: Allow workspace in the heuristic search for potential performance gains.
+    // cuBLASLt takes caller-owned workspace per matmul call.
+    cublasLtMatmulPreference_t preference;
+    CUBLAS_CHECK(cublasLtMatmulPreferenceCreate(&preference));
+    cublasLtMatmulHeuristicResult_t heuristic;
+    int returned = 0;
+    const cublasStatus_t status = cublasLtMatmulAlgoGetHeuristic(
+        ctx.cublaslt_handle(), op_desc, a_desc, b_desc, d_desc, d_desc, preference, 1, &heuristic, &returned);
+    CUBLAS_CHECK(cublasLtMatmulPreferenceDestroy(preference));
+    if (status != CUBLAS_STATUS_SUCCESS || returned == 0) {
+        fp8_destroy_matmul(op_desc, a_desc, b_desc, d_desc);
+        return false;
+    }
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+    for (int64_t i3 = 0; i3 < n_calls; ++i3) {
+        const char * a = (const char *) src0->data + (i3/r3)*nb03;
+        const uint8_t * b = src1_fp8.ptr + i3*ne12*ne11*ne10;
+        float * d = (float *) ((char *) dst->data + i3*dst->nb[3]);
+        CUBLAS_CHECK(cublasLtMatmul(ctx.cublaslt_handle(), op_desc, &alpha, a, a_desc, b, b_desc,
+            &beta, d, d_desc, d, d_desc, &heuristic.algo, nullptr, 0, stream));
+    }
+
+    fp8_destroy_matmul(op_desc, a_desc, b_desc, d_desc);
+    return true;
+}
+
+#else
+
+bool ggml_cuda_mul_mat_fp8(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+        const ggml_tensor * weight_scale) {
+    GGML_UNUSED_VARS(ctx, src0, src1, dst, weight_scale);
+    return false;
+}
+
+#endif
@@ -0,0 +1,7 @@
+#pragma once
+
+#include "common.cuh"
+
+bool ggml_cuda_mul_mat_fp8(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+        const ggml_tensor * weight_scale = nullptr);
@@ -316,6 +316,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
+        case GGML_TYPE_F8_E4M3:
+            get_rows_cuda_float((const ggml_fp8_e4m3_t *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
        case GGML_TYPE_Q1_0:
            get_rows_cuda_q<QK1_0, QR1_0, dequantize_q1_0>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
@@ -4,6 +4,7 @@

 #include "ggml-cuda/allreduce.cuh"
 #include "ggml-cuda/common.cuh"
+#include "ggml-cuda/graph.cuh"
 #include "ggml-cuda/acc.cuh"
 #include "ggml-cuda/add-id.cuh"
 #include "ggml-cuda/arange.cuh"
@@ -27,6 +28,7 @@
 #include "ggml-cuda/diagmask.cuh"
 #include "ggml-cuda/diag.cuh"
 #include "ggml-cuda/fattn.cuh"
+#include "ggml-cuda/fp8.cuh"
 #include "ggml-cuda/fwht.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
@@ -583,6 +585,7 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
    CUdeviceptr pool_addr = 0;
    size_t pool_used = 0;
    size_t pool_size = 0;
+    size_t max_size = CUDA_POOL_VMM_MAX_SIZE;
    size_t granularity;
 #if defined(GGML_USE_HIP)
    std::vector<std::pair<CUdeviceptr, size_t>> mappings;
@@ -604,7 +607,7 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
 #else
            CU_CHECK(cuMemUnmap(pool_addr, pool_size));
 #endif
-            CU_CHECK(cuMemAddressFree(pool_addr, CUDA_POOL_VMM_MAX_SIZE));
+            CU_CHECK(cuMemAddressFree(pool_addr, max_size));
        }
    }

@@ -620,7 +623,21 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
            size_t reserve_size = size - avail;
            reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);

-            GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                CUresult err = cuMemAddressReserve(&pool_addr, max_size, 0, 0, 0);
+                while (err == CUDA_ERROR_OUT_OF_MEMORY && (max_size >> 1) >= reserve_size && (max_size >> 1) >= granularity) {
+                    max_size >>= 1;
+                    err = cuMemAddressReserve(&pool_addr, max_size, 0, 0, 0);
+                }
+                CU_CHECK(err);
+                if (max_size < CUDA_POOL_VMM_MAX_SIZE) {
+                    GGML_LOG_WARN(GGML_CUDA_NAME " pool[%d]: reserved %.2f GiB of virtual memory instead of %.2f GiB\n",
+                                  device, max_size / (1024.0 * 1024.0 * 1024.0), CUDA_POOL_VMM_MAX_SIZE / (1024.0 * 1024.0 * 1024.0));
+                }
+            }
+
+            GGML_ASSERT(pool_size + reserve_size <= max_size);

            // allocate more physical memory
            CUmemAllocationProp prop = {};
@@ -630,11 +647,6 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
            CUmemGenericAllocationHandle handle;
            CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));

-            // reserve virtual address space (if not already reserved)
-            if (pool_addr == 0) {
-                CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
-            }
-
            // map at the end of the pool
            CUdeviceptr start_ptr = (CUdeviceptr)((char *)(pool_addr) + pool_size);
            CU_CHECK(cuMemMap(start_ptr, reserve_size, 0, handle, 0));
@@ -735,6 +747,124 @@ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(g
    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(ctx, device));
 }

+ggml_backend_cuda_context::ggml_backend_cuda_context(int device) :
+    device(device),
+    name(GGML_CUDA_NAME + std::to_string(device)),
+    concurrent_stream_context(std::make_unique<ggml_cuda_stream_context>()) {
+}
+
+#ifdef USE_CUDA_GRAPH
+ggml_cuda_graph * ggml_backend_cuda_context::cuda_graph(const void * first_node_ptr, int64_t time_now) {
+    // Give all cached graphs 10s to be reused after an idle period, without updating each graph.
+    if (!cuda_graphs.empty() && time_now - last_graph_use_time >= 10'000'000) {
+        next_graph_eviction_time = time_now + 10'000'000;
+    }
+    last_graph_use_time = time_now;
+
+    // Sweep every 5s during active use, evicting graphs unused for at least 10s.
+    if (time_now >= next_graph_eviction_time) {
+        next_graph_eviction_time = time_now + 5'000'000;
+        for (auto it = cuda_graphs.begin(); it != cuda_graphs.end(); ) {
+            if (it->first != first_node_ptr && time_now - it->second->last_used_time >= 10'000'000) {
+                it = cuda_graphs.erase(it);
+            } else {
+                ++it;
+            }
+        }
+    }
+
+    auto it = cuda_graphs.find(first_node_ptr);
+    if (it == cuda_graphs.end()) {
+        it = cuda_graphs.emplace(first_node_ptr, std::make_unique<ggml_cuda_graph>()).first;
+    }
+    it->second->last_used_time = time_now;
+    return it->second.get();
+}
+
+bool ggml_backend_cuda_context::any_cuda_graph_enabled() const {
+    for (const auto & [key, graph] : cuda_graphs) {
+        if (graph && graph->is_enabled()) {
+            return true;
+        }
+    }
+    return false;
+}
+
+bool ggml_backend_cuda_context::any_cuda_graph_has_instance() const {
+    for (const auto & [key, graph] : cuda_graphs) {
+        if (graph && !graph->instances.empty()) {
+            return true;
+        }
+    }
+    return false;
+}
+#endif // USE_CUDA_GRAPH
+
+#if defined(GGML_CUDA_USE_PDL)
+bool ggml_cuda_kernel_can_use_pdl(const void * kernel) {
+    static const bool env_pdl_enabled = []() {
+        const char * env = getenv("GGML_CUDA_PDL");
+        return env == nullptr || std::atoi(env) != 0;
+    }();
+    if (!env_pdl_enabled) {
+        return false;
+    }
+
+    const int device = ggml_cuda_get_device();
+
+    struct cache_key {
+        int          device;
+        const void * kernel;
+
+        bool operator==(const cache_key & other) const { return device == other.device && kernel == other.kernel; }
+    };
+
+    struct cache_key_hash {
+        // MurmurHash3 mixing function for better hash distribution (vs. just std::hash which in some implementations simply returns the identity)
+        static size_t hash_mix(size_t x) {
+            std::uint64_t       y = x;
+            const std::uint64_t m = 0xe9846af9b1a615d;
+
+            y ^= y >> 32;
+            y *= m;
+            y ^= y >> 32;
+            y *= m;
+            y ^= y >> 28;
+
+            return static_cast<size_t>(y);
+        }
+
+        size_t operator()(const cache_key & key) const {
+            // Use a nonzero seed to avoid mapping all-zero keys to zero
+            size_t h = 42;
+            h        = hash_mix(h + key.device);
+            h        = hash_mix(h + reinterpret_cast<size_t>(key.kernel));
+            return h;
+        }
+    };
+
+    static std::mutex                                          cache_mutex;
+    static std::unordered_map<cache_key, bool, cache_key_hash> cache;
+
+    const cache_key             key = { device, kernel };
+    std::lock_guard<std::mutex> lock(cache_mutex);
+    const auto                  it = cache.find(key);
+    if (it != cache.end()) {
+        return it->second;
+    }
+
+    cudaFuncAttributes attr = {};
+    CUDA_CHECK(cudaFuncGetAttributes(&attr, kernel));
+
+    // PDL device-side primitives are emitted only for PTX versions >= 90.
+    // We have to guard on a loaded kernel's PTX version so a kernel forward-JIT'ed
+    // from pre-Hopper PTX to a Hopper-or-newer GPU does not opt into PDL.
+    const bool can_use_pdl = attr.ptxVersion >= 90;
+    cache.emplace(key, can_use_pdl);
+    return can_use_pdl;
+}
+#endif // defined(GGML_CUDA_USE_PDL)
+
 // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
 // this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured

@@ -761,6 +891,11 @@ ggml_backend_cuda_context::~ggml_backend_cuda_context() {
                CUDA_CHECK(cudaFree(cublas_workspaces[i][j]));
            }
        }
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+        if (cublaslt_handles[i] != nullptr) {
+            CUBLAS_CHECK(cublasLtDestroy(cublaslt_handles[i]));
+        }
+#endif
    }
 }

@@ -1869,9 +2004,9 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const

 static void ggml_cuda_mul_mat_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    const int cc = ggml_cuda_info().devices[ctx.device].cc;
-    const ggml_prec prec = (ggml_prec) ggml_get_op_params_i32(dst, 0);
+    const ggml_prec prec = ggml_prec_get_acc(dst);
    ggml_type compute_type = src0->type;
-    if (ggml_is_quantized(compute_type)) {
+    if (ggml_is_quantized(compute_type) || compute_type == GGML_TYPE_F8_E4M3) {
        compute_type = fast_fp16_hardware_available(cc) ? GGML_TYPE_F16 : GGML_TYPE_F32;
    } else if (compute_type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc)) {
        compute_type = GGML_TYPE_F32;
@@ -2042,7 +2177,7 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
    return true;
 }

-static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
+static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor, const ggml_tensor * gate = nullptr) {
    ggml_tensor *       src0 = tensor->src[0];
    ggml_tensor *       src1 = tensor->src[1];
    const ggml_tensor * dst  = tensor;
@@ -2050,12 +2185,17 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
    const bool is_mul_mat_id = tensor->op == GGML_OP_MUL_MAT_ID;

    bool use_mul_mat_vec_f =
-        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16) &&
+        (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16 ||
+         src0->type == GGML_TYPE_F8_E4M3) &&
        src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;

    const int cc        = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
-    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, warp_size, src0->ne, src0->nb, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+    use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0, cc, warp_size, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
+
+    if (gate && !ggml_cuda_should_use_mmvf(gate->src[0], cc, warp_size, is_mul_mat_id ? src1->ne[2] : src1->ne[1])) {
+        return false;
+    }

    //we only support fusion for ncols_dst = 1
    if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
@@ -2079,11 +2219,11 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
                                   ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) &&
                                   src0->view_src;

-    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type) && !bad_padding_clear && src1->type == GGML_TYPE_F32 &&
-                             dst->type == GGML_TYPE_F32 && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    bool use_mul_mat_vec_q = ggml_cuda_should_use_mmvq(src0->type, cc, src1->ne[1]) && !bad_padding_clear &&
+                             src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;

    // fusion is not universally faster on Pascal
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    if (cc <= GGML_CUDA_CC_PASCAL) {
        return false;
    }
@@ -2151,7 +2291,7 @@ static bool ggml_cuda_match_shared_expert(const ggml_cgraph * graph, int routed_
    return true;
 }

-// The NVFP4 weight scale that scale_node (MUL) applies to the output of mm_node (MUL_MAT), or nullptr.
+// The weight scale that scale_node (MUL) applies to the output of mm_node (MUL_MAT), or nullptr.
 // Shared by ggml_cuda_try_fuse and graph_optimize so that both accept the same patterns.
 static const ggml_tensor * ggml_cuda_get_mul_mat_scale(const ggml_tensor * scale_node, const ggml_tensor * mm_node) {
    const bool scale_lhs_mm = scale_node->src[0] == mm_node;
@@ -2161,7 +2301,7 @@ static const ggml_tensor * ggml_cuda_get_mul_mat_scale(const ggml_tensor * scale
    }

    const ggml_tensor * scale = scale_lhs_mm ? scale_node->src[1] : scale_node->src[0];
-    if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+    if ((mm_node->src[0]->type != GGML_TYPE_NVFP4 && mm_node->src[0]->type != GGML_TYPE_F8_E4M3) || scale_node->type != GGML_TYPE_F32 ||
            scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != 1 ||
            !ggml_are_same_shape(scale_node, mm_node)) {
        return nullptr;
@@ -2182,7 +2322,7 @@ static const ggml_tensor * ggml_cuda_get_mul_mat_id_scale(const ggml_tensor * re
    }

    const ggml_tensor * scale = reshape->src[0];
-    if (mm_node->src[0]->type != GGML_TYPE_NVFP4 || scale_node->type != GGML_TYPE_F32 ||
+    if ((mm_node->src[0]->type != GGML_TYPE_NVFP4 && mm_node->src[0]->type != GGML_TYPE_F8_E4M3) || scale_node->type != GGML_TYPE_F32 ||
            scale->type != GGML_TYPE_F32 || !ggml_is_contiguous(scale) || ggml_nelements(scale) != mm_node->src[0]->ne[2] ||
            !ggml_are_same_shape(scale_node, mm_node)) {
        return nullptr;
@@ -2280,17 +2420,17 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
    const int cc        = ggml_cuda_info().devices[ctx.device].cc;
    const int warp_size = ggml_cuda_info().devices[ctx.device].warp_size;

-    if (ggml_cuda_should_use_mmvf(src0->type, cc, warp_size, src0->ne, src0->nb, ne11)) {
-        // The custom F16 vector kernel can be used over batched cuBLAS GEMM.
+    if (ggml_cuda_should_use_mmvf(src0, cc, warp_size, ne11)) {
+        // The custom vector kernel can be used over batched cuBLAS GEMM.
        // But this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
        ggml_cuda_mul_mat_vec_f(ctx, src0, src1, nullptr, dst);
        return;
    }
-    // A transposed vector can still use MMVQ (i.e. ne01 == 1)
+    // A transposed vector can still use MMVF (i.e. ne01 == 1)
    if (ne01 == 1 && ne11 > MMVF_MAX_BATCH_SIZE && ne2 == 1 && ne3 == 1
            && src0->type == GGML_TYPE_F32
            && ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)
-            && ggml_cuda_should_use_mmvf(src1->type, cc, warp_size, src1->ne, src1->nb, /*ne11 =*/ 1)) {
+            && ggml_cuda_should_use_mmvf(src1, cc, warp_size, /*ne11 =*/ 1)) {
        ggml_tensor dst_vec = *dst;
        dst_vec.ne[0] = ne11;
        dst_vec.ne[1] = 1;
@@ -2312,6 +2452,9 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, dst);
        return;
    }
+    if (src0->type == GGML_TYPE_F8_E4M3 && ggml_cuda_mul_mat_fp8(ctx, src0, src1, dst)) {
+        return;
+    }
    ggml_cuda_mul_mat_cublas(ctx, src0, src1, dst);
 }

@@ -2320,7 +2463,6 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
 static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int cc) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
-
    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;

    if (src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
@@ -2332,7 +2474,10 @@ static bool ggml_cuda_mul_mat_id_needs_sync(const ggml_tensor * dst, const int c
            if (dst->ne[2] <= get_mmvq_mmid_max_batch(src0->type, cc)) {
                return false;
            }
-        } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+        } else if (src0->type == GGML_TYPE_F8_E4M3 &&
+                ggml_cuda_should_use_mmvf(src0, cc, warp_size, dst->ne[2])) {
+            return false;
+        } else if (src0->type != GGML_TYPE_F8_E4M3 && GGML_CUDA_CC_IS_AMD(cc)) {
            return false;
        }
    }
@@ -2358,7 +2503,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *

    GGML_TENSOR_BINARY_OP_LOCALS

-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int cc        = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;

    // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
@@ -2371,8 +2516,12 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
                    ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
                    return;
                }
+            } else if (src0->type == GGML_TYPE_F8_E4M3 &&
+                    ggml_cuda_should_use_mmvf(src0, cc, warp_size, ne2)) {
+                ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
+                return;
            } else {
-                if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (src0->type != GGML_TYPE_F8_E4M3 && GGML_CUDA_CC_IS_AMD(cc)) {
                    ggml_cuda_mul_mat_vec_f(ctx, src0, src1, ids, dst);
                    return;
                }
@@ -2398,7 +2547,7 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
    GGML_ASSERT(nb2  % nb1  == 0);

    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
-        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+        || ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F8_E4M3 ? GGML_TYPE_F32 : src0->type;
    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
@@ -3090,7 +3239,54 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
    return res;
 }

-static void ggml_cuda_graph_update_executable(cudaGraphExec_t & instance, cudaGraph_t cuda_graph) {
+#ifdef USE_CUDA_GRAPH_CLUSTER_CHECK
+static std::vector<ggml_cuda_graph_cluster> ggml_cuda_graph_get_clusters(cudaGraph_t cuda_graph, int device) {
+    std::vector<ggml_cuda_graph_cluster> clusters;
+    static const bool runtime_affected = [] {
+        int version = 0;
+        CUDA_CHECK(cudaRuntimeGetVersion(&version));
+        // TODO: upper bound once fixed in the CUDA runtime
+        return version >= 11080;
+    }();
+    if (!runtime_affected || ggml_cuda_info().devices[device].cc < GGML_CUDA_CC_HOPPER) {
+        return clusters;
+    }
+    size_t n_nodes = 0;
+    CUDA_CHECK(cudaGraphGetNodes(cuda_graph, nullptr, &n_nodes));
+    std::vector<cudaGraphNode_t> nodes(n_nodes);
+    CUDA_CHECK(cudaGraphGetNodes(cuda_graph, nodes.data(), &n_nodes));
+    for (size_t i = 0; i < n_nodes; ++i) {
+        cudaGraphNodeType type;
+        CUDA_CHECK(cudaGraphNodeGetType(nodes[i], &type));
+        if (type != cudaGraphNodeTypeKernel) {
+            continue;
+        }
+        cudaLaunchAttributeValue val = {};
+        CUDA_CHECK(cudaGraphKernelNodeGetAttribute(nodes[i], cudaLaunchAttributeClusterDimension, &val));
+        if (val.clusterDim.x * val.clusterDim.y * val.clusterDim.z > 1) {
+            clusters.push_back({ i, val.clusterDim.x, val.clusterDim.y, val.clusterDim.z });
+        }
+    }
+    return clusters;
+}
+#endif // USE_CUDA_GRAPH_CLUSTER_CHECK
+
+static void ggml_cuda_graph_update_executable(ggml_cuda_graph_instance & instance, cudaGraph_t cuda_graph, int device) {
+#ifdef USE_CUDA_GRAPH_CLUSTER_CHECK
+    std::vector<ggml_cuda_graph_cluster> clusters = ggml_cuda_graph_get_clusters(cuda_graph, device);
+    if (instance.clusters != clusters) {
+        if (ggml_cuda_graph_debug_enabled()) {
+            GGML_LOG_DEBUG("%s: cluster dims changed, re-instantiating CUDA graph\n", __func__);
+        }
+        CUDA_CHECK(cudaGraphExecDestroy(instance));
+        CUDA_CHECK(cudaGraphInstantiate(&instance.handle, cuda_graph, nullptr, nullptr, 0));
+        instance.clusters = std::move(clusters);
+        return;
+    }
+#else
+    GGML_UNUSED(device);
+#endif // USE_CUDA_GRAPH_CLUSTER_CHECK
+
 #if CUDART_VERSION >= 12000
    cudaGraphExecUpdateResultInfo result_info;
    cudaError_t stat = cudaGraphExecUpdate(instance, cuda_graph, &result_info);
@@ -3109,8 +3305,8 @@ static void ggml_cuda_graph_update_executable(cudaGraphExec_t & instance, cudaGr
        // so instead clear error and re-instantiate
        (void)cudaGetLastError();
        CUDA_CHECK(cudaGraphExecDestroy(instance));
-        instance = nullptr;
-        CUDA_CHECK(cudaGraphInstantiate(&instance, cuda_graph, nullptr, nullptr, 0));
+        instance.handle = nullptr;
+        CUDA_CHECK(cudaGraphInstantiate(&instance.handle, cuda_graph, nullptr, nullptr, 0));
    } else {
        GGML_ASSERT(stat == cudaSuccess);
    }
@@ -3131,11 +3327,14 @@ static void ggml_cuda_graph_launch_chunk(ggml_backend_cuda_context * cuda_ctx, g
    CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_graph));

    if (idx < graph->instances.size()) {
-        ggml_cuda_graph_update_executable(graph->instances[idx], cuda_graph);
+        ggml_cuda_graph_update_executable(graph->instances[idx], cuda_graph, cuda_ctx->device);
    } else {
        GGML_ASSERT(idx == graph->instances.size());
-        graph->instances.emplace_back(nullptr);
-        CUDA_CHECK(cudaGraphInstantiate(&graph->instances[idx], cuda_graph, nullptr, nullptr, 0));
+        graph->instances.emplace_back();
+        CUDA_CHECK(cudaGraphInstantiate(&graph->instances[idx].handle, cuda_graph, nullptr, nullptr, 0));
+#ifdef USE_CUDA_GRAPH_CLUSTER_CHECK
+        graph->instances[idx].clusters = ggml_cuda_graph_get_clusters(cuda_graph, cuda_ctx->device);
+#endif // USE_CUDA_GRAPH_CLUSTER_CHECK
    }
    CUDA_CHECK(cudaGraphDestroy(cuda_graph));

@@ -4424,6 +4623,13 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                fusion_data.glu_op     = ggml_get_glu_op(glu);
                fusion_data.glu_limit  = ggml_get_op_params_f32(glu, 3);

+                if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n, gate_n)) {
+                    ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
+
                if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
                    ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
                    fused_mul_mat_vec = true;
@@ -4526,6 +4732,13 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                fusion_data.glu_op     = ggml_get_glu_op(glu);
                fusion_data.glu_limit  = ggml_get_op_params_f32(glu, 3);

+                if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n, gate_n)) {
+                    ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
+                    fused_mul_mat_vec = true;
+                    fused_node_count  = n_ops;
+                    break;
+                }
+
                if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
                    ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
                    fused_mul_mat_vec = true;
@@ -4583,7 +4796,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
            const ggml_tensor * src1 = up_n->src[1];
            const ggml_tensor * ids  = up_n->src[2];

-            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n)) {
+            if (ggml_cuda_should_fuse_mul_mat_vec_f(up_n, gate_n)) {
                ggml_cuda_mm_fusion_args_host fusion_data{};
                fusion_data.gate      = gate_n->src[0];
                fusion_data.x_bias    = up_bias_tensor;
@@ -4626,7 +4839,7 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
            const ggml_tensor * src1 = up->src[1];
            const ggml_tensor * ids  = up->src[2];

-            if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
+            if (ggml_cuda_should_fuse_mul_mat_vec_f(up, gate)) {
                ggml_cuda_mm_fusion_args_host fusion_data{};
                fusion_data.gate      = gate->src[0];
                fusion_data.glu_op    = ggml_get_glu_op(glu);
@@ -4744,6 +4957,19 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
            fusion_data.x_bias  = bias;
            fusion_data.x_scale = scale;

+            if (ggml_cuda_should_fuse_mul_mat_vec_f(mm_node)) {
+                ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, out_node, &fusion_data);
+                fused_mul_mat_vec = true;
+                fused_node_count  = n_ops;
+                break;
+            }
+
+            if (op == GGML_OP_MUL_MAT && !with_bias && src0->type == GGML_TYPE_F8_E4M3 &&
+                    ggml_get_op_params_i32(mm_node, 3) != GGML_PREC_F32 &&
+                    ggml_cuda_mul_mat_fp8(*cuda_ctx, src0, src1, scale_node, scale)) {
+                return 1;
+            }
+
            if (ggml_cuda_should_fuse_mul_mat_vec_q(mm_node)) {
                ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, out_node, &fusion_data);
                fused_mul_mat_vec = true;
@@ -5143,7 +5369,9 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,

    ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
    if (graph->is_enabled()) {
-        const bool graph_compatible = ggml_cuda_graph_check_compability(cgraph);
+        // graph->uid is set only for a compatible graph (see ggml_cuda_graph_update_required), so a reused graph needs no new check
+        const bool graph_reused     = cgraph->uid != 0 && cgraph->uid == graph->uid;
+        const bool graph_compatible = graph_reused || ggml_cuda_graph_check_compability(cgraph);
        if (graph_compatible) {
            const bool properties_changed = ggml_cuda_graph_update_required(cuda_ctx, cgraph);

@@ -5978,7 +6206,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16 && !ggml_cuda_op_mul_mat_use_fwht(op)) {
                    return false;
                }
-                if (op->op == GGML_OP_MUL_MAT_ID && (ggml_get_op_params_i32(op, 3) == GGML_PREC_BF16 || ggml_get_op_params_i32(op, 3) == GGML_PREC_F32)) {
+                const int cc_dev = ggml_cuda_info().devices[dev_ctx->device].cc;
+                const bool f32_src1_ok = GGML_CUDA_CC_IS_NVIDIA(cc_dev) && cc_dev < GGML_CUDA_CC_AMPERE;
+                const enum ggml_prec prec_src1 = ggml_prec_get_src(op, 1);
+                const bool fp8_id_mmvf = op->op == GGML_OP_MUL_MAT_ID &&
+                    a->type == GGML_TYPE_F8_E4M3 && b->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                    ggml_cuda_should_use_mmvf(a, cc_dev, ggml_cuda_info().devices[dev_ctx->device].warp_size, b->ne[2]);
+                if (prec_src1 == GGML_PREC_F32 && !fp8_id_mmvf && (op->op == GGML_OP_MUL_MAT_ID || !f32_src1_ok || a->type != GGML_TYPE_F32)) {
+                    return false;
+                }
+                if (prec_src1 == GGML_PREC_BF16 && (op->op == GGML_OP_MUL_MAT_ID || a->type != GGML_TYPE_F32)) {
                    return false;
                }
 #ifdef GGML_USE_MUSA
@@ -6006,6 +6243,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                    case GGML_TYPE_Q8_0:
                    case GGML_TYPE_MXFP4:
                    case GGML_TYPE_NVFP4:
+                    case GGML_TYPE_F8_E4M3:
                    case GGML_TYPE_Q2_K:
                    case GGML_TYPE_Q3_K:
                    case GGML_TYPE_Q4_K:
@@ -6043,6 +6281,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                    case GGML_TYPE_Q5_0:
                    case GGML_TYPE_Q5_1:
                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_F8_E4M3:
                    case GGML_TYPE_Q2_K:
                    case GGML_TYPE_Q3_K:
                    case GGML_TYPE_Q4_K:
@@ -6072,7 +6311,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            {
                return (
                           (
-                               (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 ||
+                               (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16 || op->type == GGML_TYPE_BF16 || op->type == GGML_TYPE_F8_E4M3 ||
                               op->type == GGML_TYPE_Q4_0 || op->type == GGML_TYPE_Q4_1 || op->type == GGML_TYPE_Q5_0 ||
                               op->type == GGML_TYPE_Q5_1 || op->type == GGML_TYPE_Q8_0 || op->type == GGML_TYPE_IQ4_NL) &&
                               op->src[0]->type == GGML_TYPE_F32
@@ -6093,6 +6332,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
            {
                ggml_type src0_type = op->src[0]->type;
                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F8_E4M3 && src1_type == GGML_TYPE_F8_E4M3) {
+                    return true;
+                }
+                if ((src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_F16 || src0_type == GGML_TYPE_BF16) &&
+                    src1_type == GGML_TYPE_F8_E4M3) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F8_E4M3 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
                if ((src0_type == GGML_TYPE_F32 || src0_type == GGML_TYPE_BF16 || src0_type == GGML_TYPE_F16) &&
                    (src1_type == GGML_TYPE_F32 || src1_type == GGML_TYPE_BF16 || src1_type == GGML_TYPE_F16)
                ) {
@@ -6510,12 +6759,21 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t

    {
        const auto & info = ggml_cuda_info();
+        bool has_fp8 = false;
+        bool has_fp4 = false;
        for (int id = 0; id < info.device_count; ++id) {
-            if (blackwell_mma_available(info.devices[id].cc)) {
-                features.push_back({ "BLACKWELL_NATIVE_FP4", "1"});
+            has_fp8 |= fp8_mma_hardware_available(info.devices[id].cc);
+            has_fp4 |= blackwell_mma_available(info.devices[id].cc);
+            if (has_fp8 && has_fp4) {
                break;
            }
        }
+        if (has_fp8) {
+            features.push_back({ "NATIVE_FP8", "1"});
+        }
+        if (has_fp4) {
+            features.push_back({ "BLACKWELL_NATIVE_FP4", "1"});
+        }
    }

    #undef _STRINGIFY
@@ -0,0 +1,221 @@
+#pragma once
+
+// CUDA graph capture and concurrent stream state, only needed by ggml-cuda.cu
+
+#include "common.cuh"
+
+#include <algorithm>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#ifdef USE_CUDA_GRAPH
+// cudaGraphExecUpdate keeps stale kernel cluster dims (cuBLAS sm90), re-instantiate when they change
+// TODO: add a test, fails on H100 (driver 590.48.01, CUDA runtime 12.9.79, cuBLAS 12.9.1.4) for F16 MUL_MAT with
+//       ne01 1200 <-> 1040 (k=768, n=1024) between graph updates, see https://github.com/ggml-org/llama.cpp/pull/31284
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11080
+#define USE_CUDA_GRAPH_CLUSTER_CHECK
+#endif
+
+#ifdef USE_CUDA_GRAPH_CLUSTER_CHECK
+// cluster dims of a clustered kernel node
+struct ggml_cuda_graph_cluster {
+    size_t   node;
+    uint32_t x, y, z;
+
+    bool operator==(const ggml_cuda_graph_cluster & other) const {
+        return node == other.node && x == other.x && y == other.y && z == other.z;
+    }
+};
+#endif
+
+struct ggml_cuda_graph_instance {
+    cudaGraphExec_t handle = nullptr;
+#ifdef USE_CUDA_GRAPH_CLUSTER_CHECK
+    std::vector<ggml_cuda_graph_cluster> clusters;
+#endif
+
+    operator cudaGraphExec_t() const { return handle; }
+};
+#endif // USE_CUDA_GRAPH
+
+struct ggml_cuda_graph {
+#ifdef USE_CUDA_GRAPH
+    ~ggml_cuda_graph() {
+        for (cudaGraphExec_t instance : instances) {
+            CUDA_CHECK(cudaGraphExecDestroy(instance));
+        }
+    }
+    // the ggml graph is captured as a series of CUDA graphs, so the GPU can run one chunk while the next is enqueued
+    std::vector<ggml_cuda_graph_instance> instances;
+    size_t num_nodes = 0;
+    std::vector<cudaGraphNode_t> nodes;
+    bool disable_due_to_gpu_arch = false;
+    bool warmup_complete = false;
+    uint64_t uid = 0;
+    int64_t last_used_time = 0;
+    struct node_properties {
+        ggml_tensor node;
+        void *   node_src_data_ptrs[GGML_MAX_SRC];
+        int64_t  node_src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+        size_t   node_src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
+    };
+    std::vector<node_properties> node_props;
+
+    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);
+    }
+#endif
+};
+
+struct ggml_cuda_concurrent_event {
+    std::vector<cudaEvent_t> join_events;
+    cudaEvent_t              fork_event = nullptr;
+
+    int                                          n_streams = 0;
+    std::unordered_map<const ggml_tensor *, int> stream_mapping;
+
+    // Original order of nodes in this concurrent region (before interleaving)
+    // Used to restore grouping for fusion within streams
+    std::vector<const ggml_tensor *> original_order;
+
+    const ggml_tensor * join_node;
+
+    ggml_cuda_concurrent_event() = default;
+
+    ggml_cuda_concurrent_event(const ggml_cuda_concurrent_event &) = delete;
+    ggml_cuda_concurrent_event & operator=(const ggml_cuda_concurrent_event &) = delete;
+
+    explicit ggml_cuda_concurrent_event(int n_streams) : n_streams(n_streams) {
+        join_events.resize(n_streams);
+
+        for (size_t i = 0; i < join_events.size(); ++i) {
+            CUDA_CHECK(cudaEventCreateWithFlags(&join_events[i], cudaEventDisableTiming));
+        }
+
+        CUDA_CHECK(cudaEventCreateWithFlags(&fork_event, cudaEventDisableTiming));
+    }
+
+    ggml_cuda_concurrent_event(ggml_cuda_concurrent_event && other) noexcept
+    : join_events(std::move(other.join_events))
+    , fork_event(other.fork_event)
+    , n_streams(other.n_streams)
+    , stream_mapping(std::move(other.stream_mapping))
+    , original_order(std::move(other.original_order))
+    , join_node(other.join_node) {
+        other.fork_event = nullptr;
+    }
+
+    // 1. check if any branches write to overlapping memory ranges (except the join node)
+    // 2. check whether all srcs are either within the branch or outside the nodes covered by ggml_cuda_concurrent_event
+    // we assume all nodes have the same buffer
+    bool is_valid() const {
+        std::vector<std::vector<std::pair<int64_t, int64_t>>> write_ranges;
+        write_ranges.resize(n_streams);
+
+        // get join_node's memory range to exclude from overlap checking.
+        // multiple nodes can use join_node's buffer; we synchronize on the join node.
+        const ggml_tensor * join_t     = join_node->view_src ? join_node->view_src : join_node;
+        const int64_t       join_start = (int64_t) join_t->data;
+        const int64_t       join_end   = join_start + ggml_nbytes(join_t);
+
+        for (const auto & [tensor, stream] : stream_mapping) {
+            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
+            const int64_t       t_start = (int64_t) t->data;
+            const int64_t       t_end   = t_start + ggml_nbytes(t);
+
+            // skip tensors that overlap with join_node's buffer.
+            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
+                continue;
+            }
+
+            // concurrent streams begin from 1
+            write_ranges[stream - 1].emplace_back(t_start, t_end);
+        }
+
+        for (int i = 0; i < n_streams; ++i) {
+            // sorts first by start then by end of write range
+            std::sort(write_ranges[i].begin(), write_ranges[i].end());
+        }
+
+        bool writes_overlap = false;
+        bool dependent_srcs = false;
+        for (const auto & [tensor, stream] : stream_mapping) {
+            const ggml_tensor * t = tensor->view_src ? tensor->view_src : tensor;
+            const int64_t       t_start = (int64_t) t->data;
+            const int64_t       t_end   = t_start + ggml_nbytes(t);
+
+            // skip tensors that overlap with join_node's buffer
+            if ((t_start <= join_start && join_start < t_end) || (join_start <= t_start && t_start < join_end)) {
+                continue;
+            }
+
+            // check if this buffer's write data overlaps with another stream's
+            std::pair<int64_t, int64_t> data_range = std::make_pair(t_start, t_end);
+            for (int i = 0; i < n_streams; ++i) {
+                if (i == stream - 1) {
+                    continue;
+                }
+                auto it = std::lower_bound(write_ranges[i].begin(), write_ranges[i].end(), data_range);
+
+                if (it != write_ranges[i].end()) {
+                    const std::pair<int64_t, int64_t> & other = *it;
+
+                    // std::lower_bound returns the first element where other >= data_range (lexicographically).
+                    // This guarantees other.first >= data_range.first.
+                    // Therefore, overlap occurs iff other.first < data_range.second
+                    // (i.e., the other range starts before this range ends).
+                    if (other.first < data_range.second) {
+                        GGML_LOG_DEBUG("Writes overlap for %s", tensor->name);
+                        writes_overlap = true;
+                        break;
+                    }
+                }
+            }
+
+            //check if all srcs are either in branch or don't have a branch
+            for (int i = 0; i < GGML_MAX_SRC; ++i) {
+                if (!tensor->src[i]) {
+                    continue;
+                }
+
+                auto it = stream_mapping.find(tensor->src[i]);
+
+                if (it == stream_mapping.end()) {
+                    continue;
+                }
+
+                if (it->second != stream) {
+                    dependent_srcs = true;
+                    break;
+                }
+            }
+
+            if (dependent_srcs || writes_overlap) {
+                break;
+            }
+        }
+
+        return !writes_overlap && !dependent_srcs;
+    }
+
+    ~ggml_cuda_concurrent_event() {
+        if (fork_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(fork_event));
+        }
+        for (cudaEvent_t e : join_events) {
+            if (e != nullptr) {
+                CUDA_CHECK(cudaEventDestroy(e));
+            }
+        }
+    }
+};
+
+struct ggml_cuda_stream_context {
+    std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
+
+    void reset() {
+        concurrent_events.clear();
+    }
+};
@@ -5,6 +5,14 @@
 #include "mmf.cuh"
 #include "convert.cuh"

+static __device__ __forceinline__ float2 mmvf_load_fp8x2(const ggml_fp8_e4m3_t * x, const int col2) {
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+    return ggml_cuda_cast<float2>(reinterpret_cast<const __nv_fp8x2_e4m3 *>(x)[col2]);
+#else
+    return make_float2(ggml_cuda_cast<float>(x[2*col2]), ggml_cuda_cast<float>(x[2*col2 + 1]));
+#endif
+}
+
 template <typename T, typename type_acc, int ncols_dst, int block_size, bool has_fusion = false, bool is_multi_token_id = false>
 static __global__ void mul_mat_vec_f(
        const T * x_ptr, const float * y_ptr, const int32_t * ids_ptr, const ggml_cuda_mm_fusion_args_device fusion, float * dst_ptr,
@@ -56,16 +64,24 @@ static __global__ void mul_mat_vec_f(
    bool use_gate = false;
    bool use_bias = false;
    bool use_gate_bias = false;
+    bool use_scale = false;
+    bool use_gate_scale = false;
    ggml_glu_op glu_op = ggml_glu_op::GGML_GLU_OP_SWIGLU;
    float glu_limit = 0.0f;
    const T * gate_x = nullptr;
    const float * x_bias = nullptr;
    const float * gate_bias = nullptr;
+    const float * x_scale = nullptr;
+    const float * gate_scale = nullptr;

    if constexpr (has_fusion) {
        use_gate = fusion.gate != nullptr;
        use_bias = fusion.x_bias != nullptr;
        use_gate_bias = fusion.gate_bias != nullptr;
+        if constexpr (std::is_same_v<T, ggml_fp8_e4m3_t>) {
+            use_scale = fusion.x_scale != nullptr;
+            use_gate_scale = fusion.gate_scale != nullptr && use_gate;
+        }
        glu_op = fusion.glu_op;
        glu_limit = fusion.glu_limit;

@@ -81,6 +97,12 @@ static __global__ void mul_mat_vec_f(
        } else {
            use_gate_bias = false;
        }
+        if (use_scale) {
+            x_scale = static_cast<const float *>(fusion.x_scale);
+        }
+        if (use_gate_scale) {
+            gate_scale = static_cast<const float *>(fusion.gate_scale);
+        }
    }

    if (use_gate) {
@@ -97,8 +119,6 @@ static __global__ void mul_mat_vec_f(
        }
    }

-    const float2 * y2 = (const float2 *) y;
-
    extern __shared__ char data_mmv[];
    float * buf_iw = (float *) data_mmv;
    [[maybe_unused]] float * buf_iw_gate = nullptr;
@@ -128,6 +148,7 @@ static __global__ void mul_mat_vec_f(
    }

    if constexpr (std::is_same_v<T, float>) {
+        const float2 * y2 = (const float2 *) y;
        const float2 * x2 = (const float2 *) x;
        [[maybe_unused]] const float2 * gate_x2 = nullptr;
        if constexpr (has_fusion) {
@@ -160,6 +181,7 @@ static __global__ void mul_mat_vec_f(
            }
        }
    } else if constexpr (std::is_same_v<T, half>) {
+        const float2 * y2 = (const float2 *) y;
        const half2 * x2 = (const half2 *) x;
        [[maybe_unused]] const half2 * gate_x2 = nullptr;
        if constexpr (has_fusion) {
@@ -235,6 +257,7 @@ static __global__ void mul_mat_vec_f(
 #endif // FP16_AVAILABLE
        }
    } else if constexpr (std::is_same_v<T, nv_bfloat16>) {
+        const float2 * y2 = (const float2 *) y;
 //TODO: add support for ggml_cuda_mad for hip_bfloat162
 #if defined(GGML_USE_HIP)
        const int * x2 = (const int *) x;
@@ -301,6 +324,28 @@ static __global__ void mul_mat_vec_f(
            }
        }
 #endif
+    } else if constexpr (std::is_same_v<T, ggml_fp8_e4m3_t>) {
+        const float2 * y2 = (const float2 *) y;
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = mmvf_load_fp8x2(x, col2);
+            float2 tmpx_gate = {};
+            if constexpr (has_fusion) {
+                if (use_gate) {
+                    tmpx_gate = mmvf_load_fp8x2(gate_x, col2);
+                }
+            }
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                ggml_cuda_mad(sumf[j], tmpx, tmpy);
+
+                if constexpr (has_fusion) {
+                    if (use_gate) {
+                        ggml_cuda_mad(sumf_gate[j], tmpx_gate, tmpy);
+                    }
+                }
+            }
+        }
    } else {
        static_assert(std::is_same_v<T, void>, "unsupported type");
    }
@@ -348,12 +393,18 @@ static __global__ void mul_mat_vec_f(
    float value = sumf[tid];

    if constexpr (has_fusion) {
+        if (use_scale) {
+            value *= x_scale[ids ? channel_x : 0];
+        }
        if (use_bias) {
            value += x_bias[tid*stride_col_dst + row];
        }

        if (use_gate) {
            float gate_value = sumf_gate[tid];
+            if (use_gate_scale) {
+                gate_value *= gate_scale[ids ? channel_x : 0];
+            }
            if (use_gate_bias) {
                gate_value += gate_bias[tid*stride_col_dst + row];
            }
@@ -380,7 +431,11 @@ static __global__ void mul_mat_vec_f(
    dst[tid*stride_col_dst + row] = value;

    if constexpr (!has_fusion) {
-        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, glu_op, glu_limit, gate_x, x_bias, gate_bias, sumf_gate);
+        GGML_UNUSED_VARS(use_gate, use_bias, use_gate_bias, use_scale, use_gate_scale, glu_op, glu_limit, gate_x, x_bias, gate_bias,
+            x_scale, gate_scale, sumf_gate);
+    }
+    if constexpr (!std::is_same_v<T, ggml_fp8_e4m3_t>) {
+        GGML_UNUSED_VARS(use_scale, use_gate_scale, x_scale, gate_scale);
    }
 }

@@ -395,7 +450,8 @@ static void mul_mat_vec_f_switch_fusion(

    const ggml_cuda_kernel_launch_params launch_params = {block_nums, block_dims, nbytes_shared, stream};

-    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr ||
+        fusion.x_scale != nullptr || fusion.gate_scale != nullptr;
    if constexpr (ncols_dst == 1) {
        if (has_fusion) {
            ggml_cuda_kernel_launch(mul_mat_vec_f<T, type_acc, ncols_dst, block_size, true, is_multi_token_id>, launch_params,
@@ -616,7 +672,7 @@ static void mul_mat_vec_f_cuda(
        const int64_t ids_stride, enum ggml_prec prec, cudaStream_t stream) {

    if constexpr(std::is_same_v<T, half>) {
-        if (prec == GGML_PREC_DEFAULT) {
+        if (prec == GGML_PREC_UNDEFINED) {
            mul_mat_vec_f_cuda_switch_ncols_dst<T, half>
                (x, y, ids, fusion, dst, ncols, nrows, ncols_dst, stride_row, stride_col_y, stride_col_dst,
                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
@@ -649,9 +705,11 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
    GGML_ASSERT(        nb10       == ts_src1);
    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(src0->type != GGML_TYPE_F8_E4M3 || (uintptr_t) src0->data % alignof(uint16_t) == 0);

    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
-    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+    // the fused GLU path gets the GLU node as dst, which carries no matmul precision, so it accumulates in F32
+    const enum ggml_prec prec = fast_fp16_available(cc) && !(fusion && fusion->gate) ? ggml_prec_get_acc(dst) : GGML_PREC_F32;

    const float   * src1_d =       (const float   *) src1->data;
    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -662,6 +720,7 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
    if (fusion) {
        GGML_ASSERT( !ids || dst->ne[2] == 1);
        GGML_ASSERT(  ids || dst->ne[1] == 1);
+        GGML_ASSERT((fusion->x_scale == nullptr && fusion->gate_scale == nullptr) || src0->type == GGML_TYPE_F8_E4M3);
        if (fusion->x_bias) {
            GGML_ASSERT(fusion->x_bias->type == GGML_TYPE_F32);
            GGML_ASSERT(fusion->x_bias->ne[0] == dst->ne[0]);
@@ -670,6 +729,7 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
        }
        if (fusion->gate) {
            GGML_ASSERT(fusion->gate->type == src0->type && ggml_are_same_stride(fusion->gate, src0));
+            GGML_ASSERT(src0->type != GGML_TYPE_F8_E4M3 || (uintptr_t) fusion->gate->data % alignof(uint16_t) == 0);
            fusion_local.gate = fusion->gate->data;
        }
        if (fusion->gate_bias) {
@@ -678,6 +738,18 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
            GGML_ASSERT(!ids || fusion->gate_bias->ne[1] == src0->ne[2]);
            fusion_local.gate_bias = fusion->gate_bias->data;
        }
+        if (fusion->x_scale) {
+            GGML_ASSERT(fusion->x_scale->type == GGML_TYPE_F32);
+            GGML_ASSERT(ggml_is_contiguous(fusion->x_scale));
+            GGML_ASSERT(ggml_nelements(fusion->x_scale) == (ids ? src0->ne[2] : 1));
+            fusion_local.x_scale = fusion->x_scale->data;
+        }
+        if (fusion->gate_scale) {
+            GGML_ASSERT(fusion->gate_scale->type == GGML_TYPE_F32);
+            GGML_ASSERT(ggml_is_contiguous(fusion->gate_scale));
+            GGML_ASSERT(ggml_nelements(fusion->gate_scale) == (ids ? src0->ne[2] : 1));
+            fusion_local.gate_scale = fusion->gate_scale->data;
+        }
        fusion_local.glu_op = fusion->glu_op;
        fusion_local.glu_limit = fusion->glu_limit;
    }
@@ -722,6 +794,12 @@ void ggml_cuda_mul_mat_vec_f(ggml_backend_cuda_context & ctx, const ggml_tensor
                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
                ne03,              ne3,           s03, s13,              s3,                 ids_stride, prec, ctx.stream());
        } break;
+        case GGML_TYPE_F8_E4M3: {
+            const ggml_fp8_e4m3_t * src0_d = (const ggml_fp8_e4m3_t *) src0->data;
+            mul_mat_vec_f_cuda(src0_d, src1_d, ids_d, fusion_local, dst_d, ne00, ne01, ncols_dst, s01, stride_col_y, stride_col_dst,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03, ne3, s03, s13, s3, ids_stride, prec, ctx.stream());
+        } break;
        default:
            GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
    }
@@ -743,7 +821,7 @@ void ggml_cuda_op_mul_mat_vec_f(

    const int id = ggml_cuda_get_device();
    const int cc = ggml_cuda_info().devices[id].cc;
-    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec_get_acc(dst) : GGML_PREC_F32;

    // ggml_cuda_op provides single, contiguous matrices
    const int64_t stride_row         = ne00;
@@ -788,7 +866,11 @@ void ggml_cuda_op_mul_mat_vec_f(
    GGML_UNUSED_VARS(ctx, src1, dst, src1_ddq_i, src1_ncols, src1_padded_row_size);
 }

-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11) {
+bool ggml_cuda_should_use_mmvf(const ggml_tensor * src0, int cc, int warp_size, int64_t ne11) {
+    const ggml_type type = src0->type;
+    const int64_t * src0_ne = src0->ne;
+    const size_t * src0_nb = src0->nb;
+
    if (src0_ne[0] % 2 != 0) {
        return false;
    }
@@ -798,11 +880,8 @@ bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const
        return false;
    }

-    // Pointers not aligned to the size of half2/nv_bfloat162/float2 would result in a crash:
-    for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
-        if (src0_nb[i] % (2*ts) != 0) {
-            return false;
-        }
+    if (!ggml_cuda_is_aligned(src0, 2*ts)) {
+        return false;
    }

    switch (type) {
@@ -862,6 +941,18 @@ bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const
                if (src0_small && !ggml_cuda_should_use_mmf(type, cc, warp_size, src0_ne, src0_nb, ne11, /*mul_mat_id =*/ false)) {
                    return ne11 <= MMVF_MAX_BATCH_SIZE;
                }
+                // MMF tiles 32 rows per block, so a matrix of few rows leaves most SMs idle while MMVF runs a block
+                // per row: with a long enough K, MMVF wins for the small batches of speculative decoding as long as
+                // its work (rows x batch) stays small relative to the SM count, with a budget that grows with K
+                if (src0_small && src0_ne[0] >= 1024) {
+                    const bool    blackwell = cc >= GGML_CUDA_CC_BLACKWELL;
+                    const int     nsm       = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
+                    const int64_t max_ncols = blackwell ? MMVF_MAX_BATCH_SIZE : 5;
+                    const int64_t max_k     = blackwell ? 6144 : 4096;
+                    if (ne11 <= max_ncols && src0_ne[1]*ne11*256 <= nsm*std::min(src0_ne[0], max_k)) {
+                        return true;
+                    }
+                }
                if (ampere_mma_available(cc)) {
                    return src0_small && ne11 == 1;
                }
@@ -882,6 +973,8 @@ bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const
                return ne11 <= MMVF_MAX_BATCH_SIZE;
            }
            return ne11 <= MMVF_MAX_BATCH_SIZE;
+        case GGML_TYPE_F8_E4M3:
+            return ne11 <= MMVF_MAX_BATCH_SIZE;
        default:
            return false;
    }
@@ -11,4 +11,4 @@ void ggml_cuda_op_mul_mat_vec_f(
    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
    const int64_t src1_padded_row_size, cudaStream_t stream);

-bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const int64_t * src0_ne, const size_t * src0_nb, int64_t ne11);
+bool ggml_cuda_should_use_mmvf(const ggml_tensor * src0, int cc, int warp_size, int64_t ne11);
@@ -1555,7 +1555,7 @@ void ggml_cuda_mul_mat_vec_q(
        const int64_t s11 = src1->nb[1] / ts_src1;
        const int64_t s12 = src1->nb[2] / ts_src1;
        const int64_t s13 = src1->nb[3] / ts_src1;
-        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
    }

    const int64_t s01 = src0->nb[1] / ts_src0;
@@ -557,10 +557,9 @@ static __global__ void quantize_mmq_q8_1(
 }

 void quantize_row_q8_1_cuda(
-        const float * x, const int32_t * ids, void * vy, const ggml_type type_src0,
+        const float * x, void * vy,
        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
        const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
-    GGML_ASSERT(!ids);
    GGML_ASSERT(ne0 % QK8_1 == 0);

    const uint3 ne2_fastdiv = init_fastdiv_values(ne2);
@@ -570,7 +569,6 @@ void quantize_row_q8_1_cuda(
    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(num_blocks, block_size, 0, stream);
    ggml_cuda_kernel_launch(quantize_q8_1, launch_params, x, vy, ne00, s01, s02, s03, ne0, ne1, ne2_fastdiv);
-    GGML_UNUSED(type_src0);
 }

 void quantize_mmq_q8_1_cuda(
@@ -11,14 +11,9 @@
 static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk of out-of-bounds access.");
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");

-typedef void (*quantize_cuda_t)(
-        const float * x, const int32_t * ids, void * vy,
-        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
-        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
-
 void quantize_row_q8_1_cuda(
-        const float * x, const int32_t * ids, void * vy,
-        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        const float * x, void * vy,
+        int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

 void quantize_mmq_q8_1_cuda(
@@ -257,6 +257,16 @@ static void set_rows_cuda(ggml_backend_cuda_context & ctx, const ggml_tensor * s
            nb1, nb2, nb3,
            stream
        );
+    } else if (dst->type == GGML_TYPE_F8_E4M3) {
+        set_rows_cuda(
+            src0_d, src1_d, (ggml_fp8_e4m3_t *) dst->data,
+            ne00, ne01, ne02, ne03,
+            ne10, ne11, ne12, ne13,
+            nb01, nb02, nb03,
+            nb10, nb11, nb12,
+            nb1, nb2, nb3,
+            stream
+        );
    } else if (dst->type == GGML_TYPE_Q4_0) {
        set_rows_cuda_quant<idx_t, block_q4_0, QK4_0, quantize_f32_q4_0_block>(
            src0_d, src1_d, (block_q4_0*)dst->data,
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 1, 8);
 DECL_FATTN_MMA_F16_CASE(64, 64, 1, 8);
+DECL_FATTN_MMA_F16_CASE(72, 72, 1, 8);
 DECL_FATTN_MMA_F16_CASE(80, 80, 1, 8);
 DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 16, 1);
 DECL_FATTN_MMA_F16_CASE(64, 64, 16, 1);
+DECL_FATTN_MMA_F16_CASE(72, 72, 16, 1);
 DECL_FATTN_MMA_F16_CASE(80, 80, 16, 1);
 DECL_FATTN_MMA_F16_CASE(96, 96, 16, 1);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 1);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 16, 2);
 DECL_FATTN_MMA_F16_CASE(64, 64, 16, 2);
+DECL_FATTN_MMA_F16_CASE(72, 72, 16, 2);
 DECL_FATTN_MMA_F16_CASE(80, 80, 16, 2);
 DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 16, 4);
 DECL_FATTN_MMA_F16_CASE(64, 64, 16, 4);
+DECL_FATTN_MMA_F16_CASE(72, 72, 16, 4);
 DECL_FATTN_MMA_F16_CASE(80, 80, 16, 4);
 DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 2, 4);
 DECL_FATTN_MMA_F16_CASE(64, 64, 2, 4);
+DECL_FATTN_MMA_F16_CASE(72, 72, 2, 4);
 DECL_FATTN_MMA_F16_CASE(80, 80, 2, 4);
 DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 2, 8);
 DECL_FATTN_MMA_F16_CASE(64, 64, 2, 8);
+DECL_FATTN_MMA_F16_CASE(72, 72, 2, 8);
 DECL_FATTN_MMA_F16_CASE(80, 80, 2, 8);
 DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
@@ -2,7 +2,9 @@

 #include "../fattn-mma-f16.cuh"

+DECL_FATTN_MMA_F16_CASE(40, 40, 32, 1);
 DECL_FATTN_MMA_F16_CASE(64, 64, 32, 1);
+DECL_FATTN_MMA_F16_CASE(72, 72, 32, 1);
 DECL_FATTN_MMA_F16_CASE(80, 80, 32, 1);
 DECL_FATTN_MMA_F16_CASE(96, 96, 32, 1);
 DECL_FATTN_MMA_F16_CASE(112, 112, 32, 1);
--- a/Show More
+++ b/Show More