server: Bypass API Key validation for WebUI static bundle assets (#21269)

* fix: Bypass API Key validation for static bundle assets

* refactor: All bypassed routes in `public_endpoints`

* test: Update static assets API Key test

.devops/cpu.Dockerfile

@@ -1,11 +1,13 @@
-ARG UBUNTU_VERSION=22.04
+ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libssl-dev
+    apt-get install -y gcc-14 g++-14 build-essential git cmake libssl-dev
+
+ENV CC=gcc-14 CXX=g++-14
 
 WORKDIR /app
 
@@ -34,7 +36,7 @@ RUN mkdir -p /app/full \
 FROM ubuntu:$UBUNTU_VERSION AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -55,8 +57,9 @@ RUN apt-get update \
     git \
     python3 \
     python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
+    python3-wheel \
+    && pip install --break-system-packages --upgrade setuptools \
+    && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/cuda-new.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=24.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=13.1.0
+ARG CUDA_VERSION=13.1.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
@@ -12,7 +12,9 @@ FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y gcc-14 g++-14 build-essential cmake python3 python3-pip git libssl-dev libgomp1
+
+ENV CC=gcc-14 CXX=g++-14 CUDAHOSTCXX=g++-14
 
 WORKDIR /app
 
@@ -39,7 +41,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/cuda.Dockerfile

@@ -1,6 +1,6 @@
-ARG UBUNTU_VERSION=22.04
+ARG UBUNTU_VERSION=24.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.4.0
+ARG CUDA_VERSION=12.8.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
@@ -12,7 +12,9 @@ FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libssl-dev libgomp1
+    apt-get install -y gcc-14 g++-14 build-essential cmake python3 python3-pip git libssl-dev libgomp1
+
+ENV CC=gcc-14 CXX=g++-14 CUDAHOSTCXX=g++-14
 
 WORKDIR /app
 
@@ -39,7 +41,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_CUDA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -60,7 +62,8 @@ RUN apt-get update \
     git \
     python3 \
     python3-pip \
-    && pip install --upgrade pip setuptools wheel \
+    python3-wheel \
+    && pip install --break-system-packages --upgrade setuptools \
     && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \

.devops/intel.Dockerfile

@@ -33,8 +33,25 @@ RUN mkdir -p /app/full \
 
 FROM intel/deep-learning-essentials:$ONEAPI_VERSION AS base
 
+ARG IGC_VERSION=v2.30.1
+ARG IGC_VERSION_FULL=2_2.30.1+20950
+ARG COMPUTE_RUNTIME_VERSION=26.09.37435.1
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.09.37435.1-0
+ARG IGDGMM_VERSION=22.9.0
+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 \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-ocloc-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-ocloc_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-opencl-icd-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/intel-opencl-icd_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libigdgmm12_${IGDGMM_VERSION}_amd64.deb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libze-intel-gpu1-dbgsym_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.ddeb \
+  && wget https://github.com/intel/compute-runtime/releases/download/$COMPUTE_RUNTIME_VERSION/libze-intel-gpu1_${COMPUTE_RUNTIME_VERSION_FULL}_amd64.deb \
+  && dpkg --install *.deb
+
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/musa.Dockerfile

@@ -46,7 +46,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_MUSA_RUN_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/nix/package.nix

@@ -41,6 +41,7 @@
   effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
   enableStatic ? effectiveStdenv.hostPlatform.isStatic,
   precompileMetalShaders ? false,
+  useWebUi ? true,
 }:
 
 let
@@ -164,6 +165,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
   cmakeFlags =
     [
       (cmakeBool "LLAMA_BUILD_SERVER" true)
+      (cmakeBool "LLAMA_BUILD_WEBUI" useWebUi)
       (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
       (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
       (cmakeBool "GGML_NATIVE" false)

.devops/openvino.Dockerfile

@@ -78,7 +78,7 @@ ARG http_proxy
 ARG https_proxy
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libtbb12 curl\
+    && apt-get install -y libgomp1 libtbb12 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.devops/rocm.Dockerfile

@@ -58,7 +58,7 @@ RUN mkdir -p /app/full \
 FROM ${BASE_ROCM_DEV_CONTAINER} AS base
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 curl\
+    && apt-get install -y libgomp1 curl \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
@@ -79,7 +79,7 @@ RUN apt-get update \
     git \
     python3-pip \
     python3 \
-    python3-wheel\
+    python3-wheel \
     && pip install --break-system-packages --upgrade setuptools \
     && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \

.devops/vulkan.Dockerfile

@@ -49,17 +49,20 @@ COPY --from=build /app/full /app
 
 WORKDIR /app
 
+ENV PATH="/root/.venv/bin:/root/.local/bin:${PATH}"
+
+# Flag for compatibility with pip
+ARG UV_INDEX_STRATEGY="unsafe-best-match"
 RUN apt-get update \
     && apt-get install -y \
     build-essential \
+    curl \
     git \
-    python3.13 \
-    python3.13-dev \
-    python3-pip \
-    python3-wheel \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.13 100 \
-    && pip install --break-system-packages --upgrade setuptools \
-    && pip install --break-system-packages -r requirements.txt \
+    ca-certificates \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh \
+    && uv python install 3.13 \
+    && uv venv --python 3.13 /root/.venv \
+    && uv pip install --python /root/.venv/bin/python -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.editorconfig

@@ -21,14 +21,6 @@ indent_style = tab
 [prompts/*.txt]
 insert_final_newline = unset
 
-[tools/server/public/*]
-indent_size = 2
-
-[tools/server/public/deps_*]
-trim_trailing_whitespace = unset
-indent_style = unset
-indent_size = unset
-
 [tools/server/deps_*]
 trim_trailing_whitespace = unset
 indent_style = unset
@@ -61,6 +53,14 @@ charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
 
+[tools/server/public/**]
+indent_style = unset
+indent_size = unset
+end_of_line = unset
+charset = unset
+trim_trailing_whitespace = unset
+insert_final_newline = unset
+
 [benches/**]
 indent_style = unset
 indent_size = unset

.gitattributes

@@ -0,0 +1,4 @@
+# Treat the generated single-file WebUI build as binary for diff purposes.
+# Git's pack-file delta compression still works (byte-level), but this prevents
+# git diff from printing the entire minified file on every change.
+tools/server/public/index.html -diff

.github/workflows/build.yml

@@ -150,16 +150,15 @@ jobs:
       - name: Dawn Dependency
         id: dawn-depends
         run: |
-          DAWN_VERSION="v2.0.0"
-          DAWN_OWNER="reeselevine"
+          DAWN_VERSION="v20260317.182325"
+          DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
-          curl -L -o artifact.zip \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           mkdir dawn
-          unzip artifact.zip
-          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -181,7 +180,7 @@ jobs:
           - build: 'x64'
             os: ubuntu-22.04
           - build: 'arm64'
-            os: ubuntu-22.04-arm
+            os: ubuntu-24.04-arm
           - build: 's390x'
             os: ubuntu-24.04-s390x
           - build: 'ppc64le'
@@ -207,14 +206,22 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y --no-install-recommends \
-            python3 python3-pip python3-dev \
+            python3 python3-pip python3-dev python3-wheel \
             libjpeg-dev build-essential libssl-dev \
             git-lfs
 
+      - name: Toolchain workaround (GCC 14)
+        if: ${{ contains(matrix.os, 'ubuntu-24.04') }}
+        run: |
+          sudo apt-get install -y gcc-14 g++-14
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
       - name: Python Dependencies
         id: python_depends
         run: |
-          python3 -m pip install --upgrade pip
+          export PIP_BREAK_SYSTEM_PACKAGES="1"
+          python3 -m pip install --upgrade pip setuptools
           pip3 install ./gguf-py
 
       - name: Swap Endianness
@@ -292,7 +299,15 @@ jobs:
           ctest -L main --verbose
 
   ubuntu-24-vulkan:
-    runs-on: ${{ 'ubuntu-24.04-arm' || 'ubuntu-24.04' }}
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-24.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
+
+    runs-on: ${{ matrix.os }}
 
     steps:
       - name: Clone
@@ -302,7 +317,10 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          sudo apt-get install -y glslc libvulkan-dev libssl-dev ninja-build
+          sudo apt-get update
+          sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev libssl-dev ninja-build
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
 
       - name: Configure
         id: cmake_configure
@@ -365,16 +383,15 @@ jobs:
         id: dawn-depends
         run: |
           sudo apt-get install -y libxrandr-dev libxinerama-dev libxcursor-dev mesa-common-dev libx11-xcb-dev libxi-dev
-          DAWN_VERSION="v2.0.0"
-          DAWN_OWNER="reeselevine"
+          DAWN_VERSION="v20260317.182325"
+          DAWN_OWNER="google"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
-          curl -L -o artifact.zip \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+          echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+          curl -L -o artifact.tar.gz \
+            "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
           mkdir dawn
-          unzip artifact.zip
-          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
+          tar -xvf artifact.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -408,7 +425,7 @@ jobs:
 
       - name: Fetch emdawnwebgpu
         run: |
-          DAWN_TAG="v20251027.212519"
+          DAWN_TAG="v20260317.182325"
           EMDAWN_PKG="emdawnwebgpu_pkg-${DAWN_TAG}.zip"
           echo "Downloading ${EMDAWN_PKG}"
           curl -L -o emdawn.zip \

.github/workflows/docker.yml

@@ -25,186 +25,13 @@ permissions:
   packages: write
 
 jobs:
-  push_to_registry:
-    name: Push Docker image to Docker Hub
-
-    runs-on: ${{ matrix.config.runs_on }}
-    env:
-      COMMIT_SHA: ${{ github.sha }}
-    strategy:
-      fail-fast: false
-      matrix:
-        config:
-          # Multi-stage build
-          # Note: the arm64 images are failing, which prevents the amd64 images from being built
-          # https://github.com/ggml-org/llama.cpp/issues/11888
-          #- { tag: "cpu", dockerfile: ".devops/cpu.Dockerfile", platforms: "linux/amd64,linux/arm64", full: true, light: true, server: true, free_disk_space: false }
-          - { tag: "cpu",    dockerfile: ".devops/cpu.Dockerfile",    platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "cuda cuda12", dockerfile: ".devops/cuda.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "12.4.0", ubuntu_version: "22.04" }
-          - { tag: "cuda13", dockerfile: ".devops/cuda-new.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04", cuda_version: "13.1.0", ubuntu_version: "24.04" }
-          - { tag: "musa",   dockerfile: ".devops/musa.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "intel",  dockerfile: ".devops/intel.Dockerfile",  platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "vulkan", dockerfile: ".devops/vulkan.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-          - { tag: "s390x",  dockerfile: ".devops/s390x.Dockerfile",  platforms: "linux/s390x", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04-s390x" }
-          - { tag: "rocm",   dockerfile: ".devops/rocm.Dockerfile",   platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: true,  runs_on: "ubuntu-22.04" }
-          - { tag: "openvino", dockerfile: ".devops/openvino.Dockerfile", platforms: "linux/amd64", full: true, light: true, server: true, free_disk_space: false, runs_on: "ubuntu-22.04" }
-    steps:
-      - name: Check out the repo
-        uses: actions/checkout@v6
-        with:
-          fetch-depth: 0 # preserve git history, so we can determine the build number
-
-      - name: Set up QEMU
-        if: ${{ matrix.config.tag != 's390x' }}
-        uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3
-        with:
-          image: tonistiigi/binfmt:qemu-v7.0.0-28
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
-
-      - name: Log in to Docker Hub
-        uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.repository_owner }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Determine source tag name
-        id: srctag
-        uses: ./.github/actions/get-tag-name
-        env:
-          BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-
-      - name: Determine image tag name
-        id: tag
-        shell: bash
-        run: |
-          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
-          REPO_NAME="${{ github.event.repository.name }}"
-          PREFIX="ghcr.io/${REPO_OWNER}/${REPO_NAME}:"
-
-          # list all tags possible
-          tags="${{ matrix.config.tag }}"
-          for tag in $tags; do
-              if [[ "$tag" == "cpu" ]]; then
-                  TYPE=""
-              else
-                  TYPE="-$tag"
-              fi
-              CACHETAGS="${PREFIX}buildcache${TYPE}"
-              FULLTAGS="${FULLTAGS:+$FULLTAGS,}${PREFIX}full${TYPE},${PREFIX}full${TYPE}-${{ steps.srctag.outputs.name }}"
-              LIGHTTAGS="${LIGHTTAGS:+$LIGHTTAGS,}${PREFIX}light${TYPE},${PREFIX}light${TYPE}-${{ steps.srctag.outputs.name }}"
-              SERVERTAGS="${SERVERTAGS:+$SERVERTAGS,}${PREFIX}server${TYPE},${PREFIX}server${TYPE}-${{ steps.srctag.outputs.name }}"
-          done
-          echo "cache_output_tags=$CACHETAGS" >> $GITHUB_OUTPUT
-          echo "full_output_tags=$FULLTAGS" >> $GITHUB_OUTPUT
-          echo "light_output_tags=$LIGHTTAGS" >> $GITHUB_OUTPUT
-          echo "server_output_tags=$SERVERTAGS" >> $GITHUB_OUTPUT
-          echo "cache_output_tags=$CACHETAGS"  # print out for debugging
-          echo "full_output_tags=$FULLTAGS"  # print out for debugging
-          echo "light_output_tags=$LIGHTTAGS"  # print out for debugging
-          echo "server_output_tags=$SERVERTAGS"  # print out for debugging
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
-      - name: Free Disk Space (Ubuntu)
-        if: ${{ matrix.config.free_disk_space == true }}
-        uses: ggml-org/free-disk-space@v1.3.1
-        with:
-          # this might remove tools that are actually needed,
-          # if set to "true" but frees about 6 GB
-          tool-cache: false
-
-          # all of these default to true, but feel free to set to
-          # "false" if necessary for your workflow
-          android: true
-          dotnet: true
-          haskell: true
-          large-packages: true
-          docker-images: true
-          swap-storage: true
-
-      - name: Build and push Full Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.full == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.full_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: full
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
-      - name: Build and push Light Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.light_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: light
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
-      - name: Build and push Server Docker image (tagged + versioned)
-        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
-        uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.server_output_tags }}
-          file: ${{ matrix.config.dockerfile }}
-          target: server
-          provenance: false
-          build-args: |
-            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
-            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
-          # using github experimental cache
-          #cache-from: type=gha
-          #cache-to: type=gha,mode=max
-          # return to this if the experimental github cache is having issues
-          #cache-to: type=local,dest=/tmp/.buildx-cache
-          #cache-from: type=local,src=/tmp/.buildx-cache
-          # using registry cache (no storage limit)
-          cache-from: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }}
-          cache-to: type=registry,ref=${{ steps.tag.outputs.cache_output_tags }},mode=max
-
   create_tag:
     name: Create and push git tag
-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-slim
     permissions:
       contents: write
+    outputs:
+      source_tag: ${{ steps.srctag.outputs.name }}
 
     steps:
       - name: Clone
@@ -225,3 +52,391 @@ jobs:
         run: |
           git tag ${{ steps.srctag.outputs.name }} || exit 0
           git push origin ${{ steps.srctag.outputs.name }} || exit 0
+
+  prepare_matrices:
+    name: Prepare Docker matrices
+    runs-on: ubuntu-24.04
+    outputs:
+      build_matrix: ${{ steps.matrices.outputs.build_matrix }}
+      merge_matrix: ${{ steps.matrices.outputs.merge_matrix }}
+
+    steps:
+      - name: Generate build and merge matrices
+        id: matrices
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          # Keep all build targets in one place and derive merge targets from it.
+          cat > build-matrix.json <<'JSON'
+          [
+            { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
+            { "tag": "cpu", "dockerfile": ".devops/cpu.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cpu", "dockerfile": ".devops/s390x.Dockerfile", "platforms": "linux/s390x", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-s390x" },
+            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda cuda12", "dockerfile": ".devops/cuda.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda-new.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "cuda13", "dockerfile": ".devops/cuda-new.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "musa", "dockerfile": ".devops/musa.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "intel", "dockerfile": ".devops/intel.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" },
+            { "tag": "vulkan", "dockerfile": ".devops/vulkan.Dockerfile", "platforms": "linux/arm64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04-arm" },
+            { "tag": "rocm", "dockerfile": ".devops/rocm.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": true, "runs_on": "ubuntu-24.04" },
+            { "tag": "openvino", "dockerfile": ".devops/openvino.Dockerfile", "platforms": "linux/amd64", "full": true, "light": true, "server": true, "free_disk_space": false, "runs_on": "ubuntu-24.04" }
+          ]
+          JSON
+
+          BUILD_MATRIX="$(jq -c . build-matrix.json)"
+          MERGE_MATRIX="$(jq -c '
+            reduce .[] as $entry ({}; .[$entry.tag] |= (
+              . // {
+                tag: $entry.tag,
+                arches: [],
+                full: false,
+                light: false,
+                server: false
+              }
+              | .full = (.full or ($entry.full // false))
+              | .light = (.light or ($entry.light // false))
+              | .server = (.server or ($entry.server // false))
+              | .arches += [($entry.platforms | sub("^linux/"; ""))]
+            ))
+            # Backward compatibility: s390x tags are aliases of cpu for the linux/s390x platform.
+            | if (has("cpu") and (((.cpu.arches // []) | index("s390x")) != null)) then
+                . + {
+                  s390x: {
+                    tag: "s390x",
+                    arches: ["s390x"],
+                    full: .cpu.full,
+                    light: .cpu.light,
+                    server: .cpu.server
+                  }
+                }
+              else
+                .
+              end
+            | [.[] | .arches = (.arches | unique | sort | join(" "))]
+          ' build-matrix.json)"
+
+          echo "build_matrix=$BUILD_MATRIX" >> "$GITHUB_OUTPUT"
+          echo "merge_matrix=$MERGE_MATRIX" >> "$GITHUB_OUTPUT"
+
+  push_to_registry:
+    name: Push Docker image to Docker Registry
+    needs: [prepare_matrices, create_tag]
+
+    runs-on: ${{ matrix.config.runs_on }}
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.prepare_matrices.outputs.build_matrix) }}
+    steps:
+      - name: Check out the repo
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          ref: ${{ needs.create_tag.outputs.source_tag }}
+
+      - name: Set up QEMU
+        if: ${{ contains(matrix.config.platforms, 'linux/amd64') }}
+        uses: docker/setup-qemu-action@ce360397dd3f832beb865e1373c09c0e9f86d70a # v4
+        with:
+          image: tonistiigi/binfmt:qemu-v10.2.1
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4
+
+      - name: Log in to Docker Registry
+        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Determine image metadata
+        id: meta
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
+          PREFIX="${IMAGE_REPO}:"
+          PLATFORM="${{ matrix.config.platforms }}"
+          ARCH_SUFFIX="${PLATFORM#linux/}"
+
+          # list all tags possible
+          tags="${{ matrix.config.tag }}"
+          for tag in $tags; do
+              if [[ "$tag" == "cpu" ]]; then
+                  TYPE=""
+              else
+                  TYPE="-$tag"
+              fi
+              CACHETAG="${PREFIX}buildcache${TYPE}-${ARCH_SUFFIX}"
+          done
+
+          SAFE_TAGS="$(echo "$tags" | tr ' ' '_')"
+
+          echo "image_repo=$IMAGE_REPO" >> $GITHUB_OUTPUT
+          echo "arch_suffix=$ARCH_SUFFIX" >> $GITHUB_OUTPUT
+          echo "cache_output_tag=$CACHETAG" >> $GITHUB_OUTPUT
+          echo "digest_artifact_suffix=${SAFE_TAGS}-${ARCH_SUFFIX}" >> $GITHUB_OUTPUT
+          echo "cache_output_tag=$CACHETAG"  # print out for debugging
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Free Disk Space (Ubuntu)
+        if: ${{ matrix.config.free_disk_space == true }}
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: false
+
+          # all of these default to true, but feel free to set to
+          # "false" if necessary for your workflow
+          android: true
+          dotnet: true
+          haskell: true
+          large-packages: true
+          docker-images: true
+          swap-storage: true
+
+      - name: Build and push Full Docker image by digest
+        id: build_full
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.full == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: full
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Build and push Light Docker image by digest
+        id: build_light
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.light == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: light
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Build and push Server Docker image by digest
+        id: build_server
+        if: ${{ (github.event_name == 'push' || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && matrix.config.server == true }}
+        uses: docker/build-push-action@d08e5c354a6adb9ed34480a06d141179aa583294 # v7
+        with:
+          context: .
+          platforms: ${{ matrix.config.platforms }}
+          outputs: type=image,name=${{ steps.meta.outputs.image_repo }},push-by-digest=true,name-canonical=true,push=true
+          file: ${{ matrix.config.dockerfile }}
+          target: server
+          provenance: false
+          build-args: |
+            ${{ matrix.config.ubuntu_version && format('UBUNTU_VERSION={0}', matrix.config.ubuntu_version) || '' }}
+            ${{ matrix.config.cuda_version && format('CUDA_VERSION={0}', matrix.config.cuda_version) || '' }}
+          # using github experimental cache
+          #cache-from: type=gha
+          #cache-to: type=gha,mode=max
+          # return to this if the experimental github cache is having issues
+          #cache-to: type=local,dest=/tmp/.buildx-cache
+          #cache-from: type=local,src=/tmp/.buildx-cache
+          # using registry cache (no storage limit)
+          cache-from: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }}
+          cache-to: type=registry,ref=${{ steps.meta.outputs.cache_output_tag }},mode=max
+
+      - name: Export digest metadata
+        shell: bash
+        run: |
+            set -euo pipefail
+
+            TAGS="${{ matrix.config.tag }}"
+            ARCH_SUFFIX="${{ steps.meta.outputs.arch_suffix }}"
+            DIGEST_FILE="/tmp/digests/${{ steps.meta.outputs.digest_artifact_suffix }}.tsv"
+            mkdir -p /tmp/digests
+
+            add_digest_rows() {
+                local image_type="$1"
+                local digest="$2"
+
+                if [[ -z "$digest" ]]; then
+                  echo "Missing digest for image_type=${image_type}" >&2
+                  exit 1
+                fi
+
+                for tag in $TAGS; do
+                    printf '%s\t%s\t%s\t%s\n' "$tag" "$ARCH_SUFFIX" "$image_type" "$digest" >> "$DIGEST_FILE"
+                done
+            }
+
+            if [[ "${{ matrix.config.full }}" == "true" ]]; then
+                add_digest_rows "full" "${{ steps.build_full.outputs.digest }}"
+            fi
+
+            if [[ "${{ matrix.config.light }}" == "true" ]]; then
+                add_digest_rows "light" "${{ steps.build_light.outputs.digest }}"
+            fi
+
+            if [[ "${{ matrix.config.server }}" == "true" ]]; then
+                add_digest_rows "server" "${{ steps.build_server.outputs.digest }}"
+            fi
+
+      - name: Upload digest metadata
+        uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7
+        with:
+          name: digests-${{ steps.meta.outputs.digest_artifact_suffix }}
+          path: /tmp/digests/${{ steps.meta.outputs.digest_artifact_suffix }}.tsv
+          if-no-files-found: error
+
+  merge_arch_tags:
+    name: Create shared tags from digests
+    needs: [prepare_matrices, push_to_registry, create_tag]
+    runs-on: ubuntu-24.04
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.prepare_matrices.outputs.merge_matrix) }}
+
+    steps:
+      - name: Check out the repo
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Download digest metadata
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
+        with:
+          pattern: digests-*
+          path: /tmp/digests
+          merge-multiple: true
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@4d04d5d9486b7bd6fa91e7baf45bbb4f8b9deedd # v4
+
+      - name: Log in to Docker Registry
+        uses: docker/login-action@b45d80f862d83dbcd57f89517bcf500b2ab88fb2 # v4
+        with:
+          registry: ghcr.io
+          username: ${{ github.repository_owner }}
+          password: ${{ secrets.GITHUB_TOKEN }}
+
+      - name: Create tags from digests
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+          IMAGE_REPO="ghcr.io/${REPO_OWNER}/${REPO_NAME}"
+          PREFIX="${IMAGE_REPO}:"
+          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
+          TAGS="${{ matrix.config.tag }}"
+          ARCHES="${{ matrix.config.arches }}"
+          DIGEST_GLOB="/tmp/digests/*.tsv"
+
+          if ! ls ${DIGEST_GLOB} >/dev/null 2>&1; then
+              echo "No digest metadata found in /tmp/digests" >&2
+              exit 1
+          fi
+
+          if [[ -z "$SRC_TAG" ]]; then
+              echo "Missing source tag from create_tag" >&2
+              exit 1
+          fi
+
+          find_digest() {
+              local tag_name="$1"
+              local arch="$2"
+              local image_type="$3"
+              local digest
+
+              digest="$(awk -F '\t' -v t="$tag_name" -v a="$arch" -v i="$image_type" '$1 == t && $2 == a && $3 == i { print $4; exit }' ${DIGEST_GLOB})"
+
+              # Backward compatibility: s390x tags are aliases of cpu for the linux/s390x platform.
+              if [[ -z "$digest" && "$tag_name" == "s390x" && "$arch" == "s390x" ]]; then
+                digest="$(awk -F '\t' -v t="cpu" -v a="$arch" -v i="$image_type" '$1 == t && $2 == a && $3 == i { print $4; exit }' ${DIGEST_GLOB})"
+              fi
+
+              if [[ -z "$digest" ]]; then
+                echo "Missing digest for tag=${tag_name} arch=${arch} image_type=${image_type}" >&2
+                exit 1
+              fi
+
+              echo "$digest"
+          }
+
+          create_manifest_tags() {
+              local image_type="$1"
+              local tag_name="$2"
+              local suffix="$3"
+
+              local merged_tag="${PREFIX}${image_type}${suffix}"
+              local merged_versioned_tag="${merged_tag}-${SRC_TAG}"
+
+              local refs=()
+
+              for arch in $ARCHES; do
+                  local digest
+                  digest="$(find_digest "$tag_name" "$arch" "$image_type")"
+                  refs+=("${IMAGE_REPO}@${digest}")
+              done
+
+              echo "Creating ${merged_tag} from ${refs[*]}"
+              docker buildx imagetools create --tag "${merged_tag}" "${refs[@]}"
+
+              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
+              docker buildx imagetools create --tag "${merged_versioned_tag}" "${refs[@]}"
+          }
+
+          for tag in $TAGS; do
+              if [[ "$tag" == "cpu" ]]; then
+                  TYPE=""
+              else
+                  TYPE="-$tag"
+              fi
+
+              if [[ "${{ matrix.config.full }}" == "true" ]]; then
+                  create_manifest_tags "full" "$tag" "$TYPE"
+              fi
+
+              if [[ "${{ matrix.config.light }}" == "true" ]]; then
+                  create_manifest_tags "light" "$tag" "$TYPE"
+              fi
+
+              if [[ "${{ matrix.config.server }}" == "true" ]]; then
+                  create_manifest_tags "server" "$tag" "$TYPE"
+              fi
+          done
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'

.github/workflows/python-type-check.yml

@@ -31,7 +31,7 @@ jobs:
         uses: actions/setup-python@v6
         with:
           python-version: "3.11"
-          pip-install: -r requirements/requirements-all.txt ty==0.0.24
+          pip-install: -r requirements/requirements-all.txt ty==0.0.26
       # - name: Type-check with Pyright
       #   uses: jakebailey/pyright-action@v2
       #   with:

.github/workflows/release.yml

@@ -131,17 +131,16 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.tar.gz
           name: llama-bin-macos-x64.tar.gz
 
-  ubuntu-22-cpu:
+  ubuntu-cpu:
     strategy:
       matrix:
         include:
           - build: 'x64'
             os: ubuntu-22.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
           - build: 's390x'
             os: ubuntu-24.04-s390x
-          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
-          # - build: 'arm64'
-          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -165,6 +164,13 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential libssl-dev
 
+      - name: Toolchain workaround (GCC 14)
+        if: ${{ contains(matrix.os, 'ubuntu-24.04') }}
+        run: |
+          sudo apt-get install -y gcc-14 g++-14
+          echo "CC=gcc-14" >> "$GITHUB_ENV"
+          echo "CXX=g++-14" >> "$GITHUB_ENV"
+
       - name: Build
         id: cmake_build
         run: |
@@ -194,8 +200,16 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
 
-  ubuntu-22-vulkan:
-    runs-on: ubuntu-22.04
+  ubuntu-vulkan:
+    strategy:
+      matrix:
+        include:
+          - build: 'x64'
+            os: ubuntu-22.04
+          - build: 'arm64'
+            os: ubuntu-24.04-arm
+
+    runs-on: ${{ matrix.os }}
 
     steps:
       - name: Clone
@@ -207,16 +221,23 @@ jobs:
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: ubuntu-22-vulkan
+          key: ubuntu-vulkan-${{ matrix.build }}
           evict-old-files: 1d
 
       - name: Dependencies
         id: depends
         run: |
-          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
-          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libssl-dev
+          if [[ "${{ matrix.os }}" =~ "ubuntu-22.04" ]]; then
+            wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
+            sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+            sudo apt-get update -y
+            sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libssl-dev
+          else
+            sudo apt-get update -y
+            sudo apt-get install -y gcc-14 g++-14 build-essential glslc libvulkan-dev libssl-dev ninja-build
+            echo "CC=gcc-14" >> "$GITHUB_ENV"
+            echo "CXX=g++-14" >> "$GITHUB_ENV"
+          fi
 
       - name: Build
         id: cmake_build
@@ -239,13 +260,13 @@ jobs:
         id: pack_artifacts
         run: |
           cp LICENSE ./build/bin/
-          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
         with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz
-          name: llama-bin-ubuntu-vulkan-x64.tar.gz
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
+          name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
   ubuntu-24-openvino:
     runs-on: ubuntu-24.04
@@ -977,8 +998,8 @@ jobs:
       - windows-sycl
       - windows-hip
       - ubuntu-22-rocm
-      - ubuntu-22-cpu
-      - ubuntu-22-vulkan
+      - ubuntu-cpu
+      - ubuntu-vulkan
       - ubuntu-24-openvino
       - macOS-arm64
       - macOS-x64
@@ -1061,9 +1082,11 @@ jobs:
 
             **Linux:**
             - [Ubuntu x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.tar.gz)
-            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
+            - [Ubuntu arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-arm64.tar.gz)
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
+            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
+            - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
+            - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
 
             **Windows:**

.gitignore

@@ -95,6 +95,8 @@
 # Server Web UI temporary files
 /tools/server/webui/node_modules
 /tools/server/webui/dist
+# we no longer use gz for index.html
+/tools/server/public/index.html.gz
 
 # Python
 

CMakeLists.txt

@@ -108,6 +108,7 @@ option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_TOOLS    "llama: build tools"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
+option(LLAMA_BUILD_WEBUI    "llama: build the embedded Web UI for server"  ON)
 option(LLAMA_TOOLS_INSTALL  "llama: install tools"        ${LLAMA_TOOLS_INSTALL_DEFAULT})
 option(LLAMA_TESTS_INSTALL  "llama: install tests"        ON)
 

ci/run.sh

@@ -151,35 +151,7 @@ fi
 
 if [ -n "${GG_BUILD_KLEIDIAI}" ]; then
     echo ">>===== Enabling KleidiAI support"
-
-    CANDIDATES=(
-        "armv9-a+dotprod+i8mm+sve2"
-        "armv9-a+dotprod+i8mm"
-        "armv8.6-a+dotprod+i8mm"
-        "armv8.2-a+dotprod"
-    )
-    CPU=""
-
-    for cpu in "${CANDIDATES[@]}"; do
-        if echo 'int main(){}' | ${CXX:-c++} -march="$cpu" -x c++ - -c -o /dev/null >/dev/null 2>&1; then
-            CPU="$cpu"
-            break
-        fi
-    done
-
-    if [ -z "$CPU" ]; then
-        echo "ERROR: None of the required ARM baselines (armv9/armv8.6/armv8.2 + dotprod) are supported by this compiler."
-        exit 1
-    fi
-
-    echo ">>===== Using ARM baseline: ${CPU}"
-
-    CMAKE_EXTRA="${CMAKE_EXTRA:+$CMAKE_EXTRA } \
-        -DGGML_NATIVE=OFF \
-        -DGGML_CPU_KLEIDIAI=ON \
-        -DGGML_CPU_AARCH64=ON \
-        -DGGML_CPU_ARM_ARCH=${CPU} \
-        -DBUILD_SHARED_LIBS=OFF"
+    CMAKE_EXTRA="${CMAKE_EXTRA:+$CMAKE_EXTRA } -DGGML_CPU_KLEIDIAI=ON"
 fi
 
 if [ ! -z ${GG_BUILD_BLAS} ]; then

common/arg.cpp

@@ -1079,7 +1079,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.verbose_prompt = true;
         }
-    ));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL}));
     add_opt(common_arg(
         {"--display-prompt"},
         {"--no-display-prompt"},
@@ -2807,6 +2807,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.port = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_PORT"));
+    add_opt(common_arg(
+        {"--reuse-port"},
+        string_format("allow multiple sockets to bind to the same port (default: %s)", params.reuse_port ? "enabled" : "disabled"),
+        [](common_params & params) {
+            params.reuse_port = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_REUSE_PORT"));
     add_opt(common_arg(
         {"--path"}, "PATH",
         string_format("path to serve static files from (default: %s)", params.public_path.c_str()),

common/chat-auto-parser-generator.cpp

@@ -65,7 +65,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
         data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
             foreach_function(inputs.tools, [&](const json & tool) {
                 const auto & function = tool.at("function");
-                auto         schema   = function.at("parameters");
+                auto         schema   = function.contains("parameters") ? function.at("parameters") : json::object();
                 builder.resolve_refs(schema);
             });
             parser.build_grammar(builder, data.grammar_lazy);
@@ -221,7 +221,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_json(parser_build_context
     foreach_function(inputs.tools, [&](const json & tool) {
         const auto & func   = tool.at("function");
         std::string  name   = func.at("name");
-        const auto & schema = func.at("parameters");
+        const auto & schema = func.contains("parameters") ? func.at("parameters") : json::object();
 
         // Build call_id parser based on position (if supported)
         common_peg_parser call_id_section = p.eps();
@@ -282,19 +282,11 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     common_peg_parser tool_choice = p.choice();
 
     foreach_function(inputs.tools, [&](const json & tool) {
-        const auto & func   = tool.at("function");
-        std::string  name   = func.at("name");
-        const auto & params = func.at("parameters");
-
-        if (!params.contains("properties") || !params.at("properties").is_object()) {
-            return;
-        }
-
-        const auto &          properties = params.at("properties");
+        const auto &          func       = tool.at("function");
+        std::string           name       = func.at("name");
+        const auto &          params     = func.contains("parameters") ? func.at("parameters") : json::object();
+        const auto &          properties = params.contains("properties") ? params.at("properties") : json::object();
         std::set<std::string> required;
-        if (params.contains("required") && params.at("required").is_array()) {
-            params.at("required").get_to(required);
-        }
 
         // Build parser for each argument, separating required and optional
         std::vector<common_peg_parser> required_parsers;
@@ -311,17 +303,18 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
                 }
             }
 
-            auto arg = p.tool_arg(
-                p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
-                                arguments.name_suffix) +
-                arguments.value_prefix +
-                (type == "string" ? p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
-                                                                     "tool-" + name + "-arg-" + param_name + "-schema",
-                                                                     param_schema, true)) :
-                                    p.tool_arg_json_value(p.schema(
-                                        p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
-                                        p.space()) +
-                p.tool_arg_close(p.literal(arguments.value_suffix)));
+            auto arg =
+                p.tool_arg(p.tool_arg_open(arguments.name_prefix + p.tool_arg_name(p.literal(param_name)) +
+                                           arguments.name_suffix) +
+                           arguments.value_prefix +
+                           (type == "string" ?
+                                p.tool_arg_string_value(p.schema(p.until(arguments.value_suffix),
+                                                                 "tool-" + name + "-arg-" + param_name + "-schema",
+                                                                 param_schema, true)) :
+                                p.tool_arg_json_value(p.schema(
+                                    p.json(), "tool-" + name + "-arg-" + param_name + "-schema", param_schema, false)) +
+                                    p.space()) +
+                           p.tool_arg_close(p.literal(arguments.value_suffix)));
 
             auto named_arg = p.rule("tool-" + name + "-arg-" + param_name, arg);
             if (is_required) {

common/chat-diff-analyzer.cpp

@@ -287,7 +287,7 @@ void analyze_reasoning::compare_reasoning_presence() {
             return p.literal(reasoning_content) + p.space() + p.optional(p.tag("post", (p.marker() + p.space())) + p.rest());
         });
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
+            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
         });
         // try the more aggressive parse first, if it fails, fall back to the delimiter one
         auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
@@ -297,7 +297,7 @@ void analyze_reasoning::compare_reasoning_presence() {
         if (result.result.success()) {
             if (!result.tags["pre"].empty() && !result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
-                start = trim_whitespace(result.tags["pre"]);
+                start = trim_leading_whitespace(result.tags["pre"]);
                 end   = trim_trailing_whitespace(result.tags["post"]);
             } else if (!result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
@@ -333,7 +333,7 @@ void analyze_reasoning::compare_thinking_enabled() {
     if (left_trimmed.empty() && !diff.right.empty()) {
         if (!right_trimmed.empty() && string_ends_with(comparison->output_B, right_trimmed)) {
             if (start.empty()) {
-                start = right_trimmed;
+                start = trim_leading_whitespace(diff.right);
                 mode  = reasoning_mode::TAG_BASED;
             }
         }
@@ -344,7 +344,7 @@ void analyze_reasoning::compare_thinking_enabled() {
                 if (seg.size() >= 2 && seg[seg.size() - 1].value == left_trimmed && seg[seg.size() - 2].type == segment_type::MARKER) {
                     start = seg[seg.size() - 2].value;
                 }
-                end = left_trimmed;
+                end = trim_trailing_whitespace(diff.left);
                 mode = reasoning_mode::TAG_BASED;
             }
         }
@@ -363,15 +363,23 @@ void analyze_reasoning::compare_thinking_enabled() {
             size_t len = std::min(base.size(), anchor_len);
             std::string anchor = base.substr(base.size() - len);
             auto pos = extended.rfind(anchor);
-            if (pos == std::string::npos || pos + len >= extended.size()) continue;
+            if (pos == std::string::npos || pos + len >= extended.size()) {
+                continue;
+            }
 
             std::string extra = trim_whitespace(extended.substr(pos + len));
-            if (extra.empty()) continue;
+            if (extra.empty()) {
+                continue;
+            }
 
             auto seg = prune_whitespace_segments(segmentize_markers(extra));
             if (seg.size() == 2 && seg[0].type == segment_type::MARKER && seg[1].type == segment_type::MARKER) {
-                if (start.empty()) start = seg[0].value;
-                if (end.empty())   end   = seg[1].value;
+                if (start.empty()) {
+                    start = seg[0].value;
+                }
+                if (end.empty()) {
+                    end   = seg[1].value;
+                }
                 mode = reasoning_mode::TAG_BASED;
                 break;
             }
@@ -423,7 +431,7 @@ void analyze_reasoning::compare_reasoning_scope() {
         LOG_DBG(ANSI_ORANGE "%s: Detected TOOLS_ONLY reasoning mode\n" ANSI_RESET, __func__);
 
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
+            return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
         });
         auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
         if (result.result.success()) {
@@ -516,7 +524,7 @@ analyze_content::analyze_content(const common_chat_template & tmpl, const analyz
         // Take the more promising diff
         std::string pure_content = rdiff.length() > diff_tools.left.length() ? rdiff : diff_tools.left;
         auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
-            return p.tag("pre", p.marker()) + p.space() + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
+            return p.tag("pre", p.marker() + p.space()) + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
         });
         auto result = parser_wrapped.parse_anywhere_and_extract(pure_content);
         start = result.tags["pre"];

common/chat.cpp

@@ -221,7 +221,7 @@ using chat_template_caps = jinja::caps;
 struct common_chat_templates {
     bool add_bos;
     bool add_eos;
-    bool has_explicit_template;  // Model had builtin template or template overridde was specified.
+    bool has_explicit_template;  // Model had builtin template or template overridden was specified.
     std::unique_ptr<common_chat_template> template_default;  // always set (defaults to chatml)
     std::unique_ptr<common_chat_template> template_tool_use;
 };
@@ -971,6 +971,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     auto has_tools           = inputs.tools.is_array() && !inputs.tools.empty();
     auto has_response_format = !inputs.json_schema.is_null() && inputs.json_schema.is_object();
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
+    auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto start           = p.rule("start", p.literal("<|start|>assistant"));
@@ -979,9 +980,19 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
         auto channel         = p.literal("<|channel|>") + (p.literal("commentary") | p.literal("analysis"));
         auto constrain_type  = p.chars("[A-Za-z0-9_-]", 1, -1);
 
-        auto analysis = p.rule("analysis", p.literal("<|channel|>analysis<|message|>") + p.reasoning(content) + end);
+        if (extract_reasoning) {
+            p.rule("analysis", p.literal("<|channel|>analysis<|message|>") + p.reasoning(content) + end);
+        } else {
+            p.rule("analysis", p.content(p.literal("<|channel|>analysis<|message|>") + content + end));
+        }
+
+        auto analysis = p.ref("analysis");
         auto preamble = p.rule("preamble", p.literal("<|channel|>commentary<|message|>") + p.content(content) + end);
         auto final_msg = p.rule("final", p.literal("<|channel|>final<|message|>") + p.content(content));
+
+        // Consume any unsolicited tool calls, e.g. builtin functions
+        auto unsolicited = p.rule("unsolicited", p.atomic(p.optional(channel) + p.literal(" to=") + content + end));
+
         auto any = p.rule("any", preamble | analysis);
 
         if (has_response_format) {
@@ -1025,7 +1036,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
             return p.zero_or_more(start + any) + start + (tool_call | final_msg);
         }
 
-        return p.zero_or_more(start + any) + start + final_msg;
+        return p.zero_or_more(start + any) + start + (final_msg | unsolicited);
     });
 
     data.parser = parser.save();
@@ -1263,11 +1274,12 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
     return data;
 }
 
-// LFM2 format:
-// - Reasoning: <think>{reasoning}</think> (optional, only if enable_thinking is true)
-// - Content: text after reasoning (optional)
-// - Tool calls: <|tool_call_start|>[function_name(arg1="value1", arg2="value2")]<|tool_call_end|>
-// Tool calls can appear multiple times (parallel tool calls)
+// LFM2 format: uses <|tool_list_start|>[...]<|tool_list_end|> in system prompt
+// and <|tool_call_start|>[name(arg="val")]<|tool_call_end|> for tool calls.
+// - Reasoning: <think>{reasoning}</think> (optional)
+// - Content: text before a tool call (optional)
+// - Tool calls: Python-style, e.g. [function_name(arg1="value1", arg2="value2")]
+//   Tool calls can appear multiple times (parallel tool calls supported)
 static common_chat_params common_chat_params_init_lfm2(const common_chat_template &    tmpl,
                                                        const autoparser::generation_params & inputs) {
     common_chat_params data;
@@ -1308,9 +1320,9 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
         if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
-
         auto tool_calls = p.rule("tool-calls",
-            p.trigger_rule("tool-call", p.literal(TOOL_CALL_START) +
+            p.trigger_rule("tool-call",
+                p.literal(TOOL_CALL_START) +
                 p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls) +
                 p.literal(TOOL_CALL_END)
             )
@@ -1338,6 +1350,80 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
             { COMMON_GRAMMAR_TRIGGER_TYPE_WORD, TOOL_CALL_START }
         };
     }
+    return data;
+}
+
+// LFM2.5 format: uses plain "List of tools: [...]" in system prompt, no wrapper tokens.
+// Tool calls are bare [name(arg="val")], though model may optionally emit <|tool_call_start|>.
+// - Reasoning: <think>{reasoning}</think> (optional)
+// - Content: text before a tool call (optional)
+// - Tool calls: Python-style, e.g. [function_name(arg1="value1", arg2="value2")]
+//   Tool calls can appear multiple times (parallel tool calls supported)
+static common_chat_params common_chat_params_init_lfm2_5(const common_chat_template &    tmpl,
+                                                         const autoparser::generation_params & inputs) {
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = true;
+    data.preserved_tokens  = {
+        "<|tool_call_start|>",
+        "<|tool_call_end|>",
+        "<think>",
+        "</think>",
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+
+    const std::string THINK_START     = "<think>";
+    const std::string THINK_END       = "</think>";
+
+    data.thinking_start_tag = THINK_START;
+    data.thinking_end_tag   = THINK_END;
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
+        auto end = p.end();
+
+        auto reasoning = p.eps();
+        if (extract_reasoning && inputs.enable_thinking) {
+            reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
+        }
+
+        if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
+            return generation_prompt + reasoning + p.content(p.rest()) + end;
+        }
+
+        auto tool_calls = p.rule("tool-calls",
+            p.trigger_rule("tool-call",
+                p.python_style_tool_calls(inputs.tools, inputs.parallel_tool_calls)
+            )
+        );
+
+        auto content = p.content(p.until_one_of({"<|tool_call_start|>", "["}));
+        auto maybe_start = p.optional(p.literal("<|tool_call_start|>"));
+        return generation_prompt + reasoning + content + maybe_start + tool_calls + end;
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+        data.grammar      = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto         schema   = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+        foreach_function(inputs.tools, [&](const json & tool) {
+            const std::string name = tool.at("function").at("name");
+            data.grammar_triggers.push_back({ COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "[" + name + "(" });
+        });
+    }
 
     return data;
 }
@@ -1519,14 +1605,21 @@ static std::optional<common_chat_params> try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
-    // LFM2 - uses <|tool_list_start|>/<|tool_list_end|> markers and <|tool_call_start|>[name(args)]<|tool_call_end|> format
-    // Detection: template has "<|tool_list_start|>" and "<|tool_list_end|>" markers
+    // LFM2 format detection: template uses <|tool_list_start|>[...]<|tool_list_end|> around the tool list
+    // and <|tool_call_start|>[...]<|tool_call_end|> around each tool call
     if (src.find("<|tool_list_start|>") != std::string::npos &&
         src.find("<|tool_list_end|>") != std::string::npos) {
         LOG_DBG("Using specialized template: LFM2\n");
         return common_chat_params_init_lfm2(tmpl, params);
     }
 
+    // LFM2.5 format detection: template uses plain "List of tools: [...]" with no special tokens
+    if (src.find("List of tools: [") != std::string::npos &&
+        src.find("<|tool_list_start|>") == std::string::npos) {
+        LOG_DBG("Using specialized template: LFM2.5\n");
+        return common_chat_params_init_lfm2_5(tmpl, params);
+    }
+
     // GigaChatV3 format detection
     if (src.find("<|role_sep|>") != std::string::npos &&
         src.find("<|message_sep|>") != std::string::npos &&

common/common.cpp

@@ -359,6 +359,11 @@ bool parse_cpu_mask(const std::string & mask, bool (&boolmask)[GGML_MAX_N_THREAD
 }
 
 void common_init() {
+#if defined(_WIN32)
+    SetConsoleOutputCP(CP_UTF8);
+    SetConsoleCP(CP_UTF8);
+#endif
+
     llama_log_set(common_log_default_callback, NULL);
 
 #ifdef NDEBUG
@@ -367,7 +372,7 @@ void common_init() {
     const char * build_type = " (debug)";
 #endif
 
-    LOG_INF("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
+    LOG_DBG("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
 }
 
 std::string common_params_get_system_info(const common_params & params) {
@@ -656,6 +661,97 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
     return true;
 }
 
+static inline bool glob_class_match(const char c, const char * pattern, const char * class_end) {
+    const char * class_start = pattern;
+    bool negated = false;
+
+    if (*class_start == '!') {
+        negated = true;
+        class_start++;
+    }
+
+    // If first character after negation is ']' or '-', treat it as literal
+    if (*class_start == ']' || *class_start == '-') {
+        if (class_start < class_end && *class_start == c) {
+            return !negated;
+        }
+        class_start++;
+    }
+
+    bool matched = false;
+
+    while (class_start < class_end) {
+        if (class_start + 2 < class_end && class_start[1] == '-' && class_start[2] != ']') {
+            char start_char = *class_start;
+            char end_char = class_start[2];
+            if (c >= start_char && c <= end_char) {
+                matched = true;
+                break;
+            }
+            class_start += 3;
+        } else {
+            if (*class_start == c) {
+                matched = true;
+                break;
+            }
+            class_start++;
+        }
+    }
+
+    return negated ? !matched : matched;
+}
+
+// simple glob: * matches non-/ chars, ** matches anything including /, [] matches character class
+static inline bool glob_match(const char * pattern, const char * str) {
+    if (*pattern == '\0') {
+        return *str == '\0';
+    }
+    if (pattern[0] == '*' && pattern[1] == '*') {
+        const char * p = pattern + 2;
+        if (glob_match(p, str)) return true;
+        if (*str != '\0') return glob_match(pattern, str + 1);
+        return false;
+    }
+    if (*pattern == '*') {
+        const char * p = pattern + 1;
+        for (; *str != '\0' && *str != '/'; str++) {
+            if (glob_match(p, str)) return true;
+        }
+        return glob_match(p, str);
+    }
+    if (*pattern == '?' && *str != '\0' && *str != '/') {
+        return glob_match(pattern + 1, str + 1);
+    }
+    if (*pattern == '[') {
+        const char * class_end = pattern + 1;
+        // If first character after '[' is ']' or '-', treat it as literal
+        if (*class_end == ']' || *class_end == '-') {
+            class_end++;
+        }
+        while (*class_end != '\0' && *class_end != ']') {
+            class_end++;
+        }
+        if (*class_end == ']') {
+            if (*str == '\0') return false;
+            bool matched = glob_class_match(*str, pattern + 1, class_end);
+            return matched && glob_match(class_end + 1, str + 1);
+        } else {
+            if (*str == '[') {
+                return glob_match(pattern + 1, str + 1);
+            }
+            return false;
+        }
+    }
+    if (*pattern == *str) {
+        return glob_match(pattern + 1, str + 1);
+    }
+    return false;
+}
+
+bool glob_match(const std::string & pattern, const std::string & str) {
+    return glob_match(pattern.c_str(), str.c_str());
+}
+
 //
 // Filesystem utils
 //
@@ -1152,6 +1248,9 @@ llama_context * common_init_result::context() {
 }
 
 common_sampler * common_init_result::sampler(llama_seq_id seq_id) {
+    if (seq_id < 0 || seq_id >= (int) pimpl->samplers.size()) {
+        return nullptr;
+    }
     return pimpl->samplers[seq_id].get();
 }
 

common/common.h

@@ -573,6 +573,7 @@ struct common_params {
 
     // server params
     int32_t port                = 8080;          // server listens on this network port
+    bool    reuse_port          = false;         // allow multiple sockets to bind to the same port
     int32_t timeout_read        = 600;           // http read timeout in seconds
     int32_t timeout_write       = timeout_read;  // http write timeout in seconds
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
@@ -793,6 +794,8 @@ std::string string_from(const std::vector<int> & values);
 std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
 std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
 
+bool glob_match(const std::string & pattern, const std::string & str);
+
 //
 // Filesystem utils
 //

common/download.cpp

@@ -119,6 +119,9 @@ class ProgressBar {
     static inline std::map<const ProgressBar *, int> lines;
     static inline int max_line = 0;
 
+    std::string filename;
+    size_t len = 0;
+
     static void cleanup(const ProgressBar * line) {
         lines.erase(line);
         if (lines.empty()) {
@@ -135,7 +138,23 @@ class ProgressBar {
     }
 
 public:
-    ProgressBar() = default;
+    ProgressBar(const std::string & url = "") : filename(url) {
+        if (auto pos = filename.rfind('/'); pos != std::string::npos) {
+            filename = filename.substr(pos + 1);
+        }
+        if (auto pos = filename.find('?'); pos != std::string::npos) {
+            filename = filename.substr(0, pos);
+        }
+        for (size_t i = 0; i < filename.size(); ++i) {
+            if ((filename[i] & 0xC0) != 0x80) {
+                if (len++ == 39) {
+                    filename.resize(i);
+                    filename += "…";
+                    break;
+                }
+            }
+        }
+    }
 
     ~ProgressBar() {
         std::lock_guard<std::mutex> lock(mutex);
@@ -143,11 +162,7 @@ public:
     }
 
     void update(size_t current, size_t total) {
-        if (!is_output_a_tty()) {
-            return;
-        }
-
-        if (!total) {
+        if (!total || !is_output_a_tty()) {
             return;
         }
 
@@ -159,28 +174,27 @@ public:
         }
         int lines_up = max_line - lines[this];
 
-        size_t width = 50;
+        size_t bar = 55 - len;
         size_t pct = (100 * current) / total;
-        size_t pos = (width * current) / total;
-
-        std::cout << "\033[s";
+        size_t pos = (bar * current) / total;
 
         if (lines_up > 0) {
             std::cout << "\033[" << lines_up << "A";
         }
-        std::cout << "\033[2K\r["
-            << std::string(pos, '=')
-            << (pos < width ? ">" : "")
-            << std::string(width - pos, ' ')
-            << "] " << std::setw(3) << pct << "%  ("
-            << current / (1024 * 1024) << " MB / "
-            << total / (1024 * 1024) << " MB) "
-            << "\033[u";
+        std::cout << '\r' << "Downloading " << filename << " ";
 
-        std::cout.flush();
+        for (size_t i = 0; i < bar; ++i) {
+            std::cout << (i < pos ? "—" : " ");
+        }
+        std::cout << std::setw(4) << pct << "%\033[K";
+
+        if (lines_up > 0) {
+            std::cout << "\033[" << lines_up << "B";
+        }
+        std::cout << '\r' << std::flush;
 
         if (current == total) {
-             cleanup(this);
+            cleanup(this);
         }
     }
 
@@ -208,7 +222,7 @@ static bool common_pull_file(httplib::Client & cli,
     const char * func = __func__; // avoid __func__ inside a lambda
     size_t downloaded = existing_size;
     size_t progress_step = 0;
-    ProgressBar bar;
+    ProgressBar bar(resolve_path);
 
     auto res = cli.Get(resolve_path, headers,
         [&](const httplib::Response &response) {
@@ -286,7 +300,7 @@ static int common_download_file_single_online(const std::string        & url,
     const bool file_exists = std::filesystem::exists(path);
 
     if (file_exists && skip_etag) {
-        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
+        LOG_DBG("%s: using cached file: %s\n", __func__, path.c_str());
         return 304; // 304 Not Modified - fake cached response
     }
 
@@ -294,7 +308,7 @@ static int common_download_file_single_online(const std::string        & url,
     if (file_exists) {
         last_etag = read_etag(path);
     } else {
-        LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
+        LOG_DBG("%s: no previous model file found %s\n", __func__, path.c_str());
     }
 
     auto head = cli.Head(parts.path);
@@ -328,11 +342,11 @@ static int common_download_file_single_online(const std::string        & url,
 
     if (file_exists) {
         if (etag.empty()) {
-            LOG_INF("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
+            LOG_DBG("%s: using cached file (no server etag): %s\n", __func__, path.c_str());
             return 304; // 304 Not Modified - fake cached response
         }
         if (!last_etag.empty() && last_etag == etag) {
-            LOG_INF("%s: using cached file (same etag): %s\n", __func__, path.c_str());
+            LOG_DBG("%s: using cached file (same etag): %s\n", __func__, path.c_str());
             return 304; // 304 Not Modified - fake cached response
         }
         if (remove(path.c_str()) != 0) {
@@ -368,7 +382,7 @@ static int common_download_file_single_online(const std::string        & url,
             }
         }
 
-        LOG_INF("%s: downloading from %s to %s (etag:%s)...\n",
+        LOG_DBG("%s: downloading from %s to %s (etag:%s)...\n",
                 __func__, common_http_show_masked_url(parts).c_str(),
                 path_temporary.c_str(), etag.c_str());
 
@@ -437,7 +451,7 @@ int common_download_file_single(const std::string & url,
         return -1;
     }
 
-    LOG_INF("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
+    LOG_DBG("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
     return 304; // Not Modified - fake cached response
 }
 

common/jinja/parser.cpp

@@ -539,6 +539,9 @@ private:
             statement_ptr step = slices.size() > 2 ? std::move(slices[2]) : nullptr;
             return mk_stmt<slice_expression>(start_pos, std::move(start), std::move(stop), std::move(step));
         }
+        if (slices.empty()) {
+            return mk_stmt<blank_expression>(start_pos);
+        }
         return std::move(slices[0]);
     }
 

common/jinja/runtime.cpp

@@ -771,10 +771,15 @@ value member_expression::execute_impl(context & ctx) {
     }
 
     JJ_DEBUG("Member expression on object type %s, property type %s", object->type().c_str(), property->type().c_str());
-    ensure_key_type_allowed(property);
-
     value val = mk_val<value_undefined>("object_property");
 
+    if (property->is_undefined()) {
+        JJ_DEBUG("%s", "Member expression property is undefined, returning undefined");
+        return val;
+    }
+
+    ensure_key_type_allowed(property);
+
     if (is_val<value_undefined>(object)) {
         JJ_DEBUG("%s", "Accessing property on undefined object, returning undefined");
         return val;

common/jinja/runtime.h

@@ -263,6 +263,14 @@ struct comment_statement : public statement {
 
 // Expressions
 
+// Represents an omitted expression in a computed member, e.g. `a[]`.
+struct blank_expression : public expression {
+    std::string type() const override { return "BlankExpression"; }
+    value execute_impl(context &) override {
+        return mk_val<value_undefined>();
+    }
+};
+
 struct member_expression : public expression {
     statement_ptr object;
     statement_ptr property;

common/json-schema-to-grammar.cpp

@@ -416,15 +416,30 @@ private:
                     i++;
                 } else if (c == '(') {
                     i++;
-                    if (i < length) {
-                        if (sub_pattern[i] == '?') {
+                    if (i < length && sub_pattern[i] == '?') {
+                        if (i + 1 < length && sub_pattern[i + 1] == ':') {
+                            i += 2; // skip "?:" for non-capturing group, treat as regular group
+                        } else {
+                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
                             _warnings.push_back("Unsupported pattern syntax");
+                            // skip to matching ')' to avoid UB on empty seq
+                            int depth = 1;
+                            while (i < length && depth > 0) {
+                                if (sub_pattern[i] == '\\' && i + 1 < length) {
+                                    i += 2; // skip escaped character
+                                } else {
+                                    if (sub_pattern[i] == '(') depth++;
+                                    else if (sub_pattern[i] == ')') depth--;
+                                    i++;
+                                }
+                            }
+                            continue;
                         }
                     }
                     seq.emplace_back("(" + to_rule(transform()) + ")", false);
                 } else if (c == ')') {
                     i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(') {
+                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
                         _errors.push_back("Unbalanced parentheses");
                     }
                     return join_seq();

common/ngram-map.h

@@ -51,7 +51,7 @@ struct common_ngram_map_value {
 // statistics of a n-gram
 struct common_ngram_map_key {
     size_t   key_idx;   // index of key n-gram in token-history
-    size_t   stat_idx;  // index of last token of stastistics computation (key_num, values)
+    size_t   stat_idx;  // index of last token of statistics computation (key_num, values)
 
     uint16_t key_num;   // number of occurrences of this key n-gram in token-history
     common_ngram_map_value values[COMMON_NGRAM_MAX_VALUES]; // some known values after the key

common/reasoning-budget.cpp

@@ -115,9 +115,11 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
             break;
         }
         case REASONING_BUDGET_FORCING:
-            // force_pos is advanced in apply(), not here.
-            // This ensures the first forced token isn't skipped when the sampler
-            // is initialized directly in FORCING state (e.g. COUNTING + budget=0)
+            ctx->force_pos++;
+            if (ctx->force_pos >= ctx->forced_tokens.size()) {
+                ctx->state = REASONING_BUDGET_DONE;
+                LOG_INF("reasoning-budget: forced sequence complete, done\n");
+            }
             break;
         case REASONING_BUDGET_DONE:
             break;
@@ -144,14 +146,6 @@ static void common_reasoning_budget_apply(struct llama_sampler * smpl, llama_tok
             cur_p->data[i].logit = -INFINITY;
         }
     }
-
-    // advance to next forced token (done here rather than in accept so that
-    // the first forced token isn't skipped when starting in FORCING state)
-    ctx->force_pos++;
-    if (ctx->force_pos >= ctx->forced_tokens.size()) {
-        ctx->state = REASONING_BUDGET_DONE;
-        LOG_INF("reasoning-budget: forced sequence complete, done\n");
-    }
 }
 
 static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
@@ -261,3 +255,10 @@ struct llama_sampler * common_reasoning_budget_init(
         common_reasoning_budget_state    initial_state) {
     return common_reasoning_budget_init_state(vocab, start_tokens, end_tokens, forced_tokens, budget, initial_state);
 }
+
+common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl) {
+    if (!smpl) {
+        return REASONING_BUDGET_IDLE;
+    }
+    return ((const common_reasoning_budget_ctx *)smpl->ctx)->state;
+}

common/reasoning-budget.h

@@ -51,3 +51,5 @@ struct llama_sampler * common_reasoning_budget_init(
         const std::vector<llama_token> & forced_tokens,
         int32_t                          budget,
         common_reasoning_budget_state    initial_state);
+
+common_reasoning_budget_state common_reasoning_budget_get_state(const struct llama_sampler * smpl);

common/sampling.cpp

@@ -7,6 +7,7 @@
 
 #include <algorithm>
 #include <cctype>
+#include <climits>
 #include <cmath>
 #include <cstring>
 #include <unordered_map>
@@ -109,6 +110,7 @@ struct common_sampler {
     common_params_sampling params;
 
     struct llama_sampler * grmr;
+    struct llama_sampler * rbudget;
     struct llama_sampler * chain;
 
     ring_buffer<llama_token> prev;
@@ -188,6 +190,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
     lparams.no_perf = params.no_perf;
 
     llama_sampler * grmr = nullptr;
+    llama_sampler * rbudget = nullptr;
     llama_sampler * chain = llama_sampler_chain_init(lparams);
 
     std::vector<llama_sampler *> samplers;
@@ -270,7 +273,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
             }
         }
 
-        if (grmr) {
+        if (grmr && !params.grammar_lazy) {
             try {
                 for (const auto & token : prefill_tokens) {
                     llama_sampler_accept(grmr, token);
@@ -284,15 +287,15 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
-    // reasoning budget sampler — added first so it can force tokens before other samplers
-    if (params.reasoning_budget_tokens >= 0 && !params.reasoning_budget_forced.empty()) {
-        samplers.push_back(common_reasoning_budget_init(
+    // reasoning budget sampler
+    if (!params.reasoning_budget_start.empty() && !params.reasoning_budget_end.empty()) {
+        rbudget = common_reasoning_budget_init(
             vocab,
             params.reasoning_budget_start,
             params.reasoning_budget_end,
             params.reasoning_budget_forced,
-            params.reasoning_budget_tokens,
-            prefill_tokens));
+            params.reasoning_budget_tokens < 0 ? INT_MAX : params.reasoning_budget_tokens,
+            prefill_tokens);
     }
 
     if (params.has_logit_bias()) {
@@ -380,9 +383,16 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         params.backend_sampling = false;
     }
 
+    if (rbudget && params.backend_sampling) {
+        LOG_WRN("%s: backend sampling is not compatible with reasoning budget, disabling\n", __func__);
+
+        params.backend_sampling = false;
+    }
+
     auto * result = new common_sampler {
         /* .params  = */ params,
         /* .grmr    = */ grmr,
+        /* .rbudget = */ rbudget,
         /* .chain   = */ chain,
         /* .prev    = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
         /* .cur     = */ {},
@@ -398,11 +408,27 @@ void common_sampler_free(struct common_sampler * gsmpl) {
     }
 
     llama_sampler_free(gsmpl->grmr);
+    llama_sampler_free(gsmpl->rbudget);
     llama_sampler_free(gsmpl->chain);
 
     delete gsmpl;
 }
 
+static bool grammar_should_apply(struct common_sampler * gsmpl) {
+    if (!gsmpl->grmr) {
+        return false;
+    }
+    if (!gsmpl->rbudget) {
+        return true;
+    }
+    if (gsmpl->params.grammar_lazy) {
+        // if grammar is lazy, only apply when reasoning budget is not active
+        const auto state = common_reasoning_budget_get_state(gsmpl->rbudget);
+        return state == REASONING_BUDGET_IDLE || state == REASONING_BUDGET_DONE;
+    }
+    return true;
+}
+
 void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar) {
     if (!gsmpl) {
         return;
@@ -410,6 +436,11 @@ void common_sampler_accept(struct common_sampler * gsmpl, llama_token token, boo
 
     const auto tm = gsmpl->tm();
 
+    // grammar_should_apply() checks the reasoning budget state, so calculate this before we accept
+    accept_grammar = accept_grammar && grammar_should_apply(gsmpl);
+
+    llama_sampler_accept(gsmpl->rbudget, token);
+
     if (gsmpl->grmr && accept_grammar) {
         llama_sampler_accept(gsmpl->grmr, token);
     }
@@ -431,6 +462,7 @@ struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
     return new common_sampler {
         /* .params  = */ gsmpl->params,
         /* .grmr    = */ llama_sampler_clone(gsmpl->grmr),
+        /* .rbudget = */ llama_sampler_clone(gsmpl->rbudget),
         /* .chain   = */ llama_sampler_clone(gsmpl->chain),
         /* .prev    = */ gsmpl->prev,
         /* .cur     = */ gsmpl->cur,
@@ -500,6 +532,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     llama_token id = LLAMA_TOKEN_NULL;
 
     auto & grmr  = gsmpl->grmr;
+    auto & rbudget = gsmpl->rbudget;
     auto & chain = gsmpl->chain;
     auto & cur_p = gsmpl->cur_p; // initialized by set_logits
 
@@ -511,7 +544,8 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
         if (id != LLAMA_TOKEN_NULL) {
             LOG_DBG("%s: Backend sampler selected token: '%d'. Will not run any CPU samplers\n", __func__, id);
 
-            GGML_ASSERT(!gsmpl->grmr && "using grammar in combination with backend sampling is not supported");
+            GGML_ASSERT(!gsmpl->grmr    && "using grammar in combination with backend sampling is not supported");
+            GGML_ASSERT(!gsmpl->rbudget && "using reasoning budget in combination with backend sampling is not supported");
 
             // TODO: simplify
             gsmpl->cur.resize(1);
@@ -524,7 +558,10 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 
     gsmpl->set_logits(ctx, idx);
 
-    if (grammar_first) {
+    // apply reasoning budget first
+    llama_sampler_apply(rbudget, &cur_p);
+
+    if (grammar_first && grammar_should_apply(gsmpl)) {
         llama_sampler_apply(grmr, &cur_p);
     }
 
@@ -532,7 +569,7 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
 
     id = cur_p.data[cur_p.selected].id;
 
-    if (grammar_first) {
+    if (grammar_first || !grammar_should_apply(gsmpl)) {
         return id;
     }
 
@@ -553,7 +590,12 @@ llama_token common_sampler_sample(struct common_sampler * gsmpl, struct llama_co
     // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
     gsmpl->set_logits(ctx, idx);
 
-    llama_sampler_apply(grmr,  &cur_p);
+    llama_sampler_apply(rbudget,  &cur_p);
+
+    if (grammar_should_apply(gsmpl)) {
+        llama_sampler_apply(grmr,  &cur_p);
+    }
+
     llama_sampler_apply(chain, &cur_p);
 
     GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");

convert_hf_to_gguf.py

@@ -31,10 +31,10 @@ import gguf
 from gguf.vocab import MistralTokenizerType, MistralVocab
 
 try:
-    from mistral_common.tokens.tokenizers.base import TokenizerVersion # type: 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 # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.tekken import Tekkenizer # type: ignore[import-not-found]
-    from mistral_common.tokens.tokenizers.sentencepiece import ( # type: ignore[import-not-found]
+    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]
         SentencePieceTokenizer,
     )
 

docs/build.md

@@ -728,7 +728,7 @@ To read documentation for how to build on Android, [click here](./android.md)
 
 ## WebGPU [In Progress]
 
-The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `bed1a61`.
+The WebGPU backend relies on [Dawn](https://dawn.googlesource.com/dawn). Follow the instructions [here](https://dawn.googlesource.com/dawn/+/refs/heads/main/docs/quickstart-cmake.md) to install Dawn locally so that llama.cpp can find it using CMake. The current implementation is up-to-date with Dawn commit `18eb229`.
 
 In the llama.cpp directory, build with CMake:
 

docs/docker.md

@@ -13,24 +13,30 @@ We have three Docker images available for this project:
 
 Additionally, there the following images, similar to the above:
 
-- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-cuda13`: Same as `full` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-cuda13`: Same as `light` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA 12 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-cuda13`: Same as `server` but compiled with CUDA 13 support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:full-intel`: Same as `full` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-intel`: Same as `light` but compiled with SYCL support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-intel`: Same as `server` but compiled with SYCL support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`)
-- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-vulkan`: Same as `full` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:light-vulkan`: Same as `light` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggml-org/llama.cpp:server-vulkan`: Same as `server` but compiled with Vulkan support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggml-org/llama.cpp:full-openvino`: Same as `full` but compiled with OpenVino support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:light-openvino`: Same as `light` but compiled with OpenVino support. (platforms: `linux/amd64`)
 - `ghcr.io/ggml-org/llama.cpp:server-openvino`: Same as `server` but compiled with OpenVino support. (platforms: `linux/amd64`)
+- `ghcr.io/ggml-org/llama.cpp:full-s390x`: Identical to `full`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
+- `ghcr.io/ggml-org/llama.cpp:light-s390x`: Identical to `light`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
+- `ghcr.io/ggml-org/llama.cpp:server-s390x`: Identical to `server`, an alias for the `s390x` platform. (platforms: `linux/s390x`)
 
 The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA, ROCm or MUSA library, you'll need to build the images locally for now).
 
@@ -82,7 +88,7 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
 
 The defaults are:
 
-- `CUDA_VERSION` set to `12.4.0`
+- `CUDA_VERSION` set to `12.8.1`
 - `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
 
 The resulting images, are essentially the same as the non-CUDA images:

examples/batched/batched.cpp

@@ -24,12 +24,12 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BATCHED, print_usage)) {
         return 1;
     }
 
-    common_init();
-
     // number of parallel batches
     int n_parallel = params.n_parallel;
 

examples/debug/debug.cpp

@@ -213,12 +213,12 @@ static bool run(llama_context * ctx, const common_params & params) {
 int main(int argc, char ** argv) {
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DEBUG, print_usage)) {
         return 1;
     }
 
-    common_init();
-
     llama_backend_init();
     llama_numa_init(params.numa);
 

examples/diffusion/diffusion-cli.cpp

@@ -545,11 +545,12 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DIFFUSION)) {
         return 1;
     }
 
-    common_init();
     llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();

examples/embedding/embedding.cpp

@@ -99,12 +99,12 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
         return 1;
     }
 
-    common_init();
-
     params.embedding = true;
 
     // get max number of sequences per batch

examples/eval-callback/eval-callback.cpp

@@ -37,12 +37,12 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    common_init();
-
     llama_backend_init();
     llama_numa_init(params.numa);
 

examples/idle/idle.cpp

@@ -19,12 +19,12 @@ static void print_usage(int /*argc*/, char ** argv) {
 int main(int argc, char ** argv) {
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
-    common_init();
-
     // init LLM
 
     llama_backend_init();

examples/lookahead/lookahead.cpp

@@ -43,12 +43,12 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    common_init();
-
     const int W = 15; // lookahead window
     const int N = 5;  // n-gram size
     const int G = 15; // max verification n-grams

examples/lookup/lookup-create.cpp

@@ -12,6 +12,8 @@ int main(int argc, char ** argv){
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }

examples/lookup/lookup-stats.cpp

@@ -18,12 +18,12 @@ int main(int argc, char ** argv){
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
-    common_init();
-
     const int n_draft = params.speculative.n_max;
 
     // init llama.cpp

examples/lookup/lookup.cpp

@@ -18,12 +18,12 @@ int main(int argc, char ** argv){
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
-    common_init();
-
     // max. number of additional tokens to draft if match is found
     const int n_draft = params.speculative.n_max;
 

examples/model-conversion/scripts/causal/compare-logits.py

@@ -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]
+from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
 
 def quick_logits_check(pytorch_file, llamacpp_file):
     """Lightweight sanity check before NMSE"""

examples/model-conversion/scripts/utils/check-nmse.py

@@ -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]
+from common import get_model_name_from_env_path  # type: ignore[import-not-found, ty:unresolved-import]
 
 def calculate_nmse(reference, test):
     mse = np.mean((test - reference) ** 2)

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

@@ -2,7 +2,7 @@
 
 import argparse
 import sys
-from common import compare_tokens  # type: ignore[import-not-found]
+from common import compare_tokens  # type: ignore[import-not-found, ty:unresolved-import]
 
 
 def parse_arguments():

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

@@ -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]
+from common import compare_tokens, exit_with_warning  # type: ignore[import-not-found, ty:unresolved-import]
 
 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 

examples/parallel/parallel.cpp

@@ -163,12 +163,12 @@ int main(int argc, char ** argv) {
     params.n_predict = 128;
     params.n_junk = 1;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
 
-    common_init();
-
     // number of simultaneous "clients" to simulate
     const int32_t n_clients = params.n_parallel;
 

examples/passkey/passkey.cpp

@@ -25,12 +25,12 @@ int main(int argc, char ** argv) {
     params.n_keep = 32;
     params.i_pos  = -1;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
         return 1;
     }
 
-    common_init();
-
     int n_junk = params.n_junk;
     int n_keep = params.n_keep;
     int n_grp  = params.grp_attn_n;

examples/retrieval/retrieval.cpp

@@ -117,12 +117,12 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
         return 1;
     }
 
-    common_init();
-
     // For BERT models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
     params.embedding = true;

examples/save-load-state/save-load-state.cpp

@@ -17,6 +17,8 @@ int main(int argc, char ** argv) {
 
     const std::string_view state_file = "dump_state.bin";
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
@@ -27,8 +29,6 @@ int main(int argc, char ** argv) {
         params.kv_unified = true;
     }
 
-    common_init();
-
     if (params.n_predict < 0) {
         params.n_predict = 16;
     }

examples/speculative-simple/speculative-simple.cpp

@@ -16,6 +16,8 @@ int main(int argc, char ** argv) {
 
     common_params params;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
     }
@@ -25,8 +27,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    common_init();
-
     if (params.speculative.mparams_dft.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;

examples/speculative/speculative.cpp

@@ -38,6 +38,8 @@ int main(int argc, char ** argv) {
     // needed to get candidate probs even for temp <= 0.0
     params.sampling.n_probs = 128;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
     }
@@ -47,8 +49,6 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    common_init();
-
     if (params.speculative.mparams_dft.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;

examples/sycl/build.sh

@@ -20,4 +20,4 @@ cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA
 #cmake --build . --config Release --target llama-bench
 
 #build all binary
-cmake --build . --config Release -j -v
+cmake --build . --config Release -j$((($(nproc)+1)/2)) -v

examples/sycl/run-llama2.sh

@@ -23,9 +23,9 @@ if [ $# -gt 0 ]; then
     GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} -mg $GGML_SYCL_DEVICE -sm none ${LOAD_MODE}
 
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-completion -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s 0 -c ${CONTEXT} ${LOAD_MODE}
 fi

examples/training/finetune.cpp

@@ -20,6 +20,8 @@ int main(int argc, char ** argv) {
     common_params params;
     params.escape = false;
 
+    common_init();
+
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_FINETUNE)) {
         return 1;
     }
@@ -38,7 +40,6 @@ int main(int argc, char ** argv) {
         params.cache_type_v = GGML_TYPE_F32;
     }
 
-    common_init();
     llama_backend_init();
     llama_numa_init(params.numa);
     // load the model and apply lora adapter, if any

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 8)
+set(GGML_VERSION_PATCH 10)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -166,15 +166,16 @@ if (NOT MSVC)
     option(GGML_AMX_INT8     "ggml: enable AMX-INT8"         OFF)
     option(GGML_AMX_BF16     "ggml: enable AMX-BF16"         OFF)
 endif()
-option(GGML_LASX             "ggml: enable lasx"             ON)
-option(GGML_LSX              "ggml: enable lsx"              ON)
-option(GGML_RVV              "ggml: enable rvv"              ON)
-option(GGML_RV_ZFH           "ggml: enable riscv zfh"        ON)
-option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"       ON)
-option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"     ON)
-option(GGML_RV_ZIHINTPAUSE   "ggml: enable riscv zihintpause "  ON)
-option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
-option(GGML_VXE              "ggml: enable vxe"              ${GGML_NATIVE})
+option(GGML_LASX             "ggml: enable lasx"              ON)
+option(GGML_LSX              "ggml: enable lsx"               ON)
+option(GGML_RVV              "ggml: enable rvv"               ON)
+option(GGML_RV_ZFH           "ggml: enable riscv zfh"         ON)
+option(GGML_RV_ZVFH          "ggml: enable riscv zvfh"        ON)
+option(GGML_RV_ZICBOP        "ggml: enable riscv zicbop"      ON)
+option(GGML_RV_ZIHINTPAUSE   "ggml: enable riscv zihintpause" ON)
+option(GGML_RV_ZVFBFWMA      "ggml: enable riscv zvfbfwma"    OFF)
+option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"      OFF)
+option(GGML_VXE              "ggml: enable vxe"               ${GGML_NATIVE})
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -434,6 +434,9 @@ void ggml_cann_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src = dst->src[0];
 
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
     acl_tensor_ptr acl_src = ggml_cann_create_tensor(src);
     acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
 
@@ -456,6 +459,13 @@ void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     float          p_value  = 2.0f;
     acl_scalar_ptr p_scalar = ggml_cann_create_scalar(&p_value, aclDataType::ACL_FLOAT);
     GGML_CANN_CALL_ACLNN_OP(ctx, Norm, acl_src.get(), p_scalar.get(), dims_array.get(), true, acl_div.get());
+
+    // Clamp norm to at least eps: scale = 1/fmaxf(norm, eps)
+    acl_scalar_ptr acl_min = ggml_cann_create_scalar(&eps, aclDataType::ACL_FLOAT);
+    float          flt_max = FLT_MAX;
+    acl_scalar_ptr acl_max = ggml_cann_create_scalar(&flt_max, aclDataType::ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, Clamp, acl_div.get(), acl_min.get(), acl_max.get(), acl_div.get());
+
     GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src.get(), acl_div.get(), acl_dst.get());
 }
 

ggml/src/ggml-cann/common.h

@@ -216,14 +216,16 @@ struct ggml_cann_pool_alloc {
 #ifdef USE_ACL_GRAPH
 struct ggml_graph_node_properties {
     // dst tensor
-    void *  node_address;
-    int64_t ne[GGML_MAX_DIMS];
-    size_t  nb[GGML_MAX_DIMS];
+    void *    node_address;
+    ggml_type node_type;
+    int64_t   ne[GGML_MAX_DIMS];
+    size_t    nb[GGML_MAX_DIMS];
 
     // src tensor
-    void *  src_address[GGML_MAX_SRC];
-    int64_t src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
-    size_t  src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
+    void *    src_address[GGML_MAX_SRC];
+    ggml_type src_type[GGML_MAX_SRC];
+    int64_t   src_ne[GGML_MAX_SRC][GGML_MAX_DIMS];
+    size_t    src_nb[GGML_MAX_SRC][GGML_MAX_DIMS];
 
     // op
     ggml_op node_op;
@@ -247,6 +249,10 @@ struct ggml_graph_node_properties {
             return false;
         }
 
+        if (node->type != this->node_type) {
+            return false;
+        }
+
         for (int i = 0; i < GGML_MAX_DIMS; i++) {
             if (node->ne[i] != this->ne[i]) {
                 return false;
@@ -262,6 +268,10 @@ struct ggml_graph_node_properties {
                     return false;
                 }
 
+                if (node->src[i]->type != this->src_type[i]) {
+                    return false;
+                }
+
                 for (int d = 0; d < GGML_MAX_DIMS; d++) {
                     if (node->src[i]->ne[d] != this->src_ne[i][d]) {
                         return false;
@@ -277,10 +287,7 @@ struct ggml_graph_node_properties {
             }
         }
 
-        if (node->op == GGML_OP_SCALE || node->op == GGML_OP_UNARY || node->op == GGML_OP_GLU || node->op == GGML_OP_ROPE){
-            return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
-        }
-        return true;
+        return memcmp(this->op_params, node->op_params, GGML_MAX_OP_PARAMS) == 0;
     }
 };
 
@@ -322,6 +329,7 @@ struct ggml_cann_graph {
 
             prop.node_address = node->data;
             prop.node_op      = node->op;
+            prop.node_type    = node->type;
 
             std::copy_n(node->ne, GGML_MAX_DIMS, prop.ne);
             std::copy_n(node->nb, GGML_MAX_DIMS, prop.nb);
@@ -329,10 +337,12 @@ struct ggml_cann_graph {
             for (int src = 0; src < GGML_MAX_SRC; ++src) {
                 if (node->src[src]) {
                     prop.src_address[src] = node->src[src]->data;
+                    prop.src_type[src]    = node->src[src]->type;
                     std::copy_n(node->src[src]->ne, GGML_MAX_DIMS, prop.src_ne[src]);
                     std::copy_n(node->src[src]->nb, GGML_MAX_DIMS, prop.src_nb[src]);
                 } else {
                     prop.src_address[src] = nullptr;
+                    prop.src_type[src]    = GGML_TYPE_COUNT;
                     std::fill_n(prop.src_ne[src], GGML_MAX_DIMS, 0);
                     std::fill_n(prop.src_nb[src], GGML_MAX_DIMS, 0);
                 }

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

@@ -36,10 +36,13 @@
 #include <cmath>
 #include <cstdio>
 #include <cstring>
+#include <memory>
 #include <mutex>
 #include <optional>
 #include <queue>
+#include <unordered_map>
 #include <unordered_set>
+#include <vector>
 
 #define GGML_COMMON_DECL_C
 
@@ -770,6 +773,21 @@ std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(i
 }
 
 // cann buffer
+
+/**
+ * @brief Tracks multi-threaded write progress for a single tensor.
+ *
+ * When multiple threads call set_tensor on different chunks of the same tensor,
+ * this tracker accumulates progress and defers post-processing (quantized format
+ * transform or ND-to-NZ conversion) until all data has been written.
+ */
+struct TensorSetTracker {
+    std::mutex mtx;                   ///< Protects concurrent access to this tracker
+    size_t bytes_written = 0;         ///< Accumulated bytes written so far
+    size_t total_bytes = 0;           ///< Target size (full tensor)
+    std::vector<uint8_t> host_buffer; ///< Host staging buffer for quantized tensors
+};
+
 /**
  * @brief Context for managing a CANN buffer associated with a specific device.
  *
@@ -780,6 +798,9 @@ struct ggml_backend_cann_buffer_context {
     int32_t device;             ///< The device ID associated with this buffer context.
     void *  dev_ptr = nullptr;  ///< Pointer to the device memory allocated for the buffer.
 
+    std::mutex tracker_mutex;   ///< Protects the trackers map
+    std::unordered_map<void *, std::unique_ptr<TensorSetTracker>> trackers;
+
     /**
      * @brief Constructor to initialize the CANN buffer context.
      *
@@ -792,6 +813,31 @@ struct ggml_backend_cann_buffer_context {
      * @brief Destructor to free the device memory allocated for the buffer.
      */
     ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
+
+    /**
+     * @brief Get or create a tracker for the given tensor.
+     */
+    TensorSetTracker * get_or_create_tracker(ggml_tensor * tensor) {
+        std::lock_guard<std::mutex> lock(tracker_mutex);
+        auto key = tensor->data;
+        auto it = trackers.find(key);
+        if (it == trackers.end()) {
+            auto tracker = std::make_unique<TensorSetTracker>();
+            tracker->total_bytes = ggml_nbytes(tensor);
+            auto * ptr = tracker.get();
+            trackers[key] = std::move(tracker);
+            return ptr;
+        }
+        return it->second.get();
+    }
+
+    /**
+     * @brief Remove the tracker for the given tensor.
+     */
+    void remove_tracker(ggml_tensor * tensor) {
+        std::lock_guard<std::mutex> lock(tracker_mutex);
+        trackers.erase(tensor->data);
+    }
 };
 
 // cann buffer type
@@ -1124,6 +1170,7 @@ static enum ggml_status ggml_backend_cann_buffer_init_tensor(ggml_backend_buffer
  * designed to be used with a global array, one per device.
  */
 struct ggml_cann_nz_workspace {
+    std::mutex mtx;    // Protects ptr/allocated from concurrent access
     void * ptr;        // Pointer to allocated device buffer
     size_t allocated;  // Size of currently allocated buffer in bytes
 
@@ -1190,13 +1237,15 @@ static ggml_cann_nz_workspace g_nz_workspaces[GGML_CANN_MAX_DEVICES];
  * @note The workspace buffer used in this function is managed globally and reused
  *       across calls. This reduces overhead from repeated memory allocation and deallocation.
  */
-static void weight_format_to_nz(ggml_tensor * tensor, size_t offset, int device) {
-    acl_tensor_ptr weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne, tensor->nb, 2, ACL_FORMAT_ND, offset);
+static void weight_format_to_nz(ggml_tensor * tensor, int device) {
+    acl_tensor_ptr weightTransposed = ggml_cann_create_tensor(tensor, tensor->ne, tensor->nb, 2, ACL_FORMAT_ND, 0);
     uint64_t       workspaceSize    = 0;
     aclOpExecutor * executor;
 
     // TransMatmulWeight
     ACL_CHECK(aclnnTransMatmulWeightGetWorkspaceSize(weightTransposed.get(), &workspaceSize, &executor));
+
+    std::lock_guard<std::mutex> lock(g_nz_workspaces[device].mtx);
     // Avoid frequent malloc/free of the workspace.
     g_nz_workspaces[device].realloc(workspaceSize);
 
@@ -1210,7 +1259,13 @@ static void weight_format_to_nz(ggml_tensor * tensor, size_t offset, int device)
  * @brief Set tensor data in a CANN buffer.
  *
  * This function sets tensor data in a CANN buffer, handling transformations
- * if needed based on the tensor's type.
+ * if needed based on the tensor's type. It supports multi-threaded calls
+ * where different threads write different chunks of the same tensor.
+ *
+ * For quantized tensors (Q4_0/Q8_0), data is staged in a host buffer and
+ * the format transform is deferred until all chunks are written.
+ * For NZ weight tensors, chunks are uploaded directly but the ND-to-NZ
+ * conversion is deferred until all chunks are written.
  *
  * @param buffer The CANN buffer where the tensor data will be set.
  * @param tensor Pointer to the tensor whose data will be set.
@@ -1226,26 +1281,72 @@ static void ggml_backend_cann_buffer_set_tensor(ggml_backend_buffer_t buffer,
     ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
 
     ggml_cann_set_device(ctx->device);
-    // TODO: refer to cann(#6017), it use thread's default stream.
-    // For acl, synchronous functions use this default stream.
-    // Why aclrtSynchronizeDevice?
 
     // Only check env once.
     static bool weight_to_nz = parse_bool(get_env_as_lowercase("GGML_CANN_WEIGHT_NZ").value_or("on"));
-    if (!need_transform(tensor->type)) {
+
+    bool is_quantized = need_transform(tensor->type);
+    bool is_nz        = !is_quantized && tensor->type != GGML_TYPE_BF16 && weight_to_nz &&
+                 is_matmul_weight((const ggml_tensor *) tensor);
+
+    // Plain tensor (not quantized, not NZ): direct copy, no tracking needed
+    if (!is_quantized && !is_nz) {
         ACL_CHECK(aclrtMemcpy((char *) tensor->data + offset, size, data, size, ACL_MEMCPY_HOST_TO_DEVICE));
-        if (weight_to_nz && tensor->type != GGML_TYPE_BF16
-            && is_matmul_weight((const ggml_tensor *) tensor)) {
+        return;
+    }
+
+    // Single-shot write (full tensor at once): handle directly without tracking overhead
+    if (offset == 0 && size == ggml_nbytes(tensor)) {
+        if (is_quantized) {
+            void * transform_buffer = malloc(size);
+            ggml_backend_cann_transform(tensor, data, transform_buffer);
+            ACL_CHECK(aclrtMemcpy(tensor->data, size, transform_buffer, size, ACL_MEMCPY_HOST_TO_DEVICE));
+            free(transform_buffer);
+        } else {
+            // NZ weight
             GGML_ASSERT(tensor->ne[2] == 1);
             GGML_ASSERT(tensor->ne[3] == 1);
-            weight_format_to_nz(tensor, offset, ctx->device);
+            ACL_CHECK(aclrtMemcpy(tensor->data, size, data, size, ACL_MEMCPY_HOST_TO_DEVICE));
+            weight_format_to_nz(tensor, ctx->device);
         }
-    } else {
-        void * transform_buffer = malloc(size);
-        ggml_backend_cann_transform(tensor, data, transform_buffer);
+        return;
+    }
 
-        ACL_CHECK(aclrtMemcpy((char *) tensor->data + offset, size, transform_buffer, size, ACL_MEMCPY_HOST_TO_DEVICE));
-        free(transform_buffer);
+    // Chunked write: use tracker to accumulate progress and defer transform/conversion
+    TensorSetTracker * tracker = ctx->get_or_create_tracker(tensor);
+    std::unique_lock<std::mutex> lock(tracker->mtx);
+
+    if (is_quantized) {
+        // Stage data in host buffer; transform requires full tensor data
+        if (tracker->host_buffer.empty()) {
+            tracker->host_buffer.resize(tracker->total_bytes);
+        }
+        memcpy(tracker->host_buffer.data() + offset, data, size);
+    } else {
+        // NZ weight: upload chunk to device immediately, defer conversion
+        ACL_CHECK(aclrtMemcpy((char *) tensor->data + offset, size, data, size, ACL_MEMCPY_HOST_TO_DEVICE));
+    }
+
+    tracker->bytes_written += size;
+
+    // All chunks received: perform deferred transform/conversion
+    if (tracker->bytes_written >= tracker->total_bytes) {
+        if (is_quantized) {
+            void * transform_buffer = malloc(tracker->total_bytes);
+            ggml_backend_cann_transform(tensor, tracker->host_buffer.data(), transform_buffer);
+            ACL_CHECK(aclrtMemcpy(tensor->data, tracker->total_bytes, transform_buffer, tracker->total_bytes, ACL_MEMCPY_HOST_TO_DEVICE));
+            free(transform_buffer);
+        }
+
+        if (is_nz) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
+            weight_format_to_nz(tensor, ctx->device);
+        }
+
+        // Unlock before removing tracker, as remove_tracker destroys the mutex
+        lock.unlock();
+        ctx->remove_tracker(tensor);
     }
 }
 

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

@@ -2350,11 +2350,15 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_FLASH_ATTN_BACK:
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
+            {
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_GATED_LINEAR_ATTN:
         case GGML_OP_RWKV_WKV7:
             {
-                n_tasks = n_threads;
+                const int64_t n_heads = node->src[1]->ne[1];
+                n_tasks = MIN(n_threads, n_heads);
             } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -180,44 +180,49 @@ inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
 }
 #endif
 
-#if defined(__riscv_zvfh)
-template <>
-inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
-    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
-}
-inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
-    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
-}
-inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
-    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
-}
-inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
-    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
-}
-inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
+#if defined(__riscv_v_intrinsic)
+template <> inline vfloat32m1_t madd(vfloat32m1_t a, vfloat32m1_t b, vfloat32m1_t c) {
     return __riscv_vfmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
 }
-inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
+template <> inline vfloat32m2_t madd(vfloat32m2_t a, vfloat32m2_t b, vfloat32m2_t c) {
     return __riscv_vfmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
 }
-inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
+template <> inline vfloat32m4_t madd(vfloat32m4_t a, vfloat32m4_t b, vfloat32m4_t c) {
     return __riscv_vfmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
 }
-inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
+template <> inline vfloat32m8_t madd(vfloat32m8_t a, vfloat32m8_t b, vfloat32m8_t c) {
     return __riscv_vfmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat32m1_t madd(vfloat16mf2_t a, vfloat16mf2_t b, vfloat32m1_t c) {
+    return __riscv_vfwmacc_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
+}
+template <> inline vfloat32m2_t madd(vfloat16m1_t a, vfloat16m1_t b, vfloat32m2_t c) {
+    return __riscv_vfwmacc_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
+}
+template <> inline vfloat32m4_t madd(vfloat16m2_t a, vfloat16m2_t b, vfloat32m4_t c) {
+    return __riscv_vfwmacc_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
+}
+template <> inline vfloat32m8_t madd(vfloat16m4_t a, vfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmacc_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
+#endif
+
 #if defined(__riscv_zvfbfwma)
-inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
+template <> inline vfloat32m1_t madd(vbfloat16mf2_t a, vbfloat16mf2_t b, vfloat32m1_t c) {
     return __riscv_vfwmaccbf16_vv_f32m1(c, a, b, __riscv_vsetvlmax_e32m1());
 }
-inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
+template <> inline vfloat32m2_t madd(vbfloat16m1_t a, vbfloat16m1_t b, vfloat32m2_t c) {
     return __riscv_vfwmaccbf16_vv_f32m2(c, a, b, __riscv_vsetvlmax_e32m2());
 }
-inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
+template <> inline vfloat32m4_t madd(vbfloat16m2_t a, vbfloat16m2_t b, vfloat32m4_t c) {
     return __riscv_vfwmaccbf16_vv_f32m4(c, a, b, __riscv_vsetvlmax_e32m4());
 }
+template <> inline vfloat32m8_t madd(vbfloat16m4_t a, vbfloat16m4_t b, vfloat32m8_t c) {
+    return __riscv_vfwmaccbf16_vv_f32m8(c, a, b, __riscv_vsetvlmax_e32m8());
+}
 #endif
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -272,7 +277,7 @@ inline float hsum(__m512 x) {
 }
 #endif // __AVX512F__
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
 inline float hsum(vfloat32m1_t x) {
     return __riscv_vfmv_f_s_f32m1_f32(
         __riscv_vfredusum_vs_f32m1_f32m1(x, __riscv_vfmv_v_f_f32m1(0, 1), __riscv_vsetvlmax_e32m1()));
@@ -379,19 +384,7 @@ template <> inline __m256bh load(const float *p) {
 }
 #endif
 
-#if defined(__riscv_zvfh)
-template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
-}
-template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
-}
-template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
-}
-template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
-    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
-}
+#if defined(__riscv_v_intrinsic)
 template <> inline vfloat32m1_t load(const float *p) {
     return __riscv_vle32_v_f32m1(p, __riscv_vsetvlmax_e32m1());
 }
@@ -406,6 +399,21 @@ template <> inline vfloat32m8_t load(const float *p) {
 }
 #endif
 
+#if defined(__riscv_zvfh)
+template <> inline vfloat16mf2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16mf2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16mf2());
+}
+template <> inline vfloat16m1_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m1(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m1());
+}
+template <> inline vfloat16m2_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m2(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m2());
+}
+template <> inline vfloat16m4_t load(const ggml_fp16_t *p) {
+    return __riscv_vle16_v_f16m4(reinterpret_cast<const _Float16 *>(p), __riscv_vsetvlmax_e16m4());
+}
+#endif
+
 #if defined(__riscv_zvfbfwma)
 template <> inline vbfloat16mf2_t load(const ggml_bf16_t *p) {
     return __riscv_vle16_v_bf16mf2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16mf2());
@@ -416,23 +424,14 @@ template <> inline vbfloat16m1_t load(const ggml_bf16_t *p) {
 template <> inline vbfloat16m2_t load(const ggml_bf16_t *p) {
     return __riscv_vle16_v_bf16m2(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m2());
 }
+template <> inline vbfloat16m4_t load(const ggml_bf16_t *p) {
+    return __riscv_vle16_v_bf16m4(reinterpret_cast<const __bf16*>(p), __riscv_vsetvlmax_e16m4());
+}
 #endif
 
-#if defined(__riscv_zvfh)
+#if defined(__riscv_v_intrinsic)
 template <typename T> T set_zero();
 
-template <> inline vfloat16mf2_t set_zero() {
-    return __riscv_vfmv_v_f_f16mf2(0, __riscv_vsetvlmax_e16mf2());
-}
-template <> inline vfloat16m1_t set_zero() {
-    return __riscv_vfmv_v_f_f16m1(0, __riscv_vsetvlmax_e16m1());
-}
-template <> inline vfloat16m2_t set_zero() {
-    return __riscv_vfmv_v_f_f16m2(0, __riscv_vsetvlmax_e16m2());
-}
-template <> inline vfloat16m4_t set_zero() {
-    return __riscv_vfmv_v_f_f16m4(0, __riscv_vsetvlmax_e16m4());
-}
 template <> inline vfloat32m1_t set_zero() {
     return __riscv_vfmv_v_f_f32m1(0.0f, __riscv_vsetvlmax_e32m1());
 }
@@ -449,14 +448,22 @@ template <> inline vfloat32m8_t set_zero() {
 
 #if defined(__riscv_v_intrinsic)
 template <typename T> size_t vlmax() {
-    if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
-    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
-    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
-    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
-    else if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
+    if constexpr (std::is_same_v<T, vfloat32m1_t>) { return  __riscv_vsetvlmax_e32m1(); }
     else if constexpr (std::is_same_v<T, vfloat32m2_t>) { return  __riscv_vsetvlmax_e32m2(); }
     else if constexpr (std::is_same_v<T, vfloat32m4_t>) { return  __riscv_vsetvlmax_e32m4(); }
     else if constexpr (std::is_same_v<T, vfloat32m8_t>) { return  __riscv_vsetvlmax_e32m8(); }
+    #if defined (__riscv_zvfh)
+    else if constexpr (std::is_same_v<T, vfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    #endif
+    #if defined (__riscv_zvfbfwma)
+    else if constexpr (std::is_same_v<T, vbfloat16mf2_t>) { return  __riscv_vsetvlmax_e16mf2(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m1_t>) { return  __riscv_vsetvlmax_e16m1(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m2_t>) { return  __riscv_vsetvlmax_e16m2(); }
+    else if constexpr (std::is_same_v<T, vbfloat16m4_t>) { return  __riscv_vsetvlmax_e16m4(); }
+    #endif
     return 0;
 }
 #endif
@@ -3740,7 +3747,7 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             params->ith, params->nth};
         tb.matmul(m, n);
         return true;
-#elif defined(__riscv_zvfh)
+#elif defined(__riscv_v_intrinsic)
     #if LMUL == 1
         tinyBLAS_RVV<vfloat32m1_t, vfloat32m1_t, float, float, float> tb{ params,
             k, (const float *)A, lda,
@@ -3804,23 +3811,25 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             return true;
         }
 #elif defined(__riscv_zvfbfwma)
-        #if LMUL == 1
-            tinyBLAS_RVV<vfloat32m1_t, vbfloat16mf2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
-                k, (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-        #elif LMUL == 2
-            tinyBLAS_RVV<vfloat32m2_t, vbfloat16m1_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
-                k, (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-        #else // LMUL = 4
-            tinyBLAS_RVV<vfloat32m4_t, vbfloat16m2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
-                k, (const ggml_bf16_t *)A, lda,
-                (const ggml_bf16_t *)B, ldb,
-                (float *)C, ldc};
-        #endif
-            return tb.matmul(m, n);
+        if (Btype == GGML_TYPE_BF16) {
+            #if LMUL == 1
+                tinyBLAS_RVV<vfloat32m1_t, vbfloat16mf2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                    k, (const ggml_bf16_t *)A, lda,
+                    (const ggml_bf16_t *)B, ldb,
+                    (float *)C, ldc};
+            #elif LMUL == 2
+                tinyBLAS_RVV<vfloat32m2_t, vbfloat16m1_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                    k, (const ggml_bf16_t *)A, lda,
+                    (const ggml_bf16_t *)B, ldb,
+                    (float *)C, ldc};
+            #else // LMUL = 4
+                tinyBLAS_RVV<vfloat32m4_t, vbfloat16m2_t, ggml_bf16_t, ggml_bf16_t, float> tb{ params,
+                    k, (const ggml_bf16_t *)A, lda,
+                    (const ggml_bf16_t *)B, ldb,
+                    (float *)C, ldc};
+            #endif
+                return tb.matmul(m, n);
+        }
 #endif
         return false;
     }

ggml/src/ggml-cpu/ops.cpp

@@ -9953,13 +9953,9 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * k =          (float *) dst->src[0]->data;
     float * v =          (float *) dst->src[1]->data;
@@ -10170,13 +10166,9 @@ static void ggml_compute_forward_gla_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * k = (float *) dst->src[0]->data;
     float * v = (float *) dst->src[1]->data;
@@ -10633,13 +10625,9 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    if (ith >= HEADS) {
-        return;
-    }
-
-    const int h_start = (HEADS * ith) / nth;
-    const int h_end = ((HEADS * (ith + 1)) / nth < HEADS) ?
-                (HEADS * (ith + 1)) / nth : HEADS;
+    const int h_start =  (HEADS * (ith    )) / nth;
+    const int h_end   = ((HEADS * (ith + 1)) / nth < HEADS) ?
+                         (HEADS * (ith + 1)) / nth : HEADS;
 
     float * r = (float *) dst->src[0]->data;
     float * w = (float *) dst->src[1]->data;

ggml/src/ggml-cpu/vec.h

@@ -126,7 +126,7 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         const int ggml_f16_epr = sve_register_length / 16; // running when 16
         const int ggml_f16_step = 8 * ggml_f16_epr; // choose 8 SVE registers
 
-        const int np = (n & ~(ggml_f16_step - 1));
+        int np = (n & ~(ggml_f16_step - 1));
 
         svfloat16_t sum_00 = svdup_n_f16(0.0f);
         svfloat16_t sum_01 = svdup_n_f16(0.0f);
@@ -224,71 +224,75 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         }
         GGML_F16x_VEC_REDUCE(sumf[0], sum_00, sum_01, sum_02, sum_03);
         GGML_F16x_VEC_REDUCE(sumf[1], sum_10, sum_11, sum_12, sum_13);
+        np = n;
+    #elif defined(__riscv_v_intrinsic)
+        #if defined(__riscv_zvfh)
+            size_t vl = __riscv_vsetvlmax_e32m4();
 
-    #elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
-        size_t vl = __riscv_vsetvlmax_e32m4();
+            // initialize accumulators to all zeroes
+            vfloat32m4_t vsum0_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+            vfloat32m4_t vsum0_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+            vfloat32m4_t vsum1_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+            vfloat32m4_t vsum1_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
 
-        // initialize accumulators to all zeroes
-        vfloat32m4_t vsum0_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
-        vfloat32m4_t vsum0_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
-        vfloat32m4_t vsum1_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
-        vfloat32m4_t vsum1_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+            // calculate step size
+            const size_t epr = __riscv_vsetvlmax_e16m2();
+            const size_t step = epr * 2;
+            int np = (n & ~(step - 1));
 
-        // calculate step size
-        const size_t epr = __riscv_vsetvlmax_e16m2();
-        const size_t step = epr * 2;
-        const int np = (n & ~(step - 1));
+            // unroll by 2 along the row dimension
+            for (int i = 0; i < np; i += step) {
+                vfloat16m2_t ay0 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), epr);
+                vfloat16m2_t ax0_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), epr);
+                vfloat16m2_t ax1_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), epr);
+                vsum0_0 = __riscv_vfwmacc_vv_f32m4(vsum0_0, ax0_0, ay0, epr);
+                vsum1_0 = __riscv_vfwmacc_vv_f32m4(vsum1_0, ax1_0, ay0, epr);
 
-        // unroll by 2 along the row dimension
-        for (int i = 0; i < np; i += step) {
-            vfloat16m2_t ay0 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), epr);
-            vfloat16m2_t ax0_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), epr);
-            vfloat16m2_t ax1_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), epr);
-            vsum0_0 = __riscv_vfwmacc_vv_f32m4(vsum0_0, ax0_0, ay0, epr);
-            vsum1_0 = __riscv_vfwmacc_vv_f32m4(vsum1_0, ax1_0, ay0, epr);
+                vfloat16m2_t ay1 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i + epr), epr);
+                vfloat16m2_t ax0_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i + epr), epr);
+                vfloat16m2_t ax1_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i + epr), epr);
+                vsum0_1 = __riscv_vfwmacc_vv_f32m4(vsum0_1, ax0_1, ay1, epr);
+                vsum1_1 = __riscv_vfwmacc_vv_f32m4(vsum1_1, ax1_1, ay1, epr);
+            }
 
-            vfloat16m2_t ay1 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i + epr), epr);
-            vfloat16m2_t ax0_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i + epr), epr);
-            vfloat16m2_t ax1_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i + epr), epr);
-            vsum0_1 = __riscv_vfwmacc_vv_f32m4(vsum0_1, ax0_1, ay1, epr);
-            vsum1_1 = __riscv_vfwmacc_vv_f32m4(vsum1_1, ax1_1, ay1, epr);
-        }
+            vfloat32m4_t vsum0 = __riscv_vfadd_vv_f32m4(vsum0_0, vsum0_1, vl);
+            vfloat32m4_t vsum1 = __riscv_vfadd_vv_f32m4(vsum1_0, vsum1_1, vl);
 
-        vfloat32m4_t vsum0 = __riscv_vfadd_vv_f32m4(vsum0_0, vsum0_1, vl);
-        vfloat32m4_t vsum1 = __riscv_vfadd_vv_f32m4(vsum1_0, vsum1_1, vl);
+            // leftovers
+            for (int i = np; i < n; i += vl) {
+                vl = __riscv_vsetvl_e16m2(n - i);
+                vfloat16m2_t ay = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), vl);
+                vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), vl);
+                vfloat16m2_t ax1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), vl);
 
-        // leftovers
-        for (int i = np; i < n; i += vl) {
-            vl = __riscv_vsetvl_e16m2(n - i);
-            vfloat16m2_t ay = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), vl);
-            vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), vl);
-            vfloat16m2_t ax1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), vl);
+                vsum0 = __riscv_vfwmacc_vv_f32m4(vsum0, ax0, ay, vl);
+                vsum1 = __riscv_vfwmacc_vv_f32m4(vsum1, ax1, ay, vl);
+            }
 
-            vsum0 = __riscv_vfwmacc_vv_f32m4(vsum0, ax0, ay, vl);
-            vsum1 = __riscv_vfwmacc_vv_f32m4(vsum1, ax1, ay, vl);
-        }
-
-        // reduce
-        vl = __riscv_vsetvlmax_e32m2();
-        vfloat32m2_t acc0_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum0, 0),
-                                    __riscv_vget_v_f32m4_f32m2(vsum0, 1), vl);
-        vl = __riscv_vsetvlmax_e32m1();
-        vfloat32m1_t acc0_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc0_0, 0),
-        __riscv_vget_v_f32m2_f32m1(acc0_0, 1), vl);
-        vfloat32m1_t redsum0 = __riscv_vfredusum_vs_f32m1_f32m1(
-                                    acc0_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
-
-        vl = __riscv_vsetvlmax_e32m2();
-        vfloat32m2_t acc1_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum1, 0),
-                                    __riscv_vget_v_f32m4_f32m2(vsum1, 1), vl);
-        vl = __riscv_vsetvlmax_e32m1();
-        vfloat32m1_t acc1_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc1_0, 0),
-                                    __riscv_vget_v_f32m2_f32m1(acc1_0, 1), vl);
-        vfloat32m1_t redsum1 = __riscv_vfredusum_vs_f32m1_f32m1(
-                                    acc1_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
-        sumf[0] = __riscv_vfmv_f_s_f32m1_f32(redsum0);
-        sumf[1] = __riscv_vfmv_f_s_f32m1_f32(redsum1);
+            // reduce
+            vl = __riscv_vsetvlmax_e32m2();
+            vfloat32m2_t acc0_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum0, 0),
+                                        __riscv_vget_v_f32m4_f32m2(vsum0, 1), vl);
+            vl = __riscv_vsetvlmax_e32m1();
+            vfloat32m1_t acc0_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc0_0, 0),
+            __riscv_vget_v_f32m2_f32m1(acc0_0, 1), vl);
+            vfloat32m1_t redsum0 = __riscv_vfredusum_vs_f32m1_f32m1(
+                                        acc0_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
 
+            vl = __riscv_vsetvlmax_e32m2();
+            vfloat32m2_t acc1_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum1, 0),
+                                        __riscv_vget_v_f32m4_f32m2(vsum1, 1), vl);
+            vl = __riscv_vsetvlmax_e32m1();
+            vfloat32m1_t acc1_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc1_0, 0),
+                                        __riscv_vget_v_f32m2_f32m1(acc1_0, 1), vl);
+            vfloat32m1_t redsum1 = __riscv_vfredusum_vs_f32m1_f32m1(
+                                        acc1_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+            sumf[0] = __riscv_vfmv_f_s_f32m1_f32(redsum0);
+            sumf[1] = __riscv_vfmv_f_s_f32m1_f32(redsum1);
+            np = n;
+        #else
+            const int np = 0;
+        #endif
     #else
         const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -313,21 +317,17 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
         for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
             GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
         }
-
-        // leftovers
-        for (int i = np; i < n; ++i) {
-            for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-                sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
-            }
-        }
     #endif
 #else
-    for (int i = 0; i < n; ++i) {
+    // scalar path
+    const int np = 0;
+#endif
+    // scalar and leftovers
+    for (int i = np; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
             sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
-#endif
 
     for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
         s[i] = (float)sumf[i];
@@ -532,40 +532,45 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
         svst1_f16(pg, (__fp16 *)(y + np2), hy);
     }
     np = n;
-#elif defined(__riscv_zvfh) // implies __riscv_v_intrinsic
-    const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
-    const _Float16 scale = *(const _Float16*)(&s);
+#elif defined(__riscv_v_intrinsic) // implies __riscv_v_intrinsic
+    #if defined (__riscv_zvfh)
+        const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+        const _Float16 scale = *(const _Float16*)(&s);
 
-    // calculate step size
-    const int epr = __riscv_vsetvlmax_e16m4();
-    const int step = epr * 2;
-    int np = (n & ~(step - 1));
+        // calculate step size
+        const int epr = __riscv_vsetvlmax_e16m4();
+        const int step = epr * 2;
+        int np = (n & ~(step - 1));
 
-    // unroll by 2
-    for (int i = 0; i < np; i += step) {
-        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, epr);
-        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
-        ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, epr);
-        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
-        __asm__ __volatile__ ("" ::: "memory");
+        // unroll by 2
+        for (int i = 0; i < np; i += step) {
+            vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, epr);
+            vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+            ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, epr);
+            __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+            __asm__ __volatile__ ("" ::: "memory");
 
-        vfloat16m4_t ax1 = __riscv_vle16_v_f16m4((const _Float16*)x + i + epr, epr);
-        vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
-        ay1 = __riscv_vfmacc_vf_f16m4(ay1, scale, ax1, epr);
-        __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
-        __asm__ __volatile__ ("" ::: "memory");
-    }
+            vfloat16m4_t ax1 = __riscv_vle16_v_f16m4((const _Float16*)x + i + epr, epr);
+            vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+            ay1 = __riscv_vfmacc_vf_f16m4(ay1, scale, ax1, epr);
+            __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+            __asm__ __volatile__ ("" ::: "memory");
+        }
 
-    // leftovers
-    int vl;
-    for (int i = np; i < n; i += vl) {
-        vl = __riscv_vsetvl_e16m4(n - i);
-        vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, vl);
-        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
-        ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, vl);
-        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
-    }
-    np = n;
+        // leftovers
+        int vl;
+        for (int i = np; i < n; i += vl) {
+            vl = __riscv_vsetvl_e16m4(n - i);
+            vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, vl);
+            vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+            ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, vl);
+            __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
+        }
+        np = n;
+    #else
+        // fall to scalar path
+        const int np = 0;
+    #endif
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -584,10 +589,11 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
         }
     }
 #else
+    // scalar path
     const int np = 0;
 #endif
 
-    // leftovers
+    // scalar and leftovers
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
@@ -785,7 +791,7 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
     const int ggml_f16_step = 2 * ggml_f16_epr;
 
     GGML_F16x_VEC vx =  GGML_F16x_VEC_SET1(v);
-    const int np = (n & ~(ggml_f16_step - 1));
+    int np = (n & ~(ggml_f16_step - 1));
     svfloat16_t ay1, ay2;
 
     for (int i = 0; i < np; i += ggml_f16_step) {
@@ -805,36 +811,43 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
         svfloat16_t out = svmul_f16_m(pg, hy, vx);
         svst1_f16(pg, (__fp16 *)(y + np), out);
     }
-#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfh)
-    const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
-    const _Float16 scale = *(const _Float16*)(&s);
+    np = n;
+#elif defined(__riscv_v_intrinsic)
+    #if defined(__riscv_zvfh)
+        const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+        const _Float16 scale = *(const _Float16*)(&s);
 
-    // calculate step size
-    const int epr = __riscv_vsetvlmax_e16m4();
-    const int step = epr * 2;
-    const int np = (n & ~(step - 1));
+        // calculate step size
+        const int epr = __riscv_vsetvlmax_e16m4();
+        const int step = epr * 2;
+        int np = (n & ~(step - 1));
 
-    // unroll by 2
-    for (int i = 0; i < np; i += step) {
-        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
-        ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, epr);
-        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
-        __asm__ __volatile__ ("" ::: "memory");
+        // unroll by 2
+        for (int i = 0; i < np; i += step) {
+            vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+            ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, epr);
+            __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+            __asm__ __volatile__ ("" ::: "memory");
 
-        vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
-        ay1 = __riscv_vfmul_vf_f16m4(ay1, scale, epr);
-        __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
-        __asm__ __volatile__ ("" ::: "memory");
-    }
+            vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+            ay1 = __riscv_vfmul_vf_f16m4(ay1, scale, epr);
+            __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+            __asm__ __volatile__ ("" ::: "memory");
+        }
 
-    // leftovers
-    int vl;
-    for (int i = np; i < n; i += vl) {
-        vl = __riscv_vsetvl_e16m4(n - i);
-        vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
-        ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, vl);
-        __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
-    }
+        // leftovers
+        int vl;
+        for (int i = np; i < n; i += vl) {
+            vl = __riscv_vsetvl_e16m4(n - i);
+            vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+            ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, vl);
+            __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
+        }
+        np = n;
+    #else
+        // fall to scalar path
+        const int np = 0;
+    #endif
 #elif defined(GGML_SIMD)
     const int np = (n & ~(GGML_F16_STEP - 1));
 
@@ -850,17 +863,14 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
             GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
         }
     }
-
-    // leftovers
+#else
+    // scalar path
+    const int np = 0;
+#endif
+    // scalar and leftovers
     for (int i = np; i < n; ++i) {
         y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
-#else
-    // scalar
-    for (int i = 0; i < n; ++i) {
-        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
-    }
-#endif
 }
 
 inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s);   }

ggml/src/ggml-cuda/argsort.cu

@@ -47,9 +47,11 @@ void argsort_f32_i32_cuda_cub(ggml_cuda_pool & pool,
 #ifdef STRIDED_ITERATOR_AVAILABLE
     auto offset_iterator = cuda::make_strided_iterator(cuda::make_counting_iterator(0), ncols);
 #else
-    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows + 1);
+    // offset_iterator needs to populate nrows + 1 elements, so we also have to ceildiv nrows + 1 by block_size
+    const int                 nrows_offset = nrows + 1;
+    ggml_cuda_pool_alloc<int> offsets_alloc(pool, nrows_offset);
     int *                     offset_iterator = offsets_alloc.get();
-    const dim3                offset_grid((nrows + block_size - 1) / block_size);
+    const dim3                offset_grid((nrows_offset + block_size - 1) / block_size);
     init_offsets<<<offset_grid, block_size, 0, stream>>>(offset_iterator, ncols, nrows);
 #endif
     CUDA_CHECK(cudaMemcpyAsync(temp_keys, x, ncols * nrows * sizeof(float), cudaMemcpyDeviceToDevice, stream));

ggml/src/ggml-cuda/common.cuh

@@ -800,19 +800,32 @@ static __device__ __forceinline__ float ggml_cuda_e8m0_to_fp32(uint8_t x) {
 }
 
 static __device__ __forceinline__ float ggml_cuda_ue4m3_to_fp32(uint8_t x) {
-#ifdef FP8_AVAILABLE
-    const uint32_t bits = x * (x != 0x7F && x != 0xFF); // Convert NaN to 0.0f to match CPU implementation.
-#if defined(GGML_USE_HIP) && defined(CDNA3)
-    // ROCm dose not support fp8 in software on devices with fp8 hardware,
+#if defined(GGML_USE_HIP) && defined(CDNA3) && defined(FP8_AVAILABLE) && HIP_VERSION >= 60200000
+    // ROCm does not support fp8 in software on devices with fp8 hardware,
     // but CDNA3 supports only e4m3_fnuz (no inf).
+    const uint32_t bits = x * (x != 0x7F && x != 0xFF); // Convert NaN to 0.0f to match CPU implementation.
     const __hip_fp8_e4m3_fnuz xf = *reinterpret_cast<const __hip_fp8_e4m3_fnuz *>(&bits);
-#else
-    const __nv_fp8_e4m3 xf = *reinterpret_cast<const __nv_fp8_e4m3 *>(&bits);
-#endif // defined(GGML_USE_HIP) && defined(GGML_USE_HIP)
     return static_cast<float>(xf) / 2;
 #else
-    NO_DEVICE_CODE;
-#endif // FP8_AVAILABLE
+#if defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP)
+    const uint32_t bits = x * (x != 0x7F && x != 0xFF); // Convert NaN to 0.0f to match CPU implementation.
+    const __nv_fp8_e4m3 xf = *reinterpret_cast<const __nv_fp8_e4m3 *>(&bits);
+    return static_cast<float>(xf) / 2;
+#else
+    if (x == 0 || (x == 0x7F && x != 0xFF)) { // Convert NaN to 0.0f
+        return 0.0f;
+    }
+    const int exp = (x >> 3) & 0xF;
+    const int man = x & 0x7;
+    float raw;
+    if (exp == 0) {
+        raw = ldexpf((float) man, -9);
+    } else {
+        raw = ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    }
+    return static_cast<float>(raw / 2);
+#endif // defined(FP8_AVAILABLE) && !defined(GGML_USE_HIP)
+#endif // defined(GGML_USE_HIP) && defined(CDNA3) && defined(FP8_AVAILABLE) && HIP_VERSION >= 60200000
 }
 
 __device__ __forceinline__ uint8_t ggml_cuda_float_to_fp4_e2m1(float x, float e) {

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

@@ -66,6 +66,11 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512,  8,  64, 4,  32, 256, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32, 256, 256, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 64, 256, 1,  32, 128, 128, 128, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
@@ -80,6 +85,11 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 64, 256, 1,  32, 128, 128, 128, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
@@ -89,6 +99,11 @@ 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_volta(const int DKQ, const int DV, const int ncols) {
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512,  8,  64, 4,  32, 256, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32, 256, 256,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128,  64, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 64, 256, 1,  32, 128, 128,  64, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512,  8,  64, 4,  32, 288, 256,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32, 288, 256,  64, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128,  64, 1, false);
@@ -103,6 +118,10 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
 
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 16,  64, 4,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 32, 128, 2,  32, 128, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(512, 512, 64, 256, 1,  32, 128, 128, 128, 1, false);
+
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
@@ -1552,7 +1571,7 @@ static __global__ void flash_attn_ext_f16(
 #if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)) || defined(AMD_MFMA_AVAILABLE))
 
     // Skip unused kernel variants for faster compilation:
-    if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
+    if (use_logit_softcap && !(DKQ == 128 || DKQ == 256 || DKQ == 512)) {
         NO_DEVICE_CODE;
         return;
     }
@@ -1815,6 +1834,15 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(112, 112,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(128, 128,  64)
 DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256,  64)
 
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  2,  4);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  4,  4);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  8,  4);
+extern DECL_FATTN_MMA_F16_CASE(512, 512, 16,  4);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  1,  8);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  2,  8);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  4,  8);
+extern DECL_FATTN_MMA_F16_CASE(512, 512,  8,  8);
+
 // The number of viable configurations for Deepseek is very limited:
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
 extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);

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

@@ -38,6 +38,10 @@ void ggml_cuda_flash_attn_ext_tile(ggml_backend_cuda_context & ctx, ggml_tensor
             GGML_ASSERT(V->ne[0] == K->ne[0]);
             ggml_cuda_flash_attn_ext_tile_case<256, 256>(ctx, dst);
         } break;
+        case 512: {
+            GGML_ASSERT(V->ne[0] == K->ne[0]);
+            ggml_cuda_flash_attn_ext_tile_case<512, 512>(ctx, dst);
+        } break;
         case 576: {
             GGML_ASSERT(V->ne[0] == 512);
             ggml_cuda_flash_attn_ext_tile_case<576, 512>(ctx, dst);

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

@@ -68,6 +68,10 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  64,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
@@ -124,6 +128,10 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32,  64)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  32,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  32,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  32,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  32,  64)
@@ -187,6 +195,11 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2,  32, 128)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2,  32, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 512, 1, 128,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2,  64,  64)
@@ -251,6 +264,11 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5,  32, 256)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3,  64, 128)
 
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  4, 128, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512,  8, 256, 2,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4,  64,  64)
+    GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 32, 256, 2, 128,  64)
+
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  4, 128, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512,  8, 256, 2,  64,  64)
     GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4,  64,  64)
@@ -767,7 +785,7 @@ static __global__ void flash_attn_tile(
 #ifdef GGML_USE_WMMA_FATTN
             (ncols2 != 1 && DV != 40 && DV != 72 && DV != 512) ||
 #endif // GGML_USE_WMMA_FATTN
-            (use_logit_softcap && !(DV == 128 || DV == 256))
+            (use_logit_softcap && !(DV == 128 || DV == 256 || DV == 512))
     ) {
         GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,
             max_bias, m0, m1, n_head_log2, logit_softcap,
@@ -1192,7 +1210,7 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
     const int gqa_limit = nvidia && gqa_ratio <= 4 && DV <= 256 ? 16 : INT_MAX;
     const bool use_gqa_opt = mask && max_bias == 0.0f && Q->ne[1] <= gqa_limit && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
-    if constexpr (DV == 512) {
+    if constexpr (DKQ == 576) {
         if (use_gqa_opt && gqa_ratio % 16 == 0) {
             launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
             return;
@@ -1203,7 +1221,7 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
         }
     }
 
-    if constexpr (DV <= 256) {
+    if constexpr (DKQ <= 512) {
         if (use_gqa_opt && gqa_ratio % 8 == 0) {
             launch_fattn_tile_switch_ncols1<DKQ, DV, 8, use_logit_softcap>(ctx, dst);
             return;
@@ -1214,13 +1232,15 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
             return;
         }
 
-        if (use_gqa_opt && gqa_ratio % 2 == 0) {
-            launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
+        if constexpr (DV <= 256) {
+            if (use_gqa_opt && gqa_ratio % 2 == 0) {
+                launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
+                return;
+            }
+
+            launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
             return;
         }
-
-        launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
-        return;
     }
     GGML_ABORT("fatal error");
 }
@@ -1255,4 +1275,5 @@ extern DECL_FATTN_TILE_CASE( 96,  96);
 extern DECL_FATTN_TILE_CASE(112, 112);
 extern DECL_FATTN_TILE_CASE(128, 128);
 extern DECL_FATTN_TILE_CASE(256, 256);
+extern DECL_FATTN_TILE_CASE(512, 512);
 extern DECL_FATTN_TILE_CASE(576, 512);

ggml/src/ggml-cuda/fattn.cu

@@ -135,6 +135,10 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
             GGML_ASSERT(V->ne[0] == 256);
             ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
             break;
+        case 512:
+            GGML_ASSERT(V->ne[0] == 512);
+            ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<512, 512>(ctx, dst);
+            break;
         case 576: {
             // For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
             GGML_ASSERT(V->ne[0] == 512);
@@ -340,6 +344,14 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
                 return BEST_FATTN_KERNEL_NONE;
             }
             break;
+        case 512:
+            if (V->ne[0] != K->ne[0]) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            if (!gqa_opt_applies) {
+                return BEST_FATTN_KERNEL_NONE;
+            }
+            break;
         case 576:
             if (V->ne[0] != 512) {
                 return BEST_FATTN_KERNEL_NONE;
@@ -424,7 +436,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // Use the WMMA kernel if possible:
-    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
+    if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 512 && Q->ne[0] != 576) {
         if (can_use_vector_kernel && Q->ne[1] <= 2) {
             return BEST_FATTN_KERNEL_VEC;
         }
@@ -457,7 +469,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     }
 
     // Use MFMA flash attention for CDNA (MI100+):
-    if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 576) {
+    if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 512 && Q->ne[0] != 576) {
         const int64_t eff_nq = Q->ne[1] * (gqa_opt_applies ? gqa_ratio : 1);
         // MMA vs tile crossover benchmarked on MI300X @ d32768:
         //   hsk=64  (gqa=4): MMA wins at eff >= 128 (+11%)

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

@@ -2343,7 +2343,8 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
         static_assert(MMVQ_MAX_BATCH_SIZE == MMVF_MAX_BATCH_SIZE);
         if (ne2 <= MMVQ_MAX_BATCH_SIZE) {
             if (ggml_is_quantized(src0->type)) {
-                if (ne2 <= MMVQ_MMID_MAX_BATCH_SIZE) {
+                const int mmvq_mmid_max = get_mmvq_mmid_max_batch(src0->type, cc);
+                if (ne2 <= mmvq_mmid_max) {
                     ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
                     return;
                 }
@@ -2946,14 +2947,18 @@ static bool ggml_cuda_graph_check_compability(ggml_cgraph * cgraph) {
         }
 
         // [TAG_MUL_MAT_ID_CUDA_GRAPHS]
-        if (node->op == GGML_OP_MUL_MAT_ID && (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > MMVQ_MMID_MAX_BATCH_SIZE)) {
-            // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
-            // TODO: figure out a way to enable for larger batch sizes, without hurting performance
-            // ref: https://github.com/ggml-org/llama.cpp/pull/18958
-            use_cuda_graph = false;
+        if (node->op == GGML_OP_MUL_MAT_ID) {
+            const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+            const int mmvq_mmid_max = get_mmvq_mmid_max_batch(node->src[0]->type, cc);
+            if (!ggml_is_quantized(node->src[0]->type) || node->ne[2] > mmvq_mmid_max) {
+                // under these conditions, the mul_mat_id operation will need to synchronize the stream, so we cannot use CUDA graphs
+                // TODO: figure out a way to enable for larger batch sizes, without hurting performance
+                // ref: https://github.com/ggml-org/llama.cpp/pull/18958
+                use_cuda_graph = false;
 #ifndef NDEBUG
-            GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
+                GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
 #endif
+            }
         }
 
         if (!use_cuda_graph) {
@@ -4786,9 +4791,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
                     case GGML_TYPE_MXFP4:
-#ifdef FP8_AVAILABLE
                     case GGML_TYPE_NVFP4:
-#endif // FP8_AVAILABLE
                     case GGML_TYPE_Q2_K:
                     case GGML_TYPE_Q3_K:
                     case GGML_TYPE_Q4_K:

ggml/src/ggml-cuda/mmq.cu

@@ -23,6 +23,9 @@ static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, con
         case GGML_TYPE_MXFP4:
             mul_mat_q_case<GGML_TYPE_MXFP4>(ctx, args, stream);
             break;
+        case GGML_TYPE_NVFP4:
+            mul_mat_q_case<GGML_TYPE_NVFP4>(ctx, args, stream);
+            break;
         case GGML_TYPE_Q2_K:
             mul_mat_q_case<GGML_TYPE_Q2_K>(ctx, args, stream);
             break;
@@ -273,6 +276,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -362,5 +366,4 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     }
 
     return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
-
 }

ggml/src/ggml-cuda/mmq.cuh

@@ -68,6 +68,8 @@ static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
             return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_MXFP4:
             return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_NVFP4:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_Q2_K:
             return MMQ_Q8_1_DS_LAYOUT_D2S6;
         case GGML_TYPE_Q3_K:
@@ -189,6 +191,7 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
         case GGML_TYPE_Q5_1:    return MMQ_DP4A_TXS_Q8_1;
         case GGML_TYPE_Q8_0:    return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_MXFP4:   return MMQ_DP4A_TXS_Q8_1;
+        case GGML_TYPE_NVFP4:   return MMQ_DP4A_TXS_Q8_0_16;
         case GGML_TYPE_Q2_K:    return MMQ_DP4A_TXS_Q2_K;
         case GGML_TYPE_Q3_K:    return MMQ_DP4A_TXS_Q3_K;
         case GGML_TYPE_Q4_K:    return MMQ_DP4A_TXS_Q4_K;
@@ -206,12 +209,13 @@ static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml
     }
 }
 
-#define MMQ_MMA_TILE_X_K_Q8_0 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
-#define MMQ_MMA_TILE_X_K_FP4  (2*MMQ_TILE_NE_K + 8                                       + 4)
-#define MMQ_MMA_TILE_X_K_Q8_1 (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
-#define MMQ_MMA_TILE_X_K_Q2_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K                           + 4)
-#define MMQ_MMA_TILE_X_K_Q3_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2                         + 4)
-#define MMQ_MMA_TILE_X_K_Q6_K (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K   + MMQ_TILE_NE_K/8 + 7)
+#define MMQ_MMA_TILE_X_K_Q8_0  (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
+#define MMQ_MMA_TILE_X_K_FP4   (2*MMQ_TILE_NE_K + 8                                       + 4) // MXFP4
+#define MMQ_MMA_TILE_X_K_NVFP4 (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2                         + 4) // NVFP4
+#define MMQ_MMA_TILE_X_K_Q8_1  (2*MMQ_TILE_NE_K + 2*MMQ_TILE_NE_K/QI8_0                   + 4)
+#define MMQ_MMA_TILE_X_K_Q2_K  (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K                           + 4)
+#define MMQ_MMA_TILE_X_K_Q3_K  (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/2                         + 4)
+#define MMQ_MMA_TILE_X_K_Q6_K  (2*MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI6_K   + MMQ_TILE_NE_K/8 + 7)
 
 static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding.");
@@ -220,6 +224,8 @@ static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_FP4  % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_FP4 == MMQ_MMA_TILE_X_K_Q8_1, "Wrong tile size for MXFP4");
+static_assert(MMQ_MMA_TILE_X_K_NVFP4 % 8 == 4, "Wrong padding.");
+
 
 static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
     switch (type) {
@@ -230,6 +236,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
         case GGML_TYPE_Q8_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         // tile sizes are the same for Q8_1 and FP4 for blackwell
         case GGML_TYPE_MXFP4:   return MMQ_MMA_TILE_X_K_Q8_1;
+        case GGML_TYPE_NVFP4:   return MMQ_MMA_TILE_X_K_NVFP4;
         case GGML_TYPE_Q2_K:    return MMQ_MMA_TILE_X_K_Q2_K;
         case GGML_TYPE_Q3_K:    return MMQ_MMA_TILE_X_K_Q3_K;
         case GGML_TYPE_Q4_K:    return MMQ_MMA_TILE_X_K_Q8_1;
@@ -826,6 +833,65 @@ static __device__ __forceinline__ void load_tiles_mxfp4_fp4(const char * __restr
     }
 }
 
+
+template <int mmq_y, bool need_check>
+static __device__ __forceinline__ void load_tiles_nvfp4(const char * __restrict__ x,
+                                                        int * __restrict__ x_tile,
+                                                        const int kb0,
+                                                        const int i_max,
+                                                        const int stride) {
+    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    int   * x_qs = (int   *) x_tile;
+    float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_NVFP4, mmq_y);
+    int   * x_qs = (int   *) x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+
+    constexpr int threads_per_row = MMQ_ITER_K / QK_NVFP4;
+    constexpr int rows_per_warp = warp_size / threads_per_row;
+    const int kbx = threadIdx.x % threads_per_row;
+    const int row_in_warp = threadIdx.x / threads_per_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += rows_per_warp * nwarps) {
+        int i = i0 + threadIdx.y * rows_per_warp + row_in_warp;
+
+        if constexpr (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_nvfp4 * bxi = (const block_nvfp4 *) x + kb0 + i * stride + kbx;
+        const uint32_t * __restrict__ src_qs = reinterpret_cast<const uint32_t *>(bxi->qs);
+        const int kqs = 16 * kbx;
+        const int ksc = 4 * kbx;
+
+#pragma unroll
+        for (int sub = 0; sub < QK_NVFP4 / QK_NVFP4_SUB; ++sub) {
+            const int2 q0 = get_int_from_table_16(src_qs[2 * sub + 0], kvalues_mxfp4);
+            const int2 q1 = get_int_from_table_16(src_qs[2 * sub + 1], kvalues_mxfp4);
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+            x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 0] = q0.x;
+            x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 1] = q1.x;
+            x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 2] = q0.y;
+            x_qs[i * MMQ_MMA_TILE_X_K_NVFP4 + kqs + 4 * sub + 3] = q1.y;
+            x_df[i * MMQ_MMA_TILE_X_K_NVFP4 + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]);
+#else
+            x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 0] = q0.x;
+            x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 1] = q1.x;
+            x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 2] = q0.y;
+            x_qs[i * (2 * MMQ_TILE_NE_K + 1) + kqs + 4 * sub + 3] = q1.y;
+            x_df[i * (2 * MMQ_TILE_NE_K * 2 / QI_NVFP4) + i / (QK_NVFP4_SUB / QI_NVFP4) + ksc + sub] = ggml_cuda_ue4m3_to_fp32(bxi->d[sub]);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        }
+    }
+}
+
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
@@ -1229,7 +1295,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 }
 
-// Used for Q3_K, IQ2_S, and IQ2_XS
+// Used for NVFP4, Q3_K, IQ2_S, and IQ2_XS
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
@@ -3261,6 +3327,14 @@ struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_MXFP4> {
     static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
 };
 
+template <int mmq_x, int mmq_y, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_NVFP4> {
+    static constexpr int              vdr          = VDR_NVFP4_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_nvfp4<mmq_y, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y>;
+};
+
 template <int mmq_x, int mmq_y, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q2_K> {
     static constexpr int              vdr          = VDR_Q2_K_Q8_1_MMQ;
@@ -4069,6 +4143,7 @@ extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
 extern DECL_MMQ_CASE(GGML_TYPE_Q5_1);
 extern DECL_MMQ_CASE(GGML_TYPE_Q8_0);
 extern DECL_MMQ_CASE(GGML_TYPE_MXFP4);
+extern DECL_MMQ_CASE(GGML_TYPE_NVFP4);
 extern DECL_MMQ_CASE(GGML_TYPE_Q2_K);
 extern DECL_MMQ_CASE(GGML_TYPE_Q3_K);
 extern DECL_MMQ_CASE(GGML_TYPE_Q4_K);

ggml/src/ggml-cuda/mmvq.cu

@@ -97,6 +97,197 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
+// Per-architecture maximum batch size for which MMVQ should be used for MUL_MAT_ID.
+// Returns a value <= MMVQ_MAX_BATCH_SIZE. Default is MMVQ_MAX_BATCH_SIZE.
+// Check https://github.com/ggml-org/llama.cpp/pull/20905#issuecomment-4145835627 for details
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_pascal_older(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 6;
+        case GGML_TYPE_Q4_1:    return 6;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_0:    return 6;
+        case GGML_TYPE_Q5_1:    return 6;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_turing_plus(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 7;
+        case GGML_TYPE_IQ3_S:   return 6;
+        case GGML_TYPE_IQ3_XXS: return 7;
+        case GGML_TYPE_MXFP4:   return 7;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_gcn(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 5;
+        case GGML_TYPE_IQ1_M:   return 5;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 5;
+        case GGML_TYPE_Q4_1:    return 5;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_cdna(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 5;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna1_rdna2(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_K:    return 6;
+        case GGML_TYPE_Q6_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna3(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 6;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 7;
+        case GGML_TYPE_IQ1_M:   return 7;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 7;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 5;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 7;
+        case GGML_TYPE_Q4_1:    return 7;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_0:    return 7;
+        case GGML_TYPE_Q5_1:    return 7;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 5;
+        case GGML_TYPE_Q8_0:    return 7;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+// Host function: returns the max batch size for the current arch+type at runtime.
+int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
+    // NVIDIA: Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+            return MMVQ_MAX_BATCH_SIZE;
+        }
+        if (cc >= GGML_CUDA_CC_TURING) {
+            return get_mmvq_mmid_max_batch_turing_plus(type);
+        }
+        return get_mmvq_mmid_max_batch_pascal_older(type);
+    }
+
+    // AMD
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return get_mmvq_mmid_max_batch_rdna4(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+            return get_mmvq_mmid_max_batch_rdna3(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
+            return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+        }
+        if (GGML_CUDA_CC_IS_CDNA(cc)) {
+            return get_mmvq_mmid_max_batch_cdna(type);
+        }
+        if (GGML_CUDA_CC_IS_GCN(cc)) {
+            return get_mmvq_mmid_max_batch_gcn(type);
+        }
+    }
+    return MMVQ_MAX_BATCH_SIZE;
+}
+
+// Device constexpr: returns the max batch size for the current arch+type at compile time.
+template <ggml_type type>
+static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
+#if defined(RDNA4)
+    return get_mmvq_mmid_max_batch_rdna4(type);
+#elif defined(RDNA3)
+    return get_mmvq_mmid_max_batch_rdna3(type);
+#elif defined(RDNA2) || defined(RDNA1)
+    return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+#elif defined(CDNA)
+    return get_mmvq_mmid_max_batch_cdna(type);
+#elif defined(GCN)
+    return get_mmvq_mmid_max_batch_gcn(type);
+#elif defined(__CUDA_ARCH__) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || __CUDA_ARCH__ >= GGML_CUDA_CC_ADA_LOVELACE)
+    return MMVQ_MAX_BATCH_SIZE;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+    return get_mmvq_mmid_max_batch_turing_plus(type);
+#else
+    return get_mmvq_mmid_max_batch_pascal_older(type);
+#endif
+}
+
 static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
@@ -195,7 +386,7 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-template <ggml_type type, int ncols_dst, bool has_fusion, bool is_multi_token_id = false, bool small_k = false>
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
 __launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
         const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
@@ -222,22 +413,13 @@ static __global__ void mul_mat_vec_q(
 
     const uint32_t channel_dst = blockIdx.y;
 
-    uint32_t token_idx = 0;
     uint32_t channel_x;
     uint32_t channel_y;
     uint32_t sample_dst;
 
-    if constexpr (is_multi_token_id) {
-        // Multi-token MUL_MAT_ID path, adding these in the normal path causes a perf regression for n_tokens=1 case
-        token_idx  = blockIdx.z;
-        channel_x  = ids[channel_dst + token_idx * ids_stride];
-        channel_y  = fastmodulo(channel_dst, nchannels_y);
-        sample_dst = 0;
-    } else {
-        channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-        channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-        sample_dst = blockIdx.z;
-    }
+    channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+    channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+    sample_dst = blockIdx.z;
 
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
@@ -294,9 +476,6 @@ static __global__ void mul_mat_vec_q(
     float tmp_gate[ncols_dst][rows_per_cuda_block] = {{0.0f}};
 
     const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
-    if constexpr (is_multi_token_id) {
-        y += token_idx*stride_col_y;
-    }
     const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
@@ -350,10 +529,6 @@ static __global__ void mul_mat_vec_q(
 
     dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
 
-    if constexpr (is_multi_token_id) {
-        dst += token_idx*stride_col_dst;
-    }
-
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_dst; ++j) {
@@ -413,6 +588,69 @@ static __global__ void mul_mat_vec_q(
     }
 }
 
+// Dedicated MoE multi-token kernel.
+// Grid: (ceil(nrows_x / c_rows_per_block), nchannels_dst)
+// Block: (warp_size, ncols_dst) - each warp handles one token independently.
+// No shared memory reduction needed since each warp works alone.
+template <ggml_type type, int c_rows_per_block>
+__launch_bounds__(get_mmvq_mmid_max_batch_for_device<type>()*ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mul_mat_vec_q_moe(
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids,
+        float * __restrict__ dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride) {
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+    const uint32_t token_idx   = threadIdx.y;
+    const int      row0        = c_rows_per_block*blockIdx.x;
+    const int      blocks_per_row_x = ncols_x / qk;
+    constexpr int  blocks_per_iter  = vdr * warp_size / qi;
+
+    const uint32_t channel_dst = blockIdx.y;
+
+    if (token_idx >= ncols_dst) {
+        return;
+    }
+
+    const uint32_t channel_x = ids[channel_dst + token_idx * ids_stride];
+    const uint32_t channel_y = fastmodulo(channel_dst, nchannels_y);
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + channel_y*stride_channel_y + token_idx*stride_col_y;
+    const int kbx_offset  = channel_x*stride_channel_x + row0*stride_row_x;
+
+    // partial sum for each thread
+    float tmp[c_rows_per_block] = {0.0f};
+
+    for (int kbx = threadIdx.x / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1);
+        const int kqs = vdr * (threadIdx.x % (qi/vdr));
+
+#pragma unroll
+        for (int i = 0; i < c_rows_per_block; ++i) {
+            tmp[i] += vec_dot_q_cuda(vx, &y[kby], kbx_offset + i*stride_row_x + kbx, kqs);
+        }
+    }
+
+    // Warp-level reduction only - no shared memory needed
+#pragma unroll
+    for (int i = 0; i < c_rows_per_block; ++i) {
+        tmp[i] = warp_reduce_sum<warp_size>(tmp[i]);
+    }
+
+    // Write results
+    if (threadIdx.x < c_rows_per_block && (c_rows_per_block == 1 || uint32_t(row0 + threadIdx.x) < nrows_x)) {
+        dst[channel_dst*stride_channel_dst + token_idx*stride_col_dst + row0 + threadIdx.x] = tmp[threadIdx.x];
+    }
+}
+
 template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
         const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
@@ -425,7 +663,7 @@ static std::pair<dim3, dim3> calc_launch_params(
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst, bool is_multi_token_id = false, bool small_k = false>
+template<ggml_type type, int c_ncols_dst, bool small_k = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
@@ -438,7 +676,7 @@ static void mul_mat_vec_q_switch_fusion(
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+            mul_mat_vec_q<type, c_ncols_dst, true, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
                 (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
@@ -448,12 +686,33 @@ static void mul_mat_vec_q_switch_fusion(
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false, is_multi_token_id, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
+    mul_mat_vec_q<type, c_ncols_dst, false, small_k><<<block_nums, block_dims, nbytes_shared, stream>>>
         (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
         sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
 }
 
+template <ggml_type type>
+static void mul_mat_vec_q_moe_launch(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride,
+        const int warp_size, const int nchannels_dst, cudaStream_t stream) {
+
+    constexpr int rows_per_block = 2; // 2 gives best perf based on tuning
+    const int64_t nblocks_rows = (nrows_x + rows_per_block - 1) / rows_per_block;
+    const dim3 block_nums(nblocks_rows, nchannels_dst);
+    const dim3 block_dims(warp_size, ncols_dst);
+
+    mul_mat_vec_q_moe<type, rows_per_block><<<block_nums, block_dims, 0, stream>>>(
+        vx, vy, ids, dst, ncols_x, nchannels_y, nrows_x,
+        stride_row_x, stride_col_y, stride_col_dst,
+        stride_channel_x, stride_channel_y, stride_channel_dst,
+        ncols_dst, ids_stride);
+}
+
 template <ggml_type type>
 static void mul_mat_vec_q_switch_ncols_dst(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
@@ -472,20 +731,62 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
+    const int                     cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
-    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
+    const mmvq_parameter_table_id table_id  = get_device_table_id(cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     const bool has_ids = ids != nullptr;
 
+    const auto should_use_small_k = [&](int c_ncols_dst) {
+        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
+        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
+        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+        constexpr int vdr                   = get_vdr_mmvq(type);
+        const int     blocks_per_row_x      = ncols_x / qk;
+        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
+        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+
+        constexpr std::array<ggml_type, 2> iq_slow_turing = {
+            GGML_TYPE_IQ3_XXS,
+            GGML_TYPE_IQ3_S,
+        };
+        constexpr std::array<ggml_type, 8> iq_slow_other = {
+            GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,   GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
+            GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S,   GGML_TYPE_IQ4_XS,
+        };
+        constexpr std::array<ggml_type, 3> slow_pascal = {
+            GGML_TYPE_IQ3_S,
+            GGML_TYPE_Q2_K,
+            GGML_TYPE_Q3_K,
+        };
+
+        const bool is_nvidia_turing_plus  = GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_TURING;
+        const bool is_nvidia_pascal_older = GGML_CUDA_CC_IS_NVIDIA(cc) && cc < GGML_CUDA_CC_VOLTA;
+
+        if (is_nvidia_turing_plus) {
+            if (ncols_dst == 1 &&
+                    std::find(iq_slow_turing.begin(), iq_slow_turing.end(), type) != iq_slow_turing.end()) {
+                use = false;
+            }
+        } else if ((ncols_dst == 1 && std::find(iq_slow_other.begin(), iq_slow_other.end(), type) != iq_slow_other.end()) ||
+                (is_nvidia_pascal_older && std::find(slow_pascal.begin(), slow_pascal.end(), type) != slow_pascal.end()) ||
+                GGML_CUDA_CC_IS_RDNA(cc)) {
+            use = false;
+        }
+
+        return use;
+    };
+
     if (has_ids && ncols_dst > 1) {
-        // Multi-token MUL_MAT_ID path only - single-token goes through regular path below
-        constexpr int c_ncols_dst = 1;
-        std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, ncols_dst, warp_size, table_id);
-        mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-             channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-             sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-             dims.first, dims.second, 0, ids_stride, stream);
+        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
+        mul_mat_vec_q_moe_launch<type>(
+            vx, vy, ids, dst, ncols_x, nchannels_y_fd, nrows_x,
+            stride_row_x, stride_col_y, stride_col_dst,
+            stride_channel_x, stride_channel_y, stride_channel_dst,
+            ncols_dst, ids_stride, warp_size, nchannels_dst, stream);
         return;
     }
 
@@ -493,31 +794,24 @@ static void mul_mat_vec_q_switch_ncols_dst(
         case 1: {
             constexpr int c_ncols_dst = 1;
 
-            // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
-            // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
-            constexpr int qk  = ggml_cuda_type_traits<type>::qk;
-            constexpr int qi  = ggml_cuda_type_traits<type>::qi;
-            constexpr int vdr = get_vdr_mmvq(type);
-            const int blocks_per_row_x = ncols_x / qk;
-            const int blocks_per_iter_1warp = vdr * warp_size / qi;
-            const int nwarps = calc_nwarps(type, c_ncols_dst, table_id);
-            const bool use_small_k = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+            bool use_small_k = should_use_small_k(c_ncols_dst);
+
             if (use_small_k) {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id, true);
-                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, false, true>(
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id, true);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             } else {
-                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst,
-                                                                    warp_size, table_id);
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id);
                 mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
                     vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                    sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                    dims.first, dims.second, 0, ids_stride, stream);
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
             }
         } break;
         case 2: {

ggml/src/ggml-cuda/mmvq.cuh

@@ -1,7 +1,10 @@
 #include "common.cuh"
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
-#define MMVQ_MMID_MAX_BATCH_SIZE 4 // Max. batch size for which to use MMVQ kernels for MUL_MAT_ID
+
+// Returns the maximum batch size for which MMVQ should be used for MUL_MAT_ID,
+// based on the quantization type and GPU architecture (compute capability).
+int get_mmvq_mmid_max_batch(ggml_type type, int cc);
 
 void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, const ggml_cuda_mm_fusion_args_host * fusion = nullptr);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_1-ncols2_8.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 1, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 1, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 1, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 1, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 1, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 16, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 2, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_8.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 2, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 2, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 2, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 2, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 4, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_8.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 4, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 4, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 4, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 4, 8);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu

@@ -8,4 +8,5 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
+DECL_FATTN_MMA_F16_CASE(512, 512, 8, 4);
 DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);

ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_8.cu

@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 8);
 DECL_FATTN_MMA_F16_CASE(112, 112, 8, 8);
 DECL_FATTN_MMA_F16_CASE(128, 128, 8, 8);
 DECL_FATTN_MMA_F16_CASE(256, 256, 8, 8);
+DECL_FATTN_MMA_F16_CASE(512, 512, 8, 8);

ggml/src/ggml-cuda/template-instances/fattn-tile-instance-dkq512-dv512.cu

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

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

@@ -3,7 +3,7 @@
 from glob import glob
 import os
 
-HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 576]
+HEAD_SIZES_KQ = [40, 64, 72, 80, 96, 112, 128, 256, 512, 576]
 
 TYPES_KV = ["GGML_TYPE_F16", "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", "GGML_TYPE_BF16"]
 
@@ -35,7 +35,7 @@ TYPES_MMQ = [
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
     "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
     "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",
-    "GGML_TYPE_IQ1_S", "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS", "GGML_TYPE_MXFP4"
+    "GGML_TYPE_IQ1_S", "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS", "GGML_TYPE_MXFP4", "GGML_TYPE_NVFP4"
 ]
 
 SOURCE_MMQ = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
@@ -83,6 +83,8 @@ for ncols in [8, 16, 32, 64]:
                     continue
                 if head_size_kq == 72:
                     continue
+                if head_size_kq == 512 and ncols2 not in (4, 8):
+                    continue
                 if head_size_kq != 576 and ncols2 in (16, 32):
                     continue
                 if head_size_kq == 576 and ncols2 not in (4, 16, 32):

ggml/src/ggml-cuda/template-instances/mmq-instance-nvfp4.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_NVFP4);

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1406,6 +1406,13 @@ static void ggml_backend_hexagon_buffer_set_tensor(ggml_backend_buffer_t buffer,
             repack_q8_0_q8x4x2(tensor, data, size);
             break;
 
+        case GGML_TYPE_IQ4_NL:
+            GGML_ASSERT(offset == 0);
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
+            // IQ4_NL has identical block layout to Q4_0 (ggml_half d + uint8_t qs[16])
+            repack_q4_0_q4x4x2(tensor, data, size);
+            break;
+
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
             GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
@@ -1442,6 +1449,12 @@ static void ggml_backend_hexagon_buffer_get_tensor(ggml_backend_buffer_t buffer,
             repack_q8x4x2_q8_0(data, tensor, size);
             break;
 
+        case GGML_TYPE_IQ4_NL:
+            GGML_ASSERT(offset == 0);
+            GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
+            repack_q4x4x2_q4_0(data, tensor, size);
+            break;
+
         case GGML_TYPE_MXFP4:
             GGML_ASSERT(offset == 0);
             GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
@@ -1819,6 +1832,7 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_MXFP4:
             if (src0->ne[0] % 32) {
                 return false;
@@ -1868,6 +1882,7 @@ static bool ggml_hexagon_supported_mul_mat_id(const struct ggml_hexagon_session
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_MXFP4:
             if ((src0->ne[0] % 32)) {
                 return false;
@@ -2596,8 +2611,26 @@ static void ggml_backend_hexagon_free(ggml_backend_t backend) {
     delete backend;
 }
 
+// Map weight type to its activation quantization family.
+// Types in the same family produce identical Q8 formats in VTCM and can
+// safely share quantized activation data via SKIP_QUANTIZE.
+// When adding a new quantized type, assign it the correct family here.
+static inline int act_quant_family(enum ggml_type wtype) {
+    switch (wtype) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_MXFP4:
+            return 1;  // Q8x4x2
+        default:
+            return 0;  // unknown / not quantized
+    }
+}
+
 static inline bool op_reuse_src1(const ggml_tensor * op1, const ggml_tensor * op0) {
-    return (op0 && op0->src[1] == op1->src[1] && ggml_is_quantized(op0->src[0]->type));
+    return (op0 && op0->src[1] == op1->src[1] &&
+            act_quant_family(op0->src[0]->type) == act_quant_family(op1->src[0]->type) &&
+            act_quant_family(op0->src[0]->type) != 0);
 }
 
 static inline bool is_compute_op(ggml_tensor *node)
@@ -3364,6 +3397,8 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
                   "please update hexagon_type to match ggml_type");
     static_assert((unsigned int) HTP_TYPE_MXFP4 == (unsigned int) GGML_TYPE_MXFP4,
                   "please update hexagon_type to match ggml_type");
+    static_assert((unsigned int) HTP_TYPE_IQ4_NL == (unsigned int) GGML_TYPE_IQ4_NL,
+                  "please update hexagon_type to match ggml_type");
 
     const char * str_experimental = getenv("GGML_HEXAGON_EXPERIMENTAL");
     const char * str_verbose = getenv("GGML_HEXAGON_VERBOSE");

ggml/src/ggml-hexagon/htp/flash-attn-ops.c

@@ -346,6 +346,9 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
 
     const HVX_Vector logit_cap = hvx_vec_splat_f32(factx->logit_softcap);
 
+    dma_cache m_cache;
+    dma_cache_init(&m_cache, spad_m, factx->size_m_block, DMA_CACHE_MAX_SIZE);
+
     for (uint32_t ir = ir0; ir < ir1; ++ir) {
         const uint32_t iq3 = fastdiv(ir, &factx->src0_div21);
         const uint32_t iq2 = fastdiv(ir - iq3*neq2*neq1, &factx->src0_div1);
@@ -389,9 +392,8 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             // Mask
             if (mask) {
                 const uint8_t * m_src = (const uint8_t *) (mp_base + ic_start);
-                uint8_t * m_dst = spad_m + (ib % 2) * factx->size_m_block;
                 // Mask is 1D contiguous for this row
-                dma_queue_push(dma, dma_make_ptr(m_dst, m_src), current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
+                dma_cache_push(dma, &m_cache, m_src, current_block_size * 2, current_block_size * 2, current_block_size * 2, 1);
             }
 
             // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
@@ -554,7 +556,7 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
                 // Mask
                 if (mask) {
                     const uint8_t * m_src = (const uint8_t *) (mp_base + next_ic_start);
-                    dma_queue_push(dma, dma_make_ptr(m_base, m_src), next_block_size * 2, next_block_size * 2, next_block_size * 2, 1);
+                    dma_cache_push(dma, &m_cache, m_src, next_block_size * 2, next_block_size * 2, next_block_size * 2, 1);
                 }
 
                 // FARF(HIGH, "fa %u: prefetch KVM: ir %u ib %u : iq1 %u iq2 %u iq3 %u : size_k_row %u size_v_row %u bs %u: usec %u",
@@ -684,7 +686,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->src0_spad.size_per_thread = size_q_block * 1;
     octx->src1_spad.size_per_thread = factx.size_k_block * 2;
     octx->src2_spad.size_per_thread = factx.size_v_block * 2;
-    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * 2 : 0;
+    octx->src3_spad.size_per_thread = mask ? factx.size_m_block * DMA_CACHE_MAX_SIZE : 0;
     octx->dst_spad.size_per_thread  = size_vkq_acc;
 
     octx->src0_spad.size = octx->src0_spad.size_per_thread * octx->n_threads;
@@ -705,6 +707,8 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     octx->src3_spad.data = octx->src2_spad.data + octx->src2_spad.size;
     octx->dst_spad.data  = octx->src3_spad.data + octx->src3_spad.size;
 
+    // FARF(ERROR, "fa: qrows-per-thread %u", factx.qrows_per_thread);
+
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
         worker_pool_run_func(octx->ctx->worker_pool, flash_attn_ext_f16_thread, &factx, octx->n_threads);
     }

ggml/src/ggml-hexagon/htp/hex-dma.h

@@ -143,7 +143,7 @@ static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t
     desc->desc_size  = 0; // 1D mode
     desc->src_bypass = dma_src_l2_bypass_on;
     desc->dst_bypass = dma_dst_l2_bypass_on;
-    desc->order      = 1;
+    desc->order      = 0;
     desc->done       = 0;
     desc->src        = (void *) dptr.src;
     desc->dst        = (void *) dptr.dst;
@@ -151,8 +151,12 @@ static inline bool dma_queue_push_single_1d(dma_queue * q, dma_ptr dptr, size_t
 
     q->dptr[q->push_idx] = dptr;
 
-    dmlink(q->tail, desc);
-    q->tail = (dma_descriptor_2d *) desc;
+    if (size) {
+        dmlink(q->tail, desc);
+        q->tail = (dma_descriptor_2d *) desc;
+    } else {
+        desc->done = 1;
+    }
 
     // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
     q->push_idx = (q->push_idx + 1) & q->idx_mask;
@@ -175,7 +179,7 @@ static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t
     desc->dst_bypass     = dma_dst_l2_bypass_on;
     desc->src_comp       = 0;
     desc->dst_comp       = 0;
-    desc->order          = 1;
+    desc->order          = 0;
     desc->done           = 0;
     desc->src_stride     = src_stride;
     desc->dst_stride     = dst_stride;
@@ -197,8 +201,12 @@ static inline bool dma_queue_push_single_2d(dma_queue * q, dma_ptr dptr, size_t
 
     q->dptr[q->push_idx] = dptr;
 
-    dmlink(q->tail, desc);
-    q->tail = desc;
+    if (nrows) {
+        dmlink(q->tail, desc);
+        q->tail = desc;
+    } else {
+        desc->done = 1;
+    }
 
     // FARF(ERROR, "dma-push: i %u row-size %u nrows %d dst %p src %p\n", q->push_idx, row_size, nrows, dptr.dst, dptr.src);
     q->push_idx = (q->push_idx + 1) & q->idx_mask;
@@ -215,12 +223,9 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
     dma_descriptor_2d * desc = &q->desc[q->pop_idx];
 
     // Wait for desc to complete
-    while (1) {
-        dmpoll();
-        if (desc->done) {
-            break;
-        }
+    while (!desc->done) {
         // FARF(ERROR, "dma-pop: waiting for DMA : %u\n", q->pop_idx);
+        dmpoll();
     }
 
     dptr = q->dptr[q->pop_idx];
@@ -312,6 +317,54 @@ static inline bool dma_queue_push_vtcm_to_ddr(dma_queue * q, dma_ptr dptr, size_
     return dma_queue_push(q, dptr, dst_row_size, src_row_size, dst_row_size, nrows);
 }
 
+#define DMA_CACHE_MAX_SIZE 64U
+
+typedef struct {
+    uint8_t *base;
+    uint32_t line_size;
+    uint32_t capacity;
+    uint32_t src[DMA_CACHE_MAX_SIZE];
+    uint16_t age[DMA_CACHE_MAX_SIZE];
+} dma_cache;
+
+static inline void dma_cache_init(dma_cache *c, uint8_t *base, uint32_t line_size, uint32_t capacity)
+{
+    c->capacity  = (capacity > DMA_CACHE_MAX_SIZE) ? DMA_CACHE_MAX_SIZE : capacity;
+    c->base      = base;
+    c->line_size = line_size;
+
+    for (unsigned i=0; i < c->capacity; i++) {
+        c->src[i] = 0;
+        c->age[i] = 0;
+    }
+}
+
+static inline bool dma_cache_push(dma_queue *q, dma_cache *c, const uint8_t * src, uint32_t dst_stride, uint32_t src_stride, uint32_t row_size, uint32_t nrows)
+{
+    uint32_t o_idx = 0;
+    uint16_t o_age = 0;
+    uint8_t *  dst = 0;
+
+    for (unsigned i=0; i < c->capacity; i++) {
+        if (c->src[i] == (uint32_t) src) {
+            c->age[i] = 0;
+            dst = c->base + (i * c->line_size); nrows = 0; // dummy dma
+            // FARF(ERROR, "dma-cache: found %p", src);
+        } else {
+            c->age[i]++;
+            if (c->age[i] > o_age) { o_age = c->age[i]; o_idx = i; }
+        }
+    }
+    if (!dst) {
+        // FARF(ERROR, "dma-cache: replacing #%u : age %u %p -> %p", o_idx, c->age[o_idx], (void *) c->src[o_idx], src);
+        c->age[o_idx] = 0;
+        c->src[o_idx] = (uint32_t) src;
+        dst = c->base + o_idx * c->line_size; // normal nrows dma
+    }
+
+    return dma_queue_push(q, dma_make_ptr(dst, src), dst_stride, src_stride, row_size, nrows);
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c

@@ -30,6 +30,12 @@ static const __fp16 q4_0_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
     -8, 0, -7, 0, -6, 0, -5, 0, -4, 0, -3, 0, -2, 0, -1, 0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7, 0,
 };
 
+// MXFP4 dequantization LUT: maps 4-bit index to fp16 mantissa value
+// kvalues: 0, 0.5, 1, 1.5, 2, 3, 4, 6, 0, -0.5, -1, -1.5, -2, -3, -4, -6
+static const __fp16 mxfp4_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
+    0, 0, 0.5, 0, 1, 0, 1.5, 0, 2, 0, 3, 0, 4, 0, 6, 0, 0, 0, -0.5, 0, -1, 0, -1.5, 0, -2, 0, -3, 0, -4, 0, -6, 0,
+};
+
 static const __fp16 iq4_nl_to_fp16_lut[64] __attribute__((aligned(VLEN))) = {
     -127, 0, -104, 0, -83, 0, -65, 0, -49, 0, -35, 0, -22, 0, -10, 0,
     1,    0, 13,   0, 25,  0, 38,  0, 53,  0, 69,  0, 89,  0, 113, 0,
@@ -46,7 +52,8 @@ static const int32_t weight_transpose_scatter_offsets[32] __attribute__((aligned
 
 // Scales per x4x2 logical block: 8 × sizeof(__fp16) = 16 bytes
 #define HMX_X4X2_SCALES_PER_BLK  8
-#define HMX_X4X2_DBLK_SIZE       16  // 8 * 2 bytes
+#define HMX_X4X2_DBLK_SIZE       16  // 8 * 2 bytes (fp16 scales for Q4_0/Q8_0/IQ4_NL)
+#define HMX_X4X2_MXFP4_EBLK_SIZE 8   // 8 * 1 byte  (E8M0 scales for MXFP4)
 
 static inline void swap_ptr(void **p1, void **p2) {
     void *t = *p1;
@@ -78,9 +85,11 @@ static inline size_t get_x4x2_row_stride(int weight_type, int k) {
     switch (weight_type) {
         case HTP_TYPE_Q4_0:
         case HTP_TYPE_IQ4_NL:
-            return (size_t)nb * (QK_Q4_0x4x2 / 2 + HMX_X4X2_DBLK_SIZE);  // 144 * nb
+            return (size_t) nb * (QK_Q4_0x4x2 / 2 + HMX_X4X2_DBLK_SIZE);         // 144 * nb
         case HTP_TYPE_Q8_0:
-            return (size_t)nb * (QK_Q8_0x4x2 + HMX_X4X2_DBLK_SIZE);      // 272 * nb
+            return (size_t) nb * (QK_Q8_0x4x2 + HMX_X4X2_DBLK_SIZE);             // 272 * nb
+        case HTP_TYPE_MXFP4:
+            return (size_t) nb * (QK_MXFP4x4x2 / 2 + HMX_X4X2_MXFP4_EBLK_SIZE);  // 136 * nb
         default:
             return 0;
     }
@@ -284,6 +293,87 @@ static inline HVX_Vector dequantize_x4x2_q8_0_group_hvx(
     return Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hf, v_scales));
 }
 
+// --- MXFP4 E8M0 scale conversion and dequantization ---
+//
+// HVX batch-convert 8 E8M0 bytes (one x4x2 block's scales) to __fp16[8] on stack.
+// Scalar loads from the stack array execute on the scalar pipeline, in parallel
+// with HVX vlut16/vmpy/vscatter — freeing HVX slots in the hot loop.
+// Arithmetic: fp16_bits = clamp(e - 112, 0, 30) << 10
+// e=0..112 -> 0 (underflow), e=113..142 -> valid fp16, e>=143 -> clamped to 2^15.
+
+typedef struct {
+    __fp16 v[8] __attribute__((aligned(16)));
+} mxfp4_scales_t;
+
+static inline mxfp4_scales_t mxfp4_convert_scales(const uint8_t * e8m0_8) {
+    mxfp4_scales_t s;
+    HVX_Vector     v  = hvx_vmemu(e8m0_8);
+    HVX_Vector     vh = Q6_V_lo_W(Q6_Wuh_vunpack_Vub(v));
+    vh                = Q6_Vh_vsub_VhVh(vh, Q6_Vh_vsplat_R(112));
+    vh                = Q6_Vh_vmax_VhVh(vh, Q6_V_vzero());
+    vh                = Q6_Vh_vmin_VhVh(vh, Q6_Vh_vsplat_R(30));
+    vh                = Q6_Vh_vasl_VhR(vh, 10);
+    hvx_vec_store_u(s.v, 16, vh);
+    return s;
+}
+
+static inline HVX_Vector mxfp4_extract_splat(mxfp4_scales_t scales, int idx) {
+    return hvx_vec_splat_f16(scales.v[idx]);
+}
+
+// Dequantize one x4x2 MXFP4 group (32 elements from 32 packed bytes) -> 32 FP16.
+static inline HVX_Vector dequantize_x4x2_mxfp4_group_hvx(const uint8_t *  packed_32,
+                                                         bool             upper_nibbles,
+                                                         int              sub_blk,
+                                                         const HVX_Vector vlut_cvt,
+                                                         mxfp4_scales_t   scales) {
+    HVX_Vector       vq       = hvx_vmemu(packed_32);
+    const HVX_Vector mask_h4  = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector       v_quants = upper_nibbles ? Q6_Vub_vlsr_VubR(vq, 4) : vq;
+    v_quants                  = Q6_V_vand_VV(v_quants, mask_h4);
+
+    HVX_Vector v_sc = mxfp4_extract_splat(scales, sub_blk);
+
+    v_quants            = Q6_Vb_vshuff_Vb(v_quants);
+    HVX_VectorPair vp   = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
+    HVX_Vector     v_hf = Q6_V_lo_W(vp);
+
+    return Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hf, v_sc));
+}
+
+// Batch-dequantize 4 contiguous x4x2 MXFP4 groups (4x32 = 128 packed bytes).
+static inline void dequantize_x4x2_mxfp4_x4groups_hvx(const uint8_t *  packed_128,
+                                                      bool             upper_nibbles,
+                                                      int              sub_blk_base,
+                                                      const HVX_Vector vlut_cvt,
+                                                      mxfp4_scales_t   scales,
+                                                      HVX_Vector       out[4]) {
+    HVX_Vector       vq       = hvx_vmemu(packed_128);
+    const HVX_Vector mask_h4  = Q6_Vb_vsplat_R(0x0F);
+    HVX_Vector       v_quants = upper_nibbles ? Q6_Vub_vlsr_VubR(vq, 4) : vq;
+    v_quants                  = Q6_V_vand_VV(v_quants, mask_h4);
+
+    v_quants = Q6_Vb_vshuff_Vb(v_quants);
+
+    HVX_VectorPair vp   = Q6_Wh_vlut16_VbVhR(v_quants, vlut_cvt, 0);
+    HVX_Vector     v_lo = Q6_V_lo_W(vp);
+    HVX_Vector     v_hi = Q6_V_hi_W(vp);
+
+    HVX_VectorPred q64    = Q6_Q_vsetq_R(64);
+    HVX_Vector     v_sc01 = Q6_V_vmux_QVV(q64, mxfp4_extract_splat(scales, sub_blk_base + 0),
+                                          mxfp4_extract_splat(scales, sub_blk_base + 1));
+    HVX_Vector     v_sc23 = Q6_V_vmux_QVV(q64, mxfp4_extract_splat(scales, sub_blk_base + 2),
+                                          mxfp4_extract_splat(scales, sub_blk_base + 3));
+
+    v_lo = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_lo, v_sc01));
+    v_hi = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(v_hi, v_sc23));
+
+    out[0] = v_lo;
+    out[1] = Q6_V_vror_VR(v_lo, 64);
+    out[2] = v_hi;
+    out[3] = Q6_V_vror_VR(v_hi, 64);
+}
+
 // Dequantize a tile range from x4x2 weight data (already in VTCM) to tile-major FP16.
 // Input:  vtcm_src has n_cols rows of x4x2 data, each row_stride bytes.
 // Output: vtcm_dst in tile-major FP16 layout.
@@ -295,11 +385,11 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
         int start_tile, int end_tile) {
 
     const int n_k_tiles = k_block / HMX_FP16_TILE_N_COLS;
-    const bool is_q4 = (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL);
-    const int qrow_size = is_q4 ? (k_block / 2) : k_block;
+    const int qrow_size = (weight_type == HTP_TYPE_Q8_0) ? k_block : (k_block / 2);
 
-    const HVX_Vector vlut_cvt = (weight_type == HTP_TYPE_IQ4_NL)
-        ? hvx_vmem(iq4_nl_to_fp16_lut) : hvx_vmem(q4_0_to_fp16_lut);
+    const HVX_Vector vlut_cvt = (weight_type == HTP_TYPE_IQ4_NL) ? hvx_vmem(iq4_nl_to_fp16_lut) :
+                                (weight_type == HTP_TYPE_MXFP4)  ? hvx_vmem(mxfp4_to_fp16_lut) :
+                                                                   hvx_vmem(q4_0_to_fp16_lut);
 
     // vscatter setup: write dequantized K-values directly to transposed [K][N] tile positions.
     // Each int32 element holds a K-row-pair (2 adjacent fp16 values).  word[i] at offset i*128
@@ -312,8 +402,9 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
         int ct = t / n_k_tiles;  // column tile index
         int kt = t % n_k_tiles;  // K tile index
 
-        // --- Batch-4 fast path for Q4: process 4 contiguous K-tiles with one vlut16 per row ---
-        if (is_q4 && (kt % 4 == 0) && (t + 4 <= end_tile) && ((t + 3) / n_k_tiles == ct)) {
+        // --- Batch-4 fast path for Q4_0/IQ4_NL: process 4 contiguous K-tiles with one vlut16 per row ---
+        if ((weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL) && (kt % 4 == 0) && (t + 4 <= end_tile) &&
+            ((t + 3) / n_k_tiles == ct)) {
             int blk_idx      = (kt * 32) / QK_Q4_0x4x2;
             int sub_blk_base = ((kt * 32) % QK_Q4_0x4x2) / 32;  // 0 or 4
             bool upper       = (sub_blk_base >= 4);
@@ -351,10 +442,60 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
             continue;
         }
 
+        // --- Batch-4 fast path for MXFP4: same nibble layout but E8M0 scales ---
+        if (weight_type == HTP_TYPE_MXFP4 && (kt % 4 == 0) && (t + 4 <= end_tile) && ((t + 3) / n_k_tiles == ct)) {
+            int  blk_idx      = (kt * 32) / QK_MXFP4x4x2;
+            int  sub_blk_base = ((kt * 32) % QK_MXFP4x4x2) / 32;                 // 0 or 4
+            bool upper        = (sub_blk_base >= 4);
+            int  packed_off   = blk_idx * (QK_MXFP4x4x2 / 2);                    // 128 contiguous packed bytes
+            int  e8m0_blk_off = qrow_size + blk_idx * HMX_X4X2_MXFP4_EBLK_SIZE;  // all 8 E8M0 scales
+
+            __fp16 * tile_bases[4];
+            for (int g = 0; g < 4; g++) {
+                tile_bases[g] = vtcm_dst + (t + g) * HMX_FP16_TILE_N_ELMS;
+            }
+
+            HVX_Vector v_off = v_scat_base;
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
+                int             row0 = ct * HMX_FP16_TILE_N_COLS + r;
+                int             row1 = row0 + 1;
+                const uint8_t * r0   = vtcm_src + row0 * row_stride;
+                const uint8_t * r1   = vtcm_src + row1 * row_stride;
+
+                // Batch-convert all 8 E8M0 scales once per row (stays in HVX register)
+                mxfp4_scales_t r0_e8 = mxfp4_convert_scales(r0 + e8m0_blk_off);
+
+                HVX_Vector v0[4], v1[4];
+                dequantize_x4x2_mxfp4_x4groups_hvx(r0 + packed_off, upper, sub_blk_base, vlut_cvt, r0_e8, v0);
+                if (row1 < n_cols) {
+                    mxfp4_scales_t r1_e8 = mxfp4_convert_scales(r1 + e8m0_blk_off);
+                    dequantize_x4x2_mxfp4_x4groups_hvx(r1 + packed_off, upper, sub_blk_base, vlut_cvt, r1_e8, v1);
+                } else {
+                    v1[0] = v1[1] = v1[2] = v1[3] = Q6_V_vzero();
+                }
+
+                for (int g = 0; g < 4; g++) {
+                    Q6_vscatter_QRMVwV(q_mask64, (size_t) tile_bases[g], HMX_FP16_TILE_SIZE - 1, v_off, v0[g]);
+                }
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
+                for (int g = 0; g < 4; g++) {
+                    Q6_vscatter_QRMVwV(q_mask64, (size_t) tile_bases[g], HMX_FP16_TILE_SIZE - 1, v_off, v1[g]);
+                }
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
+            }
+
+            for (int g = 0; g < 4; g++) {
+                (void) *(volatile HVX_Vector *) (tile_bases[g]);
+            }
+
+            t += 4;
+            continue;
+        }
+
         // --- Single-tile fallback ---
         __fp16 *tile_base = vtcm_dst + t * HMX_FP16_TILE_N_ELMS;
 
-        if (is_q4) {
+        if (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL) {
             int blk_idx  = (kt * 32) / QK_Q4_0x4x2;
             int sub_blk  = ((kt * 32) % QK_Q4_0x4x2) / 32;
             bool upper   = (sub_blk >= 4);
@@ -382,6 +523,39 @@ static void dequantize_x4x2_weight_to_fp16_tiles_task(
                 v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
             }
             (void) *(volatile HVX_Vector *)(tile_base);
+        } else if (weight_type == HTP_TYPE_MXFP4) {
+            int  blk_idx      = (kt * 32) / QK_MXFP4x4x2;
+            int  sub_blk      = ((kt * 32) % QK_MXFP4x4x2) / 32;
+            bool upper        = (sub_blk >= 4);
+            int  byte_off     = blk_idx * (QK_MXFP4x4x2 / 2) + (upper ? (sub_blk - 4) : sub_blk) * 32;
+            int  e8m0_blk_off = qrow_size + blk_idx * HMX_X4X2_MXFP4_EBLK_SIZE;
+
+            HVX_Vector v_off = v_scat_base;
+            for (int r = 0; r < HMX_FP16_TILE_N_ROWS; r += 2) {
+                int row0 = ct * HMX_FP16_TILE_N_COLS + r;
+                int row1 = row0 + 1;
+
+                const uint8_t * r0 = vtcm_src + row0 * row_stride;
+                const uint8_t * r1 = vtcm_src + row1 * row_stride;
+
+                // Batch-convert all 8 E8M0 scales once per row (stays in HVX register)
+                mxfp4_scales_t r0_e8 = mxfp4_convert_scales(r0 + e8m0_blk_off);
+
+                HVX_Vector v0 = dequantize_x4x2_mxfp4_group_hvx(r0 + byte_off, upper, sub_blk, vlut_cvt, r0_e8);
+                HVX_Vector v1;
+                if (row1 < n_cols) {
+                    mxfp4_scales_t r1_e8 = mxfp4_convert_scales(r1 + e8m0_blk_off);
+                    v1 = dequantize_x4x2_mxfp4_group_hvx(r1 + byte_off, upper, sub_blk, vlut_cvt, r1_e8);
+                } else {
+                    v1 = Q6_V_vzero();
+                }
+
+                Q6_vscatter_QRMVwV(q_mask64, (size_t) tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v0);
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
+                Q6_vscatter_QRMVwV(q_mask64, (size_t) tile_base, HMX_FP16_TILE_SIZE - 1, v_off, v1);
+                v_off = Q6_Vw_vadd_VwVw(v_off, v_scat_step);
+            }
+            (void) *(volatile HVX_Vector *) (tile_base);
         } else {
             // Q8_0
             int blk_idx  = (kt * 32) / QK_Q8_0x4x2;
@@ -1455,21 +1629,24 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                 {
                     qweight_fetch_task_state_t s;
 
-                    const bool is_q4 = (weight_type == HTP_TYPE_Q4_0 || weight_type == HTP_TYPE_IQ4_NL);
                     const int blk_start = kk / QK_Q4_0x4x2;
                     const int nb_sub = (k_blk_sz + QK_Q4_0x4x2 - 1) / QK_Q4_0x4x2;
-                    const int full_qrow = is_q4 ? (k / 2) : k;
+                    const int    full_qrow      = (weight_type == HTP_TYPE_Q8_0) ? k : (k / 2);
                     const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
+                    const int    scale_blk_size =
+                        (weight_type == HTP_TYPE_MXFP4) ? HMX_X4X2_MXFP4_EBLK_SIZE : HMX_X4X2_DBLK_SIZE;
 
                     s.dst         = vtcm_scratch0;
                     s.src         = w + nc * row_stride;
                     s.n_rows      = n_blk_sz;
                     s.src_stride  = row_stride;
                     s.dst_stride  = sub_row_stride;
-                    s.quant_off   = is_q4 ? (blk_start * (QK_Q4_0x4x2 / 2)) : (blk_start * QK_Q8_0x4x2);
-                    s.quant_width = is_q4 ? (nb_sub    * (QK_Q4_0x4x2 / 2)) : (nb_sub * QK_Q8_0x4x2);
-                    s.scale_off   = full_qrow + blk_start * HMX_X4X2_DBLK_SIZE;
-                    s.scale_width = nb_sub * HMX_X4X2_DBLK_SIZE;
+                    s.quant_off =
+                        (weight_type == HTP_TYPE_Q8_0) ? (blk_start * QK_Q8_0x4x2) : (blk_start * (QK_Q4_0x4x2 / 2));
+                    s.quant_width =
+                        (weight_type == HTP_TYPE_Q8_0) ? (nb_sub * QK_Q8_0x4x2) : (nb_sub * (QK_Q4_0x4x2 / 2));
+                    s.scale_off   = full_qrow + blk_start * scale_blk_size;
+                    s.scale_width = nb_sub * scale_blk_size;
 
                     // 2D DMA: quants sub-range
                     dma_queue_push(ctx->dma[0], dma_make_ptr(s.dst, s.src + s.quant_off),

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -31,6 +31,12 @@ struct htp_context {
 
     uint32_t opmask;
 
+    // Cached src1 spad position from the last quantize pass.
+    // When SKIP_QUANTIZE is set the Q8 activation data is already in VTCM
+    // at this address; the matmul must read from here instead of recomputing
+    // the offset (which depends on the current op's src0 size).
+    uint8_t * prev_src1_spad;
+
     // HMX acceleration fields (v73+, enabled by compile-time HTP_HAS_HMX)
 #ifdef HTP_HAS_HMX
     int        hmx_enabled;       // Runtime flag: HMX initialisation succeeded

ggml/src/ggml-hexagon/htp/hvx-div.h

@@ -16,8 +16,10 @@
 
 #if __HVX_ARCH__ < 79
 #define HVX_OP_MUL_F32(a, b) Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b))
+#define HVX_OP_MUL_F16(a, b) Q6_Vhf_equals_Wqf32(Q6_Wqf32_vmpy_VhfVhf(a, b))
 #else
 #define HVX_OP_MUL_F32(a, b) Q6_Vsf_vmpy_VsfVsf(a, b)
+#define HVX_OP_MUL_F16(a, b) Q6_Vhf_vmpy_VhfVhf(a, b)
 #endif
 
 // Compute div by scaler in f32. Requires first by expanding fp32 to fp16 and converting the result back to fp32.
@@ -43,46 +45,67 @@ static inline HVX_Vector hvx_div_mul_f16_const_using_f32(HVX_Vector vec1_hf, HVX
     return res;
 }
 
-#define hvx_div_scaler_f16_loop_body(dst_type, src_type, vec_store)                     \
-    do {                                                                                \
-        dst_type * restrict vdst = (dst_type *) dst;                                    \
-        src_type * restrict vsrc = (src_type *) src;                                    \
-        HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                     \
-                                                                                        \
-        const uint32_t nvec = n / VLEN_FP16;                                            \
-        const uint32_t nloe = n % VLEN_FP16;                                            \
-                                                                                        \
-        uint32_t i = 0;                                                                 \
-                                                                                        \
-        _Pragma("unroll(4)")                                                            \
-        for (; i < nvec; i++) {                                                         \
-            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
-            vdst[i] = res;                                                              \
-        }                                                                               \
-        if (nloe) {                                                                     \
-            HVX_Vector res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one); \
-            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                      \
-        }                                                                               \
+// Variant for <v79: Use pre-computed f16 reciprocal constant
+static inline HVX_Vector hvx_div_mul_f16_const_using_f16(HVX_Vector vec1_hf, HVX_Vector const_inv_hf) {
+    // Multiply by pre-computed f16 reciprocal constant
+    return HVX_OP_MUL_F16(vec1_hf, const_inv_hf);
+}
+
+#define hvx_div_scaler_f16_loop_body(dst_type, src_type, vec_store)                                    \
+    do {                                                                                               \
+        dst_type * restrict vdst = (dst_type *) dst;                                                   \
+        src_type * restrict vsrc = (src_type *) src;                                                   \
+                                                                                                       \
+        HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                                    \
+                                                                                                       \
+        const uint32_t nvec = n / VLEN_FP16;                                                           \
+        const uint32_t nloe = n % VLEN_FP16;                                                           \
+                                                                                                       \
+        uint32_t i = 0;                                                                                \
+                                                                                                       \
+        _Pragma("unroll(4)")                                                                           \
+        for (; i < nvec; i++) {                                                                        \
+            HVX_Vector res;                                                                            \
+            if (__HVX_ARCH__ < 79) {                                                                   \
+                res = hvx_div_mul_f16_const_using_f16(vsrc[i], val_vec_f16);                           \
+            } else {                                                                                   \
+                res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one);                   \
+            }                                                                                          \
+            vdst[i] = res;                                                                             \
+        }                                                                                              \
+        if (nloe) {                                                                                    \
+            HVX_Vector res;                                                                            \
+            if (__HVX_ARCH__ < 79) {                                                                   \
+                res = hvx_div_mul_f16_const_using_f16(vsrc[i], val_vec_f16);                           \
+            } else {                                                                                   \
+                res = hvx_div_mul_f16_const_using_f32(vsrc[i], val_vec_f32, hf_one);                   \
+            }                                                                                          \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                                     \
+        }                                                                                              \
     } while(0)
 
 static inline void hvx_div_scalar_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
     const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    const HVX_Vector val_vec_f16 = hvx_vec_splat_f16(1.0f / val);
     assert((uintptr_t) dst % 128 == 0);
     assert((uintptr_t) src % 128 == 0);
     hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
 }
 static inline void hvx_div_scalar_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
     const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    const HVX_Vector val_vec_f16 = hvx_vec_splat_f16(1.0f / val);
     assert((uintptr_t) dst % 128 == 0);
     hvx_div_scaler_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
 }
 static inline void hvx_div_scalar_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
     const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    const HVX_Vector val_vec_f16 = hvx_vec_splat_f16(1.0f / val);
     assert((uintptr_t) src % 128 == 0);
     hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
 }
 static inline void hvx_div_scalar_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 val, uint32_t n) {
     const HVX_Vector val_vec_f32 = hvx_vec_splat_f32(1.0f/((float)val));
+    const HVX_Vector val_vec_f16 = hvx_vec_splat_f16(1.0f / val);
     hvx_div_scaler_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
 }
 
@@ -128,13 +151,25 @@ static inline HVX_Vector hvx_vec_div_f16_using_f32(HVX_Vector vec1, HVX_Vector v
     return recip;
 }
 
+// Hybrid approach: f16 reciprocal for <v79, f32 precision for >=v79
+static inline HVX_Vector hvx_vec_hybrid_div_f16(HVX_Vector vec1, HVX_Vector vec2, HVX_Vector f32_nan_inf_mask, HVX_Vector f16_nan_inf_mask, HVX_Vector vec_hf_one_1_0) {
+#if __HVX_ARCH__ < 79
+    // For older architectures, use f16 reciprocal to avoid NaN/-inf issues
+    HVX_Vector vec2_inv = hvx_vec_inverse_f16_guard(vec2, f16_nan_inf_mask);
+    return HVX_OP_MUL_F16(vec1, vec2_inv);
+#else
+    return hvx_vec_div_f16_using_f32(vec1, vec2, f32_nan_inf_mask, vec_hf_one_1_0);
+#endif
+}
+
 #define hvx_div_f16_loop_body(dst_type, src0_type, src1_type, vec_store)                  \
     do {                                                                                  \
         dst_type * restrict vdst = (dst_type *) dst;                                      \
         src0_type * restrict vsrc0 = (src0_type *) src0;                                  \
         src1_type * restrict vsrc1 = (src1_type *) src1;                                  \
                                                                                           \
-        const HVX_Vector nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                        \
+        const HVX_Vector f32_nan_inf_mask = Q6_V_vsplat_R(0x7f800000);                    \
+        const HVX_Vector f16_nan_inf_mask = Q6_Vh_vsplat_R(0x7c00);                       \
         const HVX_Vector hf_one = Q6_Vh_vsplat_R(0x3C00);                                 \
                                                                                           \
         const uint32_t nvec = n / VLEN_FP16;                                              \
@@ -144,11 +179,15 @@ static inline HVX_Vector hvx_vec_div_f16_using_f32(HVX_Vector vec1, HVX_Vector v
                                                                                           \
         _Pragma("unroll(4)")                                                              \
         for (; i < nvec; i++) {                                                           \
-            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            HVX_Vector res = hvx_vec_hybrid_div_f16(vsrc0[i], vsrc1[i],                   \
+                                                    f32_nan_inf_mask, f16_nan_inf_mask,   \
+                                                    hf_one);                              \
             vdst[i] = res;                                                                \
         }                                                                                 \
         if (nloe) {                                                                       \
-            HVX_Vector res = hvx_vec_div_f16_using_f32(vsrc0[i], vsrc1[i], nan_inf_mask, hf_one); \
+            HVX_Vector res = hvx_vec_hybrid_div_f16(vsrc0[i], vsrc1[i],                   \
+                                                    f32_nan_inf_mask, f16_nan_inf_mask,   \
+                                                    hf_one);                              \
             vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, res);                        \
         }                                                                                 \
     } while(0)
@@ -247,5 +286,6 @@ HVX_DIV_DISPATCHER(hvx_div_f32)
 HVX_DIV_DISPATCHER(hvx_div_f16)
 
 #undef HVX_OP_MUL_F32
+#undef HVX_OP_MUL_F16
 
 #endif // HVX_DIV_H

ggml/src/ggml-hexagon/htp/main.c

@@ -1114,14 +1114,12 @@ static void proc_hmx_matmul_req(struct htp_context *     ctx,
         return;
     }
 
-    // HMX only supports F16, Q4_0, Q8_0, IQ4_NL weights.
-    // Other types (e.g. MXFP4) fall back to HVX.
+    // HMX supports F16, Q4_0, Q8_0, IQ4_NL, MXFP4 weights.
+    // Other types fall back to HVX.
     {
         uint32_t wtype = req->src0.type;
-        if (wtype != HTP_TYPE_F16  &&
-            wtype != HTP_TYPE_Q4_0 &&
-            wtype != HTP_TYPE_Q8_0 &&
-            wtype != HTP_TYPE_IQ4_NL) {
+        if (wtype != HTP_TYPE_F16 && wtype != HTP_TYPE_Q4_0 && wtype != HTP_TYPE_Q8_0 && wtype != HTP_TYPE_IQ4_NL &&
+            wtype != HTP_TYPE_MXFP4) {
             proc_matmul_req(ctx, req, bufs, n_bufs);
             return;
         }

ggml/src/ggml-hexagon/htp/matmul-ops.c

@@ -60,6 +60,16 @@ static const uint8_t __attribute__((aligned(128))) expand_x32_e8m0[128] = {
     0x00, 0x00, 0x09, 0x08, 0x00, 0x00, 0x22, 0x20, 0x24, 0x20, 0x21, 0x22, 0x20, 0x20,
 };
 
+// IQ4_NL dequantization LUT: maps 4-bit index (0-15) to int8 kvalue
+// kvalues: -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113
+static const uint8_t __attribute__((aligned(VLEN))) kvalues_iq4nl_lut[] = {
+    0x81, 0, 0x98, 0, 0xAD, 0, 0xBF, 0, 0xCF, 0, 0xDD, 0, 0xEA, 0, 0xF6, 0, 0x01, 0, 0x0D, 0, 0x19, 0, 0x26, 0,
+    0x35, 0, 0x45, 0, 0x59, 0, 0x71, 0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0,
+    0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0,
+    0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0,
+    0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0, 0,    0,
+};
+
 static const uint8_t __attribute__((aligned(VLEN))) kvalues_mxfp4_lut[] = {
     0,    0, 1,    0, 2,    0, 3, 0, 4, 0, 6, 0, 8, 0, 12, 0, 0, 0, 0xff, 0, 0xfe, 0, 0xfd, 0, 0xfc, 0,
     0xfa, 0, 0xf8, 0, 0xf4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  0, 0, 0, 0,    0, 0,    0, 0,    0, 0,    0,
@@ -68,6 +78,73 @@ static const uint8_t __attribute__((aligned(VLEN))) kvalues_mxfp4_lut[] = {
     0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  0, 0, 0, 0,    0, 0,    0, 0,    0,
 };
 
+static inline HVX_Vector_x8 hvx_vec_load_iq4nlx4x8_full(const uint8_t * restrict ptr) {
+    const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
+
+    HVX_Vector v0_1 = vptr[0];  // first 256 elements (128 bytes)
+    HVX_Vector v2_3 = vptr[1];  // ...
+    HVX_Vector v4_5 = vptr[2];  // ...
+    HVX_Vector v6_7 = vptr[3];  // ...
+
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    const HVX_Vector lut     = *(const HVX_Vector *) kvalues_iq4nl_lut;
+
+    HVX_Vector v0 = Q6_V_vand_VV(v0_1, mask_h4);  // & 0x0F
+    HVX_Vector v1 = Q6_Vub_vlsr_VubR(v0_1, 4);    // >> 4
+    HVX_Vector v2 = Q6_V_vand_VV(v2_3, mask_h4);  // & 0x0F
+    HVX_Vector v3 = Q6_Vub_vlsr_VubR(v2_3, 4);    // >> 4
+    HVX_Vector v4 = Q6_V_vand_VV(v4_5, mask_h4);  // & 0x0F
+    HVX_Vector v5 = Q6_Vub_vlsr_VubR(v4_5, 4);    // >> 4
+    HVX_Vector v6 = Q6_V_vand_VV(v6_7, mask_h4);  // & 0x0F
+    HVX_Vector v7 = Q6_Vub_vlsr_VubR(v6_7, 4);    // >> 4
+
+    v0 = Q6_Vb_vlut32_VbVbI(v0, lut, 0);
+    v1 = Q6_Vb_vlut32_VbVbI(v1, lut, 0);
+    v2 = Q6_Vb_vlut32_VbVbI(v2, lut, 0);
+    v3 = Q6_Vb_vlut32_VbVbI(v3, lut, 0);
+    v4 = Q6_Vb_vlut32_VbVbI(v4, lut, 0);
+    v5 = Q6_Vb_vlut32_VbVbI(v5, lut, 0);
+    v6 = Q6_Vb_vlut32_VbVbI(v6, lut, 0);
+    v7 = Q6_Vb_vlut32_VbVbI(v7, lut, 0);
+
+    HVX_Vector_x8 r = { v0, v1, v2, v3, v4, v5, v6, v7 };
+    return r;
+}
+
+static inline HVX_Vector_x8 hvx_vec_load_iq4nlx4x8_partial(const uint8_t * restrict ptr, uint32_t n) {
+    const HVX_Vector * restrict vptr = (const HVX_Vector *) ptr;
+
+    const uint32_t qk   = QK_Q4_0x4x2;  // 256
+    const uint32_t nb   = n / qk;
+    const uint32_t nloe = n % qk;
+
+    const HVX_Vector mask_h4 = Q6_Vb_vsplat_R(0x0F);
+    const HVX_Vector lut     = *(const HVX_Vector *) kvalues_iq4nl_lut;
+
+    HVX_Vector_x8 r;
+    uint32_t      i = 0;
+
+    #pragma unroll(2)
+    for (i = 0; i < nb; i++) {
+        HVX_Vector v   = vptr[i];                   // 256 elements (128 bytes)
+        HVX_Vector v0  = Q6_V_vand_VV(v, mask_h4);  // & 0x0F : first  128 elements
+        HVX_Vector v1  = Q6_Vub_vlsr_VubR(v, 4);    // >> 4   : second 128 elements
+        r.v[i * 2 + 0] = Q6_Vb_vlut32_VbVbI(v0, lut, 0);
+        r.v[i * 2 + 1] = Q6_Vb_vlut32_VbVbI(v1, lut, 0);
+    }
+
+    if (nloe) {
+        HVX_Vector     v      = vptr[i];                      // 256 elements (128 bytes)
+        HVX_Vector     v0     = Q6_V_vand_VV(v, mask_h4);     // & 0x0F : even 128 elements
+        HVX_Vector     v1     = Q6_Vub_vlsr_VubR(v, 4);       // >> 4   : odd  128 elements
+        HVX_VectorPair v0_1_p = Q6_W_vshuff_VVR(v1, v0, -1);  // zip even:odd:...
+        r.v[i * 2 + 0]        = Q6_Vb_vlut32_VbVbI(Q6_V_lo_W(v0_1_p), lut, 0);
+        r.v[i * 2 + 1]        = Q6_Vb_vlut32_VbVbI(Q6_V_hi_W(v0_1_p), lut, 0);
+    }
+
+    return r;
+}
+
 // q4x4x2 and q8x4x2 are the flat q4/8_0 formats where all quants are stored first followed by all scales
 
 static inline size_t q8x4x2_row_size(uint32_t ne) {
@@ -921,6 +998,293 @@ static void vec_dot_q8x4x2_q8x4x2_2x2(const int n, float * restrict s0, float *
     hvx_vec_store_u(&s1[0], 8, r0_r1_c1_sum);  // row0,col1 row1,col1
 }
 
+// ======== IQ4_NL x Q8_0 vec_dot kernels ========
+// Same structure as Q4_0 vec_dot but uses IQ4_NL LUT-based load (4-bit index -> int8 kvalue).
+// Scale format is identical to Q4_0 (fp16 scales).
+
+static void vec_dot_iq4nlx4x2_q8x4x2_1x1(const int n,
+                                         float * restrict s0,
+                                         const void * restrict vx0,
+                                         const void * restrict vy0) {
+    assert(n % 32 == 0);
+    assert((unsigned long) vx0 % 128 == 0);
+    assert((unsigned long) vy0 % 128 == 0);
+
+    const uint32_t qk = QK_Q4_0x4x2 * 4;
+
+    const uint32_t x_dblk_size = 8 * 4 * 2;                                   // 32x __fp16
+    const uint32_t x_qblk_size = qk / 2;                                      // int4
+    const uint32_t x_qrow_size = n / 2;                                       // int4 (not padded)
+
+    const uint32_t y_dblk_size = 8 * 4 * 2;                                   // 32x __fp16
+    const uint32_t y_qblk_size = qk;                                          // int8
+    const uint32_t y_qrow_size = n;                                           // int8 (not padded)
+
+    const uint8_t * restrict r0_x_q = ((const uint8_t *) vx0 + 0);            // quants first
+    const uint8_t * restrict r0_x_d = ((const uint8_t *) vx0 + x_qrow_size);  // then scales
+
+    const uint8_t * restrict y_q = ((const uint8_t *) vy0 + 0);               // quants first
+    const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size);     // then scales
+
+    HVX_Vector r0_sum = Q6_V_vzero();
+
+    const uint32_t nb   = n / qk;
+    const uint32_t nloe = n % qk;
+
+    uint32_t i = 0;
+    for (; i < nb; i++) {
+        HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
+        HVX_Vector_x8 r0_q = hvx_vec_load_iq4nlx4x8_full(r0_x_q + i * x_qblk_size);
+
+        HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
+
+        HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
+        HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
+
+        HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
+
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+    }
+
+    if (nloe) {
+        HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
+        HVX_Vector_x8 r0_q = hvx_vec_load_iq4nlx4x8_partial(r0_x_q + i * x_qblk_size, nloe);
+
+        HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
+
+        HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
+        HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
+
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
+        r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+
+        HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
+
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+    }
+
+    r0_sum = hvx_vec_reduce_sum_f32(r0_sum);
+
+    hvx_vec_store_u(s0, 4, r0_sum);
+}
+
+static void vec_dot_iq4nlx4x2_q8x4x2_2x1(const int n,
+                                         float * restrict s0,
+                                         const void * restrict vx0,
+                                         const void * restrict vx1,
+                                         const void * restrict vy0) {
+    assert(n % 32 == 0);
+    assert((unsigned long) vx0 % 128 == 0);
+    assert((unsigned long) vx1 % 128 == 0);
+    assert((unsigned long) vy0 % 128 == 0);
+
+    const uint32_t qk = QK_Q4_0x4x2 * 4;
+
+    const uint32_t x_dblk_size = 8 * 4 * 2;                                   // 32x __fp16
+    const uint32_t x_qblk_size = qk / 2;                                      // int4
+    const uint32_t x_qrow_size = n / 2;                                       // int4 (not padded)
+
+    const uint32_t y_dblk_size = 8 * 4 * 2;                                   // 32x __fp16
+    const uint32_t y_qblk_size = qk;                                          // int8
+    const uint32_t y_qrow_size = n;                                           // int8 (not padded)
+
+    const uint8_t * restrict r0_x_q = ((const uint8_t *) vx0) + 0;            // quants first
+    const uint8_t * restrict r0_x_d = ((const uint8_t *) vx0) + x_qrow_size;  // then scales
+    const uint8_t * restrict r1_x_q = ((const uint8_t *) vx1) + 0;            // quants first
+    const uint8_t * restrict r1_x_d = ((const uint8_t *) vx1) + x_qrow_size;  // then scales
+
+    const uint8_t * restrict y_q = ((const uint8_t *) vy0 + 0);               // quants first
+    const uint8_t * restrict y_d = ((const uint8_t *) vy0 + y_qrow_size);     // then scales
+
+    HVX_Vector r0_sum = Q6_V_vzero();
+    HVX_Vector r1_sum = Q6_V_vzero();
+
+    const uint32_t nb   = n / qk;
+    const uint32_t nloe = n % qk;
+
+    uint32_t i = 0;
+    for (; i < nb; i++) {
+        HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_full(y_q + i * y_qblk_size);
+        HVX_Vector_x8 r0_q = hvx_vec_load_iq4nlx4x8_full(r0_x_q + i * x_qblk_size);
+        HVX_Vector_x8 r1_q = hvx_vec_load_iq4nlx4x8_full(r1_x_q + i * x_qblk_size);
+
+        HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy_q));
+        HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy_q));
+
+        HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
+        HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+        HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
+        HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy_d)));
+
+        HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
+        HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
+
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
+    }
+
+    if (nloe) {
+        HVX_Vector_x8 vy_q = hvx_vec_load_q8x4x8_partial(y_q + i * y_qblk_size, nloe);
+        HVX_Vector_x8 r0_q = hvx_vec_load_iq4nlx4x8_partial(r0_x_q + i * x_qblk_size, nloe);
+        HVX_Vector_x8 r1_q = hvx_vec_load_iq4nlx4x8_partial(r1_x_q + i * x_qblk_size, nloe);
+
+        HVX_Vector r0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy_q, nloe));
+        HVX_Vector r1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy_q, nloe));
+
+        HVX_Vector vy_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y_d + i * y_dblk_size));
+        HVX_Vector r0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+        HVX_Vector r1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy_d)));
+        HVX_Vector r1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy_d)));
+
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
+        r0_dd                = Q6_V_vand_QV(bmask, r0_dd);
+        r1_dd                = Q6_V_vand_QV(bmask, r1_dd);
+        r0_ia                = Q6_V_vand_QV(bmask, r0_ia);
+        r1_ia                = Q6_V_vand_QV(bmask, r1_ia);
+
+        HVX_Vector r0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_ia, r0_dd);
+        HVX_Vector r1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_ia, r1_dd);
+
+        r0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_fa, r0_sum));
+        r1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_fa, r1_sum));
+    }
+
+    HVX_Vector rsum = hvx_vec_reduce_sum_f32x2(r0_sum, r1_sum);
+    hvx_vec_store_u(s0, 8, rsum);
+}
+
+static void vec_dot_iq4nlx4x2_q8x4x2_2x2(const int n,
+                                         float * restrict s0,
+                                         float * restrict s1,
+                                         const void * restrict vx0,
+                                         const void * restrict vx1,
+                                         const void * restrict vy0,
+                                         const void * restrict vy1) {
+    assert(n % 32 == 0);
+    assert((unsigned long) vx0 % 128 == 0);
+    assert((unsigned long) vx1 % 128 == 0);
+    assert((unsigned long) vy0 % 128 == 0);
+    assert((unsigned long) vy1 % 128 == 0);
+
+    const uint32_t qk = QK_Q4_0x4x2 * 4;
+
+    const uint32_t x_dblk_size = 8 * 4 * 2;  // 32x __fp16
+    const uint32_t x_qblk_size = qk / 2;     // int4
+    const uint32_t x_qrow_size = n / 2;      // int4 (not padded)
+
+    const uint32_t y_dblk_size = 8 * 4 * 2;  // 32x __fp16
+    const uint32_t y_qblk_size = qk;         // int8
+    const uint32_t y_qrow_size = n;          // int8 (not padded)
+
+    const uint8_t * restrict r0_x_q = ((const uint8_t *) vx0) + 0;
+    const uint8_t * restrict r0_x_d = ((const uint8_t *) vx0) + x_qrow_size;
+    const uint8_t * restrict r1_x_q = ((const uint8_t *) vx1) + 0;
+    const uint8_t * restrict r1_x_d = ((const uint8_t *) vx1) + x_qrow_size;
+
+    const uint8_t * restrict y0_q = ((const uint8_t *) vy0) + 0;
+    const uint8_t * restrict y0_d = ((const uint8_t *) vy0) + y_qrow_size;
+    const uint8_t * restrict y1_q = ((const uint8_t *) vy1) + 0;
+    const uint8_t * restrict y1_d = ((const uint8_t *) vy1) + y_qrow_size;
+
+    HVX_Vector r0_c0_sum = Q6_V_vzero();
+    HVX_Vector r0_c1_sum = Q6_V_vzero();
+    HVX_Vector r1_c0_sum = Q6_V_vzero();
+    HVX_Vector r1_c1_sum = Q6_V_vzero();
+
+    const uint32_t nb   = n / qk;
+    const uint32_t nloe = n % qk;
+
+    uint32_t i = 0;
+    for (; i < nb; i++) {
+        HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_full(y0_q + i * y_qblk_size);
+        HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_full(y1_q + i * y_qblk_size);
+        HVX_Vector_x8 r0_q  = hvx_vec_load_iq4nlx4x8_full(r0_x_q + i * x_qblk_size);
+        HVX_Vector_x8 r1_q  = hvx_vec_load_iq4nlx4x8_full(r1_x_q + i * x_qblk_size);
+
+        HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy0_q));
+        HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r0_q, vy1_q));
+        HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy0_q));
+        HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_full(r1_q, vy1_q));
+
+        HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
+        HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
+        HVX_Vector r0_d  = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+        HVX_Vector r1_d  = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_c0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy0_d)));
+        HVX_Vector r0_c1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy1_d)));
+        HVX_Vector r1_c0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy0_d)));
+        HVX_Vector r1_c1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy1_d)));
+
+        HVX_Vector r0_c0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_c0_ia, r0_c0_dd);
+        HVX_Vector r0_c1_fa = Q6_Vqf32_vmpy_VsfVsf(r0_c1_ia, r0_c1_dd);
+        HVX_Vector r1_c0_fa = Q6_Vqf32_vmpy_VsfVsf(r1_c0_ia, r1_c0_dd);
+        HVX_Vector r1_c1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_c1_ia, r1_c1_dd);
+
+        r0_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c0_fa, r0_c0_sum));
+        r0_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c1_fa, r0_c1_sum));
+        r1_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c0_fa, r1_c0_sum));
+        r1_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c1_fa, r1_c1_sum));
+    }
+
+    if (nloe) {
+        HVX_Vector_x8 vy0_q = hvx_vec_load_q8x4x8_partial(y0_q + i * y_qblk_size, nloe);
+        HVX_Vector_x8 vy1_q = hvx_vec_load_q8x4x8_partial(y1_q + i * y_qblk_size, nloe);
+        HVX_Vector_x8 r0_q  = hvx_vec_load_iq4nlx4x8_partial(r0_x_q + i * x_qblk_size, nloe);
+        HVX_Vector_x8 r1_q  = hvx_vec_load_iq4nlx4x8_partial(r1_x_q + i * x_qblk_size, nloe);
+
+        HVX_Vector r0_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy0_q, nloe));
+        HVX_Vector r0_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r0_q, vy1_q, nloe));
+        HVX_Vector r1_c0_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy0_q, nloe));
+        HVX_Vector r1_c1_ia = Q6_Vsf_equals_Vw(hvx_vec_rmpy_x8_partial(r1_q, vy1_q, nloe));
+
+        HVX_Vector vy0_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y0_d + i * y_dblk_size));
+        HVX_Vector vy1_d = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (y1_d + i * y_dblk_size));
+        HVX_Vector r0_d  = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r0_x_d + i * x_dblk_size));
+        HVX_Vector r1_d  = Q6_Vh_vshuff_Vh(*(const HVX_UVector *) (r1_x_d + i * x_dblk_size));
+
+        HVX_Vector r0_c0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy0_d)));
+        HVX_Vector r0_c1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r0_d, vy1_d)));
+        HVX_Vector r1_c0_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy0_d)));
+        HVX_Vector r1_c1_dd = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(Q6_Wqf32_vmpy_VhfVhf(r1_d, vy1_d)));
+
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe / 8);
+        r0_c0_dd             = Q6_V_vand_QV(bmask, r0_c0_dd);
+        r0_c1_dd             = Q6_V_vand_QV(bmask, r0_c1_dd);
+        r1_c0_dd             = Q6_V_vand_QV(bmask, r1_c0_dd);
+        r1_c1_dd             = Q6_V_vand_QV(bmask, r1_c1_dd);
+        r0_c0_ia             = Q6_V_vand_QV(bmask, r0_c0_ia);
+        r0_c1_ia             = Q6_V_vand_QV(bmask, r0_c1_ia);
+        r1_c0_ia             = Q6_V_vand_QV(bmask, r1_c0_ia);
+        r1_c1_ia             = Q6_V_vand_QV(bmask, r1_c1_ia);
+
+        HVX_Vector r0_c0_fa = Q6_Vqf32_vmpy_VsfVsf(r0_c0_ia, r0_c0_dd);
+        HVX_Vector r0_c1_fa = Q6_Vqf32_vmpy_VsfVsf(r0_c1_ia, r0_c1_dd);
+        HVX_Vector r1_c0_fa = Q6_Vqf32_vmpy_VsfVsf(r1_c0_ia, r1_c0_dd);
+        HVX_Vector r1_c1_fa = Q6_Vqf32_vmpy_VsfVsf(r1_c1_ia, r1_c1_dd);
+
+        r0_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c0_fa, r0_c0_sum));
+        r0_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r0_c1_fa, r0_c1_sum));
+        r1_c0_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c0_fa, r1_c0_sum));
+        r1_c1_sum = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(r1_c1_fa, r1_c1_sum));
+    }
+
+    HVX_Vector r0_r1_c0_sum = hvx_vec_reduce_sum_f32x2(r0_c0_sum, r1_c0_sum);
+    HVX_Vector r0_r1_c1_sum = hvx_vec_reduce_sum_f32x2(r0_c1_sum, r1_c1_sum);
+
+    hvx_vec_store_u(&s0[0], 8, r0_r1_c0_sum);
+    hvx_vec_store_u(&s1[0], 8, r0_r1_c1_sum);
+}
+
 static void vec_dot_mxfp4x4x2_q8x4x2_1x1(const int n, float * restrict s0, const void * restrict vx0, const void * restrict vy0) {
     assert(n % 32 == 0);  // min sub-block size
     assert((unsigned long) vx0 % 128 == 0);
@@ -2393,6 +2757,12 @@ static int htp_mminit_vec_dot(struct htp_matmul_context * mmctx, enum htp_data_t
             mmctx->vec_dot_2x1 = vec_dot_q8x4x2_q8x4x2_2x1;
             mmctx->vec_dot_2x2 = vec_dot_q8x4x2_q8x4x2_2x2;
             return 0;
+        case HTP_TYPE_IQ4_NL:
+            mmctx->type        = "iq4nlx4x2-f32";
+            mmctx->vec_dot_1x1 = vec_dot_iq4nlx4x2_q8x4x2_1x1;
+            mmctx->vec_dot_2x1 = vec_dot_iq4nlx4x2_q8x4x2_2x1;
+            mmctx->vec_dot_2x2 = vec_dot_iq4nlx4x2_q8x4x2_2x2;
+            return 0;
         case HTP_TYPE_MXFP4:
             mmctx->type        = "mxfp4x4x2-f32";
             mmctx->vec_dot_1x1 = vec_dot_mxfp4x4x2_q8x4x2_1x1;
@@ -2556,6 +2926,13 @@ int op_matmul(struct htp_ops_context * octx) {
         const uint32_t n_quant_jobs  = MIN(src1_nrows, octx->n_threads);
         mmctx->src1_nrows_per_thread = (src1_nrows + n_quant_jobs - 1) / n_quant_jobs;
         worker_pool_run_func(octx->ctx->worker_pool, quant_job_func, mmctx, n_quant_jobs);
+        // Cache where src1 was written so subsequent SKIP_QUANTIZE ops can find it
+        octx->ctx->prev_src1_spad = octx->src1_spad.data;
+    } else {
+        // SKIP_QUANTIZE: Q8 data lives at the address written by the previous
+        // quantize pass.  The current op may have a different src0 size (e.g.
+        // IQ4_NL vs MXFP4), so src1_spad.data computed above could be wrong.
+        octx->src1_spad.data = octx->ctx->prev_src1_spad;
     }
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {
@@ -2659,6 +3036,9 @@ int op_matmul_id(struct htp_ops_context * octx) {
         const uint32_t n_quant_jobs = MIN(src1_nrows, octx->n_threads);
         mmctx->src1_nrows_per_thread = (src1_nrows + n_quant_jobs - 1) / n_quant_jobs;
         worker_pool_run_func(octx->ctx->worker_pool, quant_job_func, mmctx, n_quant_jobs);
+        octx->ctx->prev_src1_spad = octx->src1_spad.data;
+    } else {
+        octx->src1_spad.data = octx->ctx->prev_src1_spad;
     }
 
     if (!(octx->flags & HTP_OPFLAGS_SKIP_COMPUTE)) {

ggml/src/ggml-hexagon/htp/rope-ops.c

@@ -333,8 +333,8 @@ static void rope_job_f32(unsigned int nth, unsigned int ith, void * data) {
                     //         (unsigned) HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - rctx->t_start));
                 }
 
-                // Skip DMA transactions from prev block (if any)
-                // No need to wait for these since the DMA is setup for in-order processing
+                // Skip output DMA transactions from prev block (if any)
+                // No need to wait for those here since we're explicitly waiting for the latest prefecthes below.
                 for (uint32_t d=0; d < dma_depth; d++) { dma_queue_pop_nowait(dma_queue); }
 
                 // Compute loop

ggml/src/ggml-hexagon/htp/unary-ops.c

@@ -67,34 +67,61 @@ static void hvx_fast_rms_norm_f32(const uint8_t * restrict src,
                                   uint8_t * restrict pad,
                                   const int num_elems,
                                   float     epsilon) {
+    (void)pad;
+
     const HVX_Vector * restrict v_src = (HVX_Vector *) src;
     HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
 
-    HVX_Vector sum_v     = Q6_V_vsplat_R(0x00000000);
+    const int nvec = num_elems / VLEN_FP32;    // number of full vectors
+    const int nloe = num_elems % VLEN_FP32;    // leftover elements
+
+    // Compute sum of squares for full vectors
+    HVX_Vector sum_v = Q6_V_vsplat_R(0x00000000);
     HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
 
-    int step_of_1 = num_elems >> 5;
     #pragma unroll(4)
-    for (int i = 0; i < step_of_1; i++) {
+    for (int i = 0; i < nvec; i++) {
         HVX_Vector v1 = v_src[i];
         HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
-        sum_v         = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
     }
 
-    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v)); // replicated over all lanes
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, v1);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, v2);
+    }
+
+    // Reduce HVX sum
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
 
     HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
     HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
     HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
     HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
 
+    // Scale full vectors
     HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
 
     #pragma unroll(4)
-    for (int i = 0; i < step_of_1; i++) {
+    for (int i = 0; i < nvec; i++) {
         HVX_Vector v1 = v_src[i];
         HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
-        v_dst[i]      = Q6_Vsf_equals_Vqf32(v2);
+        v_dst[i] = Q6_Vsf_equals_Vqf32(v2);
+    }
+
+    // Handle tail elements using vectorized ops with masking
+    if (nloe > 0) {
+
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * 4);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(v1, scale_v);
+        HVX_Vector result = Q6_Vsf_equals_Vqf32(v2);
+
+        // Store with masking to avoid overwriting memory beyond the tensor
+        hvx_vec_store_a(&v_dst[nvec], nloe * 4, result);
     }
 }
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -114,6 +114,8 @@ set(GGML_OPENCL_KERNELS
     gemv_noshuffle_q4_1_f32
     gemm_noshuffle_q4_1_f32
     gemv_noshuffle_general_q8_0_f32
+    gemv_noshuffle_q4_k_f32
+    gemm_noshuffle_q4_k_f32
     gemv_noshuffle_q6_k_f32
     gemm_noshuffle_q6_k_f32
     mul