ggml : revert to -lm linking instead of find_library (#22355)

* ggml : revert to -lm linking instead of find_library

`find_library(MATH_LIBRARY m)` was introduced recently, but it breaks
CUDA compilation with GGML_STATIC. I could not find any valid use case
where we would prefer `find_library` over the standard `-lm` approach.

This commit is also meant to start a discussion if there is a valid
reason to keep `find_library(MATH_LIBRARY m)`, we should clarify what
problem it was solving and find an alternative fix that does not break
CUDA with GGML_STATIC.

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* ggml : use MATH_LIBRARY only if defined

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* ggml : fix initial broken condition

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* ggml : always respect MATH_LIBRARY when defined

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

---------

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

.devops/intel.Dockerfile

@@ -1,4 +1,4 @@
-ARG ONEAPI_VERSION=2025.3.2-0-devel-ubuntu24.04
+ARG ONEAPI_VERSION=2025.3.3-0-devel-ubuntu24.04
 
 ## Build Image
 

.devops/openvino.Dockerfile

@@ -2,7 +2,19 @@ ARG OPENVINO_VERSION_MAJOR=2026.0
 ARG OPENVINO_VERSION_FULL=2026.0.0.20965.c6d6a13a886
 ARG UBUNTU_VERSION=24.04
 
-# Optional proxy build arguments - empty by default
+# Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
+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
+
+# Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
+ARG NPU_DRIVER_VERSION=v1.32.0
+ARG NPU_DRIVER_FULL=v1.32.0.20260402-23905121947
+ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+
+# Optional proxy build arguments
 ARG http_proxy=
 ARG https_proxy=
 
@@ -78,13 +90,47 @@ ARG http_proxy
 ARG https_proxy
 
 RUN apt-get update \
-    && apt-get install -y libgomp1 libtbb12 curl \
+    && apt-get install -y libgomp1 libtbb12 curl wget ocl-icd-libopencl1 \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \
     && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \
     && find /var/cache -type f -delete
 
+# Install GPU drivers
+ARG IGC_VERSION
+ARG IGC_VERSION_FULL
+ARG COMPUTE_RUNTIME_VERSION
+ARG COMPUTE_RUNTIME_VERSION_FULL
+ARG IGDGMM_VERSION
+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 \
+    && rm -rf /tmp/neo/
+
+# Install NPU drivers
+ARG NPU_DRIVER_VERSION
+ARG NPU_DRIVER_FULL
+ARG LIBZE1_VERSION
+RUN mkdir /tmp/npu/ && cd /tmp/npu/ \
+    && wget https://github.com/intel/linux-npu-driver/releases/download/${NPU_DRIVER_VERSION}/linux-npu-driver-${NPU_DRIVER_FULL}-ubuntu2404.tar.gz \
+    && tar -xf linux-npu-driver-${NPU_DRIVER_FULL}-ubuntu2404.tar.gz \
+    && dpkg --install *.deb \
+    && rm -rf /tmp/npu/
+
+RUN cd /tmp \
+    && wget https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb \
+    && dpkg --install libze1_${LIBZE1_VERSION}_amd64.deb \
+    && rm libze1_${LIBZE1_VERSION}_amd64.deb
+
 COPY --from=build /app/lib/ /app/
 
 ### Full (all binaries)

.github/pull_request_template.md

@@ -6,7 +6,7 @@
 
 <!-- You can provide more details and link related discussions here. Delete this section if not applicable -->
 
-# Requirements
+## Requirements
 
 <!-- IMPORTANT: Please do NOT delete this section, otherwise your PR may be rejected -->
 

.github/workflows/build-and-test-snapdragon.yml

@@ -0,0 +1,116 @@
+name: CI (snapdragon)
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - master
+    paths:
+      - '.github/workflows/build-and-test-snapdragon.yml'
+      - 'ggml/include/ggml-hexagon.h'
+      - 'ggml/src/ggml-hexagon/**'
+      - 'docs/backend/snapdragon/**'
+      - 'scripts/snapdragon/**'
+      - 'CMakePresets.json'
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths:
+      - '.github/workflows/build-and-test-snapdragon.yml'
+      - 'ggml/include/ggml-hexagon.h'
+      - 'ggml/src/ggml-hexagon/**'
+      - 'docs/backend/snapdragon/**'
+      - 'scripts/snapdragon/**'
+      - 'CMakePresets.json'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+jobs:
+  android-ndk-snapdragon:
+    runs-on: ubuntu-latest
+    container:
+      image: 'ghcr.io/snapdragon-toolchain/arm64-android:v0.3'
+    defaults:
+      run:
+        shell: bash
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+          lfs: false
+
+      - name: Build Llama.CPP for Snapdragon Android
+        id: build_llama_cpp_snapdragon_android
+        run: |
+          cp docs/backend/snapdragon/CMakeUserPresets.json .
+          cmake --preset arm64-android-snapdragon-release -B build
+          cmake --build build
+          cmake --install build --prefix pkg-snapdragon/llama.cpp
+
+      - name: Upload Llama.CPP Snapdragon Android Build Artifact
+        if: ${{ always() && steps.build_llama_cpp_snapdragon_android.outcome == 'success' }}
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-cpp-android-arm64-snapdragon
+          path: pkg-snapdragon/llama.cpp
+
+  test-snapdragon-qdc:
+    name: Test on QDC Android Device (${{ matrix.device }})
+    needs: [android-ndk-snapdragon]
+    runs-on: ubuntu-slim
+    strategy:
+      fail-fast: false
+      matrix:
+        device: [SM8750, SM8650, SM8850]
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+
+      - name: Download build artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-cpp-android-arm64-snapdragon
+          path: pkg-snapdragon/llama.cpp
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.x'
+          cache: pip
+
+      - name: Install system dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y curl unzip
+
+      - name: Install QDC SDK wheel
+        run: |
+          curl -fSL -o qdc_sdk.zip https://softwarecenter.qualcomm.com/api/download/software/tools/Qualcomm_Device_Cloud_SDK/All/0.2.3/qualcomm_device_cloud_sdk-0.2.3.zip
+          unzip qdc_sdk.zip -d qdc_sdk
+          pip install qdc_sdk/qualcomm_device_cloud_sdk-0.2.3-py3-none-any.whl
+
+      - name: Check QDC API key
+        id: check_secret
+        env:
+          QDC_API_KEY: ${{ secrets.QDC_API_KEY }}
+        run: echo "has-qdc-key=${{ env.QDC_API_KEY != '' }}" >> "$GITHUB_OUTPUT"
+
+      - name: Run QDC tests (${{ matrix.device }})
+        if: steps.check_secret.outputs.has-qdc-key == 'true'
+        run: |
+          python scripts/snapdragon/qdc/run_qdc_jobs.py \
+              --test       all \
+              --pkg-dir    pkg-snapdragon/llama.cpp \
+              --model-url  "https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf" \
+              --device     ${{ matrix.device }}
+        env:
+          QDC_API_KEY: ${{ secrets.QDC_API_KEY }}
+
+      - name: Cleanup
+        if: always()
+        run: rm -rf pkg-snapdragon qdc_sdk qdc_sdk.zip

.github/workflows/build-android.yml

@@ -1,26 +1,24 @@
 name: CI (android)
 
 on:
-  workflow_dispatch: # allows manual triggering
+  workflow_dispatch:
   push:
     branches:
       - master
-    paths: [
-      '.github/workflows/build-android.yml',
-      '**/CMakeLists.txt',
-      '**/.cmake',
-      '**/*.h',
-      '**/*.hpp',
-      '**/*.c',
-      '**/*.cpp'
-    ]
+    paths:
+      - '.github/workflows/build-android.yml'
+      - '**/CMakeLists.txt'
+      - '**/.cmake'
+      - '**/*.h'
+      - '**/*.hpp'
+      - '**/*.c'
+      - '**/*.cpp'
 
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: [
-      '.github/workflows/build-android.yml',
-      'examples/llama.android/**'
-    ]
+    paths:
+      - '.github/workflows/build-android.yml'
+      - 'examples/llama.android/**'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -67,35 +65,24 @@ jobs:
     defaults:
       run:
         shell: bash
-    strategy:
-      matrix:
-        include:
-          - build: 'arm64-cpu'
-            defines: '-D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF'
-          - build: 'arm64-snapdragon'
-            defines: '--preset arm64-android-snapdragon-release'
 
     steps:
       - name: Clone
-        id: checkout
         uses: actions/checkout@v6
         with:
           fetch-depth: 0
           lfs: false
 
-      - name: Build Llama.CPP for Hexagon Android
-        id: build_llama_cpp_hexagon_android
+      - name: Build
+        id: ndk_build
         run: |
-          if [[ "${{ matrix.build }}" == "arm64-snapdragon" ]]; then
-            cp docs/backend/snapdragon/CMakeUserPresets.json .
-          fi
-          cmake ${{ matrix.defines }} -B build
+          cmake -D ANDROID_ABI=arm64-v8a -D ANDROID_PLATFORM=android-31 -D CMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake -D GGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8.5-a+fp16+i8mm -G Ninja -D LLAMA_OPENSSL=OFF -D GGML_OPENMP=OFF -B build
           cmake --build build
           cmake --install build --prefix pkg-adb/llama.cpp
 
-      - name: Upload Llama.CPP Hexagon Android Build Artifact
-        if: ${{ always() && steps.build_llama_cpp_hexagon_android.outcome == 'success' }}
+      - name: Upload Android Build Artifact
+        if: ${{ always() && steps.ndk_build.outcome == 'success' }}
         uses: actions/upload-artifact@v6
         with:
-          name: llama-cpp-android-${{ matrix.build }}
+          name: llama-cpp-android-arm64-cpu
           path: pkg-adb/llama.cpp

.github/workflows/build-cross.yml

@@ -246,6 +246,7 @@ jobs:
           apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
+                  spirv-headers \
                   gcc-14-loongarch64-linux-gnu \
                   g++-14-loongarch64-linux-gnu \
                   libvulkan-dev:loong64

.github/workflows/build-openvino.yml

@@ -0,0 +1,120 @@
+name: CI (openvino)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-openvino.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-openvino.yml',
+      'ggml/src/ggml-openvino/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+  ubuntu-24-openvino:
+    name: ubuntu-24-openvino-${{ matrix.openvino_device }}
+
+    concurrency:
+      group: openvino-${{ matrix.variant }}-${{ github.head_ref || github.ref }}
+      cancel-in-progress: false
+
+    strategy:
+      matrix:
+        include:
+          - variant: cpu
+            runner: '"ubuntu-24.04"'
+            openvino_device: "CPU"
+          - variant: gpu
+            runner: '["self-hosted","Linux","Intel","OpenVINO"]'
+            openvino_device: "GPU"
+
+    runs-on: ${{ fromJSON(matrix.runner) }}
+
+    env:
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.0"
+      OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        if: runner.environment == 'github-hosted'
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
+          sudo apt-get install -y ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
+
+      - name: Use OpenVINO Toolkit Cache
+        if: runner.environment == 'github-hosted'
+        uses: actions/cache@v5
+        id: cache-openvino
+        with:
+          path: ./openvino_toolkit
+          key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
+
+      - name: Setup OpenVINO Toolkit
+        if: steps.cache-openvino.outputs.cache-hit != 'true'
+        uses: ./.github/actions/linux-setup-openvino
+        with:
+          path: ./openvino_toolkit
+          version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
+          version_full: ${{ env.OPENVINO_VERSION_FULL }}
+
+      - name: Install OpenVINO dependencies
+        run: |
+          cd ./openvino_toolkit
+          chmod +x ./install_dependencies/install_openvino_dependencies.sh
+          echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source ./openvino_toolkit/setupvars.sh
+          cmake -B build/ReleaseOV -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_OPENVINO=ON
+          time cmake --build build/ReleaseOV --config Release -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        # TODO: fix and re-enable the `test-llama-archs` test below
+        run: |
+          cd ${{ github.workspace }}
+          if [ "${{ matrix.openvino_device }}" = "GPU" ]; then
+            export GGML_OPENVINO_DEVICE=GPU
+          fi
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000

.github/workflows/build-self-hosted.yml

@@ -265,6 +265,10 @@ jobs:
   ggml-ci-intel-openvino-gpu-low-perf:
     runs-on: [self-hosted, Linux, Intel, OpenVINO]
 
+    concurrency:
+      group: openvino-gpu-${{ github.head_ref || github.ref }}
+      cancel-in-progress: false
+
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
       OPENVINO_VERSION_MAJOR: "2026.0"

.github/workflows/build-sycl.yml

@@ -0,0 +1,142 @@
+name: CI (sycl)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: [
+      '.github/workflows/build-sycl.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp'
+    ]
+
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: [
+      '.github/workflows/build-sycl.yml',
+      'ggml/src/ggml-sycl/**'
+    ]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+
+jobs:
+
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
+
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+
+    continue-on-error: true
+
+    steps:
+      - uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-sycl-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+  windows-latest-sycl:
+    runs-on: windows-2022
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: windows-latest-sycl
+          variant: ccache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat

.github/workflows/build.yml

@@ -555,186 +555,6 @@ jobs:
             -DGGML_MUSA=ON
           time cmake --build build --config Release -j $(nproc)
 
-  ubuntu-22-sycl:
-    runs-on: ubuntu-22.04
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-22-sycl
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          cmake -B build \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx
-          time cmake --build build --config Release -j $(nproc)
-
-  ubuntu-22-sycl-fp16:
-    runs-on: ubuntu-22.04
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp libssl-dev ninja-build
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: ubuntu-22-sycl-fp16
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          cmake -B build \
-            -G "Ninja" \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DGGML_SYCL=ON \
-            -DCMAKE_C_COMPILER=icx \
-            -DCMAKE_CXX_COMPILER=icpx \
-            -DGGML_SYCL_F16=ON
-          time cmake --build build --config Release -j $(nproc)
-
-  ubuntu-24-openvino:
-      name: ubuntu-24-openvino-${{ matrix.openvino_device }}
-      strategy:
-        matrix:
-          include:
-            - variant: cpu
-              runner: '"ubuntu-24.04"'
-              openvino_device: "CPU"
-            - variant: gpu
-              runner: '["self-hosted","Linux","X64","Intel"]'
-              openvino_device: "GPU"
-
-      runs-on: ${{ fromJSON(matrix.runner) }}
-
-      env:
-        # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-        OPENVINO_VERSION_MAJOR: "2026.0"
-        OPENVINO_VERSION_FULL: "2026.0.0.20965.c6d6a13a886"
-
-      steps:
-        - name: Clone
-          id: checkout
-          uses: actions/checkout@v6
-
-        - name: ccache
-          if: runner.environment == 'github-hosted'
-          uses: ggml-org/ccache-action@v1.2.21
-          with:
-            key: ubuntu-24-openvino-${{ matrix.variant }}-no-preset-v1
-            evict-old-files: 1d
-            save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-        - name: Dependencies
-          id: depends
-          run: |
-            sudo apt-get update
-            sudo apt-get install -y build-essential libssl-dev libtbb12 cmake ninja-build python3-pip
-            sudo apt-get install -y ocl-icd-opencl-dev opencl-headers opencl-clhpp-headers intel-opencl-icd
-
-        - name: Use OpenVINO Toolkit Cache
-          if: runner.environment == 'github-hosted'
-          uses: actions/cache@v5
-          id: cache-openvino
-          with:
-            path: ./openvino_toolkit
-            key: openvino-toolkit-v${{ env.OPENVINO_VERSION_FULL }}-${{ runner.os }}
-
-        - name: Setup OpenVINO Toolkit
-          if: steps.cache-openvino.outputs.cache-hit != 'true'
-          uses: ./.github/actions/linux-setup-openvino
-          with:
-            path: ./openvino_toolkit
-            version_major: ${{ env.OPENVINO_VERSION_MAJOR }}
-            version_full: ${{ env.OPENVINO_VERSION_FULL }}
-
-        - name: Install OpenVINO dependencies
-          run: |
-            cd ./openvino_toolkit
-            chmod +x ./install_dependencies/install_openvino_dependencies.sh
-            echo "Y" | sudo -E ./install_dependencies/install_openvino_dependencies.sh
-
-        - name: Build
-          id: cmake_build
-          run: |
-            source ./openvino_toolkit/setupvars.sh
-            cmake -B build/ReleaseOV -G Ninja \
-              -DCMAKE_BUILD_TYPE=Release \
-              -DGGML_OPENVINO=ON
-            time cmake --build build/ReleaseOV --config Release -j $(nproc)
-
-        - name: Test
-          id: cmake_test
-          # TODO: fix and re-enable the `test-llama-archs` test below
-          run: |
-            cd ${{ github.workspace }}
-            if [ "${{ matrix.openvino_device }}" = "GPU" ]; then
-              export GGML_OPENVINO_DEVICE=GPU
-            fi
-            ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
 
   windows-latest:
     runs-on: windows-2025
@@ -943,39 +763,6 @@ jobs:
           cmake --build build --config Release -j %NINJA_JOBS% -t ggml
           cmake --build build --config Release
 
-  windows-latest-sycl:
-    runs-on: windows-2022
-
-    defaults:
-      run:
-        shell: bash
-
-    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
-      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
-      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.21
-        with:
-          key: windows-latest-sycl
-          variant: ccache
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-
-      - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
-      # TODO: add ssl support ; we will also need to modify win-build-sycl.bat to accept user-specified args
-
-      - name: Build
-        id: cmake_build
-        run:  examples/sycl/win-build-sycl.bat
 
   windows-latest-hip:
     runs-on: windows-2022

.github/workflows/release.yml

@@ -598,15 +598,29 @@ jobs:
         shell: bash
 
     env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/b60765d1-2b85-4e85-86b6-cb0e9563a699/intel-deep-learning-essentials-2025.3.3.18_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel:intel.oneapi.win.dnnl:intel.oneapi.win.tbb.devel
       ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
 
     steps:
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
@@ -614,10 +628,6 @@ jobs:
           variant: ccache
           evict-old-files: 1d
 
-      - name: Install
-        run:  |
-          scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
-
       - name: Build
         id: cmake_build
         shell: cmd
@@ -670,6 +680,82 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+  ubuntu-24-sycl:
+    strategy:
+      matrix:
+        build: [fp32, fp16]
+        include:
+          - build: fp32
+            fp16: OFF
+          - build: fp16
+            fp16: ON
+
+    runs-on: ubuntu-24.04
+
+    env:
+      ONEAPI_ROOT: /opt/intel/oneapi/
+      ONEAPI_INSTALLER_VERSION: "2025.3.3"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: Use oneAPI Installation Cache
+        uses: actions/cache@v5
+        id: cache-sycl
+        with:
+          path: ${{ env.ONEAPI_ROOT }}
+          key: oneAPI-${{ env.ONEAPI_INSTALLER_VERSION }}-${{ runner.os }}
+
+      - name: Download & Install oneAPI
+        shell: bash
+        if: steps.cache-sycl.outputs.cache-hit != 'true'
+        run: |
+          cd /tmp
+          wget https://registrationcenter-download.intel.com/akdlm/IRC_NAS/56f7923a-adb8-43f3-8b02-2b60fcac8cab/intel-deep-learning-essentials-2025.3.3.16_offline.sh -O intel-deep-learning-essentials_offline.sh
+          sudo bash intel-deep-learning-essentials_offline.sh -s -a --silent --eula accept
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: ubuntu-24-sycl-${{ matrix.build }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Build
+        id: cmake_build
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          cmake -B build \
+            -G "Ninja" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_SYCL=ON \
+            -DCMAKE_C_COMPILER=icx \
+            -DCMAKE_CXX_COMPILER=icpx \
+            -DLLAMA_OPENSSL=OFF \
+            -DGGML_NATIVE=OFF \
+            -DGGML_SYCL_F16=${{ matrix.fp16 }}
+          time cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+          name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
+
   ubuntu-22-rocm:
     runs-on: ubuntu-22.04
 
@@ -1045,6 +1131,7 @@ jobs:
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
+      - ubuntu-24-sycl
       - android-arm64
       - macOS-cpu
       - ios-xcode-build
@@ -1133,6 +1220,8 @@ jobs:
             - [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-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
+            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
+            - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
 
             **Android:**
             - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)

.gitignore

@@ -34,7 +34,6 @@
 /.vscode/
 /nppBackup
 
-
 # Coverage
 
 /gcovr-report/
@@ -74,6 +73,7 @@
 !/models/templates
 
 # Zig
+
 /zig-out/
 /zig-cache/
 
@@ -93,6 +93,7 @@
 !/examples/sycl/*.sh
 
 # Server Web UI temporary files
+
 /tools/server/webui/node_modules
 /tools/server/webui/dist
 # we no longer use gz for index.html
@@ -106,9 +107,11 @@ __pycache__/
 poetry.toml
 
 # Nix
+
 /result
 
 # Test binaries
+
 /tests/test-backend-ops
 /tests/test-double-float
 /tests/test-grad0
@@ -124,6 +127,7 @@ poetry.toml
 /tests/test-tokenizer-1-spm
 
 # Scripts
+
 !/scripts/install-oneapi.bat
 
 # Generated by scripts
@@ -132,16 +136,24 @@ poetry.toml
 /wikitext-2-raw/
 
 # Test models for lora adapters
+
 /lora-tests
 
 # Local scripts
+
 /run-vim.sh
 /run-chat.sh
 /run-spec.sh
 /.ccache/
 
 # IDE
+
 /*.code-workspace
 /.windsurf/
 # emscripten
 a.out.*
+
+# AGENTS
+
+AGENTS.local.md
+.pi/SYSTEM.md

.pi/gg/SYSTEM.md

@@ -0,0 +1,33 @@
+You are a coding agent. Here are some very important rules that you must follow:
+
+General:
+- By very precise and concise when writing code, comments, explanations, etc.
+- PR and commit titles format: `<module> : <title>`. Lookup recents for examples
+- Don't try to build or run the code unless you are explicitly asked to do so
+
+Coding:
+- When in doubt, always refer to the CONTRIBUTING.md file of the project
+- When referencing issues or PRs in comments, use the format:
+  - C/C++ code: `// ref: <url>`
+  - Other (CMake, etc.): `# ref: <url>`
+
+Pull requests (PRs):
+- New branch names are prefixed with "gg/"
+- Before opening a pull request, ask the user to confirm the description
+- When creating a pull request, look for the repository's PR template and follow it
+- For the AI usage disclosure section, write "YES. llama.cpp + pi"
+- Always create the pull requests in draft mode
+
+Commits:
+- On every commit that you make, include a "Assisted-by: llama.cpp:local pi" tag
+- Do not explicitly set the git author in commits - rely on the default git config
+
+Resources (read on demand):
+- [CONTRIBUTING.md](CONTRIBUTING.md)
+- [Build documentation](docs/build.md)
+- [Server usage documentation](tools/server/README.md)
+- [Server development documentation](tools/server/README-dev.md)
+- [PEG parser](docs/development/parsing.md)
+- [Auto parser](docs/autoparser.md)
+- [Jinja engine](common/jinja/README.md)
+- [PR template](.github/pull_request_template.md)

CODEOWNERS

@@ -23,6 +23,7 @@
 /ci/                                    @ggerganov
 /cmake/                                 @ggerganov
 /common/                                @ggml-org/llama-common
+/common/fit.*                           @JohannesGaessler
 /common/jinja/                          @CISC
 /common/ngram-map.*                     @srogmann
 /convert_*.py                           @CISC
@@ -52,28 +53,29 @@
 /examples/speculative/                  @ggerganov
 /ggml/cmake/                            @ggerganov
 /ggml/include/                          @ggerganov
+/ggml/src/ggml-backend-meta.cpp         @JohannesGaessler
 /ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
-/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
-/ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK
+/ggml/src/ggml-cuda/fattn-wmma*         @IMbackK
+/ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
+/ggml/src/ggml-hip/                     @IMbackK
 /ggml/src/ggml-impl.h                   @ggerganov
 /ggml/src/ggml-metal/                   @ggml-org/ggml-metal
 /ggml/src/ggml-opencl/                  @ggml-org/ggml-opencl
-/ggml/src/ggml-hexagon/                 @ggml-org/ggml-hexagon
+/ggml/src/ggml-openvino/                @cavusmustafa @wine99
 /ggml/src/ggml-opt.cpp                  @JohannesGaessler
 /ggml/src/ggml-quants.*                 @ggerganov
 /ggml/src/ggml-rpc/                     @ggml-org/ggml-rpc
 /ggml/src/ggml-sycl/                    @ggml-org/ggml-sycl
 /ggml/src/ggml-threading.*              @ggerganov
-/ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-virtgpu/                 @kpouget
+/ggml/src/ggml-vulkan/                  @ggml-org/ggml-vulkan
 /ggml/src/ggml-webgpu/                  @ggml-org/ggml-webgpu
 /ggml/src/ggml-zdnn/                    @ggml-org/ggml-zdnn @Andreas-Krebbel @AlekseiNikiforovIBM
-/ggml/src/ggml-openvino/                @cavusmustafa @wine99
 /ggml/src/ggml.c                        @ggerganov
 /ggml/src/ggml.cpp                      @ggerganov
 /ggml/src/gguf.cpp                      @JohannesGaessler @Green-Sky

common/CMakeLists.txt

@@ -73,6 +73,8 @@ add_library(${TARGET}
     debug.h
     download.cpp
     download.h
+    fit.cpp
+    fit.h
     hf-cache.cpp
     hf-cache.h
     http.h

common/arg.h

@@ -25,7 +25,8 @@ struct common_arg {
     const char * value_hint_2 = nullptr; // for second arg value
     const char * env          = nullptr;
     std::string help;
-    bool is_sparam = false; // is current arg a sampling param?
+    bool is_sampling = false; // is current arg a sampling param?
+    bool is_spec = false; // is current arg a speculative decoding param?
     bool is_preset_only = false; // is current arg preset-only (not treated as CLI arg)
     void (*handler_void)   (common_params & params) = nullptr;
     void (*handler_string) (common_params & params, const std::string &) = nullptr;
@@ -74,7 +75,8 @@ struct common_arg {
     common_arg & set_examples(std::initializer_list<enum llama_example> examples);
     common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
     common_arg & set_env(const char * env);
-    common_arg & set_sparam();
+    common_arg & set_sampling();
+    common_arg & set_spec();
     common_arg & set_preset_only();
     bool in_example(enum llama_example ex);
     bool is_exclude(enum llama_example ex);

common/chat-auto-parser-generator.cpp

@@ -443,14 +443,14 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     if (!format.per_call_start.empty()) {
         auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
         if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call) + p.space());
         } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.space());
         }
         if (!format.section_start.empty()) {
             tool_calls = p.trigger_rule("tool-calls",
                                         p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end) + p.space()));
         }
     } else {
         std::string separator = ", ";  // Default

common/chat-diff-analyzer.cpp

@@ -296,7 +296,7 @@ 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.tag("post", (p.space() + 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);
@@ -306,11 +306,11 @@ 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_leading_whitespace(result.tags["pre"]);
-                end   = trim_trailing_whitespace(result.tags["post"]);
+                start = result.tags["pre"];
+                end   = result.tags["post"];
             } else if (!result.tags["post"].empty()) {
                 mode = reasoning_mode::TAG_BASED;
-                end = trim_trailing_whitespace(result.tags["post"]);
+                end = result.tags["post"];
             }
         }
     }

common/chat.cpp

@@ -397,6 +397,25 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
     return render_message_to_json(msgs, c);
 }
 
+json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
+    if (tools.empty()) {
+        return json();
+    }
+
+    auto result = json::array();
+    for (const auto & tool : tools) {
+        result.push_back({
+            { "type",     "function" },
+            { "function", {
+                { "name", tool.name },
+                { "description", tool.description },
+                { "parameters", json::parse(tool.parameters) },
+            }},
+        });
+    }
+    return result;
+}
+
 std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & tools) {
     std::vector<common_chat_tool> result;
 
@@ -432,56 +451,6 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools) {
-    if (tools.empty()) {
-        return json();
-    }
-
-    auto result = json::array();
-    for (const auto & tool : tools) {
-        result.push_back({
-            { "type",     "function" },
-            { "function",
-             {
-                  { "name", tool.name },
-                  { "description", tool.description },
-                  { "parameters", json::parse(tool.parameters) },
-              }                      },
-        });
-    }
-    return result;
-}
-
-json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
-    json delta = json::object();
-    if (!diff.reasoning_content_delta.empty()) {
-        delta["reasoning_content"] = diff.reasoning_content_delta;
-    }
-    if (!diff.content_delta.empty()) {
-        delta["content"] = diff.content_delta;
-    }
-    if (diff.tool_call_index != std::string::npos) {
-        json tool_call;
-        tool_call["index"] = diff.tool_call_index;
-        if (!diff.tool_call_delta.id.empty()) {
-            tool_call["id"]   = diff.tool_call_delta.id;
-            tool_call["type"] = "function";
-        }
-        if (!diff.tool_call_delta.name.empty() || !diff.tool_call_delta.arguments.empty()) {
-            json function = json::object();
-            if (!diff.tool_call_delta.name.empty()) {
-                function["name"] = diff.tool_call_delta.name;
-            }
-            if (!diff.tool_call_delta.arguments.empty()) {
-                function["arguments"] = diff.tool_call_delta.arguments;
-            }
-            tool_call["function"] = function;
-        }
-        delta["tool_calls"] = json::array({ tool_call });
-    }
-    return delta;
-}
-
 bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
     if (use_jinja) {
         try {
@@ -575,6 +544,26 @@ bool common_chat_templates_was_explicit(const struct common_chat_templates * tmp
     return tmpls->has_explicit_template;
 }
 
+// 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
+static bool is_lfm2_template(const std::string & src) {
+    return src.find("<|tool_list_start|>") != std::string::npos &&
+           src.find("<|tool_list_end|>")   != std::string::npos;
+}
+
+common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates) {
+    common_chat_prompt_preset asr_preset;
+    asr_preset.system = "";
+    asr_preset.user   = "Transcribe audio to text";
+
+    if (chat_templates && chat_templates->template_default && is_lfm2_template(chat_templates->template_default->source())) {
+        asr_preset.system = "Perform ASR.";
+        asr_preset.user   = "";
+    }
+
+    return asr_preset;
+}
+
 std::string common_chat_templates_source(const struct common_chat_templates * tmpls, const std::string & variant) {
     if (!variant.empty()) {
         if (variant == "tool_use") {
@@ -2084,10 +2073,7 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         return common_chat_params_init_kimi_k2(tmpl, params);
     }
 
-    // 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) {
+    if (is_lfm2_template(src)) {
         LOG_DBG("Using specialized template: LFM2\n");
         return common_chat_params_init_lfm2(tmpl, params);
     }
@@ -2396,4 +2382,3 @@ std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_tem
     GGML_ASSERT(chat_templates->template_default != nullptr);
     return chat_templates->template_default->caps.to_map();
 }
-

common/chat.h

@@ -256,14 +256,13 @@ bool common_chat_templates_support_enable_thinking(const common_chat_templates *
 // Parses a JSON array of messages in OpenAI's chat completion API format.
 std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
 
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+
 // DEPRECATED: only used in tests
 nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
 nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
-nlohmann::ordered_json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff);
-
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
 
@@ -275,3 +274,11 @@ std::optional<common_chat_params> common_chat_try_specialized_template(
         const common_chat_template &          tmpl,
         const std::string &                   src,
         autoparser::generation_params & params);
+
+// specialized per-task preset
+struct common_chat_prompt_preset {
+    std::string system;
+    std::string user;
+};
+
+common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);

common/common.cpp

@@ -3,6 +3,7 @@
 
 #include "build-info.h"
 #include "common.h"
+#include "fit.h"
 #include "log.h"
 #include "llama.h"
 #include "sampling.h"
@@ -69,7 +70,7 @@ common_time_meas::~common_time_meas() {
 // CPU utils
 //
 
-int32_t cpu_get_num_physical_cores() {
+int32_t common_cpu_get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
     std::unordered_set<std::string> siblings;
@@ -184,11 +185,11 @@ static int cpu_count_math_cpus(int n_cpu) {
 /**
  * Returns number of CPUs on system that are useful for math.
  */
-int32_t cpu_get_num_math() {
+int32_t common_cpu_get_num_math() {
 #if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
     int n_cpu = sysconf(_SC_NPROCESSORS_ONLN);
     if (n_cpu < 1) {
-        return cpu_get_num_physical_cores();
+        return common_cpu_get_num_physical_cores();
     }
     if (is_hybrid_cpu()) {
         cpu_set_t affinity;
@@ -201,7 +202,7 @@ int32_t cpu_get_num_math() {
         }
     }
 #endif
-    return cpu_get_num_physical_cores();
+    return common_cpu_get_num_physical_cores();
 }
 
 // Helper for setting process priority
@@ -262,7 +263,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 //
 
 
-void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model) {
+void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_params * role_model) {
     int32_t n_set = 0;
 
     if (cpuparams.n_threads < 0) {
@@ -270,7 +271,7 @@ void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model)
         if (role_model != nullptr) {
             cpuparams = *role_model;
         } else {
-            cpuparams.n_threads = cpu_get_num_math();
+            cpuparams.n_threads = common_cpu_get_num_math();
         }
     }
 
@@ -1147,7 +1148,7 @@ common_init_result::common_init_result(common_params & params) :
 
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
-        llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
+        common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
@@ -1382,7 +1383,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
 common_init_result::~common_init_result() = default;
 
-std::string get_model_endpoint() {
+std::string common_get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
     const char * hf_endpoint_env = getenv("HF_ENDPOINT");
@@ -1397,6 +1398,42 @@ std::string get_model_endpoint() {
     return model_endpoint;
 }
 
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
+    auto * mem = llama_get_memory(ctx);
+    if (mem == nullptr) {
+        return COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+    }
+
+    common_context_seq_rm_type res = COMMON_CONTEXT_SEQ_RM_TYPE_PART;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx);
+
+    return res;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;
@@ -1484,7 +1521,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     return cparams;
 }
 
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params) {
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params) {
     struct ggml_threadpool_params tpp;
 
     ggml_threadpool_params_init(&tpp, params.n_threads); // setup the defaults

common/common.h

@@ -54,7 +54,7 @@ struct common_control_vector_load_info;
 // CPU utils
 //
 
-struct cpu_params {
+struct common_cpu_params {
     int      n_threads                   = -1;
     bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
     bool     mask_valid                  = false;   // Default: any CPU
@@ -63,8 +63,8 @@ struct cpu_params {
     uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
 };
 
-int32_t cpu_get_num_physical_cores();
-int32_t cpu_get_num_math();
+int32_t common_cpu_get_num_physical_cores();
+int32_t common_cpu_get_num_math();
 
 //
 // Common params
@@ -274,6 +274,7 @@ struct common_params_sampling {
     std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
     std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
     std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)
+    std::string              reasoning_budget_message;         // message injected before end tag when budget exhausted
 
     bool backend_sampling = false;
 
@@ -296,35 +297,19 @@ struct common_params_model {
 
 struct common_ngram_mod;
 
-struct common_params_speculative {
-    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE; // type of speculative decoding
+// draft-model-based speculative decoding parameters
+struct common_params_speculative_draft {
+    int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
+    int32_t n_min = 0;  // minimum number of draft tokens to use for speculative decoding
 
-    // general-purpose speculative decoding parameters
+    float p_split = 0.1f;  // speculative decoding split probability
+    float p_min   = 0.75f; // minimum speculative decoding probability (greedy)
 
-    int32_t n_max   = 16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min   = 0;  // minimum number of draft tokens to use for speculative decoding
-    float   p_split = 0.1f; // speculative decoding split probability
-    float   p_min   = 0.75f; // minimum speculative decoding probability (greedy)
+    common_params_model mparams;
 
-    // ngram-based speculative decoding
+    llama_model * model = nullptr; // a llama_model that can be shared by multiple speculative contexts
 
-    uint16_t ngram_size_n     = 12; // ngram size for lookup
-    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
-    bool     use_checkpoints  =  false; // use checkpoints to rewind in token history of recurrent models
-
-    std::shared_ptr<common_ngram_mod> ngram_mod;
-
-    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding          // NOLINT
-
-    // draft-model speculative decoding
-
-    struct common_params_model mparams_dft;
-
-    llama_model * model_dft = nullptr; // a llama_model that can be shared by multiple speculative contexts
-
-    llama_context_params cparams_dft; // these are the parameters for the draft llama_context
+    llama_context_params cparams; // these are the parameters for the draft llama_context
 
     int32_t n_ctx        = 0;  // draft context size
     int32_t n_gpu_layers = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
@@ -332,25 +317,60 @@ struct common_params_speculative {
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
 
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
+    common_cpu_params cpuparams;
+    common_cpu_params cpuparams_batch;
 
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
     std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
+};
+
+struct common_params_speculative_ngram_mod {
+    int32_t n_match = 24;
+
+    int32_t n_max = 64;
+    int32_t n_min = 48;
+
+    // shared instance of the ngram container for all speculative decoding contexts
+    std::shared_ptr<common_ngram_mod> obj;
+};
+
+struct common_params_speculative_ngram_map {
+    uint16_t size_n   = 12; // ngram size for lookup
+    uint16_t size_m   = 48; // mgram size for speculative tokens
+    uint16_t min_hits = 1;  // minimum hits at ngram/mgram lookup for mgram to be proposed
+};
+
+struct common_params_speculative_ngram_cache {
+    std::string lookup_cache_static;  // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic; // path of dynamic ngram cache file for lookup decoding
+};
+
+struct common_params_speculative {
+    // TODO: become a vector in order to support "chains of speculators"
+    common_speculative_type type = COMMON_SPECULATIVE_TYPE_NONE;
+
+    common_params_speculative_draft draft;
+
+    common_params_speculative_ngram_mod ngram_mod;
+    common_params_speculative_ngram_map ngram_simple;
+    common_params_speculative_ngram_map ngram_map_k;
+    common_params_speculative_ngram_map ngram_map_k4v;
+
+    common_params_speculative_ngram_cache ngram_cache;
 
     bool has_dft() const {
-        return !mparams_dft.path.empty() || !mparams_dft.hf_repo.empty();
+        return !draft.mparams.path.empty() || !draft.mparams.hf_repo.empty();
     }
 };
 
 struct common_params_vocoder {
     struct common_params_model model;
 
-    std::string speaker_file = ""; // speaker file path                                      // NOLINT
+    std::string speaker_file; // speaker file path
 
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy            // NOLINT
+    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy
 };
 
 struct common_params_diffusion {
@@ -421,19 +441,20 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
-    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
-    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
-    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;    // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;     // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};   // how split tensors should be distributed across GPUs
+    bool    fit_params         = true;  // whether to fit unset model/context parameters to free device memory
+    bool    fit_params_print   = false; // print the estimated required memory to run the model
+    int32_t fit_params_min_ctx = 4096;  // minimum context size to set when trying to reduce memory use
 
     // margin per device in bytes for fitting parameters to free memory:
     std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
+    common_cpu_params cpuparams;
+    common_cpu_params cpuparams_batch;
 
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
@@ -567,7 +588,7 @@ struct common_params {
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
     bool    cache_prompt        = true;  // whether to enable prompt caching
-    bool    clear_idle          = true;  // save and clear idle slots upon starting a new task
+    bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
     int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
@@ -581,8 +602,6 @@ struct common_params {
     bool force_pure_content_parser = false;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
-    int reasoning_budget = -1;
-    std::string reasoning_budget_message; // message injected before end tag when budget exhausted
     bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
     int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
@@ -679,7 +698,7 @@ std::string common_params_get_system_info(const common_params & params);
 
 bool parse_cpu_range(const std::string & range, bool(&boolmask)[GGML_MAX_N_THREADS]);
 bool parse_cpu_mask(const std::string & mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
-void postprocess_cpu_params(cpu_params & cpuparams, const cpu_params * role_model = nullptr);
+void postprocess_cpu_params(common_cpu_params & cpuparams, const common_cpu_params * role_model = nullptr);
 bool set_process_priority(enum ggml_sched_priority prio);
 
 //
@@ -747,6 +766,11 @@ inline bool string_starts_with(std::string_view str, std::string_view prefix) {
            str.compare(0, prefix.size(), prefix) == 0;
 }
 
+// remove when moving to c++20
+inline bool string_starts_with(std::string_view str, char prefix) {
+    return !str.empty() && str.front() == prefix;
+}
+
 // remove when moving to c++20
 inline bool string_ends_with(std::string_view str, std::string_view suffix) {
     return str.size() >= suffix.size() &&
@@ -842,12 +866,28 @@ common_init_result_ptr common_init_from_params(common_params & params);
 
 struct llama_model_params     common_model_params_to_llama  (      common_params & params);
 struct llama_context_params   common_context_params_to_llama(const common_params & params);
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const common_cpu_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
-std::string                   get_model_endpoint();
+// model endpoint from env
+std::string common_get_model_endpoint();
+
+//
+// Context utils
+//
+
+enum common_context_seq_rm_type {
+    COMMON_CONTEXT_SEQ_RM_TYPE_NO   = 0, // seq_rm not supported (e.g. no memory module)
+    COMMON_CONTEXT_SEQ_RM_TYPE_PART = 1, // can seq_rm partial sequences
+    COMMON_CONTEXT_SEQ_RM_TYPE_FULL = 2, // can seq_rm full sequences only
+};
+
+// check if the llama_context can remove sequences
+// note: clears the memory of the context
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
+
 
 //
 // Batch utils

common/debug.cpp

@@ -1,9 +1,38 @@
 #include "debug.h"
 
+#include "common.h"
 #include "log.h"
 
 #include <cmath>
+#include <regex>
 #include <string>
+#include <vector>
+
+struct common_debug_cb_user_data::impl {
+    std::vector<uint8_t>    data;
+    std::vector<std::regex> tensor_filters;
+    bool                    abort_on_nan{false};
+};
+
+common_debug_cb_user_data::common_debug_cb_user_data() : pimpl(std::make_unique<impl>()) {}
+common_debug_cb_user_data::~common_debug_cb_user_data() = default;
+
+common_debug_cb_user_data::common_debug_cb_user_data(common_params & params, const std::vector<std::string> & filter_patterns, bool abort_on_nan)
+    : pimpl(std::make_unique<impl>())
+{
+    for (const auto & pattern : filter_patterns) {
+        try {
+            std::string anchored_pattern = "^" + pattern;
+            pimpl->tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
+        } catch (const std::regex_error & e) {
+            throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
+        }
+    }
+    pimpl->abort_on_nan = abort_on_nan;
+
+    params.cb_eval           = common_debug_cb_eval;
+    params.cb_eval_user_data = this;
+}
 
 static std::string common_ggml_ne_string(const ggml_tensor * t) {
     std::string str;
@@ -47,8 +76,7 @@ static float common_ggml_get_float_value(const uint8_t * data,
 
 #define INDENT "    "
 
-template <bool abort>
-void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+static void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n, bool abort_on_nan) {
     GGML_ASSERT(n > 0);
     float sum = 0;
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
@@ -94,7 +122,7 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
         LOG(INDENT "sum = %f\n", sum);
     }
 
-    if constexpr (abort) {
+    if (abort_on_nan) {
         if (std::isnan(sum)) {
             LOG("encountered NaN - aborting\n");
             exit(0);
@@ -112,8 +140,9 @@ void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * n
  * @param user_data user data to pass at each call back
  * @return true to receive data or continue the graph, false otherwise
  */
-template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
-    auto * cb_data = (base_callback_data *) user_data;
+bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (common_debug_cb_user_data *) user_data;
+    auto * pimpl = cb_data->pimpl.get();
 
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
@@ -122,10 +151,10 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
         return true;  // Always retrieve data
     }
 
-    bool matches_filter = cb_data->tensor_filters.empty();
+    bool matches_filter = pimpl->tensor_filters.empty();
 
     if (!matches_filter) {
-        for (const auto & filter : cb_data->tensor_filters) {
+        for (const auto & filter : pimpl->tensor_filters) {
             if (std::regex_search(t->name, filter)) {
                 matches_filter = true;
                 break;
@@ -148,20 +177,14 @@ template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, b
 
     if (!is_host) {
         auto n_bytes = ggml_nbytes(t);
-        cb_data->data.resize(n_bytes);
-        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+        pimpl->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, pimpl->data.data(), 0, n_bytes);
     }
 
     if (!ggml_is_quantized(t->type) && matches_filter) {
-        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
-        common_debug_print_tensor<abort_on_nan>(data, t->type, t->ne, t->nb, 3);
+        uint8_t * data = is_host ? (uint8_t *) t->data : pimpl->data.data();
+        common_debug_print_tensor(data, t->type, t->ne, t->nb, 3, pimpl->abort_on_nan);
     }
 
     return true;
 }
-
-// Explicit template instantiations
-template bool common_debug_cb_eval<false>(ggml_tensor *, bool, void *);
-template bool common_debug_cb_eval<true>(ggml_tensor *, bool, void *);
-template void common_debug_print_tensor<false>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);
-template void common_debug_print_tensor<true>(uint8_t *, ggml_type, const int64_t *, const size_t *, int64_t);

common/debug.h

@@ -1,43 +1,31 @@
 #pragma once
-#include "common.h"
+
+#include <memory>
 #include <string>
 #include <vector>
-#include <regex>
 
 // common debug functions and structs
 
-// Print a tensor's detailed data
-// data - the tensor's data in byte format
-// type - the tensor's quantization type
-// ne   - the tensor dimensions array
-// nb   - the tensor strides array
-// n    - the number of rows/columns to fully print
-template <bool abort_on_nan> void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n);
+struct common_params;
 
 // Intended to use as callback for ggml_backend_sched_eval_callback
 // prints tensors that are processed in the computation graph
-// by default prints all tensors, but can be configured by creating a `base_callback_data` instance with
-// non-empty filter_patterns. See examples/debug.ccp for possible usage patterns
-// The template parameter determines whether an error should be thrown whenever a NaN is encountered
+// by default prints all tensors, but can be configured by creating a `common_debug_cb_user_data` instance with
+// non-empty filter_patterns. See examples/debug.cpp for possible usage patterns
+// `common_debug_cb_user_data` contains `abort_on_nan` flag that determines whether an error should be thrown whenever a NaN is encountered
 // in a tensor (useful for stopping debug sessions on first erroneous tensor)
 // The callback data will be passed as the third parameter (user_data)
-template <bool abort_on_nan> bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
-struct base_callback_data {
-    std::vector<uint8_t>    data;
-    std::vector<std::regex> tensor_filters;
+bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);
 
-    base_callback_data() = default;
+struct common_debug_cb_user_data {
+    struct impl;
+    std::unique_ptr<impl> pimpl;
 
-    base_callback_data(common_params & params, const std::vector<std::string> & filter_patterns) {
-        for (const auto & pattern : filter_patterns) {
-            try {
-                std::string anchored_pattern = "^" + pattern;
-                tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
-            } catch (const std::regex_error & e) {
-                throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
-            }
-        }
-        params.cb_eval           = common_debug_cb_eval<false>;
-        params.cb_eval_user_data = this;
-    }
+    common_debug_cb_user_data();
+    ~common_debug_cb_user_data();
+
+    common_debug_cb_user_data(const common_debug_cb_user_data &) = delete;
+    common_debug_cb_user_data & operator=(const common_debug_cb_user_data &) = delete;
+
+    common_debug_cb_user_data(common_params & params, const std::vector<std::string> & filter_patterns, bool abort_on_nan = false);
 };

common/download.cpp

@@ -627,7 +627,7 @@ static hf_cache::hf_file find_best_model(const hf_cache::hf_files & files,
     if (!tag.empty()) {
         tags.push_back(tag);
     } else {
-        tags = {"Q4_K_M", "Q4_0"};
+        tags = {"Q4_K_M", "Q8_0"};
     }
 
     for (const auto & t : tags) {

common/fit.cpp

@@ -0,0 +1,951 @@
+#include "fit.h"
+
+#include "log.h"
+
+#include "../src/llama-ext.h"
+
+#include <array>
+#include <cassert>
+#include <stdexcept>
+#include <cinttypes>
+#include <set>
+#include <string>
+#include <vector>
+
+// this enum is only used in llama_params_fit_impl but needs to be defined outside of it to fix a Windows compilation issue
+// enum to identify part of a layer for distributing its tensors:
+enum common_layer_fraction_t {
+    LAYER_FRACTION_NONE = 0, // nothing
+    LAYER_FRACTION_ATTN = 1, // attention
+    LAYER_FRACTION_UP   = 2, // attention + up
+    LAYER_FRACTION_GATE = 3, // attention + up + gate
+    LAYER_FRACTION_MOE  = 4, // everything but sparse MoE weights
+};
+
+class common_params_fit_exception : public std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
+static std::vector<llama_device_memory_data> common_get_device_memory_data(
+        const char * path_model,
+        const llama_model_params * mparams,
+        const llama_context_params * cparams,
+        std::vector<ggml_backend_dev_t> & devs,
+        uint32_t & hp_ngl,
+        uint32_t & hp_n_ctx_train,
+        uint32_t & hp_n_expert,
+        ggml_log_level log_level) {
+    struct user_data_t {
+        struct {
+            ggml_log_callback callback;
+            void * user_data;
+        } original_logger;
+        ggml_log_level min_level; // prints below this log level go to debug log
+    };
+    user_data_t ud;
+    llama_log_get(&ud.original_logger.callback, &ud.original_logger.user_data);
+    ud.min_level = log_level;
+
+    llama_log_set([](ggml_log_level level, const char * text, void * user_data) {
+        const user_data_t * ud = (const user_data_t *) user_data;
+        const ggml_log_level level_eff = level >= ud->min_level ? level : GGML_LOG_LEVEL_DEBUG;
+        ud->original_logger.callback(level_eff, text, ud->original_logger.user_data);
+    }, &ud);
+
+    llama_model_params mparams_copy = *mparams;
+    mparams_copy.no_alloc  = true;
+    mparams_copy.use_mmap  = false;
+    mparams_copy.use_mlock = false;
+
+    llama_model * model = llama_model_load_from_file(path_model, mparams_copy);
+    if (model == nullptr) {
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to load model");
+    }
+
+    llama_context * ctx = llama_init_from_model(model, *cparams);
+    if (ctx == nullptr) {
+        llama_model_free(model);
+        llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+        throw std::runtime_error("failed to create llama_context from model");
+    }
+
+    const size_t nd = llama_model_n_devices(model);
+    std::vector<llama_device_memory_data> ret(nd + 1);
+
+    llama_memory_breakdown memory_breakdown = llama_get_memory_breakdown(ctx);
+
+    for (const auto & [buft, mb] : memory_breakdown) {
+        if (ggml_backend_buft_is_host(buft)) {
+            ret.back().mb.model   += mb.model;
+            ret.back().mb.context += mb.context;
+            ret.back().mb.compute += mb.compute;
+            continue;
+        }
+
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            continue;
+        }
+        for (size_t i = 0; i < nd; i++) {
+            if (dev == llama_model_get_device(model, i)) {
+                ret[i].mb.model   += mb.model;
+                ret[i].mb.context += mb.context;
+                ret[i].mb.compute += mb.compute;
+                break;
+            }
+        }
+    }
+
+    {
+        ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (cpu_dev == nullptr) {
+            throw std::runtime_error("no CPU backend found");
+        }
+        size_t free;
+        size_t total;
+        ggml_backend_dev_memory(cpu_dev, &free, &total);
+        ret.back().free  = free;
+        ret.back().total = total;
+    }
+    for (size_t i = 0; i < nd; i++) {
+        size_t free;
+        size_t total;
+        ggml_backend_dev_memory(llama_model_get_device(model, i), &free, &total);
+
+        // devices can return 0 bytes for free and total memory if they do not
+        // have any to report. in this case, we will use the host memory as a fallback
+        // fixes: https://github.com/ggml-org/llama.cpp/issues/18577
+        if (free == 0 && total == 0) {
+            free  = ret.back().free;
+            total = ret.back().total;
+        }
+        ret[i].free  = free;
+        ret[i].total = total;
+    }
+
+    devs.clear();
+    for (int i = 0; i < llama_model_n_devices(model); i++) {
+        devs.push_back(llama_model_get_device(model, i));
+    }
+
+    hp_ngl         = llama_model_n_layer(model);
+    hp_n_ctx_train = llama_model_n_ctx_train(model);
+    hp_n_expert    = llama_model_n_expert(model);
+
+    common_memory_breakdown_print(ctx);
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_log_set(ud.original_logger.callback, ud.original_logger.user_data);
+
+    return ret;
+}
+
+static void common_params_fit_impl(
+        const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
+        float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+    if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
+        throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
+    }
+    constexpr int64_t MiB = 1024*1024;
+    typedef std::vector<llama_device_memory_data> dmds_t;
+    const llama_model_params default_mparams = llama_model_default_params();
+
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    // step 1: get data for default parameters and check whether any changes are necessary in the first place
+
+    LOG_INF("%s: getting device memory data for initial parameters:\n", __func__);
+    const dmds_t dmds_full = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    const size_t nd = devs.size(); // number of devices
+
+    std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
+    margins.reserve(nd);
+    if (nd == 0) {
+        margins.push_back(margins_s[0]);
+    } else {
+        for (size_t id = 0; id < nd; id++) {
+            margins.push_back(margins_s[id]);
+        }
+    }
+
+    std::vector<std::string> dev_names;
+    {
+        dev_names.reserve(nd);
+        size_t max_length = 0;
+        for (const auto & dev : devs) {
+            std::string name = ggml_backend_dev_name(dev);
+            name += " (";
+            name += ggml_backend_dev_description(dev);
+            name += ")";
+            dev_names.push_back(name);
+            max_length = std::max(max_length, name.length());
+        }
+        for (std::string & dn : dev_names) {
+            dn.insert(dn.end(), max_length - dn.length(), ' ');
+        }
+    }
+
+    int64_t sum_free            = 0;
+    int64_t sum_projected_free  = 0;
+    int64_t sum_projected_used  = 0;
+    int64_t sum_projected_model = 0;
+    std::vector<int64_t> projected_free_per_device;
+    projected_free_per_device.reserve(nd);
+
+    if (nd == 0) {
+        sum_projected_used = dmds_full.back().mb.total();
+        sum_free           = dmds_full.back().total;
+        sum_projected_free = sum_free - sum_projected_used;
+        LOG_INF("%s: projected to use %" PRId64 " MiB of host memory vs. %" PRId64 " MiB of total host memory\n",
+            __func__, sum_projected_used/MiB, sum_free/MiB);
+        if (sum_projected_free >= margins[0]) {
+            LOG_INF("%s: will leave %" PRId64 " >= %" PRId64 " MiB of system memory, no changes needed\n",
+                __func__, sum_projected_free/MiB, margins[0]/MiB);
+            return;
+        }
+    } else {
+        if (nd > 1) {
+            LOG_INF("%s: projected memory use with initial parameters [MiB]:\n", __func__);
+        }
+        for (size_t id = 0; id < nd; id++) {
+            const llama_device_memory_data & dmd = dmds_full[id];
+
+            const int64_t projected_used = dmd.mb.total();
+            const int64_t projected_free = dmd.free - projected_used;
+            projected_free_per_device.push_back(projected_free);
+
+            sum_free            += dmd.free;
+            sum_projected_used  += projected_used;
+            sum_projected_free  += projected_free;
+            sum_projected_model += dmd.mb.model;
+
+            if (nd > 1) {
+                LOG_INF("%s:   - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " free vs. target of %6" PRId64 "\n",
+                    __func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, projected_free/MiB, margins[id]/MiB);
+            }
+        }
+        assert(sum_free >= 0 && sum_projected_used >= 0);
+        LOG_INF("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
+            __func__, sum_projected_used/MiB, sum_free/MiB);
+        if (nd == 1) {
+            if (projected_free_per_device[0] >= margins[0]) {
+                LOG_INF("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
+                    __func__, projected_free_per_device[0]/MiB, margins[0]/MiB);
+                return;
+            }
+        } else {
+            bool changes_needed = false;
+            for (size_t id = 0; id < nd; id++) {
+                if (projected_free_per_device[id] < margins[id]) {
+                    changes_needed = true;
+                    break;
+                }
+            }
+            if (!changes_needed) {
+                LOG_INF("%s: targets for free memory can be met on all devices, no changes needed\n", __func__);
+                return;
+            }
+        }
+    }
+
+    // step 2: try reducing memory use by reducing the context size
+
+    {
+        int64_t global_surplus = sum_projected_free;
+        if (nd == 0) {
+            global_surplus -= margins[0];
+        } else {
+            for (size_t id = 0; id < nd; id++) {
+                global_surplus -= margins[id];
+            }
+        }
+        if (global_surplus < 0) {
+            if (nd <= 1) {
+                LOG_INF("%s: cannot meet free memory target of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n",
+                    __func__, margins[0]/MiB, -global_surplus/MiB);
+            } else {
+                LOG_INF(
+                    "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
+                    __func__, -global_surplus/MiB);
+            }
+            if (cparams->n_ctx == 0) {
+                if (hp_nct > n_ctx_min) {
+                    int64_t sum_used_target = sum_free;
+                    if (nd == 0) {
+                        sum_used_target -= margins[0];
+                    } else {
+                        for (size_t id = 0; id < nd; id++) {
+                            sum_used_target -= margins[id];
+                        }
+                    }
+                    if (nd > 1) {
+                        // for multiple devices we need to be more conservative in terms of how much context we think can fit:
+                        //   - for dense models only whole layers can be assigned to devices
+                        //   - for MoE models only whole tensors can be assigned to devices, which we estimate to be <= 1/3 of a layer
+                        //   - on average we expect a waste of 0.5 layers/tensors per device
+                        //   - use slightly more than the expected average for nd devices to be safe
+                        const int64_t model_per_layer = sum_projected_model / std::min(uint32_t(mparams->n_gpu_layers), hp_ngl);
+                        sum_used_target -= (nd + 1) * model_per_layer / (hp_nex == 0 ? 2 : 6);
+                    }
+
+                    int64_t sum_projected_used_min_ctx = 0;
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    if (nd == 0) {
+                        sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
+                    } else {
+                        for (size_t id = 0; id < nd; id++) {
+                            sum_projected_used_min_ctx += dmds_min_ctx[id].mb.total();
+                        }
+                    }
+                    if (sum_used_target > sum_projected_used_min_ctx) {
+                        // linear interpolation between minimum and maximum context size:
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
+                            / (sum_projected_used - sum_projected_used_min_ctx);
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
+                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
+                        LOG_INF("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                        if (nd <= 1) {
+                            LOG_INF("%s: entire model can be fit by reducing context\n", __func__);
+                            return;
+                        }
+                        LOG_INF("%s: entire model should be fit across devices by reducing context\n", __func__);
+                    } else {
+                        const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
+                        LOG_INF("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                    }
+                } else {
+                    if (n_ctx_min == UINT32_MAX) {
+                        LOG_INF("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
+                    } else {
+                        LOG_INF("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
+                            __func__, hp_nct, n_ctx_min);
+                    }
+                }
+            } else {
+                LOG_INF("%s: context size set by user to %" PRIu32 " -> no change\n", __func__, cparams->n_ctx);
+            }
+        }
+    }
+    if (nd == 0) {
+        throw common_params_fit_exception("was unable to fit model into system memory by reducing context, abort");
+    }
+
+    if (mparams->n_gpu_layers != default_mparams.n_gpu_layers) {
+        throw common_params_fit_exception("n_gpu_layers already set by user to " + std::to_string(mparams->n_gpu_layers) + ", abort");
+    }
+    if (nd > 1) {
+        if (!tensor_split) {
+            throw common_params_fit_exception("did not provide a buffer to write the tensor_split to, abort");
+        }
+        if (mparams->tensor_split) {
+            for (size_t id = 0; id < nd; id++) {
+                if (mparams->tensor_split[id] != 0.0f) {
+                    throw common_params_fit_exception("model_params::tensor_split already set by user, abort");
+                }
+            }
+        }
+        if (mparams->split_mode == LLAMA_SPLIT_MODE_ROW) {
+            throw common_params_fit_exception("changing weight allocation for LLAMA_SPLIT_MODE_ROW not implemented, abort");
+        }
+    }
+    if (!tensor_buft_overrides) {
+        throw common_params_fit_exception("did not provide buffer to set tensor_buft_overrides, abort");
+    }
+    if (mparams->tensor_buft_overrides && (mparams->tensor_buft_overrides->pattern || mparams->tensor_buft_overrides->buft)) {
+        throw common_params_fit_exception("model_params::tensor_buft_overrides already set by user, abort");
+    }
+
+    // step 3: iteratively fill the back to front with "dense" layers
+    //   - for a dense model simply fill full layers, giving each device a contiguous slice of the model
+    //   - for a MoE model, same as dense model but with all MoE tensors in system memory
+
+    // utility function that returns a static C string matching the tensors for a specific layer index and layer fraction:
+    auto get_overflow_pattern = [&](const size_t il, const common_layer_fraction_t lf) -> const char * {
+        constexpr size_t n_strings = 1000;
+        if (il >= n_strings) {
+            throw std::runtime_error("at most " + std::to_string(n_strings) + " model layers are supported");
+        }
+        switch (lf) {
+            case LAYER_FRACTION_ATTN: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|up|gate_up|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_UP: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(gate|gate_up|down).*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_GATE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_down.*";
+                }
+                return patterns[il].c_str();
+            }
+            case LAYER_FRACTION_MOE: {
+                static std::array<std::string, n_strings> patterns;
+                if (patterns[il].empty()) {
+                    patterns[il] = "blk\\." + std::to_string(il) + "\\.ffn_(up|down|gate_up|gate)_(ch|)exps";
+                }
+                return patterns[il].c_str();
+            }
+            default:
+                GGML_ABORT("fatal error");
+        }
+    };
+
+    struct ngl_t {
+        uint32_t n_layer = 0; // number of total layers
+        uint32_t n_part  = 0; // number of partial layers, <= n_layer
+
+        // for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
+        common_layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
+
+        uint32_t n_full() const {
+            assert(n_layer >= n_part);
+            return n_layer - n_part;
+        }
+    };
+
+    const size_t ntbo = llama_max_tensor_buft_overrides();
+
+    // utility function to set n_gpu_layers and tensor_split
+    auto set_ngl_tensor_split_tbo = [&](
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts,
+            llama_model_params & mparams) {
+        mparams.n_gpu_layers = 0;
+        for (size_t id = 0; id < nd; id++) {
+            mparams.n_gpu_layers += ngl_per_device[id].n_layer;
+            if (nd > 1) {
+                tensor_split[id] = ngl_per_device[id].n_layer;
+            }
+        }
+        assert(uint32_t(mparams.n_gpu_layers) <= hp_ngl + 1);
+        uint32_t il0 = hp_ngl + 1 - mparams.n_gpu_layers; // start index for tensor buft overrides
+
+        mparams.tensor_split = tensor_split;
+
+        size_t itbo = 0;
+        for (size_t id = 0; id < nd; id++) {
+            il0 += ngl_per_device[id].n_full();
+            for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
+                if (itbo + 1 >= ntbo) {
+                    tensor_buft_overrides[itbo].pattern = nullptr;
+                    tensor_buft_overrides[itbo].buft    = nullptr;
+                    itbo++;
+                    mparams.tensor_buft_overrides = tensor_buft_overrides;
+                    throw common_params_fit_exception("llama_max_tensor_buft_overrides() == "
+                        + std::to_string(ntbo) + " is insufficient for model");
+                }
+                tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
+                tensor_buft_overrides[itbo].buft = il == il0 ? overflow_bufts[id] : ggml_backend_cpu_buffer_type();
+                itbo++;
+            }
+            il0 += ngl_per_device[id].n_part;
+        }
+        tensor_buft_overrides[itbo].pattern = nullptr;
+        tensor_buft_overrides[itbo].buft    = nullptr;
+        itbo++;
+        mparams.tensor_buft_overrides = tensor_buft_overrides;
+    };
+
+    // utility function that returns the memory use per device for given numbers of layers per device
+    auto get_memory_for_layers = [&](
+            const char * func_name,
+            const std::vector<ngl_t> & ngl_per_device,
+            const std::vector<ggml_backend_buffer_type_t> & overflow_bufts) -> std::vector<int64_t> {
+        llama_model_params mparams_copy = *mparams;
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
+
+        const dmds_t dmd_nl = common_get_device_memory_data(
+            path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        LOG_INF("%s: memory for test allocation by device:\n", func_name);
+        for (size_t id = 0; id < nd; id++) {
+            const ngl_t & n = ngl_per_device[id];
+            LOG_INF(
+                "%s: id=%zu, n_layer=%2" PRIu32 ", n_part=%2" PRIu32 ", overflow_type=%d, mem=%6" PRId64 " MiB\n",
+                func_name, id, n.n_layer, n.n_part, int(n.overflow_type), dmd_nl[id].mb.total()/MiB);
+        }
+
+        std::vector<int64_t> ret;
+        ret.reserve(nd);
+        for (size_t id = 0; id < nd; id++) {
+            ret.push_back(dmd_nl[id].mb.total());
+        }
+        return ret;
+    };
+
+    int64_t global_surplus_cpu_moe = 0;
+    if (hp_nex > 0) {
+        const static std::string pattern_moe_all = "blk\\.\\d+\\.ffn_(up|down|gate_up|gate)_(ch|)exps"; // matches all MoE tensors
+        ggml_backend_buffer_type_t cpu_buft = ggml_backend_cpu_buffer_type();
+        tensor_buft_overrides[0] = {pattern_moe_all.c_str(), cpu_buft};
+        tensor_buft_overrides[1] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+
+        LOG_INF("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data(
+            path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+        for (size_t id = 0; id < nd; id++) {
+            global_surplus_cpu_moe += dmds_cpu_moe[id].free;
+            global_surplus_cpu_moe -= int64_t(dmds_cpu_moe[id].mb.total()) + margins[id];
+        }
+
+        if (global_surplus_cpu_moe > 0) {
+            LOG_INF("%s: with only dense weights in device memory there is a total surplus of %" PRId64 " MiB\n",
+                __func__, global_surplus_cpu_moe/MiB);
+        } else {
+            LOG_INF("%s: with only dense weights in device memory there is still a total deficit of %" PRId64 " MiB\n",
+                __func__, -global_surplus_cpu_moe/MiB);
+        }
+
+        // reset
+        tensor_buft_overrides[0] = {nullptr, nullptr};
+        mparams->tensor_buft_overrides = tensor_buft_overrides;
+    }
+
+    std::vector<int64_t> targets; // maximum acceptable memory use per device
+    targets.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        targets.push_back(dmds_full[id].free - margins[id]);
+        LOG_INF("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
+    }
+
+    std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the first partial layer of a device overflows to:
+    overflow_bufts.reserve(nd);
+    for (size_t id = 0; id < nd; id++) {
+        overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
+    }
+
+    std::vector<ngl_t> ngl_per_device(nd);
+    std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts);
+
+    // optimize the number of layers per device using the method of false position:
+    //   - ngl_per_device has 0 layers for each device, lower bound
+    //   - try a "high" configuration where a device is given all unassigned layers
+    //   - interpolate the memory use / layer between low and high linearly to get a guess where it meets our target
+    //   - check memory use of our guess, replace either the low or high bound
+    //   - once we only have a difference of a single layer, stop and return the lower bound that just barely still fits
+    //   - the last device has the output layer, which cannot be a partial layer
+    if (hp_nex == 0) {
+        LOG_INF("%s: filling dense layers back-to-front:\n", __func__);
+    } else {
+        LOG_INF("%s: filling dense-only layers back-to-front:\n", __func__);
+    }
+    for (int id = nd - 1; id >= 0; id--) {
+        uint32_t n_unassigned = hp_ngl + 1;
+        for (size_t jd = id + 1; jd < nd; ++jd) {
+            assert(n_unassigned >= ngl_per_device[jd].n_layer);
+            n_unassigned -= ngl_per_device[jd].n_layer;
+        }
+
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        ngl_per_device_high[id].n_layer = n_unassigned;
+        if (hp_nex > 0) {
+            ngl_per_device_high[id].n_part = size_t(id) < nd - 1 ? ngl_per_device_high[id].n_layer : ngl_per_device_high[id].n_layer - 1;
+        }
+        if (ngl_per_device_high[id].n_layer > 0) {
+            std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
+            if (mem_high[id] > targets[id]) {
+                assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
+                uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                LOG_INF("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
+                while (delta > 1) {
+                    uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                    step_size = std::max(step_size, uint32_t(1));
+                    step_size = std::min(step_size, delta - 1);
+
+                    std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                    ngl_per_device_test[id].n_layer += step_size;
+                    if (hp_nex) {
+                        ngl_per_device_test[id].n_part += size_t(id) == nd - 1 && ngl_per_device_test[id].n_part == 0 ?
+                            step_size - 1 : step_size; // the first layer is the output layer which must always be full
+                    }
+                    const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
+
+                    if (mem_test[id] <= targets[id]) {
+                        ngl_per_device = ngl_per_device_test;
+                        mem            = mem_test;
+                        LOG_INF("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+                    } else {
+                        ngl_per_device_high = ngl_per_device_test;
+                        mem_high            = mem_test;
+                        LOG_INF("%s: set ngl_per_device_high[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device_high[id].n_layer);
+                    }
+                    delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                }
+            } else {
+                assert(ngl_per_device_high[id].n_layer == n_unassigned);
+                ngl_per_device = ngl_per_device_high;
+                mem            = mem_high;
+                LOG_INF("%s: set ngl_per_device[%d].n_layer=%" PRIu32 "\n", __func__, id, ngl_per_device[id].n_layer);
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers, %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, mem[id]/MiB, projected_margin/MiB);
+    }
+    if (hp_nex == 0 || global_surplus_cpu_moe <= 0) {
+        set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
+        return;
+    }
+
+    // step 4: for a MoE model where all dense tensors fit,
+    //     convert the dense-only layers in the back to full layers in the front until all devices are full
+    // essentially the same procedure as for the dense-only layers except front-to-back
+    // also, try fitting at least part of one more layer to reduce waste for "small" GPUs with e.g. 24 GiB VRAM
+
+    size_t id_dense_start = nd;
+    for (int id = nd - 1; id >= 0; id--) {
+        if (ngl_per_device[id].n_layer > 0) {
+            id_dense_start = id;
+            continue;
+        }
+        break;
+    }
+    assert(id_dense_start < nd);
+
+    LOG_INF("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
+    for (size_t id = 0; id <= id_dense_start && id_dense_start < nd; id++) {
+        std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
+        for (size_t jd = id_dense_start; jd < nd; jd++) {
+            const uint32_t n_layer_move = jd < nd - 1 ? ngl_per_device_high[jd].n_layer : ngl_per_device_high[jd].n_layer - 1;
+            ngl_per_device_high[id].n_layer += n_layer_move;
+            ngl_per_device_high[jd].n_layer -= n_layer_move;
+            ngl_per_device_high[jd].n_part = 0;
+        }
+        size_t id_dense_start_high = nd - 1;
+        std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
+
+        if (mem_high[id] > targets[id]) {
+            assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+            uint32_t delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
+            while (delta > 1) {
+                uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
+                step_size = std::max(step_size, uint32_t(1));
+                step_size = std::min(step_size, delta - 1);
+
+                std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+                size_t id_dense_start_test = id_dense_start;
+                uint32_t n_converted_test = 0;
+                for (;id_dense_start_test < nd; id_dense_start_test++) {
+                    const uint32_t n_convert_jd = std::min(step_size - n_converted_test, ngl_per_device_test[id_dense_start_test].n_part);
+                    ngl_per_device_test[id_dense_start_test].n_layer -= n_convert_jd;
+                    ngl_per_device_test[id_dense_start_test].n_part -= n_convert_jd;
+                    ngl_per_device_test[id].n_layer += n_convert_jd;
+                    n_converted_test += n_convert_jd;
+
+                    if (ngl_per_device_test[id_dense_start_test].n_part > 0) {
+                        break;
+                    }
+                }
+                const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
+
+                if (mem_test[id] <= targets[id]) {
+                    ngl_per_device = ngl_per_device_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                } else {
+                    ngl_per_device_high = ngl_per_device_test;
+                    mem_high            = mem_test;
+                    id_dense_start_high = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
+                        __func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
+                }
+                assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
+                delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
+            }
+        } else {
+            ngl_per_device = ngl_per_device_high;
+            mem            = mem_high;
+            id_dense_start = id_dense_start_high;
+            LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start=%zu\n",
+                __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+        }
+
+        // try to fit at least part of one more layer
+        if (ngl_per_device[id_dense_start].n_layer > (id < nd - 1 ? 0 : 1)) {
+            std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
+            size_t id_dense_start_test = id_dense_start;
+            ngl_per_device_test[id_dense_start_test].n_layer--;
+            ngl_per_device_test[id_dense_start_test].n_part--;
+            ngl_per_device_test[id].n_layer++;
+            ngl_per_device_test[id].n_part++;
+            if (ngl_per_device_test[id_dense_start_test].n_part == 0) {
+                id_dense_start_test++;
+            }
+            ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
+            std::vector<ggml_backend_buffer_type_t> overflow_bufts_test = overflow_bufts;
+            if (id < nd - 1) {
+                overflow_bufts_test[id] = ggml_backend_dev_buffer_type(devs[id + 1]);
+            }
+            LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
+            std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+            if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                ngl_per_device = ngl_per_device_test;
+                overflow_bufts = overflow_bufts_test;
+                mem            = mem_test;
+                id_dense_start = id_dense_start_test;
+                LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
+                    __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
+                LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                    ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            } else {
+                ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
+                LOG_INF("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
+                mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
+                if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
+                    ngl_per_device = ngl_per_device_test;
+                    overflow_bufts = overflow_bufts_test;
+                    mem            = mem_test;
+                    id_dense_start = id_dense_start_test;
+                    LOG_INF("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
+                        __func__, id, ngl_per_device[id].n_layer, ngl_per_device[id].n_part, id_dense_start);
+                }
+            }
+        }
+
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    // print info for devices that were not changed during the conversion from dense only to full layers:
+    for (size_t id = id_dense_start + 1; id < nd; id++) {
+        const int64_t projected_margin = dmds_full[id].free - mem[id];
+        LOG_INF(
+            "%s:   - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
+            __func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
+    }
+
+    set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
+}
+
+enum common_params_fit_status common_fit_params(
+        const char * path_model,
+        llama_model_params * mparams,
+        llama_context_params * cparams,
+        float * tensor_split,
+        llama_model_tensor_buft_override * tensor_buft_overrides,
+        size_t * margins,
+        uint32_t n_ctx_min,
+        ggml_log_level log_level) {
+    const int64_t t0_us = llama_time_us();
+    common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
+    try {
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
+        LOG_INF("%s: successfully fit params to free device memory\n", __func__);
+    } catch (const common_params_fit_exception & e) {
+        LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
+        status = COMMON_PARAMS_FIT_STATUS_FAILURE;
+    } catch (const std::runtime_error & e) {
+        LOG_ERR("%s: encountered an error while trying to fit params to free device memory: %s\n", __func__, e.what());
+        status = COMMON_PARAMS_FIT_STATUS_ERROR;
+    }
+    const int64_t t1_us = llama_time_us();
+    LOG_INF("%s: fitting params to free memory took %.2f seconds\n", __func__, (t1_us - t0_us) * 1e-6);
+    return status;
+}
+
+void common_memory_breakdown_print(const struct llama_context * ctx) {
+    //const auto & devices = ctx->get_model().devices;
+    const auto * model = llama_get_model(ctx);
+
+    std::vector<ggml_backend_dev_t> devices;
+    for (int i = 0; i < llama_model_n_devices(model); i++) {
+        devices.push_back(llama_model_get_device(model, i));
+    }
+
+    llama_memory_breakdown memory_breakdown = llama_get_memory_breakdown(ctx);
+
+    std::vector<std::array<std::string, 9>> table_data;
+    table_data.reserve(devices.size());
+    const std::string template_header = "%s: | %s | %s   %s    %s   %s   %s   %s    %s |\n";
+    const std::string template_gpu    = "%s: | %s | %s = %s + (%s = %s + %s + %s) + %s |\n";
+    const std::string template_other  = "%s: | %s | %s   %s    %s = %s + %s + %s    %s |\n";
+
+    table_data.push_back({template_header, "memory breakdown [MiB]", "total", "free", "self", "model", "context", "compute", "unaccounted"});
+
+    constexpr size_t MiB = 1024 * 1024;
+    const std::vector<std::string> desc_prefixes_strip = {"NVIDIA ", "GeForce ", "Tesla ", "AMD ", "Radeon ", "Instinct "};
+
+    // track seen buffer types to avoid double counting:
+    std::set<ggml_backend_buffer_type_t> seen_buffer_types;
+
+    // accumulative memory breakdown for each device and for host:
+    std::vector<llama_memory_breakdown_data> mb_dev(devices.size());
+    llama_memory_breakdown_data              mb_host;
+
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (ggml_backend_buft_is_host(buft)) {
+            mb_host.model   += mb.model;
+            mb_host.context += mb.context;
+            mb_host.compute += mb.compute;
+            seen_buffer_types.insert(buft);
+            continue;
+        }
+        ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (dev) {
+            int i_dev = -1;
+            for (size_t i = 0; i < devices.size(); i++) {
+                if (devices[i] == dev) {
+                    i_dev = i;
+                    break;
+                }
+            }
+            if (i_dev != -1) {
+                mb_dev[i_dev].model   += mb.model;
+                mb_dev[i_dev].context += mb.context;
+                mb_dev[i_dev].compute += mb.compute;
+                seen_buffer_types.insert(buft);
+                continue;
+            }
+        }
+    }
+
+    // print memory breakdown for each device:
+    for (size_t i = 0; i < devices.size(); i++) {
+        ggml_backend_dev_t dev = devices[i];
+        llama_memory_breakdown_data mb = mb_dev[i];
+
+        const std::string name = ggml_backend_dev_name(dev);
+        std::string desc = ggml_backend_dev_description(dev);
+        for (const std::string & prefix : desc_prefixes_strip) {
+            if (desc.length() >= prefix.length() && desc.substr(0, prefix.length()) == prefix) {
+                desc = desc.substr(prefix.length());
+            }
+        }
+
+        size_t free, total;
+        ggml_backend_dev_memory(dev, &free, &total);
+
+        const size_t self = mb.model + mb.context + mb.compute;
+        const int64_t unaccounted = static_cast<int64_t>(total) - static_cast<int64_t>(free) - static_cast<int64_t>(self);
+
+        table_data.push_back({
+            template_gpu,
+            "  - " + name + " (" + desc + ")",
+            std::to_string(total / MiB),
+            std::to_string(free / MiB),
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            std::to_string(unaccounted / static_cast<int64_t>(MiB))});
+    }
+
+    // print memory breakdown for host:
+    {
+        const size_t self = mb_host.model + mb_host.context + mb_host.compute;
+        table_data.push_back({
+            template_other,
+            "  - Host",
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb_host.model / MiB),
+            std::to_string(mb_host.context / MiB),
+            std::to_string(mb_host.compute / MiB),
+            ""}); // unaccounted
+    }
+
+    // print memory breakdown for all remaining buffer types:
+    for (const auto & buft_mb : memory_breakdown) {
+        ggml_backend_buffer_type_t          buft = buft_mb.first;
+        const llama_memory_breakdown_data & mb   = buft_mb.second;
+        if (seen_buffer_types.count(buft) == 1) {
+            continue;
+        }
+        const std::string name = ggml_backend_buft_name(buft);
+        const size_t self = mb.model + mb.context + mb.compute;
+        table_data.push_back({
+            template_other,
+            "  - " + name,
+            "", // total
+            "", // free
+            std::to_string(self / MiB),
+            std::to_string(mb.model / MiB),
+            std::to_string(mb.context / MiB),
+            std::to_string(mb.compute / MiB),
+            ""}); // unaccounted
+        seen_buffer_types.insert(buft);
+    }
+
+    for (size_t j = 1; j < table_data[0].size(); j++) {
+        size_t max_len = 0;
+        for (const auto & td : table_data) {
+            max_len = std::max(max_len, td[j].length());
+        }
+        for (auto & td : table_data) {
+            td[j].insert(j == 1 ? td[j].length() : 0, max_len - td[j].length(), ' ');
+        }
+    }
+    for (const auto & td : table_data) {
+        LOG_INF(td[0].c_str(),
+            __func__, td[1].c_str(), td[2].c_str(), td[3].c_str(), td[4].c_str(), td[5].c_str(),
+            td[6].c_str(), td[7].c_str(), td[8].c_str());
+    }
+}
+
+void common_fit_print(
+        const char * path_model,
+        llama_model_params * mparams,
+        llama_context_params * cparams) {
+    std::vector<ggml_backend_dev_t> devs;
+    uint32_t hp_ngl = 0; // hparams.n_gpu_layers
+    uint32_t hp_nct = 0; // hparams.n_ctx_train
+    uint32_t hp_nex = 0; // hparams.n_expert
+
+    auto dmd = common_get_device_memory_data(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, GGML_LOG_LEVEL_ERROR);
+    GGML_ASSERT(dmd.size() == devs.size() + 1);
+
+    for (size_t id = 0; id < devs.size(); id++) {
+        printf("%s ",  ggml_backend_dev_name(devs[id]));
+        printf("%zu ", dmd[id].mb.model/1024/1024);
+        printf("%zu ", dmd[id].mb.context/1024/1024);
+        printf("%zu ", dmd[id].mb.compute/1024/1024);
+        printf("\n");
+    }
+
+    printf("Host ");
+    printf("%zu ", dmd.back().mb.model/1024/1024);
+    printf("%zu ", dmd.back().mb.context/1024/1024);
+    printf("%zu ", dmd.back().mb.compute/1024/1024);
+    printf("\n");
+}

common/fit.h

@@ -0,0 +1,32 @@
+#pragma once
+
+#include "ggml.h"
+
+enum common_params_fit_status {
+    COMMON_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit
+    COMMON_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit
+    COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
+};
+
+// fits mparams and cparams to free device memory (assumes system memory is unlimited)
+//   - returns true if the parameters could be successfully modified to fit device memory
+//   - this function is NOT thread safe because it modifies the global llama logger state
+//   - only parameters that have the same value as in llama_default_model_params are modified
+//     with the exception of the context size which is modified if and only if equal to 0
+enum common_params_fit_status common_fit_params(
+                               const char   * path_model,
+                struct llama_model_params   * mparams,
+                struct llama_context_params * cparams,
+                                      float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
+    struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
+                                     size_t * margins,               // margins of memory to leave per device in bytes
+                                   uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+                        enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
+
+// print estimated memory to stdout
+void common_fit_print(
+                               const char   * path_model,
+                struct llama_model_params   * mparams,
+                struct llama_context_params * cparams);
+
+void common_memory_breakdown_print(const struct llama_context * ctx);

common/hf-cache.cpp

@@ -230,7 +230,7 @@ static nl::json api_get(const std::string & url,
 static std::string get_repo_commit(const std::string & repo_id,
                                    const std::string & token) {
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/refs", token);
 
         if (!json.is_object() ||
@@ -308,7 +308,7 @@ hf_files get_repo_files(const std::string & repo_id,
     hf_files files;
 
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/tree/" + commit + "?recursive=true", token);
 
         if (!json.is_array()) {

common/jinja/caps.cpp

@@ -1,4 +1,3 @@
-#include "log.h"
 #include "value.h"
 #include "runtime.h"
 #include "caps.h"

common/jinja/runtime.h

@@ -106,10 +106,16 @@ struct statement {
     size_t pos; // position in source, for debugging
     virtual ~statement() = default;
     virtual std::string type() const { return "Statement"; }
+
     // execute_impl must be overridden by derived classes
-    virtual value execute_impl(context &) { throw std::runtime_error("cannot exec " + type()); }
+    virtual value execute_impl(context &) { throw_exec_error(); }
     // execute is the public method to execute a statement with error handling
     value execute(context &);
+
+private:
+    [[noreturn]] void throw_exec_error() const {
+        throw std::runtime_error("cannot exec " + type());
+    }
 };
 
 // Type Checking Utilities
@@ -143,7 +149,7 @@ struct program : public statement {
     program() = default;
     explicit program(statements && body) : body(std::move(body)) {}
     std::string type() const override { return "Program"; }
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw std::runtime_error("Cannot execute program directly, use jinja::runtime instead");
     }
 };
@@ -195,7 +201,7 @@ struct break_statement : public statement {
         }
     };
 
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw break_statement::signal();
     }
 };
@@ -209,7 +215,7 @@ struct continue_statement : public statement {
         }
     };
 
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw continue_statement::signal();
     }
 };
@@ -509,7 +515,7 @@ struct slice_expression : public expression {
         chk_type<expression>(this->step_expr);
     }
     std::string type() const override { return "SliceExpression"; }
-    value execute_impl(context &) override {
+    [[noreturn]] value execute_impl(context &) override {
         throw std::runtime_error("must be handled by MemberExpression");
     }
 };

common/jinja/value.cpp

@@ -590,6 +590,10 @@ static bool string_endswith(const std::string & str, const std::string & suffix)
     return str.compare(str.length() - suffix.length(), suffix.length(), suffix) == 0;
 }
 
+[[noreturn]] static value string_join_not_implemented(const func_args &) {
+    throw not_implemented_exception("String join builtin not implemented");
+}
+
 const func_builtins & value_string_t::get_builtins() const {
     static const func_builtins builtins = {
         {"default", default_value},
@@ -851,9 +855,7 @@ const func_builtins & value_string_t::get_builtins() const {
             res->val_str.mark_input_based_on(val_input->as_string());
             return res;
         }},
-        {"join", [](const func_args &) -> value {
-            throw not_implemented_exception("String join builtin not implemented");
-        }},
+        {"join", string_join_not_implemented},
     };
     return builtins;
 }
@@ -884,6 +886,9 @@ const func_builtins & value_bool_t::get_builtins() const {
     return builtins;
 }
 
+[[noreturn]] static value array_unique_not_implemented(const func_args &) {
+    throw not_implemented_exception("Array unique builtin not implemented");
+}
 
 const func_builtins & value_array_t::get_builtins() const {
     static const func_builtins builtins = {
@@ -1084,13 +1089,14 @@ const func_builtins & value_array_t::get_builtins() const {
             std::reverse(arr.begin(), arr.end());
             return is_val<value_tuple>(val) ? mk_val<value_tuple>(std::move(arr)) : mk_val<value_array>(std::move(arr));
         }},
-        {"unique", [](const func_args &) -> value {
-            throw not_implemented_exception("Array unique builtin not implemented");
-        }},
+        {"unique", array_unique_not_implemented},
     };
     return builtins;
 }
 
+[[noreturn]] static value object_join_not_implemented(const func_args &) {
+    throw not_implemented_exception("object join not implemented");
+}
 
 const func_builtins & value_object_t::get_builtins() const {
     if (!has_builtins) {
@@ -1183,9 +1189,7 @@ const func_builtins & value_object_t::get_builtins() const {
             });
             return result;
         }},
-        {"join", [](const func_args &) -> value {
-            throw not_implemented_exception("object join not implemented");
-        }},
+        {"join", object_join_not_implemented},
     };
     return builtins;
 }

common/jinja/value.h

@@ -129,27 +129,25 @@ struct value_t {
     // Note: only for debugging and error reporting purposes
     virtual std::string type() const { return ""; }
 
-    virtual int64_t as_int() const { throw std::runtime_error(type() + " is not an int value"); }
-    virtual double as_float() const { throw std::runtime_error(type() + " is not a float value"); }
-    virtual string as_string() const { throw std::runtime_error(type() + " is not a string value"); }
-    virtual bool as_bool() const { throw std::runtime_error(type() + " is not a bool value"); }
-    virtual const std::vector<value> & as_array() const { throw std::runtime_error(type() + " is not an array value"); }
-    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value invoke(const func_args &) const { throw std::runtime_error(type() + " is not a function value"); }
+    virtual int64_t as_int() const { throw_type_error("is not an int value"); }
+    virtual double as_float() const { throw_type_error("is not a float value"); }
+    virtual string as_string() const { throw_type_error("is not a string value"); }
+    virtual bool as_bool() const { throw_type_error("is not a bool value"); }
+    virtual const std::vector<value> & as_array() const { throw_type_error("is not an array value"); }
+    virtual const std::vector<std::pair<value, value>> & as_ordered_object() const { throw_type_error("is not an object value"); }
+    virtual value invoke(const func_args &) const { throw_type_error("is not a function value"); }
     virtual bool is_none() const { return false; }
     virtual bool is_undefined() const { return false; }
-    virtual const func_builtins & get_builtins() const {
-        throw std::runtime_error("No builtins available for type " + type());
-    }
+    virtual const func_builtins & get_builtins() const { throw_type_error("has no builtins"); }
 
-    virtual bool has_key(const value &) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual void insert(const value & /* key */, const value & /* val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const value & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const value & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const std::string & /* key */, value & /* default_val */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(const std::string & /* key */) { throw std::runtime_error(type() + " is not an object value"); }
-    virtual value & at(int64_t /* idx */, value & /* default_val */) { throw std::runtime_error(type() + " is not an array value"); }
-    virtual value & at(int64_t /* idx */) { throw std::runtime_error(type() + " is not an array value"); }
+    virtual bool has_key(const value &) { throw_type_error("is not an object value"); }
+    virtual void insert(const value & /* key */, const value & /* val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const value & /* key */, value & /* default_val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const value & /* key */) { throw_type_error("is not an object value"); }
+    virtual value & at(const std::string & /* key */, value & /* default_val */) { throw_type_error("is not an object value"); }
+    virtual value & at(const std::string & /* key */) { throw_type_error("is not an object value"); }
+    virtual value & at(int64_t /* idx */, value & /* default_val */) { throw_type_error("is not an array value"); }
+    virtual value & at(int64_t /* idx */) { throw_type_error("is not an array value"); }
 
     virtual bool is_numeric() const { return false; }
     virtual bool is_hashable() const { return false; }
@@ -163,6 +161,11 @@ struct value_t {
     // Note: only for debugging purposes
     virtual std::string as_repr() const { return as_string().str(); }
 
+private:
+    [[noreturn]] void throw_type_error(const char* expected) const {
+        throw std::runtime_error(type() + " " + expected);
+    }
+
 protected:
     virtual bool equivalent(const value_t &) const = 0;
     virtual bool nonequal(const value_t & other) const { return !equivalent(other); }

common/preset.cpp

@@ -43,7 +43,7 @@ static std::set<std::string> get_remote_preset_whitelist(const std::map<std::str
     for (const auto & it : key_to_opt) {
         const std::string & key = it.first;
         const common_arg & opt = it.second;
-        if (allowed_options.find(key) != allowed_options.end() || opt.is_sparam) {
+        if (allowed_options.find(key) != allowed_options.end() || opt.is_sampling) {
             allowed_keys.insert(key);
             // also add variant keys (args without leading dashes and env vars)
             for (const auto & arg : opt.get_args()) {

common/sampling.cpp

@@ -1,10 +1,12 @@
 #include "sampling.h"
 
 #include "common.h"
-#include "ggml.h"
+#include "fit.h"
 #include "log.h"
 #include "reasoning-budget.h"
 
+#include "ggml.h"
+
 #include <algorithm>
 #include <cctype>
 #include <climits>
@@ -511,7 +513,7 @@ void common_perf_print(const struct llama_context * ctx, const struct common_sam
         LOG_INF("%s: unaccounted time = %10.2f ms / %5.1f %%      (total - sampling - prompt eval - eval) / (total)\n", __func__, t_unacc_ms, t_unacc_pc);
         LOG_INF("%s:    graphs reused = %10d\n", __func__, data.n_reused);
 
-        llama_memory_breakdown_print(ctx);
+        common_memory_breakdown_print(ctx);
     }
 }
 

common/speculative.cpp

@@ -61,18 +61,26 @@ static bool common_speculative_are_compatible(
     LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
 
     if (vocab_type_tgt != vocab_type_dft) {
-        LOG_DBG("%s: draft model vocab type must match target model to use speculation but ", __func__);
-        LOG_DBG("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
+        LOG_WRN("%s: draft model vocab type must match target model to use speculation but "
+                "vocab_type_dft = %d while vocab_type_tgt = %d\n", __func__, vocab_type_dft, vocab_type_tgt);
         return false;
     }
 
-    if (
-        llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
-        llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
-        llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
-        llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)
-    ) {
-        LOG_DBG("%s: draft model special tokens must match target model to use speculation\n", __func__);
+    if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
+        (llama_vocab_get_add_bos(vocab_tgt) && llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft))) {
+        LOG_WRN("%s: draft model bos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_dft),
+                llama_vocab_bos(vocab_tgt), llama_vocab_bos(vocab_dft));
+        return false;
+    }
+
+    if (llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
+        (llama_vocab_get_add_eos(vocab_tgt) && llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft))) {
+        LOG_WRN("%s: draft model eos tokens must match target model to use speculation. add: %d - %d, id: %d - %d)\n",
+                __func__,
+                llama_vocab_get_add_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_dft),
+                llama_vocab_eos(vocab_tgt), llama_vocab_eos(vocab_dft));
         return false;
     }
 
@@ -143,6 +151,9 @@ struct common_speculative_state {
             llama_tokens & result) = 0;
 
     virtual void accept(uint16_t n_accepted) = 0;
+
+    virtual int32_t n_max(const common_params_speculative & params) const = 0;
+    virtual int32_t n_min(const common_params_speculative & params) const = 0;
 };
 
 struct common_speculative_checkpoint {
@@ -164,8 +175,8 @@ struct common_speculative_state_draft : public common_speculative_state {
     llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
     llama_context * ctx_dft;
 
+    bool use_ckpt = false;
     struct common_speculative_checkpoint ckpt;
-    bool use_checkpoint;
 
     common_sampler * smpl;
 
@@ -180,11 +191,11 @@ struct common_speculative_state_draft : public common_speculative_state {
             llama_context * ctx_tgt,
             llama_context * ctx_dft,
             const std::vector<std::pair<std::string, std::string>> & replacements,
-            bool use_checkpoint)
+            bool use_ckpt)
         : common_speculative_state(type)
         , ctx_tgt(ctx_tgt)
         , ctx_dft(ctx_dft)
-        , use_checkpoint(use_checkpoint)
+        , use_ckpt(use_ckpt)
     {
         batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
         smpl = nullptr;
@@ -239,7 +250,7 @@ struct common_speculative_state_draft : public common_speculative_state {
     }
 
     void begin(const llama_tokens & prompt) override {
-        if (use_checkpoint && ckpt.size() > 0) {
+        if (use_ckpt && ckpt.size() > 0) {
             // delete checkpoint
             LOG_DBG("%s: delete checkpoint, prompt.size=%zu, pos_min=%d, pos_max=%d, n_tokens=%" PRId64 ", size=%.3f MiB\n",
                     __func__, prompt.size(), ckpt.pos_min, ckpt.pos_max, ckpt.n_tokens, (float) ckpt.data.size() / 1024 / 1024);
@@ -288,6 +299,8 @@ struct common_speculative_state_draft : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
+        const auto & sparams = params.draft;
+
         auto * spec = this;
 
         auto & batch      = spec->batch;
@@ -301,7 +314,7 @@ struct common_speculative_state_draft : public common_speculative_state {
         int reuse_i = 0; // index of part to be reused in prompt_dft
         int reuse_n = 0; // length of part to be reused in prompt_dft
 
-        const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max;
+        const int n_ctx = llama_n_ctx(ctx_dft) - sparams.n_max;
 
         llama_tokens prompt_cnv;
         if (!spec->vocab_cmpt) {
@@ -351,7 +364,7 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         LOG_DBG("%s: reuse_i = %d, reuse_n = %d, #prompt_dft = %zu, #prompt_cur = %zu\n",
                 __func__, reuse_i, reuse_n, prompt_dft.size(), prompt_cur.size());
-        if (use_checkpoint && ckpt.ckpt_size == 0 && reuse_n > 0) {
+        if (use_ckpt && ckpt.ckpt_size == 0 && reuse_n > 0) {
             LOG_DBG("%s: no checkpoint available, no reuse, (reuse_i=%d, reuse_n=%d) -> (0, 0)\n",
                     __func__, reuse_i, reuse_n);
             reuse_i = 0;
@@ -359,10 +372,10 @@ struct common_speculative_state_draft : public common_speculative_state {
         }
 
         result.clear();
-        result.reserve(params.n_max);
+        result.reserve(sparams.n_max);
 
-        bool needs_ckpt = use_checkpoint && prompt_dft.size() > 0;
-        if (reuse_n == 0 || (use_checkpoint && reuse_i > 0)) {
+        bool needs_ckpt = use_ckpt && prompt_dft.size() > 0;
+        if (reuse_n == 0 || (use_ckpt && reuse_i > 0)) {
             llama_memory_clear(mem_dft, false);
             prompt_dft.clear();
         } else {
@@ -372,7 +385,7 @@ struct common_speculative_state_draft : public common_speculative_state {
                 for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
                     result.push_back(prompt_dft[i]);
 
-                    if (params.n_max <= (int) result.size()) {
+                    if (sparams.n_max <= (int) result.size()) {
                         break;
                     }
                 }
@@ -400,7 +413,7 @@ struct common_speculative_state_draft : public common_speculative_state {
             }
 
             if (reuse_n < (int) prompt_dft.size() || do_restore) {
-                if (use_checkpoint) {
+                if (use_ckpt) {
                     if (ckpt.n_tokens > (int64_t) prompt_dft.size()) {
                         LOG_INF("%s: checkpoint is too large, prompt_tgt.size=%zu, ckpt.n_tokens=%" PRId64 ", reuse_n=%d, prompt_dft.size=%zu\n",
                                 __func__, prompt_tgt.size(), ckpt.n_tokens, reuse_n, prompt_dft.size());
@@ -465,7 +478,7 @@ struct common_speculative_state_draft : public common_speculative_state {
         common_sampler_reset(smpl);
 
         // sample n_draft tokens from the draft model
-        for (int i = 0; i < params.n_max; ++i) {
+        for (int i = 0; i < sparams.n_max; ++i) {
             common_batch_clear(batch);
 
             common_sampler_sample(smpl, ctx_dft, 0, true);
@@ -484,12 +497,12 @@ struct common_speculative_state_draft : public common_speculative_state {
 
             result.push_back(id);
 
-            if (params.n_max <= (int) result.size()) {
+            if (sparams.n_max <= (int) result.size()) {
                 break;
             }
 
             // only collect very high-confidence draft tokens
-            if (cur_p->data[0].p < params.p_min) {
+            if (cur_p->data[0].p < sparams.p_min) {
                 break;
             }
 
@@ -510,10 +523,14 @@ struct common_speculative_state_draft : public common_speculative_state {
             detokenized = replace_to_tgt(detokenized);
             LOG_DBG("draft->main detokenized string: '%s'\n", detokenized.c_str());
             result = common_tokenize(ctx_tgt, detokenized, false, true);
-            if (result.size() > (size_t)params.n_max) {
-                result.resize(params.n_max);
+            if (result.size() > (size_t) sparams.n_max) {
+                result.resize(sparams.n_max);
             }
         }
+
+        if (result.size() < (size_t) sparams.n_min) {
+            result.clear();
+        }
     }
 
     void accept(uint16_t n_accepted) override {
@@ -521,6 +538,14 @@ struct common_speculative_state_draft : public common_speculative_state {
         GGML_UNUSED(n_accepted);
     }
 
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.draft.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.draft.n_min;
+    }
+
     std::string replace_to_dft(const std::string & input) const {
         std::string result = input;
 
@@ -573,6 +598,14 @@ struct common_speculative_state_eagle3 : public common_speculative_state {
         // noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.draft.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.draft.n_min;
+    }
 };
 
 // state of self-speculation (simple implementation, not ngram-map)
@@ -602,19 +635,27 @@ struct common_speculative_state_ngram_simple : public common_speculative_state {
         // noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return config.size_mgram;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return config.size_mgram;
+    }
 };
 
 struct common_speculative_state_ngram_map_k : public common_speculative_state {
     // draft ngram map for speculative decoding without draft model
-    common_ngram_map map;
+    common_ngram_map config;
 
     common_speculative_state_ngram_map_k(
             enum common_speculative_type type,
-            common_ngram_map map)
-        : common_speculative_state(type), map(std::move(map)) {}
+            common_ngram_map config)
+        : common_speculative_state(type), config(std::move(config)) {}
 
     void begin(const llama_tokens & prompt) override {
-        common_ngram_map_begin(map, prompt);
+        common_ngram_map_begin(config, prompt);
     }
 
     void draft(
@@ -622,12 +663,20 @@ struct common_speculative_state_ngram_map_k : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        common_ngram_map_draft(map, prompt_tgt, id_last, result);
+        common_ngram_map_draft(config, prompt_tgt, id_last, result);
         GGML_UNUSED(params);
     }
 
     void accept(uint16_t n_accepted) override {
-        common_ngram_map_accept(map, n_accepted);
+        common_ngram_map_accept(config, n_accepted);
+    }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return config.size_value;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return config.size_value;
     }
 };
 
@@ -684,7 +733,7 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
             const llama_tokens & prompt_tgt,
             llama_token id_last,
             llama_tokens & result) override {
-        GGML_UNUSED(params);
+        const auto & sparams = params.ngram_mod;
 
         n_draft_last = 0;
 
@@ -704,16 +753,16 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
             i_last = cur_len - n;
         }
 
-        result.resize(n + params.n_max);
+        result.resize(n + sparams.n_max);
         for (size_t i = 0; i < n - 1; ++i) {
             result[i] = prompt_tgt[cur_len - n + 1 + i];
         }
         result[n - 1] = id_last;
 
-        for (int i = 0; i < params.n_max; ++i) {
+        for (int i = 0; i < sparams.n_max; ++i) {
             const llama_token token = mod.get(result.data() + i);
             if (token == common_ngram_mod::EMPTY) {
-                if (i < params.n_min) {
+                if (i < sparams.n_min) {
                     result.clear();
                     return;
                 }
@@ -749,12 +798,21 @@ struct common_speculative_state_ngram_mod : public common_speculative_state {
 
                     mod.reset();
                     n_low = 0;
+                    i_last = 0;
                 }
             } else {
                 n_low = 0;
             }
         }
     }
+
+    int32_t n_max(const common_params_speculative & params) const override {
+        return params.ngram_mod.n_max;
+    }
+
+    int32_t n_min(const common_params_speculative & params) const override {
+        return params.ngram_mod.n_min;
+    }
 };
 
 struct common_speculative_state_ngram_cache : public common_speculative_state {
@@ -848,6 +906,14 @@ struct common_speculative_state_ngram_cache : public common_speculative_state {
         // TODO: noop
         GGML_UNUSED(n_accepted);
     }
+
+    int32_t n_max(const common_params_speculative & /*params*/) const override {
+        return n_draft;
+    }
+
+    int32_t n_min(const common_params_speculative & /*params*/) const override {
+        return 0;
+    }
 };
 
 struct common_speculative {
@@ -856,11 +922,13 @@ struct common_speculative {
     common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
 };
 
-static common_ngram_map get_common_ngram_map(const common_speculative_config & config) {
-    uint16_t size_key   = config.params.ngram_size_n;
-    uint16_t size_value = config.params.ngram_size_m;
-    bool     key_only   = (config.type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
-    uint16_t min_hits   = config.params.ngram_min_hits;
+static common_ngram_map get_common_ngram_map(
+        common_speculative_type type,
+        const common_params_speculative_ngram_map & config) {
+    uint16_t size_key   = config.size_n;
+    uint16_t size_value = config.size_m;
+    bool     key_only   = type == COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K;
+    uint16_t min_hits   = config.min_hits;
 
     return common_ngram_map(size_key, size_value, key_only, min_hits);
 }
@@ -912,50 +980,14 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
-common_speculative_compat_type common_speculative_is_compat(llama_context * ctx_tgt) {
-    auto * mem = llama_get_memory(ctx_tgt);
-    if (mem == nullptr) {
-        return COMMON_SPECULATIVE_COMPAT_TYPE_NO;
-    }
-
-    common_speculative_compat_type res = COMMON_SPECULATIVE_COMPAT_TYPE_FULL;
-
-    llama_memory_clear(mem, true);
-
-    // eval 2 tokens to check if the context is compatible
-    std::vector<llama_token> tmp;
-    tmp.push_back(0);
-    tmp.push_back(0);
-
-    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
-    if (ret != 0) {
-        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
-        res = COMMON_SPECULATIVE_COMPAT_TYPE_NO;
-        goto done;
-    }
-
-    // try to remove the last tokens
-    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
-        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
-        res = COMMON_SPECULATIVE_COMPAT_TYPE_CKPT;
-        goto done;
-    }
-
-done:
-    llama_memory_clear(mem, true);
-    llama_synchronize(ctx_tgt);
-
-    return res;
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt) {
     llama_context * ctx_dft = nullptr;
-    if (params.model_dft) {
-        ctx_dft = llama_init_from_model(params.model_dft, params.cparams_dft);
+    if (params.draft.model) {
+        ctx_dft = llama_init_from_model(params.draft.model, params.draft.cparams);
         if (ctx_dft == nullptr) {
             LOG_ERR("%s", "failed to create draft context\n");
             return nullptr;
@@ -965,7 +997,7 @@ common_speculative * common_speculative_init(
     // Compute the implementations to use based on the config and their order of preference
     std::vector<common_speculative_config> configs = {}; // list of speculative configs to try
     {
-        bool has_draft = !params.mparams_dft.path.empty();
+        bool has_draft = !params.draft.mparams.path.empty();
         bool has_draft_eagle3 = false; // TODO PR-18039: if params.speculative.eagle3
 
         bool has_ngram_cache   = (params.type == COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
@@ -988,16 +1020,17 @@ common_speculative * common_speculative_init(
             configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
         }
         if (has_ngram_mod) {
-            // shared instance for all speculative decoding contexts
-            if (!params.ngram_mod) {
-                params.ngram_mod = std::make_shared<common_ngram_mod>(params.ngram_size_n, 4*1024*1024);
+            auto & sparams = params.ngram_mod;
 
-                LOG_INF("%s: initialized ngram_mod with n=%d, size=%zu (%.3f MB)\n", __func__,
-                        params.ngram_size_n, params.ngram_mod->size(),
-                        (float)(params.ngram_mod->size_bytes())/1024/1024);
+            if (!sparams.obj) {
+                sparams.obj = std::make_shared<common_ngram_mod>(sparams.n_match, 4*1024*1024);
 
-                if (params.ngram_size_n < 16) {
-                    LOG_WRN("%s: ngram_mod n=%d is too small - poor quality is possible, see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, params.ngram_size_n);
+                LOG_INF("%s: initialized ngram_mod with n_match=%d, size=%zu (%.3f MB)\n", __func__,
+                        sparams.n_match, sparams.obj->size(), (float)(sparams.obj->size_bytes())/1024/1024);
+
+                if (sparams.n_match < 16) {
+                    LOG_WRN("%s: ngram_mod n_match=%d is too small - poor quality is possible, "
+                            "see: https://github.com/ggml-org/llama.cpp/pull/19164\n", __func__, sparams.n_match);
                 }
             }
 
@@ -1022,11 +1055,13 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NONE:
                 break;
             case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                const bool use_ckpt = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+
                 impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
-                    /* .ctx_tgt       = */ ctx_tgt,
-                    /* .ctx_dft       = */ ctx_dft,
-                    /* .replacements  = */ params.replacements,
-                    /* .use_checkpoint= */ params.use_checkpoints // TODO: this should be based on the draft model!
+                    /* .ctx_tgt      = */ ctx_tgt,
+                    /* .ctx_dft      = */ ctx_dft,
+                    /* .replacements = */ params.draft.replacements,
+                    /* .use_ckpt     = */ use_ckpt
                 ));
                 break;
             }
@@ -1035,18 +1070,18 @@ common_speculative * common_speculative_init(
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE: {
-                common_ngram_map ngram_map = get_common_ngram_map(config);
+                common_ngram_map ngram_map = get_common_ngram_map(config.type, config.params.ngram_simple);
 
                 uint16_t ngram_size_key   = ngram_map.size_key;
                 uint16_t mgram_size_value = ngram_map.size_value;
 
                 auto config_simple = common_ngram_simple_config {
-                    /* .size_ngram      = */ ngram_size_key,
-                    /* .size_mgram      = */ mgram_size_value
+                    /* .size_ngram = */ ngram_size_key,
+                    /* .size_mgram = */ mgram_size_value
                 };
                 auto state = std::make_unique<common_speculative_state_ngram_simple>(
-                    /* .type            = */ config.type,
-                    /* .state           = */ config_simple
+                    /* .type  = */ config.type,
+                    /* .state = */ config_simple
                 );
                 impls.push_back(std::move(state));
                 break;
@@ -1055,18 +1090,17 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V: {
                 impls.push_back(std::make_unique<common_speculative_state_ngram_map_k>(
                     (config.type),
-                    get_common_ngram_map(config)
+                    get_common_ngram_map(config.type, config.params.ngram_map_k)
                 ));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_MOD: {
-                GGML_ASSERT(config.params.ngram_mod);
-                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod));
+                GGML_ASSERT(config.params.ngram_mod.obj);
+                impls.push_back(std::make_unique<common_speculative_state_ngram_mod>(config.type, *config.params.ngram_mod.obj));
                 break;
             }
             case COMMON_SPECULATIVE_TYPE_NGRAM_CACHE: {
-                auto state = create_state_ngram_cache(
-                        params.lookup_cache_static, params.lookup_cache_dynamic, config);
+                auto state = create_state_ngram_cache(params.ngram_cache.lookup_cache_static, params.ngram_cache.lookup_cache_dynamic, config);
                 impls.push_back(std::make_unique<common_speculative_state_ngram_cache>(state));
                 break;
             }
@@ -1124,6 +1158,15 @@ llama_tokens common_speculative_draft(
             impl->n_call_draft++;
         }
 
+        {
+            const int n_min = impl->n_min(params);
+
+            if (!result.empty() && (int) result.size() < n_min) {
+                LOG_DBG("%s: ignoring small draft: %d < %d\n", __func__, (int) result.size(), n_min);
+                result.clear();
+            }
+        }
+
         if (!result.empty()) {
             LOG_DBG("%s: called impl %s, hist size = %zu, call_count = %zu, gen = %zu\n", __func__,
                     common_speculative_type_to_str(impl.get()->type).c_str(), prompt_tgt.size(),
@@ -1133,7 +1176,7 @@ llama_tokens common_speculative_draft(
             impl->n_gen_drafts++;
             impl->n_gen_tokens += result.size();
 
-            break; // We have a draft, so break out of the loop and return it.
+            break; // we have a draft, so break out of the loop and return it.
         }
     }
 
@@ -1161,6 +1204,32 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted) {
     }
 }
 
+int32_t common_speculative_n_max(const common_speculative * spec, const common_params_speculative & params) {
+    if (spec == nullptr) {
+        return 0;
+    }
+
+    int32_t n_max = 0;
+    for (const auto & impl : spec->impls) {
+        n_max = std::max(n_max, impl->n_max(params));
+    }
+
+    return n_max;
+}
+
+int32_t common_speculative_n_min(const common_speculative * spec, const common_params_speculative & params) {
+    if (spec == nullptr) {
+        return 0;
+    }
+
+    int32_t n_min = 0;
+    for (const auto & impl : spec->impls) {
+        n_min = std::max(n_min, impl->n_min(params));
+    }
+
+    return n_min;
+}
+
 void common_speculative_print_stats(const common_speculative * spec) {
     if (spec == nullptr) {
         return;

common/speculative.h

@@ -14,16 +14,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
-enum common_speculative_compat_type {
-    COMMON_SPECULATIVE_COMPAT_TYPE_NO   = 0,
-    COMMON_SPECULATIVE_COMPAT_TYPE_FULL = 1,
-    COMMON_SPECULATIVE_COMPAT_TYPE_CKPT = 2,
-};
-
-// check if the llama_context is compatible for speculative decoding
-// note: clears the memory of the context
-common_speculative_compat_type common_speculative_is_compat(llama_context * ctx_tgt);
-
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);
@@ -43,6 +33,9 @@ llama_tokens common_speculative_draft(
 // informs the speculative decoder that n_accepted tokens were accepted by the target model
 void common_speculative_accept(common_speculative * spec, uint16_t n_accepted);
 
+int32_t common_speculative_n_max(const common_speculative * spec, const common_params_speculative & params);
+int32_t common_speculative_n_min(const common_speculative * spec, const common_params_speculative & params);
+
 // print statistics about the speculative decoding
 void common_speculative_print_stats(const common_speculative * spec);
 

convert_hf_to_gguf.py

@@ -272,6 +272,22 @@ class ModelBase:
 
         return tensors
 
+    @staticmethod
+    def _scale_is_trivial(scale: Tensor) -> bool:
+        return scale.numel() <= 1 and abs(float(scale.float().sum()) - 1.0) < 1e-6
+
+    def _write_scale_tensor(self, scale_name: str, scale: Tensor):
+        if not self._scale_is_trivial(scale):
+            scale_f32 = scale.float().numpy().flatten()
+            logger.info(f"  + {scale_name} (per-tensor scale, shape [{scale_f32.size}])")
+            self.gguf_writer.add_tensor(scale_name, scale_f32)
+
+    def _write_scales_tensor(self, scale_name: str, scales: list[float]):
+        if not np.allclose(scales, 1.0, atol=1e-6):
+            scale_vals = np.array(scales, dtype=np.float32)
+            logger.info(f"  + {scale_name} (per-expert scale, shape [{len(scales)}])")
+            self.gguf_writer.add_tensor(scale_name, scale_vals)
+
     def dequant_model(self):
         # If all quantized tensors were already handled (e.g. pure NVFP4), skip
         if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
@@ -494,7 +510,7 @@ class ModelBase:
                         s = self.model_tensors[name]
                         self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
                         tensors_to_remove.append(name)
-                    if name.endswith((".k_scale", ".v_scale")):
+                    if name.endswith((".input_scale", ".k_scale", ".v_scale")):
                         tensors_to_remove.append(name)
             elif quant_method is not None:
                 raise NotImplementedError(f"Quant method is not yet supported: {quant_method!r}")
@@ -602,10 +618,6 @@ class ModelBase:
         raw = np.concatenate([d_grouped, qs_grouped], axis=-1).reshape(out_features, n_super * 36)
         return raw, [out_features, n_super * 64]
 
-    @staticmethod
-    def _nvfp4_scale2_is_trivial(scale2: Tensor) -> bool:
-        return scale2.numel() <= 1 and abs(float(scale2.float().sum()) - 1.0) < 1e-6
-
     def _repack_nvfp4(self, name: str, weight: Tensor, scale: Tensor, scale2: Tensor, input_scale: Tensor):
         if "language_model." in name:
             name = name.replace("language_model.", "")
@@ -616,19 +628,8 @@ class ModelBase:
         logger.info(f"Repacked {new_name} with shape {shape} and quantization NVFP4")
         self.gguf_writer.add_tensor(new_name, raw, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
 
-        # Emit per-tensor scale2 as a separate F32 tensor when non-trivial
-        if not self._nvfp4_scale2_is_trivial(scale2):
-            scale2_f32 = scale2.float().numpy().flatten()
-            scale_name = new_name.replace(".weight", ".scale")
-            logger.info(f"  + {scale_name} (per-tensor NVFP4 scale2, shape [{scale2_f32.size}])")
-            self.gguf_writer.add_tensor(scale_name, scale2_f32)
-
-        # Emit per-tensor input_scale as a separate F32 tensor when non-trivial
-        if not self._nvfp4_scale2_is_trivial(input_scale):
-            input_scale_f32 = input_scale.float().numpy().flatten()
-            input_scale_name = new_name.replace(".weight", ".input_scale")
-            logger.info(f"  + {input_scale_name} (per-tensor NVFP4 input_scale, shape [{input_scale_f32.size}])")
-            self.gguf_writer.add_tensor(input_scale_name, input_scale_f32)
+        self._write_scale_tensor(new_name.replace(".weight", ".scale"), scale2)
+        self._write_scale_tensor(new_name.replace(".weight", ".input_scale"), input_scale)
 
     def _generate_nvfp4_tensors(self):
         # Per-layer expert merging to avoid holding all experts in memory
@@ -719,21 +720,11 @@ class ModelBase:
         logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
         self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
 
-        # Emit per-expert scale2 tensor if any expert has non-trivial scale2
         scales.sort(key=lambda x: x[0])
-        scale_vals = np.array([s[1] for s in scales], dtype=np.float32)
-        if not np.allclose(scale_vals, 1.0, atol=1e-6):
-            scale_name = new_name.replace(".weight", ".scale")
-            logger.info(f"  + {scale_name} (per-expert NVFP4 scale2, shape [{len(scales)}])")
-            self.gguf_writer.add_tensor(scale_name, scale_vals)
+        self._write_scales_tensor(new_name.replace(".weight", ".scale"), [s[1] for s in scales])
 
-        # Emit per-expert input_scale tensor if any expert has non-trivial input_scale
         input_scales.sort(key=lambda x: x[0])
-        input_scale_vals = np.array([s[1] for s in input_scales], dtype=np.float32)
-        if not np.allclose(input_scale_vals, 1.0, atol=1e-6):
-            input_scale_name = new_name.replace(".weight", ".input_scale")
-            logger.info(f"  + {input_scale_name} (per-expert NVFP4 input_scale, shape [{len(input_scales)}])")
-            self.gguf_writer.add_tensor(input_scale_name, input_scale_vals)
+        self._write_scales_tensor(new_name.replace(".weight", ".input_scale"), [s[1] for s in input_scales])
 
         del experts, merged
 
@@ -746,7 +737,12 @@ class ModelBase:
 
         if (not quant_algo or not quant_layers) and quant_config_file.is_file():
             with open(quant_config_file, "r", encoding="utf-8") as f:
-                quant_config = json.load(f).get("quantization") or {}
+                hf_quant_config = json.load(f)
+                quant_config = hf_quant_config.get("quantization") or {}
+                producer = hf_quant_config.get("producer") or {}
+                producer_name = (producer.get("name") or "").lower()
+                if quant_method is None:
+                    self.hparams.setdefault("quantization_config", {})["quant_method"] = producer_name
                 quant_algo = quant_config.get("quant_algo", quant_algo)
                 quant_layers = quant_config.get("quantized_layers", quant_layers) or {}
 
@@ -1850,20 +1846,28 @@ class TextModel(ModelBase):
             with open(module_path, encoding="utf-8") as f:
                 modules = json.load(f)
             for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
+                if mod["type"].endswith("Pooling"):
                     pooling_path = mod["path"]
                     break
 
+        mode_mapping = {
+            "mean": gguf.PoolingType.MEAN,
+            "cls": gguf.PoolingType.CLS,
+            "lasttoken": gguf.PoolingType.LAST,
+        }
+
         # get pooling type
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
+            if pooling.get("pooling_mode_mean_tokens"):
                 pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
+            elif pooling.get("pooling_mode_cls_token"):
                 pooling_type = gguf.PoolingType.CLS
-            elif pooling["pooling_mode_lasttoken"]:
+            elif pooling.get("pooling_mode_lasttoken"):
                 pooling_type = gguf.PoolingType.LAST
+            elif (pooling_mode := pooling.get("pooling_mode")) in mode_mapping:
+                pooling_type = mode_mapping[pooling_mode]
             else:
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
@@ -7180,7 +7184,7 @@ class EmbeddingGemma(Gemma3Model):
                 with open(modules_file, encoding="utf-8") as modules_json_file:
                     mods = json.load(modules_json_file)
                 for mod in mods:
-                    if mod["type"] == "sentence_transformers.models.Dense":
+                    if mod["type"].endswith("Dense"):
                         mod_path = mod["path"]
                         # check if model.safetensors file for Dense layer exists
                         model_tensors_file = self.dir_model / mod_path / "model.safetensors"
@@ -11847,7 +11851,7 @@ class LLaDAMoEModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("HunYuanDenseV1ForCausalLM", "HunYuanVLForConditionalGeneration")
+@ModelBase.register("HunYuanDenseV1ForCausalLM")
 class HunYuanModel(TextModel):
     model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
 
@@ -11986,28 +11990,58 @@ class HunYuanModel(TextModel):
 
 
 @ModelBase.register("HunYuanVLForConditionalGeneration")
-class HunyuanOCRVisionModel(MmprojModel):
+class HunyuanVLVisionModel(MmprojModel):
+    # Handles both HunyuanOCR and HunyuanVL, which share the HF architecture name
+    # "HunYuanVLForConditionalGeneration" and the `vit.perceive.*` vision layout.
+    # Each variant maps to a different projector type in clip.cpp so image
+    # preprocessing follows the correct code path.
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         assert self.hparams_vision is not None
-        # HunyuanOCR uses max_image_size instead of image_size
+        # HunyuanOCR / HunyuanVL uses max_image_size instead of image_size
         if "image_size" not in self.hparams_vision:
             self.hparams_vision["image_size"] = self.hparams_vision.get("max_image_size", 2048)
 
+    @staticmethod
+    def is_ocr_variant(hparams: dict) -> bool:
+        """Return True for HunyuanOCR, False for HunyuanVL.
+
+        The projector's output dim must equal the text model's hidden_size by
+        construction (that's what "projector" means). HunyuanOCR pairs a 1B text
+        backbone (hidden=1024); HunyuanVL pairs a 4B one (hidden=3072). So the
+        ViT -> LLM projection dim is a hard architectural signature, not a
+        magic number.
+        """
+        vision_out = int((hparams.get("vision_config") or {}).get("out_hidden_size", 0))
+        return vision_out == 1024
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         assert self.hparams_vision is not None
-        hparams = self.hparams_vision
-        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANOCR)
-        self.gguf_writer.add_vision_use_gelu(True)
-        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("rms_norm_eps", 1e-5))
-        self.gguf_writer.add_vision_spatial_merge_size(hparams.get("spatial_merge_size", 2))
-        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
-        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+        vcfg = self.hparams_vision
+
+        if self.is_ocr_variant(self.global_config):
+            # --- HunyuanOCR ---
+            self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANOCR)
+            self.gguf_writer.add_vision_use_gelu(True)
+            self.gguf_writer.add_vision_attention_layernorm_eps(vcfg.get("rms_norm_eps", 1e-5))
+            self.gguf_writer.add_vision_spatial_merge_size(vcfg.get("spatial_merge_size", 2))
+            self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
+            self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+            return
+
+        # --- HunyuanVL ---
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.HUNYUANVL)
+        self.gguf_writer.add_vision_use_gelu(str(vcfg["hidden_act"]).lower() == "gelu")
+        self.gguf_writer.add_vision_attention_layernorm_eps(float(vcfg["rms_norm_eps"]))
+        self.gguf_writer.add_vision_spatial_merge_size(int(vcfg["spatial_merge_size"]))
+        self.gguf_writer.add_vision_min_pixels(int(self.preprocessor_config["min_pixels"]))
+        self.gguf_writer.add_vision_max_pixels(int(self.preprocessor_config["max_pixels"]))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if not name.startswith("vit."):
-            return  # skip text tensors
+            return
         # strip CLS token (row 0) from position embeddings so resize_position_embeddings works
         if "position_embedding" in name:
             data_torch = data_torch[1:]  # [n_patches+1, n_embd] -> [n_patches, n_embd]
@@ -12015,11 +12049,66 @@ class HunyuanOCRVisionModel(MmprojModel):
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         # force conv weights to F32 or F16 to avoid BF16 IM2COL issues on Metal
+        # Both HunyuanOCR and HunyuanVL emit the ViT -> LLM projection as mm.0/mm.2.
         if ("mm.0." in new_name or "mm.2." in new_name) and new_name.endswith(".weight"):
             return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
 
+@ModelBase.register("HunYuanVLForConditionalGeneration")
+class HunyuanVLTextModel(HunYuanModel):
+    # The "HunYuanVLForConditionalGeneration" HF architecture covers both HunyuanOCR
+    # and HunyuanVL. HunyuanOCR reuses the HunYuan-Dense text backbone (standard RoPE),
+    # while HunyuanVL introduces a new LLM arch with XD-RoPE. Detect the variant from
+    # the config and pick the matching GGUF architecture.
+    model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
+
+    @staticmethod
+    def _is_ocr_config(hparams: dict) -> bool:
+        # OCR pairs a 1B text backbone (hidden=1024) with a ViT projector that
+        # outputs 1024-d; HunyuanVL uses 3072-d. Keep in sync with
+        # HunyuanVLVisionModel.is_ocr_variant.
+        return int((hparams.get("vision_config") or {}).get("out_hidden_size", 0)) == 1024
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        raw_hparams = kwargs.get("hparams") or ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        if self._is_ocr_config(raw_hparams):
+            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_DENSE
+        else:
+            self.model_arch = gguf.MODEL_ARCH.HUNYUAN_VL
+        super().__init__(dir_model, *args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # Only emit XD-RoPE metadata for the HunyuanVL backbone; HunyuanOCR uses
+        # the HunYuan-Dense arch which already handles standard rope in super().
+        if self.model_arch != gguf.MODEL_ARCH.HUNYUAN_VL:
+            return
+
+        if self.rope_parameters.get("rope_type") != "xdrope":
+            return
+
+        # defaults for HunyuanVL. The C++ side later computes:
+        #   freq_base = rope_theta * alpha ** (head_dim / (head_dim - 2))
+        self.gguf_writer.add_rope_freq_base(float(self.rope_parameters["rope_theta"]))
+        self.gguf_writer.add_rope_scaling_alpha(float(self.rope_parameters["alpha"]))
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        self.gguf_writer.add_rope_scaling_factor(float(self.rope_parameters.get("factor", 1)))
+
+        ctx_len = int(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_rope_scaling_orig_ctx_len(ctx_len)
+        self.gguf_writer.add_context_length(ctx_len)
+
+        self.gguf_writer.add_rope_dimension_sections(list(self.rope_parameters["xdrope_section"]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # Skip vision tensors — they are written by HunyuanVLVisionModel
+        if name.startswith("vit."):
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("SmolLM3ForCausalLM")
 class SmolLM3Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.SMOLLM3

docs/backend/OPENVINO.md

@@ -244,7 +244,6 @@ build\ReleaseOV\bin\llama-cli.exe -m "C:\models\Llama-3.2-1B-Instruct-Q4_0.gguf"
 - `-fa 1` is required when running llama-bench with the OpenVINO backend.
   - `GGML_OPENVINO_STATEFUL_EXECUTION=1 GGML_OPENVINO_DEVICE=GPU ./llama-bench -fa 1`
 - `llama-server` with OpenVINO backend supports only one chat session/thread, when `GGML_OPENVINO_STATEFUL_EXECUTION=1` is enabled.
-- For Intel GPU, NPU detection in containers, GPU, NPU user-space drivers/libraries must be present inside the image. We will include in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile)
 
 > [!NOTE]
 > The OpenVINO backend is actively under development. Fixes are underway, and this document will continue to be updated as issues are resolved.
@@ -274,8 +273,6 @@ docker build --build-arg http_proxy=$http_proxy --build-arg https_proxy=$https_p
 Run llama.cpp with OpenVINO backend Docker container.
 Save sample models in `~/models` as [shown above](#3-download-sample-model). It will be mounted to the container in the examples below.
 
-> [!NOTE]
-> Intel GPU, NPU detection in containers will be included in a future PR. Until then, you can use this reference Dockerfile: [openvino.Dockerfile](https://github.com/ravi9/llama.cpp/blob/ov-docker-update/.devops/openvino.Dockerfile).
 
 ```bash
 #  Run Docker container

docs/backend/SYCL.md

@@ -31,6 +31,8 @@ SYCL cross-platform capabilities enable support for other vendor GPUs as well.
 
 ## Recommended Release
 
+### Windows
+
 The following releases are verified and recommended:
 
 |Commit ID|Tag|Release|Verified  Platform| Update date|
@@ -39,9 +41,22 @@ The following releases are verified and recommended:
 |3bcd40b3c593d14261fb2abfabad3c0fb5b9e318|b4040 |[llama-b4040-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b4040/llama-b4040-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1| 2024-11-19|
 |fb76ec31a9914b7761c1727303ab30380fd4f05c|b3038 |[llama-b3038-bin-win-sycl-x64.zip](https://github.com/ggml-org/llama.cpp/releases/download/b3038/llama-b3038-bin-win-sycl-x64.zip) |Arc A770/Linux/oneAPI 2024.1<br>MTL Arc GPU/Windows 11/oneAPI 2024.1||
 
+### Ubuntu 24.04
+
+The release packages for Ubuntu 24.04 x64 (FP32/FP16) only include the binary files of the llama.cpp SYCL backend. They require the target machine to have pre-installed Intel GPU drivers and oneAPI packages that are the same version as the build package. To get the version and installation info, refer to release.yml: ubuntu-24-sycl -> Download & Install oneAPI.
+
+It is recommended to use them with Intel Docker.
+
+The packages for FP32 and FP16 would have different accuracy and performance on LLMs. Please choose it acording to the test result.
 
 ## News
 
+- 2026.04
+
+  - Optimize mul_mat by reorder feature for data type: Q4_K, Q5_K, Q_K, Q8_0.
+  - Fused MoE.
+  - Upgrate CI and built package for oneAPI 2025.3.3, support Ubuntu 24.04 built package.
+
 - 2026.03
   - Support Flash-Attention: less memory usage, performance impact depends on LLM.
 
@@ -229,6 +244,7 @@ Upon a successful installation, SYCL is enabled for the available intel devices,
 
 |Verified release|
 |-|
+|2025.3.3 |
 |2025.2.1|
 |2025.1|
 |2024.1|
@@ -339,6 +355,12 @@ Choose one of following methods to run.
 ./examples/sycl/test.sh
 ```
 
+- Run llama-server:
+
+```sh
+./examples/sycl/start-svr.sh -m PATH/MODEL_FILE
+```
+
 2. Command line
 Launch inference
 
@@ -627,10 +649,18 @@ Choose one of following methods to run.
 
 1. Script
 
+- Run test:
+
 ```
 examples\sycl\win-test.bat
 ```
 
+- Run llama-server:
+
+```
+examples\sycl\win-start-svr.bat -m PATH\MODEL_FILE
+```
+
 2. Command line
 
 Launch inference

docs/backend/snapdragon/README.md

@@ -249,18 +249,27 @@ build: 6a8cf8914 (6733)
   ```
 
 - `GGML_HEXAGON_PROFILE=1`
-  Generates a host-side profile for the ggml-hexagon Ops.
+  Enables Op profiling:
 
-- `GGML_HEXAGON_OPMASK=0x0`
-  Allows enabling specific stages of the processing pipeline:
+  - `1` Basic profile with per-op `usecs` and `cycles` counters
+  - `2` Extended profile with per-op `usecs`, `cycles` and default PMU counter data
+  - `0x1,...,0x8` Extended profile with per-op `usecs`, `cycles` and custom PMU counter data
+
+  The logging output can be either saved into a file for post-processing or it can be piped directly into the post-processing tool to generate the report.
+  Examples:
+
+      `GGML_HEXAGON_PROFILE=1 llama-completion ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+
+- `GGML_HEXAGON_OPSTAGE=0x0`
+  Allows enabling specific stages of the Op processing pipeline:
 
   - `0x1` Enable Op Queue (i.e., queuing Ops into NPU)
   - `0x2` Enable Op Compute (MUL_MAT, etc.)
 
   Examples:
 
-      `GGML_HEXAGON_OPMASK=0x1 llama-completion ...` - Ops are enqueued but NPU-side processing is stubbed out
-      `GGML_HEXAGON_OPMASK=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
+      `GGML_HEXAGON_OPSTAGE=0x1 llama-completion ...` - Ops are enqueued to the NPU but dma & compute are disabled
+      `GGML_HEXAGON_OPSTAGE=0x3 llama-completion ...` - Full queuing and processing of Ops (default)
 
 - `GGML_HEXAGON_OPFILTER=regex`
   Allows filtering (disabling) Ops that match the regex pattern:

docs/ops.md

@@ -26,7 +26,7 @@ Legend:
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -60,7 +60,7 @@ Legend:
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
@@ -105,7 +105,7 @@ Legend:
 |                              SQR | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |

examples/debug/debug.cpp

@@ -202,10 +202,14 @@ static bool run(llama_context * ctx, const common_params & params) {
     print_tokenized_prompt(ctx, tokens, params.prompt);
 
     if (params.save_logits) {
-        output_data output {ctx, model, params};
-        std::filesystem::path model_path{params.model.path};
-        std::string model_name{model_path.stem().string()};
-        save_output_data(output, model_name, params.logits_output_dir);
+        try {
+            output_data output {ctx, model, params};
+            std::filesystem::path model_path{params.model.path};
+            std::string model_name{model_path.stem().string()};
+            save_output_data(output, model_name, params.logits_output_dir);
+        } catch (const std::exception & e) {
+            LOG_ERR("%s : error saving logits: %s\n", __func__, e.what());
+        }
     }
 
     return true;
@@ -223,7 +227,7 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    std::optional<base_callback_data> cb_data;
+    std::optional<common_debug_cb_user_data> cb_data;
     if (!params.save_logits) {
         cb_data.emplace(params, params.tensor_filter);
     }

examples/eval-callback/eval-callback.cpp

@@ -3,7 +3,6 @@
 #include "debug.h"
 #include "log.h"
 #include "llama.h"
-#include "llama-cpp.h"
 
 #include <clocale>
 #include <string>
@@ -38,7 +37,7 @@ static bool run(llama_context * ctx, const common_params & params) {
 int main(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
-    base_callback_data cb_data;
+    common_debug_cb_user_data cb_data;
 
     common_params params;
 
@@ -53,7 +52,7 @@ int main(int argc, char ** argv) {
 
     // pass the callback to the backend scheduler
     // it will be executed for each node during the graph computation
-    params.cb_eval = common_debug_cb_eval<false>;
+    params.cb_eval = common_debug_cb_eval;
     params.cb_eval_user_data = &cb_data;
     params.warmup = false;
 

examples/gen-docs/gen-docs.cpp

@@ -73,12 +73,12 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
     auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;
-    std::vector<common_arg *> sparam_options;
+    std::vector<common_arg *> sampling_options;
     std::vector<common_arg *> specific_options;
     for (auto & opt : ctx_arg.options) {
         // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
-        if (opt.is_sparam) {
-            sparam_options.push_back(&opt);
+        if (opt.is_sampling) {
+            sampling_options.push_back(&opt);
         } else if (opt.in_example(ctx_arg.ex)) {
             specific_options.push_back(&opt);
         } else {
@@ -93,7 +93,7 @@ static void write_help(std::ostringstream & ss, const md_file & md) {
     ss << "### Common params\n\n";
     write_table(ss, common_options);
     ss << "\n\n### Sampling params\n\n";
-    write_table(ss, sparam_options);
+    write_table(ss, sampling_options);
     ss << "\n\n### " << md.specific_section_header << "\n\n";
     write_table(ss, specific_options);
 

examples/lookup/lookup-create.cpp

@@ -37,9 +37,9 @@ int main(int argc, char ** argv){
 
     common_ngram_cache ngram_cache;
     common_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
-    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.lookup_cache_static.c_str());
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.speculative.ngram_cache.lookup_cache_static.c_str());
 
-    common_ngram_cache_save(ngram_cache, params.speculative.lookup_cache_static);
+    common_ngram_cache_save(ngram_cache, params.speculative.ngram_cache.lookup_cache_static);
 
     return 0;
 }

examples/lookup/lookup-stats.cpp

@@ -24,7 +24,7 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.speculative.draft.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -49,18 +49,18 @@ int main(int argc, char ** argv){
     {
         const int64_t t_start_draft_us = ggml_time_us();
 
-        if (!params.speculative.lookup_cache_static.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.ngram_cache.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.speculative.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 

examples/lookup/lookup.cpp

@@ -25,7 +25,7 @@ int main(int argc, char ** argv){
     }
 
     // max. number of additional tokens to draft if match is found
-    const int n_draft = params.speculative.n_max;
+    const int n_draft = params.speculative.draft.n_max;
 
     // init llama.cpp
     llama_backend_init();
@@ -54,18 +54,18 @@ int main(int argc, char ** argv){
         const int64_t t_start_draft_us = ggml_time_us();
         common_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
 
-        if (!params.speculative.lookup_cache_static.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_static.empty()) {
             try {
-                ngram_cache_static = common_ngram_cache_load(params.speculative.lookup_cache_static);
+                ngram_cache_static = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_static);
             } catch (std::ifstream::failure const &) {
-                LOG_ERR("failed to open static lookup cache: %s", params.speculative.lookup_cache_static.c_str());
+                LOG_ERR("failed to open static lookup cache: %s", params.speculative.ngram_cache.lookup_cache_static.c_str());
                 exit(1);
             }
         }
 
-        if (!params.speculative.lookup_cache_dynamic.empty()) {
+        if (!params.speculative.ngram_cache.lookup_cache_dynamic.empty()) {
             try {
-                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.lookup_cache_dynamic);
+                ngram_cache_dynamic = common_ngram_cache_load(params.speculative.ngram_cache.lookup_cache_dynamic);
             } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
         }
 
@@ -213,7 +213,7 @@ int main(int argc, char ** argv){
 
     // Update dynamic ngram cache with context ngram cache and save it to disk:
     common_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.lookup_cache_dynamic);
+    common_ngram_cache_save(ngram_cache_dynamic, params.speculative.ngram_cache.lookup_cache_dynamic);
 
     LOG("\n\n");
 

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

@@ -25,7 +25,11 @@ MODEL_NAME="${MODEL_NAME:-$(basename "$MODEL_PATH")}"
 OUTPUT_DIR="${OUTPUT_DIR:-../../models}"
 TYPE="${OUTTYPE:-f16}"
 METADATA_OVERRIDE="${METADATA_OVERRIDE:-}"
-CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+if [[ -n "$MMPROJ" ]]; then
+    CONVERTED_MODEL="${OUTPUT_DIR}/mmproj-${MODEL_NAME}.gguf"
+else
+    CONVERTED_MODEL="${OUTPUT_DIR}/${MODEL_NAME}.gguf"
+fi
 
 echo "Model path: ${MODEL_PATH}"
 echo "Model name: ${MODEL_NAME}"
@@ -38,6 +42,7 @@ if [[ -n "$DEBUG" ]]; then
 else
     CMD_ARGS=("python")
 fi
+
 CMD_ARGS+=("../../convert_hf_to_gguf.py" "--verbose")
 CMD_ARGS+=("${MODEL_PATH}")
 CMD_ARGS+=("--outfile" "${CONVERTED_MODEL}")
@@ -50,7 +55,3 @@ CMD_ARGS+=("--outtype" "${TYPE}")
 echo ""
 echo "The environment variable CONVERTED_MODEL can be set to this path using:"
 echo "export CONVERTED_MODEL=$(realpath ${CONVERTED_MODEL})"
-if [[ -n "$MMPROJ" ]]; then
-    mmproj_file="${OUTPUT_DIR}/mmproj-$(basename "${CONVERTED_MODEL}")"
-    echo "The mmproj model was created in $(realpath "$mmproj_file")"
-fi

examples/speculative-simple/speculative-simple.cpp

@@ -8,8 +8,24 @@
 #include <clocale>
 #include <cstdio>
 #include <cstring>
+#include <cinttypes>
 #include <string>
 #include <vector>
+#include <utility>
+
+struct spec_checkpoint {
+    int64_t n_tokens = 0;
+
+    std::vector<uint8_t> data;
+
+    size_t size() const {
+        return data.size();
+    }
+
+    bool empty() const {
+        return data.empty();
+    }
+};
 
 int main(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
@@ -27,7 +43,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.speculative.mparams_dft.path.empty()) {
+    if (params.speculative.draft.mparams.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -46,6 +62,14 @@ int main(int argc, char ** argv) {
     model_tgt = llama_init_tgt->model();
     ctx_tgt   = llama_init_tgt->context();
 
+    // check if the context supports partial sequence removal
+    const auto ctx_seq_rm = common_context_can_seq_rm(ctx_tgt);
+    const bool use_ckpt = (ctx_seq_rm == COMMON_CONTEXT_SEQ_RM_TYPE_FULL);
+
+    if (use_ckpt) {
+        LOG_INF("speculative decoding will use checkpoints (context does not support partial sequence removal)\n");
+    }
+
     const llama_vocab * vocab = llama_model_get_vocab(model_tgt);
 
     // load the draft model
@@ -53,7 +77,7 @@ int main(int argc, char ** argv) {
 
     // TODO: simplify this logic
     {
-        const auto & params_spec = params.speculative;
+        const auto & params_spec = params.speculative.draft;
 
         auto params_dft = params;
 
@@ -61,15 +85,15 @@ int main(int argc, char ** argv) {
         params_dft.n_ctx        = params_spec.n_ctx;
         params_dft.n_batch      = llama_n_ctx_seq(ctx_tgt);
         params_dft.devices      = params_spec.devices;
-        params_dft.model        = params_spec.mparams_dft;
+        params_dft.model        = params_spec.mparams;
         params_dft.n_gpu_layers = params_spec.n_gpu_layers;
 
         if (params_spec.cpuparams.n_threads > 0) {
-            params_dft.cpuparams.n_threads       = params.speculative.cpuparams.n_threads;
-            params_dft.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+            params_dft.cpuparams.n_threads       = params.speculative.draft.cpuparams.n_threads;
+            params_dft.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
         }
 
-        params_dft.tensor_buft_overrides = params.speculative.tensor_buft_overrides;
+        params_dft.tensor_buft_overrides = params.speculative.draft.tensor_buft_overrides;
 
         auto mparams_dft = common_model_params_to_llama(params_dft);
 
@@ -79,8 +103,8 @@ int main(int argc, char ** argv) {
             return 1;
         }
 
-        params.speculative.model_dft = model_dft.get();
-        params.speculative.cparams_dft = common_context_params_to_llama(params_dft);
+        params.speculative.draft.model = model_dft.get();
+        params.speculative.draft.cparams = common_context_params_to_llama(params_dft);
     }
 
     // Tokenize the prompt
@@ -119,7 +143,7 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // target model sampling context
-    struct common_sampler * smpl = common_sampler_init(model_tgt, params.sampling);
+    common_sampler_ptr smpl(common_sampler_init(model_tgt, params.sampling));
 
     // eval the prompt
     llama_decode(ctx_tgt, llama_batch_get_one(inp.data(), inp.size() - 1));
@@ -142,21 +166,49 @@ int main(int argc, char ** argv) {
 
     llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, 1);
 
+    size_t n_draft = 0;
+
+    llama_tokens draft;
+    spec_checkpoint spec_ckpt;
+
     const auto t_enc_end = ggml_time_us();
 
     const auto t_dec_start = ggml_time_us();
 
     while (true) {
-        // optionally, generate draft tokens that can be appended to the target batch
+        // generate or reuse draft tokens
         //
         // this is the most important part of the speculation. the more probable tokens that are provided here
         // the better the performance will be. in theory, this computation can be performed asynchronously and even
         // offloaded to a remote device. it doesn't even have to be based on an LLM. instead, it can provide tokens
         // from a cache or lookup tables.
         //
-        llama_tokens draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last);
+        if (draft.empty()) {
+            // generate a new draft
+            draft = common_speculative_draft(spec, params_spec, prompt_tgt, id_last);
 
-        //LOG_DBG("draft: %s\n", string_from(ctx_dft, draft).c_str());
+            // save the original draft size
+            n_draft = draft.size();
+
+            // save a checkpoint of the target context before evaluating the draft
+            // this allows us to restore the state if partial draft acceptance occurs
+            if (!draft.empty() && use_ckpt) {
+                const size_t ckpt_size = llama_state_seq_get_size_ext(ctx_tgt, 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                spec_ckpt.data.resize(ckpt_size);
+
+                const size_t n = llama_state_seq_get_data_ext(ctx_tgt, spec_ckpt.data.data(), ckpt_size, 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+                GGML_ASSERT(n == ckpt_size);
+
+                spec_ckpt.n_tokens = (int64_t) prompt_tgt.size();
+                LOG_DBG("created speculative checkpoint (n_tokens = %" PRId64 ", size = %.3f MiB)\n",
+                        spec_ckpt.n_tokens, (float) spec_ckpt.data.size() / 1024 / 1024);
+            }
+        } else {
+            // we have a previous (partial) draft to reuse from checkpoint restoration
+            if (use_ckpt) {
+                GGML_ASSERT(!spec_ckpt.empty());
+            }
+        }
 
         // always have a token to evaluate from before - id_last
         common_batch_clear(batch_tgt);
@@ -164,11 +216,6 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on [id_last, draft0, draft1, ..., draftN-1]
         {
-            // do not waste time on small drafts
-            if (draft.size() < (size_t) params_spec.n_min) {
-                draft.clear();
-            }
-
             for (size_t i = 0; i < draft.size(); ++i) {
                 common_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
             }
@@ -178,6 +225,12 @@ int main(int argc, char ** argv) {
             llama_decode(ctx_tgt, batch_tgt);
         }
 
+        // only save the sampler sampler state if we use checkpoints
+        common_sampler_ptr smpl_save;
+        if (use_ckpt) {
+            smpl_save.reset(common_sampler_clone(smpl.get()));
+        }
+
         // sample from the full target batch and return the accepted tokens based on the target sampler
         //
         // for each token to be accepted, the sampler would have to sample that same token
@@ -185,14 +238,38 @@ int main(int argc, char ** argv) {
         // available logits from the batch and sample the next token until we run out of logits or the sampler
         // disagrees with the draft
         //
-        const auto ids = common_sampler_sample_and_accept_n(smpl, ctx_tgt, draft);
+        auto ids = common_sampler_sample_and_accept_n(smpl.get(), ctx_tgt, draft);
 
         //LOG_DBG("ids: %s\n", string_from(ctx_tgt, ids).c_str());
 
         GGML_ASSERT(ids.size() > 0); // there will always be at least one accepted token
 
+        // check for partial draft acceptance:
+        // if the context doesn't support partial sequence removal, restore the checkpoint
+        // and make the accepted tokens the new partial draft for the next iteration
+        if (use_ckpt && ids.size() - 1 < draft.size()) {
+            LOG_DBG("partial acceptance: %zu < %zu, restoring checkpoint\n", ids.size() - 1, draft.size());
+
+            draft = std::move(ids);
+
+            const size_t n = llama_state_seq_set_data_ext(ctx_tgt, spec_ckpt.data.data(), spec_ckpt.size(), 0, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+            GGML_ASSERT(n == spec_ckpt.size());
+
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, spec_ckpt.n_tokens, -1);
+
+            prompt_tgt.resize(spec_ckpt.n_tokens);
+            smpl = std::move(smpl_save);
+
+            n_past = (int) prompt_tgt.size();
+
+            continue;
+        }
+
+        common_speculative_accept(spec, ids.size() - 1);
+
+        // full acceptance: consume the draft and commit accepted tokens
         n_past    += ids.size() - 1;
-        n_drafted += draft.size(); // note: we ignore the discarded small drafts
+        n_drafted += n_draft; // note: we ignore the discarded small drafts
         n_accept  += ids.size() - 1;
         n_predict += ids.size();
 
@@ -222,6 +299,9 @@ int main(int argc, char ** argv) {
 
         LOG_DBG("accepted %d/%d draft tokens, the last target token is: (%d)\n", (int) ids.size() - 1, (int) draft.size(), id_last);
 
+        // clear the draft since it has been consumed
+        draft.clear();
+
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
@@ -243,7 +323,7 @@ int main(int argc, char ** argv) {
     LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_INF("\n");
-    LOG_INF("n_draft   = %d\n", params_spec.n_max);
+    LOG_INF("n_draft   = %d\n", params_spec.draft.n_max);
     LOG_INF("n_predict = %d\n", n_predict);
     LOG_INF("n_drafted = %d\n", n_drafted);
     LOG_INF("n_accept  = %d\n", n_accept);
@@ -254,11 +334,10 @@ int main(int argc, char ** argv) {
 
     LOG_INF("\n");
     LOG_INF("target:\n\n");
-    common_perf_print(ctx_tgt, smpl);
+    common_perf_print(ctx_tgt, smpl.get());
 
     llama_batch_free(batch_tgt);
 
-    common_sampler_free(smpl);
     common_speculative_free(spec);
 
     llama_backend_free();

examples/speculative/speculative.cpp

@@ -49,7 +49,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    if (params.speculative.mparams_dft.path.empty()) {
+    if (params.speculative.draft.mparams.path.empty()) {
         LOG_ERR("%s: --model-draft is required\n", __func__);
         return 1;
     }
@@ -58,7 +58,7 @@ int main(int argc, char ** argv) {
     const int n_seq_dft = params.n_parallel;
 
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
-    const float p_draft_split = params.speculative.p_split;
+    const float p_draft_split = params.speculative.draft.p_split;
 
     std::default_random_engine rng(params.sampling.seed == LLAMA_DEFAULT_SEED ? std::random_device()() : params.sampling.seed);
     std::uniform_real_distribution<> u_dist;
@@ -80,15 +80,15 @@ int main(int argc, char ** argv) {
     ctx_tgt   = llama_init_tgt->context();
 
     // load the draft model
-    params.devices = params.speculative.devices;
-    params.model = params.speculative.mparams_dft;
-    params.n_gpu_layers = params.speculative.n_gpu_layers;
-    if (params.speculative.cpuparams.n_threads > 0) {
-        params.cpuparams.n_threads = params.speculative.cpuparams.n_threads;
+    params.devices = params.speculative.draft.devices;
+    params.model = params.speculative.draft.mparams;
+    params.n_gpu_layers = params.speculative.draft.n_gpu_layers;
+    if (params.speculative.draft.cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.speculative.draft.cpuparams.n_threads;
     }
 
-    params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
-    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+    params.cpuparams_batch.n_threads = params.speculative.draft.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.draft.tensor_buft_overrides;
 
     auto llama_init_dft = common_init_from_params(params);
 
@@ -183,7 +183,7 @@ int main(int argc, char ** argv) {
     //GGML_ASSERT(n_vocab == llama_vocab_n_tokens(model_dft));
 
     // how many tokens to draft each time
-    int n_draft = params.speculative.n_max;
+    int n_draft = params.speculative.draft.n_max;
 
     int n_predict = 0;
     int n_drafted = 0;

examples/sycl/start-svr.sh

@@ -0,0 +1,124 @@
+#!/bin/bash
+
+#  MIT license
+#  Copyright (C) 2024 Intel Corporation
+#  SPDX-License-Identifier: MIT
+
+Help() {
+  cat << EOF
+Usage: $(basename "$0") [OPTIONS]
+
+This script processes files with specified options.
+
+Options:
+  -h, --help    Display this help message and exit.
+  -c, --context <value>    Set context length. Bigger need more memory.
+  -p, --promote <value>    Prompt to start generation with.
+  -m, --model   <value>    Full model file path.
+  -mg,--main-gpu <value>   Set main GPU ID (0 - n) for single GPU mode.
+  -sm,--split-mode <value> How to split the model across multiple GPUs, one of:
+                            - none: use one GPU only
+                            - layer (default): split layers and KV across GPUs
+                            - row: split rows across GPUs
+  -ngl,--n-gpu-layers <value>  Max. number of layers to store in VRAM (default: -1)
+  -lv,--log-verbosity <value>  Set the verbosity threshold. Messages with a higher verbosity will be
+                               ignored. Values:
+                                - 0: generic output
+                                - 1: error
+                                - 2: warning
+                                - 3: info
+                                - 4: debug
+
+
+EOF
+}
+
+BIN_FILE=./build/bin/llama-server
+SEED=0
+GPUS_SETTING=""
+
+MODEL_FILE=../models/Qwen3.5-4B-Q4_0.gguf
+NGL=99
+CONTEXT=4096
+GGML_SYCL_DEVICE=-1
+SPLIT_MODE=layer
+LOG_VERBOSE=3
+while [[ $# -gt 0 ]]; do
+    case "$1" in
+        -c|--context)
+            CONTEXT=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -m|--model)
+            MODEL_FILE="$2"
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -mg|--main-gpu)
+            GGML_SYCL_DEVICE=$2
+            SPLIT_MODE=none
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -sm|--split-mode)
+            SPLIT_MODE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -ngl|--n-gpu-layers)
+            NGL=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -lv|--log-verbosity)
+            LOG_VERBOSE=$2
+            # Shift twice to consume both the option flag and its value
+            shift
+            shift
+            ;;
+        -h|--help)
+            Help
+            exit 0
+            ;;
+        *)
+            # Handle unknown options or stop processing options
+            echo "Invalid option: $1"
+            # Optional: exit script or shift to treat remaining as positional args
+            exit 1
+            ;;
+    esac
+done
+
+
+
+source /opt/intel/oneapi/setvars.sh
+
+#export GGML_SYCL_DEBUG=1
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+#support malloc device memory more than 4GB.
+export UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
+echo "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=${UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS}"
+
+if [ $GGML_SYCL_DEVICE -ne -1 ]; then
+    echo "Use $GGML_SYCL_DEVICE as main GPU"
+    #use signle GPU only
+    GPUS_SETTING="-mg $GGML_SYCL_DEVICE -sm ${SPLIT_MODE}"
+    export ONEAPI_DEVICE_SELECTOR="level_zero:${$GGML_SYCL_DEVICE}"
+    echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
+else
+    echo "Use all Intel GPUs, including iGPU & dGPU"
+    GPUS_SETTING="-sm ${SPLIT_MODE}"
+ fi
+
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap --host 0.0.0.0 --port 8000
+
+

examples/sycl/test.sh

@@ -38,7 +38,7 @@ SEED=0
 GPUS_SETTING=""
 
 INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
-MODEL_FILE=models/llama-2-7b.Q4_0.gguf
+MODEL_FILE=../models/llama-2-7b.Q4_0.gguf
 NGL=99
 CONTEXT=4096
 GGML_SYCL_DEVICE=-1
@@ -122,9 +122,10 @@ if [ $GGML_SYCL_DEVICE -ne -1 ]; then
     export ONEAPI_DEVICE_SELECTOR="level_zero:${$GGML_SYCL_DEVICE}"
     echo "ONEAPI_DEVICE_SELECTOR=${ONEAPI_DEVICE_SELECTOR}"
 else
-   echo "Use all Intel GPUs, including iGPU & dGPU"
+    echo "Use all Intel GPUs, including iGPU & dGPU"
+    GPUS_SETTING="-sm ${SPLIT_MODE}"
  fi
 
-echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
-ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
+echo "run cmd: ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE}  --mmap "
+ZES_ENABLE_SYSMAN=1 ${BIN_FILE} -m ${MODEL_FILE} -no-cnv -p "${INPUT_PROMPT}" -n 200 -e -ngl ${NGL} -s ${SEED} -c ${CONTEXT} ${GPUS_SETTING} -lv ${LOG_VERBOSE} --mmap
 

examples/sycl/win-start-svr.bat

@@ -0,0 +1,179 @@
+::  MIT license
+::  Copyright (C) 2024 Intel Corporation
+::  SPDX-License-Identifier: MIT
+
+@echo off
+setlocal EnableExtensions EnableDelayedExpansion
+
+set "BIN_FILE=.\build\bin\llama-server.exe"
+set "SEED=0"
+set "GPUS_SETTING="
+
+set "MODEL_FILE=..\models\Qwen3.5-4B-Q4_0.gguf"
+set "NGL=99"
+set "CONTEXT=4096"
+set "GGML_SYCL_DEVICE=-1"
+set "SPLIT_MODE=layer"
+set "LOG_VERBOSE=3"
+
+if "%~1"=="" goto after_args
+
+:parse_args
+if "%~1"=="" goto after_args
+
+if /I "%~1"=="-c" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--context" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-m" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--model" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-mg" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--main-gpu" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-sm" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--split-mode" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-ngl" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--n-gpu-layers" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-lv" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--log-verbosity" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-h" goto help
+if /I "%~1"=="--help" goto help
+
+echo Invalid option: %~1
+exit /b 1
+
+:missing_value
+echo Missing value for option: %~1
+exit /b 1
+
+:help
+echo Usage: %~n0 [OPTIONS]
+echo.
+echo This script processes files with specified options.
+echo.
+echo Options:
+echo   -h, --help    Display this help message and exit.
+echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
+echo   -m, --model   ^<value^>    Full model file path.
+echo   -mg,--main-gpu ^<value^>   Set main GPU ID (0 - n) for single GPU mode.
+echo   -sm,--split-mode ^<value^> How to split the model across multiple GPUs, one of:
+echo                             - none: use one GPU only
+echo                             - layer (default): split layers and KV across GPUs
+echo                             - row: split rows across GPUs
+echo   -ngl,--n-gpu-layers ^<value^>  Max. number of layers to store in VRAM (default: -1)
+echo   -lv,--log-verbosity ^<value^>  Set the verbosity threshold. Messages with a higher verbosity will be
+echo                                ignored. Values:
+echo                                 - 0: generic output
+echo                                 - 1: error
+echo                                 - 2: warning
+echo                                 - 3: info
+echo                                 - 4: debug
+exit /b 0
+
+:after_args
+
+REM In Windows CMD, source is not available; call oneAPI setvars if present.
+if exist "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" (
+  call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" >nul
+) else (
+  echo Warning: oneAPI setvars.bat not found. Continuing without environment setup.
+)
+
+REM Support malloc device memory more than 4GB.
+set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
+echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%
+
+if not "%GGML_SYCL_DEVICE%"=="-1" (
+  echo Use %GGML_SYCL_DEVICE% as main GPU
+  REM Use single GPU only.
+  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
+  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
+  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+) else (
+  echo Use all Intel GPUs, including iGPU ^& dGPU
+  set "GPUS_SETTING=-sm %SPLIT_MODE%"
+)
+
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+set "ZES_ENABLE_SYSMAN=1"
+%BIN_FILE% -m "%MODEL_FILE%" -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap --host 0.0.0.0 --port 8000
+
+endlocal
+

examples/sycl/win-test.bat

@@ -2,10 +2,200 @@
 ::  Copyright (C) 2024 Intel Corporation
 ::  SPDX-License-Identifier: MIT
 
-set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
-@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-:: support malloc device memory more than 4GB.
-set UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1
-set LOAD_MODE="--mmap"
-.\build\bin\llama-completion.exe -m models\llama-2-7b.Q4_0.gguf -no-cnv -p %INPUT2% -n 400 -e -ngl 99 -s 0 %LOAD_MODE%
+@echo off
+setlocal EnableExtensions EnableDelayedExpansion
+
+REM MIT license
+REM Copyright (C) 2024 Intel Corporation
+REM SPDX-License-Identifier: MIT
+
+set "BIN_FILE=.\build\bin\llama-completion.exe"
+set "SEED=0"
+set "GPUS_SETTING="
+
+set "INPUT_PROMPT=Building a website can be done in 10 simple steps:^nStep 1:"
+set "MODEL_FILE=..\models\llama-2-7b.Q4_0.gguf"
+set "NGL=99"
+set "CONTEXT=4096"
+set "GGML_SYCL_DEVICE=-1"
+set "SPLIT_MODE=layer"
+set "LOG_VERBOSE=3"
+
+if "%~1"=="" goto after_args
+
+:parse_args
+if "%~1"=="" goto after_args
+
+if /I "%~1"=="-c" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--context" (
+  if "%~2"=="" goto missing_value
+  set "CONTEXT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-p" (
+  if "%~2"=="" goto missing_value
+  set "INPUT_PROMPT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--promote" (
+  if "%~2"=="" goto missing_value
+  set "INPUT_PROMPT=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-m" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--model" (
+  if "%~2"=="" goto missing_value
+  set "MODEL_FILE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-mg" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--main-gpu" (
+  if "%~2"=="" goto missing_value
+  set "GGML_SYCL_DEVICE=%~2"
+  set "SPLIT_MODE=none"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-sm" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--split-mode" (
+  if "%~2"=="" goto missing_value
+  set "SPLIT_MODE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-ngl" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--n-gpu-layers" (
+  if "%~2"=="" goto missing_value
+  set "NGL=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-lv" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+if /I "%~1"=="--log-verbosity" (
+  if "%~2"=="" goto missing_value
+  set "LOG_VERBOSE=%~2"
+  shift
+  shift
+  goto parse_args
+)
+
+if /I "%~1"=="-h" goto help
+if /I "%~1"=="--help" goto help
+
+echo Invalid option: %~1
+exit /b 1
+
+:missing_value
+echo Missing value for option: %~1
+exit /b 1
+
+:help
+echo Usage: %~n0 [OPTIONS]
+echo.
+echo This script processes files with specified options.
+echo.
+echo Options:
+echo   -h, --help    Display this help message and exit.
+echo   -c, --context ^<value^>    Set context length. Bigger need more memory.
+echo   -p, --promote ^<value^>    Prompt to start generation with.
+echo   -m, --model   ^<value^>    Full model file path.
+echo   -mg,--main-gpu ^<value^>   Set main GPU ID (0 - n) for single GPU mode.
+echo   -sm,--split-mode ^<value^> How to split the model across multiple GPUs, one of:
+echo                             - none: use one GPU only
+echo                             - layer (default): split layers and KV across GPUs
+echo                             - row: split rows across GPUs
+echo   -ngl,--n-gpu-layers ^<value^>  Max. number of layers to store in VRAM (default: -1)
+echo   -lv,--log-verbosity ^<value^>  Set the verbosity threshold. Messages with a higher verbosity will be
+echo                                ignored. Values:
+echo                                 - 0: generic output
+echo                                 - 1: error
+echo                                 - 2: warning
+echo                                 - 3: info
+echo                                 - 4: debug
+exit /b 0
+
+:after_args
+
+REM In Windows CMD, source is not available; call oneAPI setvars if present.
+if exist "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" (
+  call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" >nul
+) else (
+  echo Warning: oneAPI setvars.bat not found. Continuing without environment setup.
+)
+
+REM Support malloc device memory more than 4GB.
+set "UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=1"
+echo UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS=%UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS%
+
+if not "%GGML_SYCL_DEVICE%"=="-1" (
+  echo Use %GGML_SYCL_DEVICE% as main GPU
+  REM Use single GPU only.
+  set "GPUS_SETTING=-mg %GGML_SYCL_DEVICE% -sm %SPLIT_MODE%"
+  set "ONEAPI_DEVICE_SELECTOR=level_zero:%GGML_SYCL_DEVICE%"
+  echo ONEAPI_DEVICE_SELECTOR=%ONEAPI_DEVICE_SELECTOR%
+) else (
+  echo Use all Intel GPUs, including iGPU ^& dGPU
+  set "GPUS_SETTING=-sm %SPLIT_MODE%"
+)
+
+echo run cmd: ZES_ENABLE_SYSMAN=1 %BIN_FILE% -m %MODEL_FILE% -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+set "ZES_ENABLE_SYSMAN=1"
+%BIN_FILE% -m "%MODEL_FILE%" -no-cnv -p "%INPUT_PROMPT%" -n 200 -e -ngl %NGL% -s %SEED% -c %CONTEXT% %GPUS_SETTING% -lv %LOG_VERBOSE% --mmap
+
+endlocal
+

ggml/CMakeLists.txt

@@ -1,17 +1,11 @@
 cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 
-# ref: https://cmake.org/cmake/help/latest/policy/CMP0194.html
-# MSVC is not a valid assembler for the ASM language.
-# Set to NEW to avoid a warning on CMake 4.1+ with MSVC.
-if (POLICY CMP0194)
-    cmake_policy(SET CMP0194 NEW)
-endif()
 project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 9)
-set(GGML_VERSION_PATCH 11)
+set(GGML_VERSION_MINOR 10)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -219,7 +213,7 @@ set   (GGML_CUDA_COMPRESSION_MODE "size" CACHE STRING
 set_property(CACHE GGML_CUDA_COMPRESSION_MODE PROPERTY STRINGS "none;speed;balance;size")
 
 option(GGML_HIP                             "ggml: use HIP"                                   OFF)
-option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
+option(GGML_HIP_GRAPHS                      "ggml: use HIP graph"                              ON)
 option(GGML_HIP_RCCL                        "ggml: use ROCm Collective Comm. Library"         OFF)
 option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)

ggml/src/CMakeLists.txt

@@ -470,11 +470,10 @@ endforeach()
 
 target_link_libraries(ggml-base PRIVATE Threads::Threads)
 
-find_library(MATH_LIBRARY m)
-if (MATH_LIBRARY)
-    if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT})
-        target_link_libraries(ggml-base PRIVATE m)
-    endif()
+if (DEFINED MATH_LIBRARY)
+    target_link_libraries(ggml-base PRIVATE ${MATH_LIBRARY})
+elseif (NOT WIN32 AND NOT DEFINED ENV{ONEAPI_ROOT})
+    target_link_libraries(ggml-base PRIVATE m)
 endif()
 
 if (CMAKE_SYSTEM_NAME MATCHES "Android")

ggml/src/ggml-backend-meta.cpp

@@ -1133,7 +1133,7 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
         if (t_ij->view_src != nullptr && ggml_backend_buffer_is_meta(t_ij->view_src->buffer)) {
             t_ij->view_src = ggml_backend_meta_buffer_simple_tensor(tensor->view_src, j);
             if (t_ij->view_offs > 0 && split_dim >= 0 && split_dim < GGML_MAX_DIMS) {
-                GGML_ASSERT(ne[split_dim] != 0 && tensor->ne[split_dim] != 0);
+                GGML_ASSERT(tensor->ne[split_dim] != 0);
                 const int split_dim_view_src = ggml_backend_meta_get_split_state(tensor->view_src, /*assume_sync =*/ true).axis;
                 GGML_ASSERT(split_dim_view_src >= 0 && split_dim_view_src < GGML_MAX_DIMS);
 
@@ -1170,6 +1170,28 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
 
         simple_tensors.push_back(t_ij);
     }
+
+    // If one of the sources has a zero-sized slice, disable the computation:
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (tensor->src[i] == nullptr || !ggml_backend_buffer_is_meta(tensor->src[i]->buffer)) {
+            continue;
+        }
+
+        const ggml_backend_meta_split_state split_state_src = ggml_backend_meta_get_split_state(tensor->src[i], /*assume_sync =*/ true);
+        if (split_state_src.axis < 0 || split_state_src.axis >= GGML_MAX_DIMS) {
+            continue;
+        }
+        for (size_t j = 0; j < n_simple_bufs; j++) {
+            int64_t ne_sum = 0;
+            for (size_t s = 0; s < split_state_src.n_segments; s++) {
+                ne_sum += split_state_src.ne[s*n_simple_bufs + j];
+            }
+            if (ne_sum == 0) {
+                simple_tensors[j]->flags &= ~GGML_TENSOR_FLAG_COMPUTE;
+            }
+        }
+    }
+
     buf_ctx->simple_tensors[tensor] = simple_tensors;
 
     return GGML_STATUS_SUCCESS;
@@ -1183,40 +1205,57 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
 
     if (split_state.n_segments != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
-        GGML_ASSERT(offset == 0);
-        GGML_ASSERT(size == ggml_nbytes(tensor));
         GGML_ASSERT(tensor->ne[3] == 1);
+
         size_t offset_data = 0;
         std::vector<size_t> simple_offsets(n_bufs, 0);
         if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
             GGML_ASSERT(tensor->ne[2] == 1);
+
+            const size_t row_stride = tensor->nb[1];
+            GGML_ASSERT(offset % row_stride == 0);
+            GGML_ASSERT(size   % row_stride == 0);
+            const int64_t r_start = offset / row_stride;
+            const int64_t r_count = size   / row_stride;
+            GGML_ASSERT(r_start + r_count <= tensor->ne[1]);
+
             const int64_t blck_size = ggml_blck_size(tensor->type);
             for (size_t s = 0; s < split_state.n_segments; s++) {
                 for (size_t j = 0; j < n_bufs; j++) {
                     ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
                     GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
                     const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
-                    ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data, simple_offsets[j], nbytes,
-                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data,
+                        simple_offsets[j] + r_start * simple_tensor->nb[1], nbytes,
+                        r_count, simple_tensor->nb[1], tensor->nb[1]);
                     offset_data       += nbytes;
                     simple_offsets[j] += nbytes;
                 }
             }
-            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            GGML_ASSERT(offset_data*r_count == size);
             return;
         }
         GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+
+        const size_t row_stride = tensor->nb[2];
+        GGML_ASSERT(offset % row_stride == 0);
+        GGML_ASSERT(size   % row_stride == 0);
+        const int64_t r_start = offset / row_stride;
+        const int64_t r_count = size   / row_stride;
+        GGML_ASSERT(r_start + r_count <= tensor->ne[2]);
+
         for (size_t s = 0; s < split_state.n_segments; s++) {
             for (size_t j = 0; j < n_bufs; j++) {
                 ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
                 const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
-                ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data, simple_offsets[j], nbytes,
-                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                ggml_backend_tensor_set_2d(simple_tensor, (const char *) data + offset_data,
+                    simple_offsets[j] + r_start * simple_tensor->nb[2], nbytes,
+                    r_count, simple_tensor->nb[2], tensor->nb[2]);
                 offset_data       += nbytes;
                 simple_offsets[j] += nbytes;
             }
         }
-        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        GGML_ASSERT(offset_data*r_count == size);
         return;
     }
 
@@ -1273,40 +1312,57 @@ static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, co
 
     if (split_state.n_segments != 1) {
         GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
-        GGML_ASSERT(offset == 0);
-        GGML_ASSERT(size == ggml_nbytes(tensor));
         GGML_ASSERT(tensor->ne[3] == 1);
+
         size_t offset_data = 0;
         std::vector<size_t> simple_offsets(n_bufs, 0);
         if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
             GGML_ASSERT(tensor->ne[2] == 1);
+
+            const size_t row_stride = tensor->nb[1];
+            GGML_ASSERT(offset % row_stride == 0);
+            GGML_ASSERT(size   % row_stride == 0);
+            const int64_t r_start = offset / row_stride;
+            const int64_t r_count = size   / row_stride;
+            GGML_ASSERT(r_start + r_count <= tensor->ne[1]);
+
             const int64_t blck_size = ggml_blck_size(tensor->type);
             for (size_t s = 0; s < split_state.n_segments; s++) {
                 for (size_t j = 0; j < n_bufs; j++) {
                     const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
                     GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
                     const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
-                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
-                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data,
+                        simple_offsets[j] + r_start * simple_tensor->nb[1], nbytes,
+                        r_count, simple_tensor->nb[1], tensor->nb[1]);
                     offset_data       += nbytes;
                     simple_offsets[j] += nbytes;
                 }
             }
-            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            GGML_ASSERT(offset_data*r_count == size);
             return;
         }
         GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+
+        const size_t row_stride = tensor->nb[2];
+        GGML_ASSERT(offset % row_stride == 0);
+        GGML_ASSERT(size   % row_stride == 0);
+        const int64_t r_start = offset / row_stride;
+        const int64_t r_count = size   / row_stride;
+        GGML_ASSERT(r_start + r_count <= tensor->ne[2]);
+
         for (size_t s = 0; s < split_state.n_segments; s++) {
             for (size_t j = 0; j < n_bufs; j++) {
                 const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
                 const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
-                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
-                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data,
+                    simple_offsets[j] + r_start * simple_tensor->nb[2], nbytes,
+                    r_count, simple_tensor->nb[2], tensor->nb[2]);
                 offset_data       += nbytes;
                 simple_offsets[j] += nbytes;
             }
         }
-        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        GGML_ASSERT(offset_data*r_count == size);
         return;
     }
 
@@ -1442,26 +1498,32 @@ struct ggml_backend_meta_context {
     struct backend_config {
         ggml_backend_t backend;
 
-        std::vector<cgraph_config> cgraphs;
-        std::vector<ggml_tensor *> nodes;
-        ggml_backend_buffer_ptr    buf;
+        std::vector<cgraph_config>           cgraphs;
+        std::vector<ggml_tensor *>           nodes;
+        std::vector<ggml_backend_buffer_ptr> bufs;
 
-        backend_config(ggml_backend_t backend) : backend(backend) {}
+        backend_config(ggml_backend_t backend, const size_t n_reduce_steps) : backend(backend) {
+            bufs.resize(n_reduce_steps);
+        }
     };
     std::string                 name;
     std::vector<backend_config> backend_configs;
     ggml_context_ptr            ctx;
     std::vector<ggml_cgraph *>  cgraphs_aux;
     std::vector<ggml_tensor *>  nodes_aux;
+    size_t                      n_reduce_steps;
     int                         max_nnodes    = 0;
     size_t                      max_tmp_size  = 0;
     size_t                      max_subgraphs = 0;
+    size_t                      n_subgraphs   = 0;
+    uint64_t                    uid           = 0;
 
     void *                               comm_ctx       = nullptr;
     ggml_backend_comm_allreduce_tensor_t comm_allreduce = nullptr;
 
     ggml_backend_meta_context(ggml_backend_dev_t meta_dev, const char * params) {
         const size_t n_devs = ggml_backend_meta_dev_n_devs(meta_dev);
+        n_reduce_steps = std::ceil(std::log2(n_devs));
         name = "Meta(";
         std::vector<ggml_backend_t> simple_backends;
         backend_configs.reserve(n_devs);
@@ -1473,7 +1535,7 @@ struct ggml_backend_meta_context {
             }
             name += ggml_backend_dev_name(simple_dev);
             simple_backends.push_back(ggml_backend_dev_init(simple_dev, params));
-            backend_configs.emplace_back(simple_backends.back());
+            backend_configs.emplace_back(simple_backends.back(), n_reduce_steps);
         }
         name += ")";
 
@@ -1503,10 +1565,6 @@ struct ggml_backend_meta_context {
             ggml_backend_free(bc.backend);
         }
     }
-
-    size_t n_reduce_steps() const {
-        return std::ceil(std::log2(backend_configs.size()));
-    }
 };
 
 static const char * ggml_backend_meta_get_name(ggml_backend_t backend) {
@@ -1616,6 +1674,9 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     const size_t n_backends = ggml_backend_meta_n_backends(backend);
     ggml_backend_meta_context * backend_ctx = (ggml_backend_meta_context *) backend->context;
 
+    // If the previous cgraph had a defined UID it can be used to skip rebuilding the subgraphs per simple backend.
+    const bool needs_rebuild = (cgraph->uid == 0) || (cgraph->uid != backend_ctx->uid);
+
     bool max_nnodes_raised = false;
     if (cgraph->n_nodes > backend_ctx->max_nnodes) {
         for (size_t j = 0; j < n_backends; j++) {
@@ -1625,173 +1686,216 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         backend_ctx->max_nnodes = cgraph->n_nodes;
         max_nnodes_raised = true;
+        assert(needs_rebuild);
     }
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
 
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
-                // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
-                bcj.nodes[i] = node;
-                continue;
+    if (needs_rebuild) {
+        size_t n_subgraphs  = 0;
+        size_t max_tmp_size = 0;
+
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
+                    // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
+                    bcj.nodes[i] = node;
+                    continue;
+                }
+                bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
+                GGML_ASSERT(bcj.nodes[i]);
             }
-            bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
-            GGML_ASSERT(bcj.nodes[i]);
         }
-    }
 
-    size_t n_subgraphs  = 0;
-    size_t max_tmp_size = 0;
-    {
-        // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-        auto get_i_delayed = [&](const int i) -> int {
-            int id = i; // i_delayed
-            int idr = i; // i_delayed return, last safe return value
+        {
+            // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
+            auto get_i_delayed = [&](const int i) -> int {
+                int id = i; // i_delayed
+                int idr = i; // i_delayed return, last safe return value
 
-            ggml_tensor * node = cgraph->nodes[id];
-            int32_t n_used = ggml_node_get_use_count(cgraph, id);
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
-                        ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                ggml_tensor * node = cgraph->nodes[id];
+                int32_t n_used = ggml_node_get_use_count(cgraph, id);
+
+                // Skip MIRRORED nodes that don't consume node
+                auto skip_unrelated = [&]() {
+                    while (id + 1 < cgraph->n_nodes) {
+                        ggml_tensor * next = cgraph->nodes[id+1];
+                        if (ggml_backend_meta_get_split_state(next, false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                            break;
+                        }
+                        bool safe = true;
+                        for (int s = 0; s < GGML_MAX_SRC; s++) {
+                            if (next->src[s] == nullptr) {
+                                continue;
+                            }
+                            if (next->src[s] == node) {
+                                safe = false;
+                                break;
+                            }
+                            if (ggml_backend_meta_get_split_state(next->src[s], false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                                safe = false;
+                                break;
+                            }
+                        }
+                        if (!safe) {
+                            break;
+                        }
+                        id++;
+                    }
+                };
+
+                skip_unrelated();
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
+                            ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+                // Chain of MULs with MIRRORED src[1]
+                while (true) {
+                    skip_unrelated();
+                    if (id + 1 >= cgraph->n_nodes) {
+                        return idr;
+                    }
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_MUL && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    } else {
+                        break;
+                    }
+                }
+
+                if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                for (int32_t k = 0; k < n_used; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
+                            next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
+                            ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_MUL && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-
-            if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                for (int32_t k = 0; k < n_used - 2; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
+                    id++;
+                }
+                idr = id;
                 return idr;
-            }
-            for (int32_t k = 0; k < n_used; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
-                        next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
-                        ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            for (int32_t k = 0; k < n_used - 2; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            idr = id;
-            return idr;
-        };
+            };
 
-        int i_start = 0;
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                continue;
-            }
-            const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
-            if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
-            }
-            const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
-            if (!new_subgraph) {
-                continue;
-            }
+            int i_start = 0;
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    continue;
+                }
+                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
+                if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                    max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
+                }
+                const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
+                if (!new_subgraph) {
+                    continue;
+                }
 
-            i = get_i_delayed(i);
+                i = get_i_delayed(i);
 
+                for (size_t j = 0; j < n_backends; j++) {
+                    auto & bcj = backend_ctx->backend_configs[j];
+                    bcj.cgraphs[n_subgraphs].offset = i_start;
+                }
+                n_subgraphs++;
+                i_start = i + 1;
+            }
+            GGML_ASSERT(i_start == cgraph->n_nodes);
+        }
+
+        backend_ctx->uid         = cgraph->uid;
+        backend_ctx->n_subgraphs = n_subgraphs;
+
+        if (max_tmp_size > backend_ctx->max_tmp_size) {
             for (size_t j = 0; j < n_backends; j++) {
                 auto & bcj = backend_ctx->backend_configs[j];
-                bcj.cgraphs[n_subgraphs].offset = i_start;
+                for (size_t i = 0; i < backend_ctx->n_reduce_steps; i++) {
+                    bcj.bufs[i].reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
+                }
             }
-            n_subgraphs++;
-            i_start = i + 1;
+            backend_ctx->max_tmp_size = max_tmp_size;
         }
-        GGML_ASSERT(i_start == cgraph->n_nodes);
-    }
 
-    if (max_tmp_size > backend_ctx->max_tmp_size) {
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            bcj.buf.reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
-        }
-        backend_ctx->max_tmp_size = max_tmp_size;
-    }
-
-
-    if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
-        backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
-        const size_t n_reduce_steps = backend_ctx->n_reduce_steps();
-        const size_t n_nodes_per_device = 2 * n_reduce_steps; // tmp + ADD per step
-        const size_t n_cgraphs_per_device = n_reduce_steps;    // 1 ADD graph per step
-        const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
-        const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
-        const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
-        ggml_init_params params = {
-            /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        backend_ctx->ctx.reset(ggml_init(params));
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            for (size_t i = 0; i < n_subgraphs; i++) {
-                bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+        if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
+            backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
+            const size_t n_nodes_per_device = 3 * backend_ctx->n_reduce_steps; // tmp + ADD (+zeroing) graph per step and device
+            const size_t n_cgraphs_per_device = 2 * backend_ctx->n_reduce_steps; // ADD ( + zeroing) graph per step and device
+            const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
+            const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
+            const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
+            ggml_init_params params = {
+                /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
+                /*.mem_buffer =*/ nullptr,
+                /*.no_alloc   =*/ true,
+            };
+            backend_ctx->ctx.reset(ggml_init(params));
+            for (size_t j = 0; j < n_backends; j++) {
+                auto & bcj = backend_ctx->backend_configs[j];
+                for (size_t i = 0; i < n_subgraphs; i++) {
+                    bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+                }
+            }
+            backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
+                backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
+            }
+            backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
+                backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
             }
         }
-        backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
-            backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
-        }
-        backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
-            backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
-        }
-    }
 
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
-        for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
-            ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
-            const size_t i_node_start = bcj.cgraphs[i_graph].offset;
-            const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
-            cgraph_ij->n_nodes = i_node_stop - i_node_start;
-            ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
-            for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
-                ggml_tensor * node_ij = bcj.nodes[i_node];
-                cgraph_ij->nodes[i_node - i_node_start] = node_ij;
-                const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
-                const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
-                cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
+                ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
+                const size_t i_node_start = bcj.cgraphs[i_graph].offset;
+                const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
+                cgraph_ij->n_nodes = i_node_stop - i_node_start;
+                ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
+                for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
+                    ggml_tensor * node_ij = bcj.nodes[i_node];
+                    cgraph_ij->nodes[i_node - i_node_start] = node_ij;
+                    const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
+                    const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
+                    cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+                }
+                cgraph_ij->uid = ggml_graph_next_uid();
             }
         }
     }
@@ -1799,11 +1903,6 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     size_t iga = 0; // i graph aux
     size_t ina = 0; // i node aux
 
-    // FIXME usage_counts
-    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
-        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
-        return ret;
-    };
     auto get_node_aux = [&](ggml_tensor * t) -> ggml_tensor * {
         ggml_tensor * ret = backend_ctx->nodes_aux[ina++];
         memset(ret, 0, sizeof(ggml_tensor));
@@ -1815,75 +1914,110 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         return ret;
     };
+    auto set_tmp_data = [&](ggml_tensor * tensor, const size_t j, const size_t i_buf) {
+        auto & bcj = backend_ctx->backend_configs[j];
+        ggml_backend_buffer_ptr & buf_ptr = bcj.bufs[i_buf];
+        if (!buf_ptr || ggml_backend_buffer_get_size(buf_ptr.get()) < backend_ctx->max_tmp_size) {
+            buf_ptr.reset(ggml_backend_alloc_buffer(bcj.backend, backend_ctx->max_tmp_size));
+        }
+        tensor->buffer = buf_ptr.get();
+        tensor->data   = ggml_backend_buffer_get_base(buf_ptr.get());
+    };
+    // FIXME usage_counts
+    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
+        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
+        return ret;
+    };
 
     // Preferentially use backend-specific allreduce_tensor_async (e.g. NCCL for CUDA), use a generic fallback if unavailable:
     auto allreduce_fallback = [&](size_t i) -> ggml_status {
         std::vector<ggml_cgraph *> step_cgraphs(n_backends, nullptr);
 
-        for (size_t offset_j = 1; offset_j < n_backends; offset_j *= 2) {
+        // Zero out nodes that were disabled due to having a zero-sized slice:
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            ggml_tensor * node = bcj.cgraphs[i].cgraph_main->nodes[bcj.cgraphs[i].cgraph_main->n_nodes - 1];
+            if (node->flags & GGML_TENSOR_FLAG_COMPUTE) {
+                continue;
+            }
+            ggml_tensor * node_zero = get_node_aux(node);
+            node_zero->op = GGML_OP_SCALE; // FIXME 0.0f * NaN == NaN
+            node_zero->src[0] = node;
+            ggml_set_op_params_f32(node_zero, 0, 0.0f);
+            node_zero->data = node->data;
+            node_zero->flags |= GGML_TENSOR_FLAG_COMPUTE;
+
+            step_cgraphs[j] = get_cgraph_aux();
+            step_cgraphs[j]->nodes[0] = node_zero;
+            step_cgraphs[j]->n_nodes = 1;
+            const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, step_cgraphs[j]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+        }
+        std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
+
+        auto push_data = [&](const size_t j_src, const size_t j_dst, const size_t i_buf) {
+            assert(step_cgraphs[j_dst] == nullptr);
+            auto & bcj_src = backend_ctx->backend_configs[j_src];
+            auto & bcj_dst = backend_ctx->backend_configs[j_dst];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            GGML_ASSERT(ggml_is_contiguous(node_src));
+            GGML_ASSERT(ggml_is_contiguous(node_dst));
+
+            ggml_tensor * node_tmp = get_node_aux(node_dst);
+            set_tmp_data(node_tmp, j_dst, i_buf);
+
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_tmp);
+
+            ggml_tensor * node_red = get_node_aux(node_dst);
+            node_red->view_src = node_dst->view_src == nullptr ? node_dst : node_dst->view_src;
+            node_red->view_offs = node_dst->view_offs;
+            node_red->op = GGML_OP_ADD;
+            node_red->src[0] = node_dst;
+            node_red->src[1] = node_tmp;
+            node_red->flags |= GGML_TENSOR_FLAG_COMPUTE;
+            ggml_backend_view_init(node_red);
+
+            ggml_cgraph * cgraph_aux = get_cgraph_aux();
+            cgraph_aux->nodes[0] = node_red;
+            cgraph_aux->n_nodes = 1;
+            step_cgraphs[j_dst] = cgraph_aux;
+        };
+
+        size_t offset_j = n_backends/2;
+        while ((offset_j & (offset_j - 1)) != 0) {
+            offset_j--;
+        }
+        const size_t offset_j_max = offset_j;
+        size_t i_buf = 0;
+
+        // If n_backends is not a power of 2, fold in the excess prior to butterfly reduction:
+        for (size_t j_src = 2*offset_j_max; j_src < n_backends; j_src++) {
+            const size_t j_dst = j_src - 2*offset_j_max;
+            push_data(j_src, j_dst, i_buf);
+            const ggml_status status = ggml_backend_graph_compute_async(backend_ctx->backend_configs[j_dst].backend, step_cgraphs[j_dst]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+            i_buf = 1;
+        }
+
+        // Butterfly reduction:
+        for (; offset_j >= 1; offset_j /= 2) {
             std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
 
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 const size_t j_other = j ^ offset_j;
-                if (j_other > j) {
+                if (j_other >= n_backends) {
                     continue;
                 }
-
-                auto & bcj1 = backend_ctx->backend_configs[j];
-                auto & bcj2 = backend_ctx->backend_configs[j_other];
-
-                ggml_tensor * node1 = bcj1.cgraphs[i].cgraph_main->nodes[bcj1.cgraphs[i].cgraph_main->n_nodes - 1];
-                ggml_tensor * node2 = bcj2.cgraphs[i].cgraph_main->nodes[bcj2.cgraphs[i].cgraph_main->n_nodes - 1];
-                GGML_ASSERT(ggml_is_contiguous(node1));
-                GGML_ASSERT(ggml_is_contiguous(node2));
-
-                // Tmp tensors to receive P2P copies
-                ggml_tensor * node_tmp_1 = get_node_aux(node1);
-                node_tmp_1->buffer = bcj1.buf.get();
-                node_tmp_1->data = ggml_backend_buffer_get_base(bcj1.buf.get());
-
-                ggml_tensor * node_tmp_2 = get_node_aux(node2);
-                node_tmp_2->buffer = bcj2.buf.get();
-                node_tmp_2->data = ggml_backend_buffer_get_base(bcj2.buf.get());
-
-                // 2 P2P copies: exchange full buffers
-                ggml_backend_tensor_copy_async(bcj1.backend, bcj2.backend, node1, node_tmp_2);
-                ggml_backend_tensor_copy_async(bcj2.backend, bcj1.backend, node2, node_tmp_1);
-
-                // Local ADD: node1 += tmp1 (in-place via view)
-                ggml_tensor * node_red_1 = get_node_aux(node1);
-                node_red_1->view_src = node1->view_src == nullptr ? node1 : node1->view_src;
-                node_red_1->view_offs = node1->view_offs;
-                node_red_1->op = GGML_OP_ADD;
-                node_red_1->src[0] = node1;
-                node_red_1->src[1] = node_tmp_1;
-                node_red_1->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_1);
-
-                // Local ADD: node2 += tmp2 (in-place via view)
-                ggml_tensor * node_red_2 = get_node_aux(node2);
-                node_red_2->view_src = node2->view_src == nullptr ? node2 : node2->view_src;
-                node_red_2->view_offs = node2->view_offs;
-                node_red_2->op = GGML_OP_ADD;
-                node_red_2->src[0] = node2;
-                node_red_2->src[1] = node_tmp_2;
-                node_red_2->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_2);
-
-                // Build 1-node cgraphs for the ADD ops
-                ggml_cgraph * cgraph_aux_1 = get_cgraph_aux();
-                cgraph_aux_1->nodes[0] = node_red_1;
-                cgraph_aux_1->n_nodes = 1;
-                step_cgraphs[j] = cgraph_aux_1;
-
-                ggml_cgraph * cgraph_aux_2 = get_cgraph_aux();
-                cgraph_aux_2->nodes[0] = node_red_2;
-                cgraph_aux_2->n_nodes = 1;
-                step_cgraphs[j_other] = cgraph_aux_2;
+                push_data(j, j_other, i_buf);
             }
 
-            // Execute local ADDs for this step
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 if (step_cgraphs[j] == nullptr) {
                     continue;
                 }
@@ -1893,12 +2027,25 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                     return status;
                 }
             }
+            i_buf++;
         }
+        assert(i_buf == backend_ctx->n_reduce_steps);
+
+        // If n_backends is not a power of 2, copy back the reduced tensors to the excess:
+        for (size_t j = 2*offset_j_max; j < n_backends; j++) {
+            auto & bcj_src = backend_ctx->backend_configs[j - 2*offset_j_max];
+            auto & bcj_dst = backend_ctx->backend_configs[j];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_dst);
+        }
+
         return GGML_STATUS_SUCCESS;
     };
 
 
-    for (size_t i = 0; i < n_subgraphs; i++) {
+    for (size_t i = 0; i < backend_ctx->n_subgraphs; i++) {
         for (size_t j = 0; j < n_backends; j++) {
             auto & bcj = backend_ctx->backend_configs[j];
             const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, bcj.cgraphs[i].cgraph_main);
@@ -1907,7 +2054,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
             }
         }
 
-        if (n_backends > 1 && i < n_subgraphs - 1) {
+        if (n_backends > 1 && i < backend_ctx->n_subgraphs - 1) {
             bool backend_allreduce_success = false;
             if (backend_ctx->comm_ctx) {
                 std::vector<ggml_tensor *> nodes;

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -25,6 +25,7 @@
 #include "ggml-impl.h"
 #include "ggml.h"
 
+
 #include <aclnnop/aclnn_add.h>
 #include <aclnnop/aclnn_add_rms_norm.h>
 #include <aclnnop/aclnn_addcdiv.h>
@@ -45,7 +46,9 @@
 #include <aclnnop/aclnn_fused_infer_attention_score_v2.h>
 #include <aclnnop/aclnn_ger.h>
 #include <aclnnop/aclnn_group_norm.h>
+#include <aclnnop/aclnn_gather_v2.h>
 #include <aclnnop/aclnn_grouped_matmul_v3.h>
+#include <aclnnop/aclnn_scatter.h>
 #include <aclnnop/aclnn_gt_scalar.h>
 #include <aclnnop/aclnn_im2col.h>
 #include <aclnnop/aclnn_index_copy.h>
@@ -62,6 +65,7 @@
 #include <aclnnop/aclnn_permute.h>
 #include <aclnnop/aclnn_pow.h>
 #include <aclnnop/aclnn_pow_tensor_tensor.h>
+#include <aclnnop/aclnn_recurrent_gated_delta_rule.h>
 #include <aclnnop/aclnn_reduce_sum.h>
 #include <aclnnop/aclnn_reflection_pad1d.h>
 #include <aclnnop/aclnn_repeat.h>
@@ -69,11 +73,15 @@
 #include <aclnnop/aclnn_rms_norm.h>
 #include <aclnnop/aclnn_roll.h>
 #include <aclnnop/aclnn_softmax.h>
+#include <aclnnop/aclnn_softmax_cross_entropy_with_logits.h>
 #include <aclnnop/aclnn_sub.h>
 #include <aclnnop/aclnn_sum.h>
 #include <aclnnop/aclnn_threshold.h>
 #include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triangular_solve.h>
 #include <aclnnop/aclnn_triu.h>
+#include <aclnnop/aclnn_logical_not.h>
+#include <aclnnop/aclnn_masked_fill_scalar.h>
 #include <aclnnop/aclnn_upsample_nearest_2d.h>
 #include <aclnnop/aclnn_weight_quant_batch_matmul_v2.h>
 #include <aclnnop/aclnn_zero.h>
@@ -151,6 +159,107 @@ void ggml_cann_op_unary_gated(std::function<void(ggml_backend_cann_context &, ac
     GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst.get(), acl_src1.get());
 }
 
+// Fused SwiGLU using aclnnSwiGlu: splits input along innermost dim, applies
+// SiLU to left half, multiplies by right half.
+//
+// Falls back to the generic two-kernel path when src[1] != nullptr (two
+// independent halves) or swapped != 0 (reversed activation order), as
+// aclnnSwiGlu only handles the single interleaved tensor in standard order.
+//
+// CANN tiling for SwiGlu requires (storageShapeDim + viewDims) to be even.
+// aclCreateTensor always uses storageShapeDim=1, so viewDims must be odd.
+// We use a 3D view (1+3=4, even) to satisfy this constraint while preserving
+// correct split semantics along the innermost (ne[0]) dimension.
+void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    auto silu_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Silu, acl_src, acl_dst);
+    };
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    if (dst->src[1] != nullptr || swapped != 0) {
+        ggml_cann_op_unary_gated(silu_fn, ctx, dst);
+        return;
+    }
+
+    // aclnnSwiGlu requires the split dim (src->ne[0]) to be even; fall back otherwise.
+    if (dst->src[0]->ne[0] % 2 != 0) {
+        ggml_cann_op_unary_gated(silu_fn, ctx, dst);
+        return;
+    }
+
+    ggml_tensor * src0 = dst->src[0];
+    size_t elem_size = ggml_element_size(src0);
+
+    // src0 GGML: [2*ne0, ne1, ne2, ne3] → 3D view [2*ne0, ne1, ne2*ne3]
+    // CANN reversed: [ne2*ne3, ne1, 2*ne0], split along CANN dim 2 (last).
+    int64_t ne0_x2   = src0->ne[0];
+    int64_t ne1      = src0->ne[1];
+    int64_t ne23     = src0->ne[2] * src0->ne[3];
+    int64_t src3d_ne[] = { ne0_x2, ne1, ne23 };
+    size_t  src3d_nb[] = { (size_t)src0->nb[0], (size_t)src0->nb[1], (size_t)src0->nb[2] };
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0->data, ggml_cann_type_mapping(src0->type),
+                                                     elem_size, src3d_ne, src3d_nb, 3);
+
+    // dst GGML: [ne0, ne1, ne2, ne3] → 3D view [ne0, ne1, ne2*ne3]
+    int64_t ne0      = dst->ne[0];
+    int64_t dst3d_ne[] = { ne0, ne1, ne23 };
+    size_t  dst3d_nb[] = { (size_t)dst->nb[0], (size_t)dst->nb[1], (size_t)dst->nb[2] };
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ggml_cann_type_mapping(dst->type),
+                                                     elem_size, dst3d_ne, dst3d_nb, 3);
+
+    // CANN tensor [ne23, ne1, 2*ne0]: split along CANN dim 2 (last) = 2*ne0.
+    GGML_CANN_CALL_ACLNN_OP(ctx, SwiGlu, acl_src.get(), (int64_t)2, acl_dst.get());
+}
+
+// Fused GeGLU using aclnnGeGluV3: splits input along ne[0] (CANN last dim),
+// activates the LEFT half with GELU, multiplies by right half.
+// approximate: 0=tanh, 1=none(erf). activateLeft=true matches GGML convention.
+// outGelu is a required-but-discard output buffer.
+//
+// Falls back to the generic two-kernel path when src[1] != nullptr (two
+// independent halves) or swapped != 0 (reversed activation order), as
+// aclnnGeGluV3 only handles the single interleaved tensor in standard order.
+void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate) {
+    auto gelu_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, Gelu, acl_src, acl_dst);
+    };
+
+    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
+    if (dst->src[1] != nullptr || swapped != 0) {
+        ggml_cann_op_unary_gated(gelu_fn, ctx, dst);
+        return;
+    }
+
+    // aclnnGeGluV3 requires the split dim (src->ne[0]) to be even; fall back otherwise.
+    if (dst->src[0]->ne[0] % 2 != 0) {
+        ggml_cann_op_unary_gated(gelu_fn, ctx, dst);
+        return;
+    }
+
+    ggml_tensor * src0 = dst->src[0];
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+
+    // Allocate a temporary buffer for the required outGelu output (same shape as dst).
+    // Build contiguous strides since the pool allocation is a fresh buffer.
+    size_t  elem_size    = ggml_element_size(dst);
+    int64_t ne[GGML_MAX_DIMS] = { dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3] };
+    size_t  nb[GGML_MAX_DIMS];
+    nb[0] = elem_size;
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        nb[i] = nb[i - 1] * ne[i - 1];
+    }
+    size_t gelu_out_size = nb[GGML_MAX_DIMS - 1] * ne[GGML_MAX_DIMS - 1];
+    ggml_cann_pool_alloc gelu_out_alloc(ctx.pool(), gelu_out_size);
+
+    acl_tensor_ptr acl_gelu_out = ggml_cann_create_tensor(
+        gelu_out_alloc.get(), ggml_cann_type_mapping(dst->type), elem_size, ne, nb, GGML_MAX_DIMS);
+    // V3 adds activateLeft param; true → Gelu(left)*right, matching GGML convention.
+    // GGML dim 0 → CANN last dim (index GGML_MAX_DIMS-1 = 3 for 4D tensor).
+    GGML_CANN_CALL_ACLNN_OP(ctx, GeGluV3, acl_src.get(), (int64_t)(GGML_MAX_DIMS - 1), approximate, true,
+                             acl_dst.get(), acl_gelu_out.get());
+}
+
 /**
  * @brief Repeats elements of a tensor along each dimension according to the
  * specified repeat array.
@@ -445,28 +554,33 @@ void ggml_cann_l2_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_cann_pool_alloc temp_buffer_allocator(ctx.pool(), n_bytes);
     void *               buffer = temp_buffer_allocator.get();
 
-    int64_t div_ne[] = { 1, src->ne[1], src->ne[2], src->ne[3] };
-    size_t  div_nb[GGML_MAX_DIMS];
-    div_nb[0] = sizeof(float);
+    int64_t norm_ne[] = { 1, src->ne[1], src->ne[2], src->ne[3] };
+    size_t  norm_nb[GGML_MAX_DIMS];
+    norm_nb[0] = sizeof(float);
     for (int i = 1; i < GGML_MAX_DIMS; ++i) {
-        div_nb[i] = div_nb[i - 1] * div_ne[i - 1];
+        norm_nb[i] = norm_nb[i - 1] * norm_ne[i - 1];
     }
-    acl_tensor_ptr acl_div = ggml_cann_create_tensor(buffer, ACL_FLOAT, type_size, div_ne, div_nb, GGML_MAX_DIMS);
+    acl_tensor_ptr acl_norm = ggml_cann_create_tensor(buffer, ACL_FLOAT, sizeof(float), norm_ne, norm_nb, GGML_MAX_DIMS);
 
     std::vector<int64_t> norm_dims  = { 3 };
     acl_int_array_ptr    dims_array = ggml_cann_create_int_array(norm_dims.data(), norm_dims.size());
 
     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());
+    GGML_CANN_CALL_ACLNN_OP(ctx, Norm, acl_src.get(), p_scalar.get(), dims_array.get(), true, acl_norm.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_pool_alloc clamp_buffer_allocator(ctx.pool());
+    acl_tensor_ptr       acl_clamped;
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src.get(), acl_div.get(), acl_dst.get());
+    if (eps > 0.0f) {
+        void *         clamp_buf  = clamp_buffer_allocator.alloc(n_bytes);
+        acl_clamped               = ggml_cann_create_tensor(clamp_buf, ACL_FLOAT, sizeof(float), norm_ne, norm_nb, GGML_MAX_DIMS);
+        acl_scalar_ptr eps_scalar = ggml_cann_create_scalar(&eps, aclDataType::ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, ClampMin, acl_norm.get(), eps_scalar.get(), acl_clamped.get());
+    }
+
+    aclTensor * acl_div_input = acl_clamped ? acl_clamped.get() : acl_norm.get();
+    GGML_CANN_CALL_ACLNN_OP(ctx, Div, acl_src.get(), acl_div_input, acl_dst.get());
 }
 
 void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -482,56 +596,30 @@ void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor *
     logits_nb[1]              = logits_nb[0] * logits_ne[0];
     acl_tensor_ptr acl_logits = ggml_cann_create_tensor(src0->data, ACL_FLOAT, sizeof(float), logits_ne, logits_nb, 2);
 
-    size_t               log_softmax_type_size = sizeof(float);
-    int64_t              log_softmax_n_bytes   = nr * nc * log_softmax_type_size;
-    ggml_cann_pool_alloc log_softmax_allocator(ctx.pool(), log_softmax_n_bytes);
-    void *               log_softmax_buffer = log_softmax_allocator.get();
-
-    int64_t log_softmax_ne[] = { nc, nr };
-    size_t  log_softmax_nb[2];
-    log_softmax_nb[0]              = log_softmax_type_size;
-    log_softmax_nb[1]              = log_softmax_nb[0] * log_softmax_ne[0];
-    acl_tensor_ptr acl_log_softmax = ggml_cann_create_tensor(log_softmax_buffer, ACL_FLOAT, log_softmax_type_size,
-                                                             log_softmax_ne, log_softmax_nb, 2);
-
-    GGML_CANN_CALL_ACLNN_OP(ctx, LogSoftmax, acl_logits.get(), 1, acl_log_softmax.get());
-
     int64_t labels_ne[] = { nc, nr };
     size_t  labels_nb[2];
     labels_nb[0]              = ggml_type_size(src1->type);
     labels_nb[1]              = labels_nb[0] * labels_ne[0];
     acl_tensor_ptr acl_labels = ggml_cann_create_tensor(src1->data, ACL_FLOAT, sizeof(float), labels_ne, labels_nb, 2);
 
-    size_t               mul_type_size = sizeof(float);
-    int64_t              mul_n_bytes   = nr * nc * mul_type_size;
-    ggml_cann_pool_alloc mul_allocator(ctx.pool(), mul_n_bytes);
-    void *               mul_buffer = mul_allocator.get();
+    size_t               loss_per_sample_type_size = sizeof(float);
+    int64_t              loss_per_sample_n_bytes   = nr * loss_per_sample_type_size;
+    ggml_cann_pool_alloc loss_per_sample_allocator(ctx.pool(), loss_per_sample_n_bytes);
+    void *               loss_per_sample_buffer = loss_per_sample_allocator.get();
 
-    int64_t mul_ne[] = { nc, nr };
-    size_t  mul_nb[2];
-    mul_nb[0]                     = mul_type_size;
-    mul_nb[1]                     = mul_nb[0] * mul_ne[0];
-    acl_tensor_ptr acl_mul_result = ggml_cann_create_tensor(mul_buffer, ACL_FLOAT, mul_type_size, mul_ne, mul_nb, 2);
+    int64_t loss_per_sample_ne[] = { nr };
+    size_t  loss_per_sample_nb[1];
+    loss_per_sample_nb[0] = loss_per_sample_type_size;
+    acl_tensor_ptr acl_loss_per_sample = ggml_cann_create_tensor(
+        loss_per_sample_buffer, ACL_FLOAT, loss_per_sample_type_size, loss_per_sample_ne, loss_per_sample_nb, 1);
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, Mul, acl_log_softmax.get(), acl_labels.get(), acl_mul_result.get());
+    size_t               backprop_n_bytes = nr * nc * sizeof(float);
+    ggml_cann_pool_alloc backprop_allocator(ctx.pool(), backprop_n_bytes);
+    void *               backprop_buffer = backprop_allocator.get();
+    acl_tensor_ptr acl_backprop = ggml_cann_create_tensor(backprop_buffer, ACL_FLOAT, sizeof(float), logits_ne, logits_nb, 2);
 
-    size_t               sum_per_sample_type_size = sizeof(float);
-    int64_t              sum_per_sample_n_bytes   = nr * sum_per_sample_type_size;
-    ggml_cann_pool_alloc sum_per_sample_allocator(ctx.pool(), sum_per_sample_n_bytes);
-    void *               sum_per_sample_buffer = sum_per_sample_allocator.get();
-
-    int64_t sum_per_sample_ne[] = { nr };
-    size_t  sum_per_sample_nb[1];
-    sum_per_sample_nb[0]              = sum_per_sample_type_size;
-    acl_tensor_ptr acl_sum_per_sample = ggml_cann_create_tensor(
-        sum_per_sample_buffer, ACL_FLOAT, sum_per_sample_type_size, sum_per_sample_ne, sum_per_sample_nb, 1);
-
-    std::vector<int64_t> sum_dims   = { 1 };
-    acl_int_array_ptr    dims_array = ggml_cann_create_int_array(sum_dims.data(), sum_dims.size());
-    bool                 keep_dims  = false;
-
-    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_mul_result.get(), dims_array.get(), keep_dims, ACL_FLOAT,
-                            acl_sum_per_sample.get());
+    GGML_CANN_CALL_ACLNN_OP(ctx, SoftmaxCrossEntropyWithLogits, acl_logits.get(), acl_labels.get(),
+                            acl_loss_per_sample.get(), acl_backprop.get());
 
     size_t               total_sum_type_size = sizeof(float);
     int64_t              total_sum_n_bytes   = 1 * total_sum_type_size;
@@ -547,11 +635,12 @@ void ggml_cann_cross_entropy_loss(ggml_backend_cann_context & ctx, ggml_tensor *
 
     std::vector<int64_t> total_sum_dims    = { 0 };
     acl_int_array_ptr total_sum_dims_array = ggml_cann_create_int_array(total_sum_dims.data(), total_sum_dims.size());
+    bool              keep_dims            = false;
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_sum_per_sample.get(), total_sum_dims_array.get(), keep_dims, ACL_FLOAT,
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_loss_per_sample.get(), total_sum_dims_array.get(), keep_dims, ACL_FLOAT,
                             acl_total_sum.get());
 
-    float          value        = -1.0f / static_cast<float>(nr);
+    float          value        = 1.0f / static_cast<float>(nr);
     acl_scalar_ptr scale_factor = ggml_cann_create_scalar(&value, aclDataType::ACL_FLOAT);
     acl_tensor_ptr acl_dst =
         ggml_cann_create_tensor(dst->data, ACL_FLOAT, sizeof(float), total_sum_ne, total_sum_nb, 1);
@@ -589,6 +678,33 @@ void ggml_cann_group_norm(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
                             acl_mean_out.get(), acl_rstd_out.get());
 }
 
+void ggml_cann_set(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+    ggml_tensor * src1 = dst->src[1];
+
+    size_t nb1     = ((int32_t *) dst->op_params)[0];
+    size_t nb2     = ((int32_t *) dst->op_params)[1];
+    size_t nb3     = ((int32_t *) dst->op_params)[2];
+    size_t offset  = ((int32_t *) dst->op_params)[3];
+    bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
+
+    size_t param_nb[] = { ggml_element_size(src0), nb1, nb2, nb3 };
+
+    // Create a view of dst at the target offset with src1's dimensions
+    acl_tensor_ptr acl_dst  = ggml_cann_create_tensor(dst, src1->ne, param_nb, GGML_MAX_DIMS, ACL_FORMAT_ND, offset);
+    acl_tensor_ptr acl_src1 = ggml_cann_create_tensor(src1);
+
+    if (!inplace) {
+        // First copy src0 to dst entirely
+        size_t cpy_size = ggml_nbytes(dst);
+        ACL_CHECK(
+            aclrtMemcpyAsync(dst->data, cpy_size, src0->data, cpy_size, ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // Copy src1 into the target region of dst
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCopy, acl_dst.get(), acl_src1.get());
+}
+
 void ggml_cann_acc(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
@@ -652,6 +768,113 @@ void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     aclnn_reduce_sum(ctx, dst, reduce_dims, 4);
 }
 
+void ggml_cann_cumsum(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    // GGML cumsum operates along dim 0 (innermost / ne[0]).
+    // ggml_cann_create_tensor reverses dimensions to [ne3,ne2,ne1,ne0],
+    // so GGML dim 0 maps to CANN dim 3 (the last dim of the 4-D tensor).
+    GGML_CANN_CALL_ACLNN_OP(ctx, Cumsum, acl_src.get(), (int64_t)3,
+                            ggml_cann_type_mapping(dst->type), acl_dst.get());
+}
+
+void ggml_cann_solve_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];  // A: [N, N, B2, B3] lower triangular
+    ggml_tensor * src1 = dst->src[1];  // B: [K, N, B2, B3]
+
+    acl_tensor_ptr acl_a = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_b = ggml_cann_create_tensor(src1);
+    acl_tensor_ptr acl_x = ggml_cann_create_tensor(dst);
+
+    // mOut: triangular copy of A (required output), same shape as A.
+    const size_t a_bytes = ggml_nbytes(src0);
+    ggml_cann_pool_alloc m_alloc(ctx.pool(), a_bytes);
+    acl_tensor_ptr acl_m = ggml_cann_create_tensor(
+        m_alloc.get(), ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
+
+    // Solve AX = B: upper=false (lower tri), transpose=false, unitriangular=false.
+    GGML_CANN_CALL_ACLNN_OP(ctx, TriangularSolve,
+        acl_b.get(), acl_a.get(), false, false, false,
+        acl_x.get(), acl_m.get());
+}
+
+void ggml_cann_diag(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+
+    GGML_ASSERT(src->ne[1] == 1);
+
+    const int64_t N       = src->ne[0];
+    const int64_t n_batch = src->ne[2] * src->ne[3];
+    const size_t  nb_f32  = sizeof(float);
+
+    // Fill dst with zeros.
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    {
+        float          zero = 0.0f;
+        acl_scalar_ptr acl_zero = ggml_cann_create_scalar(&zero, ACL_FLOAT);
+        GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst.get(), acl_zero.get());
+    }
+
+    // Copy src vector onto the diagonal of dst via strided views.
+    // src viewed as [N, n_batch], contiguous strides.
+    int64_t ne_vec[2]      = { N, n_batch };
+    size_t  nb_src_vec[2]  = { nb_f32, N * nb_f32 };
+    // dst diagonal view: stride (N+1)*4 steps along the diagonal.
+    size_t  nb_dst_diag[2] = { (N + 1) * nb_f32, N * N * nb_f32 };
+
+    acl_tensor_ptr acl_src_vec  = ggml_cann_create_tensor(src->data, ACL_FLOAT, nb_f32, ne_vec, nb_src_vec, 2);
+    acl_tensor_ptr acl_dst_diag = ggml_cann_create_tensor(dst->data, ACL_FLOAT, nb_f32, ne_vec, nb_dst_diag, 2);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceCopy, acl_dst_diag.get(), acl_src_vec.get());
+}
+
+void ggml_cann_fill(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    float c = ggml_get_op_params_f32(dst, 0);
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+    acl_scalar_ptr acl_c   = ggml_cann_create_scalar(&c, ACL_FLOAT);
+    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceFillScalar, acl_dst.get(), acl_c.get());
+}
+
+void ggml_cann_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src = dst->src[0];
+
+    const int64_t S       = src->ne[0];
+    const int64_t n_batch = src->ne[2] * src->ne[3];
+    const size_t  nb_f32  = sizeof(float);
+
+    int64_t ne3d[3] = { S, S, n_batch };
+    size_t  nb3d[3] = { nb_f32, S * nb_f32, S * S * nb_f32 };
+
+    const ggml_tri_type ttype = (ggml_tri_type) ggml_get_op_params_i32(dst, 0);
+
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src->data, ACL_FLOAT, nb_f32, ne3d, nb3d, 3);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ACL_FLOAT, nb_f32, ne3d, nb3d, 3);
+
+    switch (ttype) {
+        case GGML_TRI_TYPE_LOWER:
+            // Tril(-1): preserve row > col (strict lower), zero upper + diagonal.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Tril, acl_src.get(), (int64_t)-1, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_UPPER_DIAG:
+            // Triu(0): preserve row <= col (upper + diagonal), zero strict lower.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Triu, acl_src.get(), (int64_t)0, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_UPPER:
+            // Triu(1): preserve row < col (strict upper), zero lower + diagonal.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Triu, acl_src.get(), (int64_t)1, acl_dst.get());
+            break;
+        case GGML_TRI_TYPE_LOWER_DIAG:
+            // Tril(0): preserve row >= col (lower + diagonal), zero strict upper.
+            GGML_CANN_CALL_ACLNN_OP(ctx, Tril, acl_src.get(), (int64_t)0, acl_dst.get());
+            break;
+        default:
+            GGML_ABORT("unsupported tri type");
+    }
+}
+
 void ggml_cann_upsample_nearest2d(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor *  src     = dst->src[0];
     acl_tensor_ptr acl_src = ggml_cann_create_tensor(src, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
@@ -1695,152 +1918,90 @@ void ggml_cann_softmax(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     aclnn_softmax(ctx, softmax_tensor.get(), 3, acl_dst.get());
 }
 
-/**
- * @brief Performs index select operation on a 4D tensor using the CANN backend.
- *
- * This function applies the `IndexSelect` operation along a specific dimension
- * of the source tensor (`src_buffer`) using the indices from the index tensor (`index`).
- * It iterates over the last two dimensions of the source tensor, creates the corresponding
- * CANN tensors for the source, index, and output slices, and executes the `IndexSelect`
- * operation for each slice.
- *
- * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer containing the 4D input tensor data.
- * @param src_ne The dimensions of the source tensor.
- * @param src_nb The strides (byte offsets) of the source tensor.
- * @param dst_buffer The destination buffer where the output tensor data will be written.
- * @param dst_ne The dimensions of the destination tensor.
- * @param dst_nb The strides (byte offsets) of the destination tensor.
- * @param index The index tensor specifying the indices to select from the source tensor.
- * @param type The data type of the source and destination tensors.
- */
-static void aclnn_index_select_4d(ggml_backend_cann_context & ctx,
-                                  void *                      src_buffer,
-                                  int64_t *                   src_ne,
-                                  size_t *                    src_nb,
-                                  void *                      dst_buffer,
-                                  int64_t *                   dst_ne,
-                                  size_t *                    dst_nb,
-                                  ggml_tensor *               index,
-                                  ggml_type                   type) {
-    for (int64_t i = 0; i < src_ne[3]; i++) {
-        for (int64_t j = 0; j < src_ne[2]; j++) {
-            // src
-            acl_tensor_ptr acl_src_tensor =
-                ggml_cann_create_tensor((char *) src_buffer + i * src_nb[3] + j * src_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), src_ne, src_nb, 2);
-
-            // index
-            acl_tensor_ptr acl_index = ggml_cann_create_tensor(
-                (char *) index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
-                ggml_cann_type_mapping(index->type), ggml_element_size(index), index->ne, index->nb, 1);
-
-            // out
-            acl_tensor_ptr acl_out =
-                ggml_cann_create_tensor((char *) dst_buffer + i * dst_nb[3] + j * dst_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), dst_ne, dst_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, IndexSelect, acl_src_tensor.get(), 0, acl_index.get(), acl_out.get());
-        }
-    }
-}
-
-/**
- * @brief Performs inplace index copy operation on a 4D tensor using the CANN backend.
- *
- * This function applies the `IndexCopy` operation along a specific dimension of the
- * destination tensor (`dst_buffer`) by copying elements from the source tensor (`src_buffer`)
- * to positions specified by the index tensor (`index`).
- * It iterates over the last two dimensions of the tensors, creates the corresponding
- * CANN tensors for source, index, and destination slices, and performs the index copy
- * operation for each slice.
- *
- * @param ctx The context for CANN backend operations.
- * @param src_buffer The source buffer containing the 4D input tensor data to be copied.
- * @param src_ne The dimensions of the source tensor.
- * @param src_nb The strides (byte offsets) of the source tensor.
- * @param dst_buffer The destination buffer where values will be copied to.
- * @param dst_ne The dimensions of the destination tensor.
- * @param dst_nb The strides (byte offsets) of the destination tensor.
- * @param index The index tensor specifying target positions in the destination tensor.
- * @param type The data type of the source and destination tensors.
- */
-static void aclnn_index_copy_4d(ggml_backend_cann_context & ctx,
-                                void *                      src_buffer,
-                                int64_t *                   src_ne,
-                                size_t *                    src_nb,
-                                void *                      dst_buffer,
-                                int64_t *                   dst_ne,
-                                size_t *                    dst_nb,
-                                ggml_tensor *               index,
-                                ggml_type                   type) {
-    for (int64_t i = 0; i < src_ne[3]; i++) {
-        for (int64_t j = 0; j < src_ne[2]; j++) {
-            // src
-            acl_tensor_ptr acl_src_tensor =
-                ggml_cann_create_tensor((char *) src_buffer + i * src_nb[3] + j * src_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), src_ne, src_nb, 2);
-
-            // index
-            acl_tensor_ptr acl_index = ggml_cann_create_tensor(
-                (char *) index->data + (i % index->ne[2]) * index->nb[2] + (j % index->ne[1]) * index->nb[1],
-                ggml_cann_type_mapping(index->type), ggml_element_size(index), index->ne, index->nb, 1);
-
-            // out
-            acl_tensor_ptr acl_out =
-                ggml_cann_create_tensor((char *) dst_buffer + i * dst_nb[3] + j * dst_nb[2],
-                                        ggml_cann_type_mapping(type), ggml_type_size(type), dst_ne, dst_nb, 2);
-            GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_out.get(), 0, acl_index.get(), acl_src_tensor.get());
-        }
-    }
-}
 
 void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // src
+    ggml_tensor * src0 = dst->src[0];  // weight
     ggml_tensor * src1 = dst->src[1];  // index
 
     GGML_ASSERT(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16
                 || dst->type == GGML_TYPE_BF16);
 
+    // n_idx: number of row indices per (i2, i3) batch slice.
+    // ggml guarantees: src0->ne[2] == src1->ne[1], src0->ne[3] == src1->ne[2], src1->ne[3] == 1.
+    const int64_t n_idx = src1->ne[0];
+
+    // Gather all (i2, i3) batch slices from src into dst.
+    // ggml_cann_create_tensor reverses dims, so ACL sees [ne1, ne0].
+    // GatherV2 with dim=0 gathers along ACL dim-0 == ggml ne[1] (the vocabulary / row axis).
+    // nb: the 4 strides of the source buffer (nb[0..1] for the 2D slice shape,
+    //     nb[2..3] for computing per-batch-slice base pointer offsets).
+    auto gather_batched = [&](void * src_base, aclDataType acl_type, size_t type_size,
+                              const size_t * nb) {
+        int64_t src_ne[2]  = { src0->ne[0], src0->ne[1] };
+        size_t  src_nb_2d[2] = { nb[0], nb[1] };
+        int64_t dst_ne[2]  = { src0->ne[0], n_idx };
+        size_t  dst_nb_2d[2] = { dst->nb[0], dst->nb[1] };
+        int64_t idx_ne[1]  = { n_idx };
+        size_t  idx_nb[1]  = { (size_t)ggml_element_size(src1) };
+
+        for (int64_t i3 = 0; i3 < src0->ne[3]; i3++) {
+            for (int64_t i2 = 0; i2 < src0->ne[2]; i2++) {
+                acl_tensor_ptr acl_src = ggml_cann_create_tensor(
+                    (char *)src_base + i3 * nb[3] + i2 * nb[2],
+                    acl_type, type_size, src_ne, src_nb_2d, 2);
+                acl_tensor_ptr acl_idx = ggml_cann_create_tensor(
+                    (char *)src1->data + i3 * src1->nb[2] + i2 * src1->nb[1],
+                    ggml_cann_type_mapping(src1->type), (size_t)ggml_element_size(src1),
+                    idx_ne, idx_nb, 1);
+                acl_tensor_ptr acl_dst = ggml_cann_create_tensor(
+                    (char *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2],
+                    acl_type, type_size, dst_ne, dst_nb_2d, 2);
+                GGML_CANN_CALL_ACLNN_OP(ctx, GatherV2, acl_src.get(), 0, acl_idx.get(), acl_dst.get());
+            }
+        }
+    };
+
     switch (src0->type) {
         case GGML_TYPE_BF16:
         case GGML_TYPE_F16:
         case GGML_TYPE_F32:
             if (src0->type == dst->type) {
-                aclnn_index_select_4d(ctx, src0->data, src0->ne, src0->nb, dst->data, dst->ne, dst->nb, src1,
-                                      dst->type);
+                gather_batched(src0->data,
+                               ggml_cann_type_mapping(src0->type), ggml_type_size(src0->type),
+                               src0->nb);
             } else {
-                acl_tensor_ptr       acl_src0 = ggml_cann_create_tensor(src0);
-                ggml_cann_pool_alloc src_buffer_allocator(ctx.pool(), ggml_nelements(src0) * ggml_element_size(dst));
-                void *               src_trans_buffer = src_buffer_allocator.get();
-                size_t               src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = dst->nb[0];
+                // Cast src0 to dst type, then gather.
+                ggml_cann_pool_alloc src_cast_allocator(ctx.pool(),
+                                                        ggml_nelements(src0) * ggml_element_size(dst));
+                size_t src_cast_nb[GGML_MAX_DIMS];
+                src_cast_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                    src_cast_nb[i] = src_cast_nb[i - 1] * src0->ne[i - 1];
                 }
-                acl_tensor_ptr src_trans_tensor =
-                    ggml_cann_create_tensor(src_trans_buffer, ggml_cann_type_mapping(dst->type),
-                                            ggml_type_size(dst->type), src0->ne, src_trans_nb, GGML_MAX_DIMS);
-                aclnn_cast(ctx, acl_src0.get(), src_trans_tensor.get(), ggml_cann_type_mapping(dst->type));
-                aclnn_index_select_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb, dst->data, dst->ne, dst->nb, src1,
-                                      dst->type);
+                acl_tensor_ptr acl_src0     = ggml_cann_create_tensor(src0);
+                acl_tensor_ptr acl_src_cast = ggml_cann_create_tensor(
+                    src_cast_allocator.get(), ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                    src0->ne, src_cast_nb, GGML_MAX_DIMS);
+                aclnn_cast(ctx, acl_src0.get(), acl_src_cast.get(), ggml_cann_type_mapping(dst->type));
+
+                gather_batched(src_cast_allocator.get(),
+                               ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                               src_cast_nb);
             }
             break;
         case GGML_TYPE_Q8_0:
             {
-                // add 1 dim for bcast mul.
+                // Dequantize Q8_0 to dst type, then gather.
                 size_t  weight_nb[GGML_MAX_DIMS + 1], scale_nb[GGML_MAX_DIMS + 1], dequant_nb[GGML_MAX_DIMS + 1];
                 int64_t weight_ne[GGML_MAX_DIMS + 1], scale_ne[GGML_MAX_DIMS + 1], *dequant_ne;
-                int64_t scale_offset = 0;
-                // [3,4,5,64] -> [3,4,5,2,32]
-                weight_ne[0]         = QK8_0;
-                weight_ne[1]         = src0->ne[0] / QK8_0;
-                weight_nb[0]         = sizeof(int8_t);
-                weight_nb[1]         = weight_nb[0] * weight_ne[0];
+                weight_ne[0] = QK8_0;
+                weight_ne[1] = src0->ne[0] / QK8_0;
+                weight_nb[0] = sizeof(int8_t);
+                weight_nb[1] = weight_nb[0] * weight_ne[0];
                 for (int i = 2; i < GGML_MAX_DIMS + 1; i++) {
                     weight_ne[i] = src0->ne[i - 1];
                     weight_nb[i] = weight_nb[i - 1] * weight_ne[i - 1];
                 }
-                // [3,4,5,64] -> [3,4,5,2,1]
                 scale_ne[0] = 1;
                 scale_ne[1] = src0->ne[0] / QK8_0;
                 scale_nb[0] = sizeof(uint16_t);
@@ -1849,31 +2010,33 @@ void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
                     scale_ne[i] = src0->ne[i - 1];
                     scale_nb[i] = scale_nb[i - 1] * scale_ne[i - 1];
                 }
-                // [3,4,5,64] -> [3,4,5,2,32]
                 dequant_ne    = weight_ne;
                 dequant_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS + 1; i++) {
                     dequant_nb[i] = dequant_nb[i - 1] * dequant_ne[i - 1];
                 }
-                scale_offset = ggml_nelements(src0) * sizeof(int8_t);
-                ggml_cann_pool_alloc dequant_buffer_allocator(ctx.pool(),
-                                                              ggml_nelements(src0) * ggml_type_size(dst->type));
-                acl_tensor_ptr       acl_weight_tensor = ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t),
-                                                                                 weight_ne, weight_nb, GGML_MAX_DIMS + 1);
-                acl_tensor_ptr       acl_scale_tensor =
-                    ggml_cann_create_tensor(src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
-                                            GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
-                acl_tensor_ptr dequant_tensor =
-                    ggml_cann_create_tensor(dequant_buffer_allocator.get(), ggml_cann_type_mapping(dst->type),
-                                            ggml_type_size(dst->type), dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
-                aclnn_mul(ctx, acl_weight_tensor.get(), acl_scale_tensor.get(), dequant_tensor.get());
-                dequant_nb[0] = ggml_type_size(dst->type);
+                const int64_t scale_offset = ggml_nelements(src0) * sizeof(int8_t);
+                ggml_cann_pool_alloc dequant_allocator(ctx.pool(),
+                                                       ggml_nelements(src0) * ggml_type_size(dst->type));
+                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(src0->data, ACL_INT8, sizeof(int8_t),
+                                                                     weight_ne, weight_nb, GGML_MAX_DIMS + 1);
+                acl_tensor_ptr acl_scale  = ggml_cann_create_tensor(
+                    src0->data, ACL_FLOAT16, sizeof(uint16_t), scale_ne, scale_nb,
+                    GGML_MAX_DIMS + 1, ACL_FORMAT_ND, scale_offset);
+                acl_tensor_ptr acl_dequant = ggml_cann_create_tensor(
+                    dequant_allocator.get(), ggml_cann_type_mapping(dst->type),
+                    ggml_type_size(dst->type), dequant_ne, dequant_nb, GGML_MAX_DIMS + 1);
+                aclnn_mul(ctx, acl_weight.get(), acl_scale.get(), acl_dequant.get());
+
+                // Reinterpret dequant buffer as 4D [src0->ne] with contiguous strides.
                 dequant_ne    = src0->ne;
+                dequant_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
                     dequant_nb[i] = dequant_nb[i - 1] * src0->ne[i - 1];
                 }
-                aclnn_index_select_4d(ctx, dequant_buffer_allocator.get(), dequant_ne, dequant_nb, dst->data, dst->ne,
-                                      dst->nb, src1, dst->type);
+                gather_batched(dequant_allocator.get(),
+                               ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                               dequant_nb);
                 break;
             }
         default:
@@ -1883,31 +2046,70 @@ void ggml_cann_get_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
 }
 
 void ggml_cann_set_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // src
-    ggml_tensor * src1 = dst->src[1];  // index
+    ggml_tensor * src0 = dst->src[0];  // source values
+    ggml_tensor * src1 = dst->src[1];  // row indices
+
+    // n_idx: number of source rows to scatter per batch slice.
+    // ggml guarantees: src0->ne[1] == src1->ne[0].
+    const int64_t n_idx = src1->ne[0];
+
+    // Copy n_idx rows of src [ne0, n_idx] into dst [ne0, ne1] at positions given by a 1D index.
+    // ggml_cann_create_tensor reverses dims, so ACL sees [ne1, ne0] for dst.
+    // InplaceIndexCopy with dim=0 copies along ACL dim-0 == ggml ne[1] (the row axis).
+    // src_nb: the 4 strides of the source buffer (nb[0..1] for the 2D slice shape,
+    //         nb[2..3] for computing per-batch-slice base pointer offsets).
+    auto scatter_batched = [&](void * src_base, aclDataType acl_type, size_t type_size,
+                               const size_t * src_nb) {
+        int64_t d_ne[2]    = { dst->ne[0], dst->ne[1] };
+        size_t  d_nb[2]    = { dst->nb[0], dst->nb[1] };
+        int64_t s_ne[2]    = { dst->ne[0], n_idx };
+        size_t  s_nb_2d[2] = { src_nb[0], src_nb[1] };
+        int64_t i_ne[1]    = { n_idx };
+        size_t  i_nb[1]    = { (size_t)ggml_element_size(src1) };
+
+        for (int64_t i3 = 0; i3 < dst->ne[3]; i3++) {
+            for (int64_t i2 = 0; i2 < dst->ne[2]; i2++) {
+                acl_tensor_ptr acl_dst = ggml_cann_create_tensor(
+                    (char *)dst->data + i3 * dst->nb[3] + i2 * dst->nb[2],
+                    acl_type, type_size, d_ne, d_nb, 2);
+                acl_tensor_ptr acl_idx = ggml_cann_create_tensor(
+                    (char *)src1->data + (i3 % src1->ne[2]) * src1->nb[2] + (i2 % src1->ne[1]) * src1->nb[1],
+                    ggml_cann_type_mapping(src1->type), (size_t)ggml_element_size(src1),
+                    i_ne, i_nb, 1);
+                acl_tensor_ptr acl_src = ggml_cann_create_tensor(
+                    (char *)src_base + i3 * src_nb[3] + i2 * src_nb[2],
+                    acl_type, type_size, s_ne, s_nb_2d, 2);
+                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceIndexCopy, acl_dst.get(), 0, acl_idx.get(), acl_src.get());
+            }
+        }
+    };
 
     switch (dst->type) {
         case GGML_TYPE_F32:
-            {
-                aclnn_index_copy_4d(ctx, src0->data, src0->ne, src0->nb, dst->data, dst->ne, dst->nb, src1, dst->type);
-                break;
-            }
+            scatter_batched(src0->data,
+                            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                            src0->nb);
+            break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
             {
-                acl_tensor_ptr       acl_src0 = ggml_cann_create_tensor(src0);
-                ggml_cann_pool_alloc src_buffer_allocator(ctx.pool(), ggml_nelements(src0) * sizeof(uint16_t));
-                void *               src_trans_buffer = src_buffer_allocator.get();
-                size_t               src_trans_nb[GGML_MAX_DIMS];
-                src_trans_nb[0] = sizeof(uint16_t);
+                // Cast src0 (F32) to dst type first.
+                ggml_cann_pool_alloc src_cast_allocator(ctx.pool(),
+                                                        ggml_nelements(src0) * ggml_type_size(dst->type));
+                size_t src_cast_nb[GGML_MAX_DIMS];
+                src_cast_nb[0] = ggml_type_size(dst->type);
                 for (int i = 1; i < GGML_MAX_DIMS; i++) {
-                    src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
+                    src_cast_nb[i] = src_cast_nb[i - 1] * src0->ne[i - 1];
                 }
-                acl_tensor_ptr src_trans_tensor = ggml_cann_create_tensor(
-                    src_trans_buffer, ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type), src0->ne, src_trans_nb, GGML_MAX_DIMS);
-                aclnn_cast(ctx, acl_src0.get(), src_trans_tensor.get(), ggml_cann_type_mapping(dst->type));
-                aclnn_index_copy_4d(ctx, src_trans_buffer, src0->ne, src_trans_nb, dst->data, dst->ne, dst->nb, src1,
-                                    dst->type);
+                acl_tensor_ptr acl_src0     = ggml_cann_create_tensor(src0);
+                acl_tensor_ptr acl_src_cast = ggml_cann_create_tensor(
+                    src_cast_allocator.get(), ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                    src0->ne, src_cast_nb, GGML_MAX_DIMS);
+                aclnn_cast(ctx, acl_src0.get(), acl_src_cast.get(), ggml_cann_type_mapping(dst->type));
+
+                scatter_batched(src_cast_allocator.get(),
+                                ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+                                src_cast_nb);
                 break;
             }
         default:
@@ -3268,29 +3470,50 @@ void ggml_cann_pad_reflect_1d(ggml_backend_cann_context & ctx, ggml_tensor * dst
     int64_t           paddingsArray[2] = { opts[0], opts[1] };
     acl_int_array_ptr paddings         = ggml_cann_create_int_array(paddingsArray, 2);
 
-    for (int64_t i = 0; i < src0->ne[3]; i++) {
-        acl_tensor_ptr acl_src =
-            ggml_cann_create_tensor((char *) src0->data + i * src0->ne[3], ggml_cann_type_mapping(src0->type),
-                                    ggml_element_size(src0), src0->ne, src0->nb, 3);
+    // Collapsing ne[2]*ne[3] into a single batch dimension requires that dim3
+    // is contiguous with respect to dim2 in both src and dst.
+    GGML_ASSERT(src0->nb[3] == src0->nb[2] * src0->ne[2]);
+    GGML_ASSERT(dst->nb[3]  == dst->nb[2]  * dst->ne[2]);
 
-        acl_tensor_ptr acl_dst =
-            ggml_cann_create_tensor((char *) dst->data + i * src0->ne[3], ggml_cann_type_mapping(dst->type),
-                                    ggml_element_size(dst), dst->ne, dst->nb, 3);
+    int64_t src_ne_3d[3] = { src0->ne[0], src0->ne[1], src0->ne[2] * src0->ne[3] };
+    int64_t dst_ne_3d[3] = { dst->ne[0],  dst->ne[1],  dst->ne[2]  * dst->ne[3]  };
 
-        GGML_CANN_CALL_ACLNN_OP(ctx, ReflectionPad1d, acl_src.get(), paddings.get(), acl_dst.get());
-    }
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0->data, ggml_cann_type_mapping(src0->type),
+                                                     ggml_element_size(src0), src_ne_3d, src0->nb, 3);
+
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst->data, ggml_cann_type_mapping(dst->type),
+                                                     ggml_element_size(dst), dst_ne_3d, dst->nb, 3);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReflectionPad1d, acl_src.get(), paddings.get(), acl_dst.get());
 }
 
 void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     ggml_tensor * src0 = dst->src[0];
     ggml_tensor * src1 = dst->src[1];
 
+    // Write element-wise equality (0 or 1) into a temporary buffer to avoid
+    // modifying src0 in-place.  Use the same type as src0 so ReduceSum can
+    // consume it directly without a type cast.
+    ggml_cann_pool_alloc eq_alloc(ctx.pool(), ggml_nelements(src0) * ggml_element_size(src0));
+    size_t eq_nb[GGML_MAX_DIMS];
+    eq_nb[0] = ggml_element_size(src0);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        eq_nb[i] = eq_nb[i - 1] * src0->ne[i - 1];
+    }
+    acl_tensor_ptr acl_eq = ggml_cann_create_tensor(
+        eq_alloc.get(), ggml_cann_type_mapping(src0->type), ggml_element_size(src0),
+        src0->ne, eq_nb, GGML_MAX_DIMS);
+
     acl_tensor_ptr acl_self  = ggml_cann_create_tensor(src0);
     acl_tensor_ptr acl_other = ggml_cann_create_tensor(src1);
+    GGML_CANN_CALL_ACLNN_OP(ctx, EqTensor, acl_self.get(), acl_other.get(), acl_eq.get());
 
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceEqTensor, acl_self.get(), acl_other.get());
-
-    ggml_cann_sum(ctx, dst);
+    // Sum the 0/1 values into dst.
+    acl_tensor_ptr    acl_dst    = ggml_cann_create_tensor(dst);
+    int64_t           dims[4]    = { 0, 1, 2, 3 };
+    acl_int_array_ptr dims_arr   = ggml_cann_create_int_array(dims, 4);
+    GGML_CANN_CALL_ACLNN_OP(ctx, ReduceSum, acl_eq.get(), dims_arr.get(), true,
+                            ggml_cann_type_mapping(dst->type), acl_dst.get());
 }
 
 void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
@@ -3306,6 +3529,27 @@ void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, GtScalar, acl_src.get(), alpha.get(), acl_dst.get());
 }
 
+void ggml_cann_softplus(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * src0 = dst->src[0];
+
+    acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
+    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
+
+    float          beta_val      = 1.0f;
+    float          threshold_val = 20.0f;
+    acl_scalar_ptr beta          = ggml_cann_create_scalar(&beta_val,      ACL_FLOAT);
+    acl_scalar_ptr threshold     = ggml_cann_create_scalar(&threshold_val, ACL_FLOAT);
+
+    GGML_CANN_CALL_ACLNN_OP(ctx, Softplus, acl_src.get(), beta.get(), threshold.get(), acl_dst.get());
+}
+
+void ggml_cann_geglu_quick(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
+    auto gelu_quick_fn = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
+        GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
+    };
+    ggml_cann_op_unary_gated(gelu_quick_fn, ctx, dst);
+}
+
 /**
  * @brief Performs expert-specific matrix multiplication (MoE) with
  * floating-point precision using the CANN backend.
@@ -3892,46 +4136,65 @@ void ggml_cann_flash_attn_ext(ggml_backend_cann_context & ctx, ggml_tensor * dst
 }
 
 static void ggml_cann_out_prod_fp(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];  // weight
-    ggml_tensor * src1 = dst->src[1];  // input
+    ggml_tensor * src0 = dst->src[0];  // weight  [ne00=m, ne01=K, ne02, ne03]
+    ggml_tensor * src1 = dst->src[1];  // input   [ne10=n, ne11=K, ne12, ne13]
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceZero, acl_dst.get());
+    // dst[i,j] = sum_k src0[i,k] * src1[j,k]  i.e. dst = src0 @ src1^T.
+    //
+    // ggml_cann_create_tensor reverses dimension order, so ACL sees:
+    //   acl_src0 slice:   ggml[m,K]  ->  ACL[K,m]
+    //   acl_src1 slice:   ggml[n,K]  ->  ACL[K,n]
+    //   acl_dst  slice:   ggml[m,n]  ->  ACL[n,m]
+    //
+    // Build a transposed view of src1 by swapping ne[0]/ne[1]:
+    //   src1_t:  ggml[K,n] (swapped strides)  ->  ACL[n,K]
+    //
+    // Matmul(src1_t [n,K], src0 [K,m]) = [n,m] = acl_dst  ✓
+    //
+    // The outer batch loop is kept because src0 may have fewer batch slices than
+    // dst (ne02 <= ne2, ne03 <= ne3): this is a strided-broadcast not supported
+    // by standard CANN Matmul broadcasting.
+
+    const aclDataType src0_acl_type = ggml_cann_type_mapping(src0->type);
+    const aclDataType src1_acl_type = ggml_cann_type_mapping(src1->type);
+    const aclDataType dst_acl_type  = ggml_cann_type_mapping(dst->type);
+    const size_t      src0_type_sz  = ggml_type_size(src0->type);
+    const size_t      src1_type_sz  = ggml_type_size(src1->type);
+    const size_t      dst_type_sz   = ggml_type_size(dst->type);
 
     const int64_t dps2 = ne2 / ne02;
     const int64_t dps3 = ne3 / ne03;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t i02 = i2 / dps2;
             const int64_t i03 = i3 / dps3;
 
-            const int64_t  i12 = i2;
-            const int64_t  i13 = i3;
-            acl_tensor_ptr accumulator =
-                ggml_cann_create_tensor((char *) dst->data + i2 * nb2 + i3 * nb3, ggml_cann_type_mapping(dst->type),
-                                        ggml_type_size(dst->type), dst->ne, dst->nb, 2);
+            // src0 2D slice at [i02, i03]: ggml [m, K] -> ACL [K, m]
+            int64_t src0_ne[2] = { ne00, ne01 };
+            size_t  src0_nb[2] = { nb00, nb01 };
+            acl_tensor_ptr acl_src0_s = ggml_cann_create_tensor(
+                (char *) src0->data + i02 * nb02 + i03 * nb03,
+                src0_acl_type, src0_type_sz, src0_ne, src0_nb, 2);
 
-            // The outer product needs to be accumulated in this dimension.
-            for (int64_t i1 = 0; i1 < ne11; i1++) {
-                acl_tensor_ptr acl_input = ggml_cann_create_tensor(
-                    (char *) src1->data + i1 * nb11 + i12 * nb12 + i13 * nb13, ggml_cann_type_mapping(src0->type),
-                    ggml_type_size(src0->type), src1->ne, src1->nb, 1);
+            // src1 transposed 2D slice at [i2, i3]: swap ne/nb -> ggml[K,n] -> ACL[n,K]
+            int64_t src1_t_ne[2] = { ne11, ne10 };
+            size_t  src1_t_nb[2] = { nb11, nb10 };
+            acl_tensor_ptr acl_src1_t = ggml_cann_create_tensor(
+                (char *) src1->data + i2 * nb12 + i3 * nb13,
+                src1_acl_type, src1_type_sz, src1_t_ne, src1_t_nb, 2);
 
-                acl_tensor_ptr acl_weight = ggml_cann_create_tensor(
-                    (char *) src0->data + i1 * nb01 + i02 * nb02 + i03 * nb03, ggml_cann_type_mapping(src0->type),
-                    ggml_type_size(src0->type), src0->ne, src0->nb, 1);
+            // dst 2D slice at [i2, i3]: ggml [m, n] -> ACL [n, m]
+            int64_t dst_ne[2] = { ne0, ne1 };
+            size_t  dst_nb[2] = { nb0, nb1 };
+            acl_tensor_ptr acl_dst_s = ggml_cann_create_tensor(
+                (char *) dst->data + i2 * nb2 + i3 * nb3,
+                dst_acl_type, dst_type_sz, dst_ne, dst_nb, 2);
 
-                ggml_cann_pool_alloc output_allocator(ctx.pool());
-                void *               output_buffer = output_allocator.alloc(ggml_nbytes(dst));
-                acl_tensor_ptr       acl_out = ggml_cann_create_tensor(output_buffer, ggml_cann_type_mapping(dst->type),
-                                                                       ggml_type_size(dst->type), dst->ne, dst->nb, 2);
-
-                GGML_CANN_CALL_ACLNN_OP(ctx, Ger, acl_input.get(), acl_weight.get(), acl_out.get());
-                float       alpha_value = 1.0f;
-                aclScalar * alpha       = aclCreateScalar(&alpha_value, ACL_FLOAT);
-                GGML_CANN_CALL_ACLNN_OP(ctx, InplaceAdd, accumulator.get(), acl_out.get(), alpha);
-            }
+            // Matmul(src1_t [n,K], src0 [K,m]) = [n,m] = acl_dst_s  ✓
+            GGML_CANN_CALL_ACLNN_OP(ctx, Matmul,
+                acl_src1_t.get(), acl_src0_s.get(), acl_dst_s.get(), (int8_t) 1);
         }
     }
 }
@@ -4170,3 +4433,4 @@ void ggml_cann_gated_linear_attn(ggml_backend_cann_context & ctx, ggml_tensor *
         }
     }
 }
+

ggml/src/ggml-cann/aclnn_ops.h

@@ -32,6 +32,9 @@
 #include <aclnnop/aclnn_cat.h>
 #include <aclnnop/aclnn_clamp.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_cumsum.h>
+#include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triu.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_gelu.h>
 #include <aclnnop/aclnn_gelu_v2.h>
@@ -47,6 +50,9 @@
 #include <aclnnop/aclnn_sign.h>
 #include <aclnnop/aclnn_silu.h>
 #include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_softplus.h>
+#include <aclnnop/aclnn_swi_glu.h>
+#include <aclnnop/aclnn_geglu.h>
 #include <aclnnop/aclnn_slice.h>
 #include <aclnnop/aclnn_sqrt.h>
 #include <aclnnop/aclnn_tanh.h>
@@ -69,6 +75,9 @@
  */
 void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate);
+
 /**
  * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
  *          backend.
@@ -325,6 +334,48 @@ void ggml_cann_sum_rows(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 void ggml_cann_sum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+/**
+ * @brief   Computes the cumulative sum of a ggml tensor along dim 0 using the
+ *          CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_CUMSUM`.
+ */
+void ggml_cann_cumsum(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Computes a triangular mask (tril/triu) of a square ggml tensor
+ *          using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_TRI`.
+ */
+void ggml_cann_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Solves a triangular linear system AX=B using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_SOLVE_TRI`.
+ */
+void ggml_cann_solve_tri(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Creates a diagonal matrix from a vector using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_DIAG`.
+ */
+void ggml_cann_diag(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
+/**
+ * @brief   Fills a tensor with a constant scalar value using the CANN backend.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor. dst->op is `GGML_OP_FILL`.
+ */
+void ggml_cann_fill(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
  *          the CANN backend.
@@ -461,6 +512,9 @@ void ggml_cann_timestep_embedding(ggml_backend_cann_context & ctx, ggml_tensor *
 // @see ggml_cann_dup.
 void ggml_cann_cpy(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
+// @see ggml_cann_acc, but copies src1 into dst instead of adding.
+void ggml_cann_set(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+
 /**
  * @brief   Computes the softmax activation with optional masking.
  *
@@ -813,6 +867,8 @@ void ggml_cann_count_equal(ggml_backend_cann_context & ctx, ggml_tensor * dst);
  *            dst->op is expected to be `GGML_OP_STEP`.
  */
 void ggml_cann_step(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_softplus(ggml_backend_cann_context & ctx, ggml_tensor * dst);
+void ggml_cann_geglu_quick(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 /**
  * @brief   Performs the Flash Attention extended operator using the CANN backend.

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

@@ -1428,6 +1428,22 @@ static bool ggml_backend_cann_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
     return false;
 }
 
+/**
+ * @brief Set a region of a tensor's device memory to a specified value.
+ *
+ * @param buffer The CANN buffer containing the tensor.
+ * @param tensor Pointer to the tensor whose memory will be set.
+ * @param value The value to which each byte in the region will be set.
+ * @param offset Byte offset within the tensor's data to start setting.
+ * @param size Number of bytes to set.
+ */
+static void ggml_backend_cann_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context * ctx = (ggml_backend_cann_buffer_context *) buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    ACL_CHECK(aclrtMemset((char *) tensor->data + offset, size, value, size));
+}
+
 /**
  * @brief Clear a CANN buffer by setting all its memory to a specified value.
  *
@@ -1454,7 +1470,7 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cann_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_cann_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
     /* .set_tensor_2d   = */ NULL,
@@ -1835,6 +1851,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
                 case GGML_UNARY_OP_STEP:
                     ggml_cann_step(ctx, dst);
                     break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    ggml_cann_softplus(ctx, dst);
+                    break;
                 default:
                     return false;
             }
@@ -1845,20 +1864,16 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
                     GGML_CANN_CALL_OP_UNARY_GATED(Relu);
                     break;
                 case GGML_GLU_OP_GEGLU:
+                    ggml_cann_geglu(ctx, dst, 0);  // approximate=0 → tanh
+                    break;
                 case GGML_GLU_OP_GEGLU_ERF:
-                    // aclnnGelu internally uses the erf-based approximation.
-                    GGML_CANN_CALL_OP_UNARY_GATED(Gelu);
+                    ggml_cann_geglu(ctx, dst, 1);  // approximate=1 → erf
                     break;
                 case GGML_GLU_OP_SWIGLU:
-                    GGML_CANN_CALL_OP_UNARY_GATED(Silu);
+                    ggml_cann_swiglu(ctx, dst);
                     break;
                 case GGML_GLU_OP_GEGLU_QUICK:
-                    {
-                        auto lambda = [](ggml_backend_cann_context & ctx, aclTensor * acl_src, aclTensor * acl_dst) {
-                            GGML_CANN_CALL_ACLNN_OP(ctx, GeluV2, acl_src, 0, acl_dst);
-                        };
-                        ggml_cann_op_unary_gated(lambda, ctx, dst);
-                    }
+                    ggml_cann_geglu_quick(ctx, dst);
                     break;
                 default:
                     return false;
@@ -1920,6 +1935,9 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_CPY:
             ggml_cann_cpy(ctx, dst);
             break;
+        case GGML_OP_SET:
+            ggml_cann_set(ctx, dst);
+            break;
         case GGML_OP_CONT:
             ggml_cann_dup(ctx, dst);
             break;
@@ -1989,6 +2007,21 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
         case GGML_OP_SSM_CONV:
             ggml_cann_ssm_conv(ctx, dst);
             break;
+        case GGML_OP_CUMSUM:
+            ggml_cann_cumsum(ctx, dst);
+            break;
+        case GGML_OP_TRI:
+            ggml_cann_tri(ctx, dst);
+            break;
+        case GGML_OP_FILL:
+            ggml_cann_fill(ctx, dst);
+            break;
+        case GGML_OP_DIAG:
+            ggml_cann_diag(ctx, dst);
+            break;
+        case GGML_OP_SOLVE_TRI:
+            ggml_cann_solve_tri(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -2324,6 +2357,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
     if (use_cann_graph) {
         // If no matching graph is found, the graph needs to be recaptured.
         graph_capture_required = !cann_ctx->graph_lru_cache.find_and_move_to_front(cgraph);
+
         if (graph_capture_required) {
             // If no matching graph is found, add a new ACL graph.
             ggml_cann_graph * new_graph = ggml_cann_graph::create_from_cgraph(cgraph);
@@ -2382,6 +2416,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 case GGML_UNARY_OP_SGN:
                 case GGML_UNARY_OP_STEP:
                 case GGML_UNARY_OP_GELU_ERF:
+                case GGML_UNARY_OP_SOFTPLUS:
                     return true;
                 default:
                     return false;
@@ -2572,6 +2607,7 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
+        case GGML_OP_SET:
         case GGML_OP_GROUP_NORM:
             return true;
         case GGML_OP_PAD:
@@ -2649,6 +2685,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
             }
         case GGML_OP_SSM_CONV:
             return true;
+        case GGML_OP_CUMSUM:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_TRI:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_FILL:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_DIAG:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SOLVE_TRI:
+            return op->src[0]->type == GGML_TYPE_F32;
         default:
             return false;
     }

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -2005,12 +2005,12 @@ void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const v
     const int lda = KB * sizeof(TA);
     //const int ldb = KB * sizeof(TB);
 
-    static thread_local packed_B_t Tile0[TILE_N * TILE_K];
-    static thread_local packed_B_t Tile1[TILE_N * TILE_K];
-    static thread_local int8_t Tile23[TILE_M * TILE_K];
+    alignas(64) static thread_local packed_B_t Tile0[TILE_N * TILE_K];
+    alignas(64) static thread_local packed_B_t Tile1[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile23[TILE_M * TILE_K];
 
-    static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
-    static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
+    alignas(64) static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
+    alignas(64) static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
 
     // double buffering C to interleave avx512 and amx
     int32_t * C_cur = TileC0;
@@ -2187,21 +2187,21 @@ void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const v
     const int m1 = std::max(M - TILE_M, 0);
     //const int lda = KB * sizeof(TA);
 
-    static thread_local int8_t Tile0[TILE_N * TILE_K];
-    static thread_local int8_t Tile1[TILE_N * TILE_K];
-    static thread_local int8_t Tile23[TILE_M * TILE_K];
+    alignas(64) static thread_local int8_t Tile0[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile1[TILE_N * TILE_K];
+    alignas(64) static thread_local int8_t Tile23[TILE_M * TILE_K];
 
     // mat mul result for each group
-    static thread_local int32_t Tile4[TILE_M * TILE_N];
-    static thread_local int32_t Tile5[TILE_M * TILE_N];
-    static thread_local int32_t Tile6[TILE_M * TILE_N];
-    static thread_local int32_t Tile7[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile4[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile5[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile6[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Tile7[TILE_M * TILE_N];
 
     // sum of each QK_K block, contains 8 groups, int32
-    static thread_local int32_t Sumi4[TILE_M * TILE_N];
-    static thread_local int32_t Sumi5[TILE_M * TILE_N];
-    static thread_local int32_t Sumi6[TILE_M * TILE_N];
-    static thread_local int32_t Sumi7[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi4[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi5[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi6[TILE_M * TILE_N];
+    alignas(64) static thread_local int32_t Sumi7[TILE_M * TILE_N];
 
     const int k_group_size = std::is_same<TB, block_q6_K>::value ? 16 : 32;
     for (int i = 0; i < KB; ++i) {

ggml/src/ggml-cpu/arch-fallback.h

@@ -83,7 +83,6 @@
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
-#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4

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

@@ -151,8 +151,6 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q1_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-    float sumf = 0.0f;
-
 #if defined(__ARM_NEON)
     float32x4_t sumv = vdupq_n_f32(0.0f);
 
@@ -212,31 +210,13 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
     }
 
-    sumf = vaddvq_f32(sumv);
+    *s = vaddvq_f32(sumv);
 #else
-    // Scalar fallback
-    for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-
-        // Process 4 Q8_0 blocks
-        for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
-            int sumi = 0;
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
-
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi += xi * y[i*4 + k].qs[j];
-            }
-            sumf += d0 * d1 * sumi;
-        }
-    }
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
-    *s = sumf;
 }
 
 

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -274,6 +274,18 @@ static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const
 }
 #endif
 #elif defined(__SSSE3__)
+static inline __m128i bytes_from_bits_16(const uint8_t * x) {
+    uint16_t x16;
+    memcpy(&x16, x, sizeof(uint16_t));
+
+    const __m128i shuf_mask = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    __m128i bytes = _mm_shuffle_epi8(_mm_set1_epi16((short) x16), shuf_mask);
+    const __m128i bit_mask = _mm_set_epi64x(0x7fbfdfeff7fbfdfe, 0x7fbfdfeff7fbfdfe);
+    bytes = _mm_or_si128(bytes, bit_mask);
+
+    return _mm_cmpeq_epi8(bytes, _mm_set1_epi64x(-1));
+}
+
 // horizontally add 4x4 floats
 static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
     __m128 res_0 =_mm_hadd_ps(a, b);
@@ -540,6 +552,152 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    const __m256i ones_8 = _mm256_set1_epi8(1);
+    const __m256i ones_16 = _mm256_set1_epi16(1);
+    const __m256i byte_shuf = _mm256_setr_epi8(
+            0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+            2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
+    const __m256i bit_masks = _mm256_setr_epi8(
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128,
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128);
+    const __m256i zero = _mm256_setzero_si256();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const uint32_t * GGML_RESTRICT qs32 = (const uint32_t *) x[ib].qs;
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+        __m256 acc_block;
+        {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[0].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[0]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), _mm256_cvtepi32_ps(s32));
+        }
+        for (int K = 1; K < 4; ++K) {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[K].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[K]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[K].d)), _mm256_cvtepi32_ps(s32), acc_block);
+        }
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d0), acc_block, acc);
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+        __m256 acc_block;
+        {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[0]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), q);
+        }
+        for(int K = 1; K < 4; ++K) {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[(K) * 4]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_add_ps(acc_block, _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(K)].d)), q));
+        }
+#undef Q1_AVX_BLOCK
+
+        acc = _mm256_add_ps(acc, _mm256_mul_ps(_mm256_set1_ps(d0), acc_block));
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__SSSE3__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const __m128 d0 = _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+#define Q1_SSSE3_BLOCK(QS_OFF, Y_IDX, ACC) \
+        { \
+            const __m128i bit_mask_0 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 0]); \
+            const __m128i bit_mask_1 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 2]); \
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[0]); \
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[16]); \
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero); \
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero); \
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0); \
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1); \
+            const __m128i sum_0 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_0), ones_16); \
+            const __m128i sum_1 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_1), ones_16); \
+            const __m128 q = _mm_cvtepi32_ps(_mm_add_epi32(sum_0, sum_1)); \
+            (ACC) = _mm_add_ps((ACC), _mm_mul_ps(_mm_mul_ps(d0, _mm_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(Y_IDX)].d))), q)); \
+        }
+        Q1_SSSE3_BLOCK(0,  0, acc_0)
+        Q1_SSSE3_BLOCK(4,  1, acc_1)
+        Q1_SSSE3_BLOCK(8,  2, acc_2)
+        Q1_SSSE3_BLOCK(12, 3, acc_3)
+#undef Q1_SSSE3_BLOCK
+    }
+
+    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
@@ -2142,9 +2300,8 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 #if defined __AVX2__
 
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i m15 = _mm256_set1_epi8(15);
 
     __m256 acc = _mm256_setzero_ps();
 
@@ -2156,53 +2313,45 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
         const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m256i scales_16 = _mm256_cvtepi8_epi16(scales);
+        const __m256i q8sclsub = _mm256_slli_epi32(_mm256_madd_epi16(q8sums, scales_16), 5);
 
         __m256i sumi = _mm256_setzero_si256();
 
         int is = 0;
 
         for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
             const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
             const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
             const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
 
-            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
-            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
+            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m3), 4);
+            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(12)), 2);
+            const __m256i q4h_2 = _mm256_and_si256(q4bitsH, _mm256_set1_epi8(48));
+            const __m256i q4h_3 = _mm256_srli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(-64)), 2);
 
-            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m15), q4h_0);
+            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m15), q4h_1);
+            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m15), q4h_2);
+            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m15), q4h_3);
 
             const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
             const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
 
-            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
-
             __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
             __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
             __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
             __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
 
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
 
             p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
             p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
@@ -2214,6 +2363,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         }
 
+        sumi = _mm256_sub_epi32(sumi, q8sclsub);
         acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
     }
 

ggml/src/ggml-cpu/quants.c

@@ -137,22 +137,28 @@ void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     float sumf = 0.0;
 
     for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         float sumi = 0.0f;
 
         for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
             int sumi_block = 0;
 
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
+            const uint8_t * GGML_RESTRICT bits = &x[i].qs[k * 4];
+            const int8_t  * GGML_RESTRICT qy   = yb->qs;
 
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi_block += xi * y[i*4 + k].qs[j];
+            for (int b = 0; b < 4; ++b, qy += 8) {
+                const unsigned mask = bits[b];
+                sumi_block += ((mask & 0x01) ? qy[0] : -qy[0])
+                           +  ((mask & 0x02) ? qy[1] : -qy[1])
+                           +  ((mask & 0x04) ? qy[2] : -qy[2])
+                           +  ((mask & 0x08) ? qy[3] : -qy[3])
+                           +  ((mask & 0x10) ? qy[4] : -qy[4])
+                           +  ((mask & 0x20) ? qy[5] : -qy[5])
+                           +  ((mask & 0x40) ? qy[6] : -qy[6])
+                           +  ((mask & 0x80) ? qy[7] : -qy[7]);
             }
 
             sumi += d1 * sumi_block;

ggml/src/ggml-cpu/vec.h

@@ -1036,12 +1036,12 @@ inline static float ggml_gelu_quick_f32(float x) {
     return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
 }
 
-//inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
-//    const uint16_t * i16 = (const uint16_t *) x;
-//    for (int i = 0; i < n; ++i) {
-//        y[i] = ggml_table_gelu_quick_f16[i16[i]];
-//    }
-//}
+inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+    const uint16_t * i16 = (const uint16_t *) x;
+    for (int i = 0; i < n; ++i) {
+        y[i] = ggml_table_gelu_quick_f16[i16[i]];
+    }
+}
 
 #ifdef GGML_GELU_QUICK_FP16
 inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
@@ -1060,13 +1060,6 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
 }
 #endif
 
-inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
-    for (int i = 0; i < n; ++i) {
-        float v = GGML_CPU_FP16_TO_FP32(x[i]);
-        y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
-    }
-}
-
 // Sigmoid Linear Unit (SiLU) function
 inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));

ggml/src/ggml-cuda/common.cuh

@@ -269,10 +269,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
-#if defined(TURING_MMA_AVAILABLE)
-#define LDMATRIX_TRANS_AVAILABLE
-#endif // defined(TURING_MMA_AVAILABLE)
-
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);

ggml/src/ggml-cuda/concat.cu

@@ -1,96 +1,79 @@
 #include "concat.cuh"
 
 // contiguous kernels
-static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+template <int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE) concat_f32_cont(const float * x,
+                                                                                 const float * y,
+                                                                                 float *       dst,
+                                                                                 int64_t       ne00,
+                                                                                 int64_t       ne01,
+                                                                                 int64_t       ne02,
+                                                                                 int64_t       ne0,
+                                                                                 int64_t       ne1,
+                                                                                 int64_t       ne2) {
+    static_assert(dim >= 0 && dim <= 2, "dim must be in [0, 2]");
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
+    const int64_t n = ne0 * ne1 * ne2;
 
-    if (nidx < ne00) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            (nidx - ne00) +
-            blockIdx.y * (ne0 - ne00) +
-            blockIdx.z * (ne0 - ne00) * gridDim.y;
-        dst[offset_dst] = y[offset_src];
+    for (int64_t i = (int64_t) blockIdx.x * blockDim.x + threadIdx.x; i < n; i += (int64_t) blockDim.x * gridDim.x) {
+        if constexpr (dim == 0) {
+            const int64_t row = i / ne0;
+            const int64_t i0  = i - row * ne0;
+
+            if (i0 < ne00) {
+                dst[i] = x[row * ne00 + i0];
+            } else {
+                dst[i] = y[row * (ne0 - ne00) + (i0 - ne00)];
+            }
+        } else if constexpr (dim == 1) {
+            const int64_t dst_plane  = ne0 * ne1;
+            const int64_t src0_plane = ne0 * ne01;
+            const int64_t src1_plane = dst_plane - src0_plane;
+            const int64_t i2         = i / dst_plane;
+            const int64_t i01        = i - i2 * dst_plane;
+
+            if (i01 < src0_plane) {
+                dst[i] = x[i2 * src0_plane + i01];
+            } else {
+                dst[i] = y[i2 * src1_plane + (i01 - src0_plane)];
+            }
+        } else {
+            const int64_t src0_size = ne0 * ne1 * ne02;
+
+            if (i < src0_size) {
+                dst[i] = x[i];
+            } else {
+                dst[i] = y[i - src0_size];
+            }
+        }
     }
 }
 
-static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
+static void concat_f32_cuda(const float * x,
+                            const float * y,
+                            float *       dst,
+                            int64_t       ne00,
+                            int64_t       ne01,
+                            int64_t       ne02,
+                            int64_t       ne0,
+                            int64_t       ne1,
+                            int64_t       ne2,
+                            int           dim,
+                            cudaStream_t  stream) {
+    const int64_t n          = ne0 * ne1 * ne2;
+    const int     num_blocks = (n + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
 
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.y < (unsigned)ne01) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * ne01;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            (blockIdx.y - ne01) * ne0 +
-            blockIdx.z * ne0 * (gridDim.y - ne01);
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-
-    if (blockIdx.z < (unsigned)ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-        dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-        dst[offset_dst] = y[offset_src];
-    }
-}
-
-static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
     if (dim == 0) {
-        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
+        concat_f32_cont<0>
+            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
     if (dim == 1) {
-        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
+        concat_f32_cont<1>
+            <<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
         return;
     }
-    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+    concat_f32_cont<2><<<num_blocks, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne00, ne01, ne02, ne0, ne1, ne2);
 }
 
 // non-contiguous kernel (slow)

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

@@ -305,12 +305,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
-    // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes.
+    // The minimum granularity is 16 bytes.
+    constexpr int h2_per_chunk = 16/sizeof(half2);
+    const int chunks_per_row = D2 / h2_per_chunk;
     if constexpr (use_cp_async) {
+        static_assert(warp_size == 32, "bad warp_size");
         static_assert(!oob_check, "OOB check not compatible with cp_async");
         constexpr int preload = 64;
-        constexpr int h2_per_chunk = 16/sizeof(half2);
-        const int chunks_per_row = D2 / h2_per_chunk;
 
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
@@ -348,11 +349,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         // 6: max  1*16= 16 bytes,   8 half
         ggml_cuda_unroll<6>{}(load);
     } else {
-        // TODO use ggml_cuda_memcpy_1
+        const half2 zero[4] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
         auto load = [&] __device__ (const int n) {
-            const int stride_k = warp_size >> n;
-            const int k0_start = stride_k == warp_size ? 0 : D2 - D2 % (2*stride_k);
-            const int k0_stop  =                             D2 - D2 % (1*stride_k);
+            const int stride_k = 32 >> n;
+            const int k0_start = stride_k == 32 ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
+            const int k0_stop  =                      chunks_per_row - chunks_per_row % (1*stride_k);
             const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
@@ -371,15 +372,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    tile_KV[i*stride_tile + k] = !oob_check || i < i_sup ? KV[i*stride_KV + k] : make_half2(0.0f, 0.0f);
+                    ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
+                        !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
                 }
             }
         };
-        // 1: max 32* 4=128 bytes,  64 half
-        // 2: max 16* 4= 64 bytes,  32 half
-        // 3: max  8* 4= 32 bytes,  16 half
-        // 4: max  4* 4= 16 bytes,   8 half
-        ggml_cuda_unroll<4>{}(load);
+        // 1: max 32*16=512 bytes, 256 half
+        // 2: max 16*16=256 bytes, 128 half
+        // 3: max  8*16=128 bytes,  64 half
+        // 4: max  4*16= 64 bytes,  32 half
+        // 5: max  2*16= 32 bytes,  16 half
+        // 6: max  1*16= 16 bytes,   8 half
+        ggml_cuda_unroll<6>{}(load);
     }
 }
 
@@ -862,11 +866,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-    T_A_VKQ A_identity;
-    make_identity_mat(A_identity);
-#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
     for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
@@ -897,29 +896,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
-#if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-#elif defined(AMD_MFMA_AVAILABLE)
-                // MFMA A register layout: A_mat[i=lane%16][k=4*(lane/16)+reg].
-                // Normal load gives A_mat[seq][dv] but we need A_mat[dv][seq] = V^T.
-                // Load with transposed addressing: 4 strided half loads.
-                {
-                    const half2 * xs0 = tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2;
-                    const half * xs0_h = (const half *) xs0;
-                    const int stride_h = stride_tile_V * 2; // stride in half units
-                    half * A_h = (half *) A.x;
-#pragma unroll
-                    for (int l = 0; l < 4; ++l) {
-                        A_h[l] = xs0_h[(4*(threadIdx.x / 16) + l) * stride_h + threadIdx.x % 16];
-                    }
-                }
-#else
-                // TODO: Try to transpose tile_V when loading gmem to smem.
-                // Use mma to transpose T_A_VKQ for RDNA.
-                T_A_VKQ A_trans;
-                load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-                mma(A, A_trans, A_identity);
-#endif // defined(LDMATRIX_TRANS_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {

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

@@ -368,15 +368,21 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
     }
 
     ~ggml_cuda_pool_leg() {
+        clear_pool();
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    void clear_pool() {
         ggml_cuda_set_device(device);
         for (int i = 0; i < MAX_BUFFERS; ++i) {
             ggml_cuda_buffer & b = buffer_pool[i];
             if (b.ptr != nullptr) {
                 CUDA_CHECK(cudaFree(b.ptr));
                 pool_size -= b.size;
+                b.ptr  = nullptr;
+                b.size = 0;
             }
         }
-        GGML_ASSERT(pool_size == 0);
     }
 
     void * alloc(size_t size, size_t * actual_size) override {
@@ -421,7 +427,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
         size_t look_ahead_size = (size_t) (1.05 * size);
         look_ahead_size = 256 * ((look_ahead_size + 255)/256);
         ggml_cuda_set_device(device);
-        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
+        cudaError_t err = ggml_cuda_device_malloc(&ptr, look_ahead_size, device);
+        if (err == cudaErrorMemoryAllocation) {
+            (void)cudaGetLastError();
+            const size_t cached_bytes = pool_size;
+            GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: alloc of %.2f MiB failed, flushing %.2f MiB of cached buffers and retrying\n",
+                           device, look_ahead_size/1024.0/1024.0, cached_bytes/1024.0/1024.0);
+            CUDA_CHECK(cudaDeviceSynchronize());
+            clear_pool();
+            err = ggml_cuda_device_malloc(&ptr, look_ahead_size, device);
+            if (err == cudaSuccess) {
+                GGML_LOG_DEBUG(GGML_CUDA_NAME " pool[%d]: retry succeeded\n", device);
+            }
+        }
+        CUDA_CHECK(err);
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
@@ -1203,6 +1222,13 @@ static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct gg
     // For small tensors, simply reduce them as FP32.
     // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
     if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+        for (size_t i = 0; i < n_backends; ++i) {
+            if ((tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+                ggml_cuda_set_device(cuda_ctx->device);
+                CUDA_CHECK(cudaMemsetAsync(tensors[i]->data, 0, ggml_nbytes(tensors[i]), cuda_ctx->stream()));
+            }
+        }
         NCCL_CHECK(ncclGroupStart());
         for (size_t i = 0; i < n_backends; ++i) {
             ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
@@ -1224,7 +1250,11 @@ static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct gg
         tmp[i].alloc(ne);
 
         ggml_cuda_set_device(cuda_ctx->device);
-        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        if (tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) {
+            to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        } else {
+            CUDA_CHECK(cudaMemsetAsync(tmp[i].get(), 0, ne * sizeof(nv_bfloat16), cuda_ctx->stream()));
+        }
         CUDA_CHECK(cudaGetLastError());
     }
 
@@ -3562,6 +3592,30 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
         return true;
     }
 
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_UNARY && ops.begin()[1] == GGML_OP_SQR
+     && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_RELU) {
+        const ggml_tensor * unary = cgraph->nodes[node_idx];
+        const ggml_tensor * sqr   = cgraph->nodes[node_idx+1];
+
+        if (ggml_get_unary_op(unary) != GGML_UNARY_OP_RELU) {
+            return false;
+        }
+
+        if (unary->type != GGML_TYPE_F32 && unary->type != GGML_TYPE_F16) {
+            return false;
+        }
+
+        if (unary->type != sqr->type) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(unary->src[0])) {
+            return false;
+        }
+
+        return true;
+    }
+
     if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SCALE && ops.begin()[1] == GGML_OP_UNARY && ops.begin()[2] == GGML_OP_SCALE
      && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_TANH) {
         const ggml_tensor *scale  = cgraph->nodes[node_idx];
@@ -4070,6 +4124,12 @@ static void ggml_cuda_graph_evaluate_and_capture(ggml_backend_cuda_context * cud
                         continue;
                     }
 
+                    if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_UNARY, GGML_OP_SQR }, { GGML_UNARY_OP_RELU })) {
+                        ggml_cuda_op_relu_sqr(*cuda_ctx, node, cgraph->nodes[i+1]);
+                        i++;
+                        continue;
+                    }
+
                     if (ggml_cuda_can_fuse(cgraph, i, { GGML_OP_SCALE, GGML_OP_UNARY, GGML_OP_SCALE }, { GGML_UNARY_OP_TANH })) {
                         i += 2;
                         ggml_cuda_op_softcap(*cuda_ctx, cgraph->nodes[i], node);

ggml/src/ggml-cuda/mma.cuh

@@ -86,17 +86,12 @@ namespace ggml_cuda_mma {
     //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
     //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
 
-    static constexpr bool is_i_major(const data_layout dl) {
-        return dl == DATA_LAYOUT_I_MAJOR ||
-               dl == DATA_LAYOUT_I_MAJOR_MIRRORED;
-    }
-
     static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
         return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
         return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -113,7 +108,6 @@ namespace ggml_cuda_mma {
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
-            if (I == 64 && J ==  2) return true;
             if (I == 16 && J ==  8) return true;
             if (I == 32 && J ==  4) return true;
             if (I == 16 && J == 16) return true;
@@ -122,7 +116,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+            if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 16;
             } else if constexpr (I == 16 && J == 8) {
                 return threadIdx.x % 16;
@@ -139,8 +133,8 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-                return (2 * ((threadIdx.x / 16) % 2) + l);
+            if constexpr (I == 16 && J == 4) {
+                return threadIdx.x / 16;
             } else if constexpr (I == 16 && J == 8) {
                 return 2 * (threadIdx.x / 16) + l;
             } else if constexpr (I == 32 && J == 4) {
@@ -154,7 +148,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#elif defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
@@ -283,7 +277,7 @@ namespace ggml_cuda_mma {
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -407,7 +401,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     };
 
     template <int I_, int J_>
@@ -701,57 +695,12 @@ namespace ggml_cuda_mma {
     }
 #endif // defined(TURING_MMA_AVAILABLE)
 
-    static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
-#if defined(RDNA4)
-        const int row = t.get_i(0);
-        const int left_right = t.get_j(0) / 4;
-        const int up_down = row / 8;
-        const int idx = row % 8;
-        reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
-#else
-        GGML_UNUSED_VARS(t);
-        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
-    }
-
     template <int I, int J, typename T, data_layout dl>
     static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
-#if defined(AMD_MFMA_AVAILABLE)
-        if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        } else {
-            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-        }
-#elif defined(AMD_WMMA_AVAILABLE)
-        // All wmma layout has contiguous data when i-major.
-        if constexpr (is_i_major(dl)) {
-            // the data must be aligned to 16 bytes when bigger than ggml_cuda_get_max_cpy_bytes()
-            constexpr int aligned_copy_bytes = ggml_cuda_get_max_cpy_bytes();
-            if constexpr (sizeof(t.x) > aligned_copy_bytes) {
-                static_assert(sizeof(t.x) % aligned_copy_bytes == 0, "bad type size");
-                constexpr int aligned_copy_count = sizeof(t.x)/aligned_copy_bytes;
-#pragma unroll
-                for (int i = 0; i < aligned_copy_count; ++i) {
-                    ggml_cuda_memcpy_1<aligned_copy_bytes>(t.x + t.ne/aligned_copy_count*i, xs0 + t.get_i(0) * stride + t.get_j(t.ne/aligned_copy_count*i));
-                }
-            } else {
-                ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-            }
-        } else {
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        }
-#else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
             t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
         }
-#endif // defined(AMD_MFMA_AVAILABLE)
     }
 
     template <typename T>
@@ -764,26 +713,37 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(t, xs0, stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
-    template <typename T>
+    template <typename T, data_layout dl>
     static __device__ __forceinline__ void load_ldmatrix(
-            tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) {
+            tile<16, 4, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<8>(t.x + 2, xs0 + t.get_i(0)*stride + 2);
+#else
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 4, "bad ne");
+        ggml_cuda_memcpy_1<4>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -796,19 +756,26 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
-#else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#if 1
-        // TODO: more generic handling
-        static_assert(sizeof(T) == 4, "bad type size");
+#elif defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 32, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<16>(t.x + 4, xs0 + t.get_i(0)*stride + 4);
 #else
-        load_generic(t, xs0, stride);
-#endif // 1
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -827,23 +794,30 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void load_ldmatrix(
             tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
-        int * xi = (int * ) t.x;
+        int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
+#elif defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        half * xh = (half *) t.x;
+#pragma unroll
+        for (int l = 0; l < t.ne; ++l) {
+            xh[2*l + 0] = ((const half *) xs0)[(2*t.get_j(l) + 0)*(2*stride) + t.get_i(l)];
+            xh[2*l + 1] = ((const half *) xs0)[(2*t.get_j(l) + 1)*(2*stride) + t.get_i(l)];
+        }
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
@@ -1218,73 +1192,27 @@ namespace ggml_cuda_mma {
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * acc = (int32x4_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
-                                                      B.x[0],
-                                                      acc[0],
-                                                      0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1],
-                                                      B.x[1],
-                                                      acc[0],
-                                                      0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
-
 #elif defined(AMD_WMMA_AVAILABLE)
-
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
-
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
 #endif // RDNA4
-
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1297,19 +1225,10 @@ namespace ggml_cuda_mma {
         using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
         int32x16_t * acc = (int32x16_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
-                                                     B.x[0],
-                                                     acc[0],
-                                                     0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1],
-                                                     B.x[1],
-                                                     acc[0],
-                                                     0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
 
 #else
@@ -1329,7 +1248,7 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void mma(
             tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1344,12 +1263,12 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     static __device__ __forceinline__ void mma(
             tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1364,41 +1283,35 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <data_layout dl_d, data_layout dl_ab>
     static __device__ __forceinline__ void mma(
             tile<16, 16, int, dl_d> & D, const tile<16, 4, int, dl_ab> & A, const tile<16, 4, int, dl_ab> & B) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * acc = (int32x4_t *) D.x;
+#if defined(CDNA4) || defined(CDNA3)
+        const int64_t xA = uint32_t(A.x[0]);
+        const int64_t xB = uint32_t(B.x[0]);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(xA, xB, acc[0], 0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA1)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+#endif // defined(CDNA4) || defined(CDNA3)
+#elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
 #endif // RDNA4
 #else
         GGML_UNUSED(D);

ggml/src/ggml-cuda/mmq.cuh

@@ -104,7 +104,7 @@ struct tile_x_sizes {
 };
 
 static int get_mmq_x_max_host(const int cc) {
-    return (amd_mfma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
+    return (turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
         GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
@@ -114,9 +114,9 @@ static int get_mmq_x_max_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_x_max_device() {
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     return 128;
-#else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#else // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
 #if defined(GGML_USE_HIP)
     return 64;
@@ -1054,13 +1054,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             float dB;
             const int j = j0 + tile_C::get_j(0);
@@ -1295,13 +1295,13 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]);
@@ -1435,57 +1435,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1510,13 +1460,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -1742,74 +1692,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const half2 * y_ds = (const half2 *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x/2 : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y/2;
-            const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0
-                                              : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y
-                                                             : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x);
-
-            tile_C Cm;
-            if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                tile_A A1;
-                A1.x[0] = 0x01010101;
-                A1.x[1] = 0x01010101;
-                mma(Cm, A1, B[0]);
-            }
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C Cd;
-                mma(Cd, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]);
-                    float tmp = Cd.x[l]*dm.x;
-                    if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                        tmp -= Cm.x[l]*dm.y;
-                    }
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB;
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1834,13 +1717,13 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y;
@@ -2573,59 +2456,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * x_sc = (const int   *) x_df + MMQ_TILE_NE_K/QI6_K;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -2651,13 +2482,13 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -3647,10 +3478,10 @@ template <ggml_type type, int mmq_x, bool need_check>
 static __global__ void mul_mat_q(
         const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
         const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-        const int ncols_x, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
-        const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
-        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        const int ncols_max) {
+        const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst, const int stride_row_x, const int ncols_y, const int stride_col_dst,
+        const uint3 channel_ratio, const uint3 nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const uint3 sample_ratio, const uint3 nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        const uint3 ntx) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -3664,8 +3495,7 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
-    const int ntx = (ncols_max + mmq_x - 1) / mmq_x; // Number of tiles x
-    const int nty = (nrows_x   + mmq_y - 1) / mmq_y; // Number of tiles y
+    const uint32_t nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y
 
     // Initialize the ids for writing back data with just the index.
     // For regular matrix multiplications this is never changed.
@@ -3686,8 +3516,9 @@ static __global__ void mul_mat_q(
     // On non-CDNA AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
 #if (defined(GGML_USE_HIP) && !defined(CDNA)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
-        const int wt = blockIdx.z / nchannels_y;
-        const int zt = blockIdx.z - wt*nchannels_y;
+        const uint2 tmp2 = fast_div_modulo(blockIdx.z, nchannels_y);
+        const int wt = tmp2.x;
+        const int zt = tmp2.y;
         const int jt = blockIdx.y;
         const int it = blockIdx.x;
 
@@ -3730,40 +3561,40 @@ static __global__ void mul_mat_q(
         const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
         const int tile_y_max_j = col_diff - jt*mmq_x - 1;
 
-        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+        const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, need_check, fixup>
             (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst,
-             tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
+             tile_x_max_i, tile_y_max_j, 0, blocks_per_ne00.z);
         return;
     }
 #endif // (defined(GGML_USE_HIP) && !defined(CDNA4) && !defined(CDNA3)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
-    constexpr int ITER_K = get_iter_k(type);
-
-    const     int64_t blocks_per_ne00 = ncols_x / qk;
-    constexpr int     blocks_per_iter = ITER_K / qk;
+    constexpr int ITER_K          = get_iter_k(type);
+    constexpr int blocks_per_iter = ITER_K / qk;
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int64_t kbc      = (int64_t) blockIdx.x     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
-    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int kbc      = int64_t(blockIdx.x)    *(nsamples_y.z*nchannels_y.z*ntx.z*nty*blocks_per_ne00.z) / gridDim.x;
+    int kbc_stop = int64_t(blockIdx.x + 1)*(nsamples_y.z*nchannels_y.z*ntx.z*nty*blocks_per_ne00.z) / gridDim.x;
 
-    kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
-    kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
+    kbc      -= fastmodulo(kbc,      blocks_per_ne00) % blocks_per_iter;
+    kbc_stop -= fastmodulo(kbc_stop, blocks_per_ne00) % blocks_per_iter;
 
     // kb0 == k index when doing the matrix multiplication for an output tile.
-    int kb0_start = kbc % blocks_per_ne00;
-    int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
-    while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
-        int tmp = kbc;
-        const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-        tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-        const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
-        tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
-        const int zt = tmp / (ntx*blocks_per_ne00);
-        tmp -= zt * (ntx*blocks_per_ne00);
-        const int jt = tmp / blocks_per_ne00;
+    int kb0_start = fastmodulo(kbc, blocks_per_ne00);
+    int kb0_stop  = min(blocks_per_ne00.z, uint32_t(kb0_start + kbc_stop - kbc));
+    while (kbc < kbc_stop && kb0_stop == int(blocks_per_ne00.z)) {
+        int tmp = fastdiv(kbc, blocks_per_ne00);
+        uint2 tmp2 = fast_div_modulo(tmp, ntx);
+        const int jt = tmp2.y;
+        tmp = tmp2.x;
+        tmp2 = fast_div_modulo(tmp, nchannels_y);
+        const int zt = tmp2.y;
+        tmp = tmp2.x;
+        tmp2 = fast_div_modulo(tmp, nsamples_y);
+        const int wt = tmp2.y;
+        const int it = tmp2.x;
 
         // Defaults for regular matrix multiplication:
         int col_low    = 0;
@@ -3781,11 +3612,11 @@ static __global__ void mul_mat_q(
             offset_dst = 0;
 
             if (jt*mmq_x >= col_diff) {
-                kbc += blocks_per_ne00;
-                kbc -= kbc % blocks_per_ne00;
+                kbc += blocks_per_ne00.z;
+                kbc -= fastmodulo(kbc, blocks_per_ne00);
 
                 kb0_start = 0;
-                kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+                kb0_stop  = min(blocks_per_ne00.z, uint32_t(kbc_stop - kbc));
 
                 continue;
             }
@@ -3810,32 +3641,34 @@ static __global__ void mul_mat_q(
         const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
         const int tile_y_max_j = col_diff - jt*mmq_x - 1;
 
-        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+        const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
         constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         mul_mat_q_process_tile<type, mmq_x, need_check, fixup>
             (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, stride_row_x, ncols_y, stride_col_dst,
              tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 
-        kbc += blocks_per_ne00;
-        kbc -= kbc % blocks_per_ne00;
+        kbc += blocks_per_ne00.z;
+        kbc -= fastmodulo(kbc, blocks_per_ne00);
 
         kb0_start = 0;
-        kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+        kb0_stop  = min(blocks_per_ne00.z, uint32_t(kbc_stop - kbc));
     }
 
     if (kbc >= kbc_stop) {
         return;
     }
 
-    int tmp = kbc;
-    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
-    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
-    const int zt = tmp / (ntx*blocks_per_ne00);
-    tmp -= zt * (ntx*blocks_per_ne00);
-    const int jt = tmp / blocks_per_ne00;
+    int tmp = fastdiv(kbc, blocks_per_ne00);
+    uint2 tmp2 = fast_div_modulo(tmp, ntx);
+    const int jt = tmp2.y;
+    tmp = tmp2.x;
+    tmp2 = fast_div_modulo(tmp, nchannels_y);
+    const int zt = tmp2.y;
+    tmp = tmp2.x;
+    tmp2 = fast_div_modulo(tmp, nsamples_y);
+    const int wt = tmp2.y;
+    const int it = tmp2.x;
 
     // Defaults for regular matrix multiplication:
     int col_low    = 0;
@@ -3877,7 +3710,7 @@ static __global__ void mul_mat_q(
     const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
     const int tile_y_max_j = col_diff - jt*mmq_x - 1;
 
-    const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+    const int offset_x = fastdiv(wt, sample_ratio)*stride_sample_x + fastdiv(zt, channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
     constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     mul_mat_q_process_tile<type, mmq_x, need_check, fixup>
@@ -3886,46 +3719,37 @@ static __global__ void mul_mat_q(
 }
 
 template <ggml_type type, int mmq_x, bool need_check>
-static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
-                                                const int32_t * expert_bounds,
-                                                float * __restrict__ dst,
-                                                const float * __restrict__ tmp_last_tile,
-                                                const int    ncols_x,
-                                                const int    nrows_x,
-                                                const int    ncols_dst,
-                                                const size_t stride_col_dst,
-                                                const int    nchannels_y,
-                                                const size_t stride_channel_dst,
-                                                const int    nsamples_y,
-                                                const size_t stride_sample_dst,
-                                                const int    ncols_max) {
-    constexpr int     mmq_y           = get_mmq_y_device();
-    constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
-    constexpr int     ITER_K          = get_iter_k(type);
+__launch_bounds__(ggml_cuda_get_physical_warp_size()*mmq_get_nwarps_device()/2, 1)
+static __global__ void mul_mat_q_stream_k_fixup(
+        const int32_t * __restrict__ ids_dst, const int32_t * __restrict__ expert_bounds, float * __restrict__ dst,
+        float * __restrict__ tmp_last_tile, const uint3 blocks_per_ne00, const int nrows_x, const int ncols_dst,
+        const int stride_col_dst, const uint3 nchannels_y, const int stride_channel_dst, const uint3 nsamples_y,
+        const int stride_sample_dst, const uint3 ntx) {
+    constexpr int mmq_y           = get_mmq_y_device();
+    constexpr int qk              = ggml_cuda_type_traits<type>::qk;
+    constexpr int ITER_K          = get_iter_k(type);
+    constexpr int blocks_per_iter = ITER_K / qk;
 
-    constexpr int     blocks_per_iter = ITER_K / qk;
-    const     int64_t blocks_per_ne00 = ncols_x / qk;
-
-    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int nwarps = mmq_get_nwarps_device()/2;
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
-    float sum[mmq_x*mmq_y / (nwarps*warp_size)] = {0.0f};
+    float sum[mmq_x / nwarps] = {0.0f};
+    const int i = blockIdx.y*warp_size + threadIdx.x;
 
-    const int ntx  = (ncols_max + mmq_x - 1) / mmq_x;
-    const int nty  = (nrows_x   + mmq_y - 1) / mmq_y;
+    const int nty = (nrows_x + mmq_y - 1) / mmq_y;
 
     const int bidx0 = blockIdx.x;
 
     // kbc == k block continuous, current index in continuous ijk space.
-    int64_t kbc0      = (int64_t) bidx0     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
-    int64_t kbc0_stop = (int64_t)(bidx0 + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int kbc0      = int64_t(blockIdx.x)    *(nsamples_y.z*nchannels_y.z*ntx.z*nty*blocks_per_ne00.z) / gridDim.x;
+    int kbc0_stop = int64_t(blockIdx.x + 1)*(nsamples_y.z*nchannels_y.z*ntx.z*nty*blocks_per_ne00.z) / gridDim.x;
 
-    kbc0      -= (kbc0      % blocks_per_ne00) % blocks_per_iter;
-    kbc0_stop -= (kbc0_stop % blocks_per_ne00) % blocks_per_iter;
+    kbc0      -= fastmodulo(kbc0,      blocks_per_ne00) % blocks_per_iter;
+    kbc0_stop -= fastmodulo(kbc0_stop, blocks_per_ne00) % blocks_per_iter;
 
     const bool did_not_have_any_data   = kbc0 == kbc0_stop;
-    const bool wrote_beginning_of_tile = kbc0 % blocks_per_ne00 == 0;
-    const bool did_not_write_last      = kbc0/blocks_per_ne00 == kbc0_stop/blocks_per_ne00 && kbc0_stop % blocks_per_ne00 != 0;
+    const bool wrote_beginning_of_tile = fastmodulo(kbc0, blocks_per_ne00) == 0;
+    const bool did_not_write_last      = fastdiv(kbc0, blocks_per_ne00) == fastdiv(kbc0_stop, blocks_per_ne00) && fastmodulo(kbc0_stop, blocks_per_ne00) != 0;
     if (did_not_have_any_data || wrote_beginning_of_tile || did_not_write_last) {
         return;
     }
@@ -3934,11 +3758,11 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
 
     // Iterate over previous blocks and sum up partial sums written to fixup buffer.
     // All CUDA blocks that get here must have a previous block that needs a fixup.
-    int64_t bidx = bidx0 - 1;
-    int64_t kbc_stop = kbc0;
+    int bidx = bidx0 - 1;
+    int kbc_stop = kbc0;
     while(true) {
-        int64_t kbc = bidx*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
-        kbc -= (kbc % blocks_per_ne00) % blocks_per_iter;
+        int kbc = int64_t(bidx)*(nsamples_y.z*nchannels_y.z*ntx.z*nty*blocks_per_ne00.z) / gridDim.x;
+        kbc -= fastmodulo(kbc, blocks_per_ne00) % blocks_per_iter;
 
         if (kbc == kbc_stop) { // Did not have any data.
             bidx--;
@@ -3948,20 +3772,16 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
 
         any_fixup = true;
 
+
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
             const int j = j0 + threadIdx.y;
 
-#pragma unroll
-            for (int i0 = 0; i0 < mmq_y; i0 += warp_size) {
-                const int i = i0 + threadIdx.x;
-
-                sum[(j0/nwarps) * (mmq_y/warp_size) + i0/warp_size] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
-            }
+            sum[j0/nwarps] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
         }
 
         // If this block started in a previous tile we are done and don't need to combine additional partial results.
-        if (kbc % blocks_per_ne00 == 0 || kbc/blocks_per_ne00 < kbc0/blocks_per_ne00) {
+        if (fastmodulo(kbc, blocks_per_ne00) == 0 || fastdiv(kbc, blocks_per_ne00) < fastdiv(kbc0, blocks_per_ne00)) {
             break;
         }
         bidx--;
@@ -3972,14 +3792,16 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
         return;
     }
 
-    int tmp = kbc0;
-    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
-    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
-    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
-    const int zt = tmp / (ntx*blocks_per_ne00);
-    tmp -= zt * (ntx*blocks_per_ne00);
-    const int jt = tmp / blocks_per_ne00;
+    int tmp = fastdiv(kbc0, blocks_per_ne00);
+    uint2 tmp2 = fast_div_modulo(tmp, ntx);
+    const int jt = tmp2.y;
+    tmp = tmp2.x;
+    tmp2 = fast_div_modulo(tmp, nchannels_y);
+    const int zt = tmp2.y;
+    tmp = tmp2.x;
+    tmp2 = fast_div_modulo(tmp, nsamples_y);
+    const int wt = tmp2.y;
+    const int it = tmp2.x;
 
     if (!ids_dst) {
         const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y;
@@ -3987,6 +3809,9 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
 
         const int i_max = nrows_x   - it*mmq_y - 1;
         const int j_max = ncols_dst - jt*mmq_x - 1;
+        if (need_check && i > i_max) {
+            return;
+        }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -3996,16 +3821,7 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
                 return;
             }
 
-#pragma unroll
-            for (int i0 = 0; i0 < mmq_y; i0 += warp_size) {
-                const int i = i0 + threadIdx.x;
-
-                if (need_check && i > i_max) {
-                    continue;
-                }
-
-                dst[j*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/warp_size) + i0/warp_size];
-            }
+            dst[j*stride_col_dst + i] += sum[j0/nwarps];
         }
         return;
     }
@@ -4025,6 +3841,9 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
 
     const int i_max = nrows_x  - it*mmq_y - 1;
     const int j_max = col_diff - jt*mmq_x - 1;
+    if (need_check && i > i_max) {
+        return;
+    }
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -4034,16 +3853,7 @@ static __global__ void mul_mat_q_stream_k_fixup(const int32_t * ids_dst,
             return;
         }
 
-#pragma unroll
-        for (int i0 = 0; i0 < mmq_y; i0 += warp_size) {
-            const int i = i0 + threadIdx.x;
-
-            if (need_check && i > i_max) {
-                continue;
-            }
-
-            dst[ids_dst_shared[j]*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/warp_size) + i0/warp_size];
-        }
+        dst[ids_dst_shared[j]*stride_col_dst + i] += sum[j0/nwarps];
     }
 }
 
@@ -4091,29 +3901,44 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
     const int channel_ratio = args.nchannels_y / args.nchannels_x;
     const int sample_ratio  = args.nsamples_y  / args.nsamples_x;
 
+    const uint3 blocks_per_ne00_fd = init_fastdiv_values(args.ncols_x / ggml_cuda_type_traits<type>::qk);
+    const uint3 ntx_fd             = init_fastdiv_values(ntx);
+    const uint3 nchannels_y_fd     = init_fastdiv_values(args.nchannels_y);
+    const uint3 nsamples_y_fd      = init_fastdiv_values(args.nsamples_y);
+    const uint3 channel_ratio_fd   = init_fastdiv_values(channel_ratio);
+    const uint3 sample_ratio_fd    = init_fastdiv_values(sample_ratio);
+
     if (!args.use_stream_k) {
         if (args.nrows_x % mmq_y == 0) {
             constexpr bool need_check = false;
             mul_mat_q<type, mmq_x, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
-                 args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
-                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
-                 args.ncols_max);
+                 blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
+                 channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 ntx_fd);
         } else {
             constexpr bool need_check = true;
             mul_mat_q<type, mmq_x, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
                 (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
-                 args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
-                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
-                 args.ncols_max);
+                 blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
+                 channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+                 ntx_fd);
         }
         return;
     }
 
-    const dim3 block_nums_stream_k(nsm, 1, 1);
-    const bool fixup_needed = ntx*nty*ntzw % nsm != 0;
+    // For the stream-k kernel it is possible to run it with tiling by setting the number of CUDA blocks equal to the number of tiles.
+    // This is worthwhile if the efficiency of tiling is high and skipping the fixup kernel is more important.
+    const int ntiles_dst = ntx * nty * ntzw;
+    const int tiles_nwaves = (ntiles_dst + nsm - 1) / nsm;
+    const int tiles_efficiency_percent = 100 * ntiles_dst / (nsm*tiles_nwaves);
+    const dim3 block_nums_stream_k(GGML_CUDA_CC_IS_NVIDIA(cc) && tiles_efficiency_percent >= 90 ? ntiles_dst : nsm, 1, 1);
+
+    GGML_ASSERT(ntiles_dst * blocks_per_ne00_fd.z < (1 << 30)); // Assert that variable kbc will not overflow.
+
+    const bool fixup_needed = ntiles_dst % block_nums_stream_k.x != 0;
 
     ggml_cuda_pool & pool = ctx.pool(id);
     ggml_cuda_pool_alloc<float> tmp_fixup(pool);
@@ -4121,40 +3946,45 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
         tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y);
     }
 
+    const dim3 block_nums_fixup(block_nums_stream_k.x, mmq_y/warp_size, 1);
+    const dim3 block_dims_fixup(block_dims.x, block_dims.y/2, block_dims.z);
+
     if (args.nrows_x % mmq_y == 0) {
         constexpr bool need_check = false;
         mul_mat_q<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
-             args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
-             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
-             args.ncols_max);
+             blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
+             channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             ntx_fd);
 
         if (!fixup_needed) {
             return;
         }
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
-            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
-             args.ncols_max);
+        CUDA_CHECK(cudaGetLastError());
+        mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_fixup, block_dims_fixup, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst,
+             args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst,
+             ntx_fd);
     } else {
         constexpr bool need_check = true;
         mul_mat_q<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
             (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
-             args.ncols_x, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
-             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
-             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
-             args.ncols_max);
+             blocks_per_ne00_fd, args.nrows_x, args.ncols_dst, args.stride_row_x, args.ncols_y, args.nrows_dst,
+             channel_ratio_fd, nchannels_y_fd, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio_fd, nsamples_y_fd, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst,
+             ntx_fd);
 
         if (!fixup_needed) {
             return;
         }
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
-            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_dst,
-             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst,
-             args.ncols_max);
+        CUDA_CHECK(cudaGetLastError());
+        mul_mat_q_stream_k_fixup<type, mmq_x, need_check><<<block_nums_fixup, block_dims_fixup, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, blocks_per_ne00_fd, args.nrows_x, args.ncols_dst,
+             args.nrows_dst, nchannels_y_fd, args.stride_channel_dst, nsamples_y_fd, args.stride_sample_dst,
+             ntx_fd);
     }
 }
 

ggml/src/ggml-cuda/unary.cu

@@ -65,6 +65,11 @@ static __device__ __forceinline__ float op_sqr(float x) {
     return x * x;
 }
 
+static __device__ __forceinline__ float op_relu_sqr(float x) {
+    const float r = fmaxf(x, 0.0f);
+    return r * r;
+}
+
 static __device__ __forceinline__ float op_sqrt(float x) {
     return sqrtf(x);
 }
@@ -615,3 +620,21 @@ void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary
             GGML_ABORT("Unsupported unary op for fused unary+mul");
     }
 }
+
+/* fused relu + sqr */
+
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node) {
+    const ggml_tensor * src = relu_node->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src));
+    GGML_ASSERT(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
+    GGML_ASSERT(src->type == sqr_node->type);
+
+    const int k = ggml_nelements(src);
+    if (src->type == GGML_TYPE_F16) {
+        unary_cuda<op_relu_sqr>((const half *)src->data, (half *)sqr_node->data, k, stream);
+    } else {
+        unary_cuda<op_relu_sqr>((const float *)src->data, (float *)sqr_node->data, k, stream);
+    }
+}

ggml/src/ggml-cuda/unary.cuh

@@ -91,6 +91,8 @@ void ggml_cuda_op_xielu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_unary_mul(ggml_backend_cuda_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);
 
+void ggml_cuda_op_relu_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * relu_node, ggml_tensor * sqr_node);
+
 __device__ __forceinline__ float ggml_cuda_op_silu_single(float x) {
     return x / (1.0f + expf(-x));
 }

ggml/src/ggml-cuda/vendors/hip.h

@@ -33,7 +33,6 @@
 #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
-#define NCCL_CHECK(fn) {ncclResult_t err = fn; if(err != ncclSuccess) { GGML_ABORT("RCCL Failure RCCL returned: %i\n", err); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
@@ -59,6 +58,7 @@
 #define cudaDeviceProp hipDeviceProp_t
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
+#define cudaErrorMemoryAllocation hipErrorOutOfMemory
 #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
 #define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags hipEventCreateWithFlags

ggml/src/ggml-cuda/vendors/musa.h

@@ -42,6 +42,7 @@
 #define cudaDeviceProp musaDeviceProp
 #define cudaDeviceSynchronize musaDeviceSynchronize
 #define cudaError_t musaError_t
+#define cudaErrorMemoryAllocation musaErrorMemoryAllocation
 #define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
 #define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
 #define cudaEventCreateWithFlags musaEventCreateWithFlags

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

@@ -12,9 +12,12 @@
 #include <cstddef>
 #include <stdexcept>
 #include <string>
+#include <sstream>
+#include <iomanip>
 #include <unordered_set>
 #include <unordered_map>
 #include <regex>
+#include <queue>
 
 #ifdef _WIN32
 #    include <sal.h>
@@ -41,18 +44,26 @@
 #include "htp_iface.h"
 #include "htp-drv.h"
 
+using intvec  = std::vector<int>;
+using uintvec = std::vector<unsigned int>;
+using u32vec  = std::vector<uint32_t>;
+
 static size_t opt_ndev         = 1;
 static size_t opt_nhvx         = 0; // use all
 static int    opt_arch         = 0; // autodetect
 static int    opt_etm          = 0;
 static int    opt_verbose      = 0;
-static int    opt_profile      = 0;
+static int    opt_profile      = 0; // profiling mode (0-disabled, 1-basic, 2-pmu)
 static int    opt_hostbuf      = 1; // hostbuf ON by default
 static int    opt_use_hmx      = 1; // when set, enable HMX; when 0, use HVX only
 
+// Default PMU events, if profiling with PMU (mode=2) is enabled
+// See https://docs.qualcomm.com/doc/80-N2040-60/topic/pmu-events.html
+//     https://docs.qualcomm.com/doc/80-N2040-61/topic/hvx-pmu-events.html
+static u32vec opt_pmu_evt { 0x3, 0x111, 0x100, 0x105, 0x240, 0x256, 0x7D, 0x8C };
+
 // Enable all stages by default
-static int opt_opmask   = HTP_OPMASK_QUEUE | HTP_OPMASK_COMPUTE;
-static int opt_opsync   = 0;  // synchronous ops
+static int opt_opstage  = HTP_OPSTAGE_QUEUE | HTP_OPSTAGE_COMPUTE;
 static int opt_opbatch  = 1024; // max number of ops in a batch
 static int opt_opqueue  = 16;   // max number of pending batches
 static std::regex* opt_opfilter = NULL; // regex of ops to not claim
@@ -104,19 +115,26 @@ static void ggml_hexagon_dump_op_supp(const std::string &sess_name, const struct
 }
 
 static void ggml_hexagon_dump_op_prof(const std::string &sess_name, const ggml_tensor * op,
-                                      uint32_t op_usec, uint32_t op_cycles, uint32_t op_pkts, uint64_t call_usec) {
+                                      uint32_t op_usec, uint32_t op_cycles, const uint32_t pmu[]) {
     if (!opt_profile) return;
 
     op_desc desc(op);
-    GGML_LOG_DEBUG("ggml-hex: %s profile-op %s: %s : %s : %s : %s : %s : op-usec %u op-cycles %u op-pkts %u (%f) call-usec %llu\n", sess_name.c_str(),
-                ggml_op_desc(op), desc.names, desc.dims, desc.types, desc.strides, desc.buffs,
-                op_usec, op_cycles, op_pkts, (float) op_cycles / op_pkts, (unsigned long long) call_usec);
+
+    char pmu_str[256] = "";
+    if (opt_profile > 1) {
+        static_assert(HTP_PROF_PMU_NCNT == 8, "current implementation assumes 8 PMU counters");
+        sprintf(pmu_str, " pmu [%u,%u,%u,%u,%u,%u,%u,%u]",
+                pmu[0], pmu[1], pmu[2], pmu[3], pmu[4], pmu[5], pmu[6], pmu[7]);
+    }
+
+    GGML_LOG_DEBUG("ggml-hex: %s profile-op %s: %s : %s : %s : %s : usec %u cycles %u%s\n", sess_name.c_str(),
+            ggml_op_desc(op), desc.names, desc.dims, desc.types, desc.strides, op_usec, op_cycles, pmu_str);
 }
 
 // ** backend sessions
 
 struct ggml_hexagon_opbatch;
-struct ggml_hexagon_opshm;
+struct ggml_hexagon_opqueue;
 
 struct ggml_hexagon_session {
     std::string      name;
@@ -132,8 +150,8 @@ struct ggml_hexagon_session {
     bool             valid_iface;
 
     std::atomic<int>      op_pending;
-    ggml_hexagon_opbatch *op_batch;
-    ggml_hexagon_opshm   *op_shm;
+    ggml_hexagon_opbatch* op_batch;
+    ggml_hexagon_opqueue* op_queue;
 
     ggml_backend_buffer_type buffer_type        = {};
     ggml_backend_buffer_type repack_buffer_type = {};
@@ -1521,65 +1539,14 @@ static ggml_backend_buffer_type_i ggml_backend_hexagon_repack_buffer_type_interf
 
 // Backend session implementation
 
-struct ggml_hexagon_opshm {
-    ggml_hexagon_shared_buffer *sbuf;
-
-    std::vector<bool> block_mask;
-    size_t            block_size;
-
-    uint8_t * base()     const { return this->sbuf->base; }
-    int       fd()       const { return this->sbuf->fd;   }
-    size_t    n_blocks() const { return this->block_mask.size(); }
-
-    ggml_hexagon_opshm(ggml_hexagon_session *sess, size_t max_batch, size_t max_pending) {
-        size_t n_bufs    = HTP_OP_MAX_BUFS;
-        size_t n_ops     = max_batch;
-        size_t n_tensors = n_ops + n_ops * HTP_OP_MAX_INPUTS;
-
-        block_mask.resize(max_pending, true);
-
-        block_size = sizeof(htp_buf_desc) * n_bufs    +
-                     sizeof(htp_tensor)   * n_tensors +
-                     sizeof(htp_op_desc)  * n_ops;
-
-        sbuf = new ggml_hexagon_shared_buffer(sess, block_size * block_mask.size(), true /* pinned */);
-
-        if (opt_verbose) {
-            GGML_LOG_INFO("ggml-hex: %s allocated shared buf %zu : block-size %zu max-batch %zu max-pending %zu\n",
-                    sess->c_name(), (size_t) sbuf->size, block_size, max_batch, max_pending);
-        }
-    }
-
-    ~ggml_hexagon_opshm() {
-        delete sbuf;
-    }
-
-    uint8_t * allocate() {
-        auto it = std::find(block_mask.begin(), block_mask.end(), true);
-        if (it == block_mask.end())
-            return nullptr;
-
-        unsigned int i = std::distance(block_mask.begin(), it);
-        uint8_t*  addr = sbuf->base + (i * block_size);
-        block_mask[i]  = false;
-
-        HEX_VERBOSE("ggml-hex: %s allocated op shm #%u %p\n", sbuf->sess->c_name(), i, (void*) addr);
-        return addr;
-    }
-
-    void release(uint8_t * addr) {
-        int i = (addr - sbuf->base) / block_size;
-        block_mask[i] = true;
-        HEX_VERBOSE("ggml-hex: %s released op shm #%u %p\n", sbuf->sess->c_name(), i, (void*) addr);
-    }
-};
-
 struct ggml_hexagon_opbatch {
-    const char* name;
+    ggml_hexagon_session*            sess;
 
-    std::vector<htp_buf_desc> buffers;
-    std::vector<htp_tensor>   tensors;
-    std::vector<htp_op_desc>  ops;
+    std::vector<const ggml_tensor*>  ops;       // pointers to original ops
+
+    std::vector<htp_buf_desc>        h_bufs;    // htp buffer descriptors
+    std::vector<htp_tensor>          h_tens;    // htp tensor descriptors
+    std::vector<htp_op_desc>         h_ops;     // htp op descriptors
 
     std::unordered_map<int, int>                b_map; // buffer fd   to index
     std::unordered_map<const ggml_tensor*, int> t_map; // tensor ptr  to index
@@ -1606,19 +1573,21 @@ struct ggml_hexagon_opbatch {
         d_map.clear();
     }
 
-    ggml_hexagon_opbatch(ggml_hexagon_session *sess, size_t max_batch) {
-        name = sess->c_name();
+    ggml_hexagon_opbatch(ggml_hexagon_session *sess, size_t batch_size) {
+        this->sess = sess;
 
         n_bufs_max = HTP_OP_MAX_BUFS;
-        n_ops_max  = max_batch;
+        n_ops_max  = batch_size;
         n_tens_max = n_ops_max + n_ops_max * HTP_OP_MAX_INPUTS;
 
         b_vmem_max = HTP_OP_MAX_VMEM;
 
-        buffers.resize(n_bufs_max);
-        tensors.resize(n_tens_max);
         ops.resize(n_ops_max);
 
+        h_bufs.resize(n_bufs_max);
+        h_tens.resize(n_tens_max);
+        h_ops.resize(n_ops_max);
+
         b_map.reserve(n_bufs_max);
         t_map.reserve(n_tens_max);
         d_map.reserve(n_tens_max);
@@ -1640,7 +1609,7 @@ struct ggml_hexagon_opbatch {
 
         b_map.insert({sbuf->fd, bi});
 
-        htp_buf_desc &b = buffers[bi];
+        htp_buf_desc &b = h_bufs[bi];
         b.base = (uint64_t) sbuf->base;
         b.fd   = sbuf->fd;
         b.size = sbuf->size;
@@ -1664,7 +1633,7 @@ struct ggml_hexagon_opbatch {
         // First lookup by tensor data
         auto range = d_map.equal_range(t->data);
         for (auto it = range.first; it != range.second; ++it) {
-            htp_tensor * h = &tensors[it->second];
+            htp_tensor * h = &h_tens[it->second];
             if (same_shape(h, t)) { return it->second; }
         }
 
@@ -1682,7 +1651,7 @@ struct ggml_hexagon_opbatch {
         uint64_t t_offset = (uint8_t *) t->data - sbuf->base;
         size_t   t_size   = ggml_nbytes(t);
 
-        htp_tensor &h = tensors[ti];
+        htp_tensor &h = h_tens[ti];
         h.bi    = add_buffer(sbuf);
         h.data  = t_offset;
         h.size  = t_size;
@@ -1737,65 +1706,170 @@ struct ggml_hexagon_opbatch {
     // assumes that fit_op() was called first and returned true
     void add_op(htp_op_code opcode, const struct ggml_tensor * t) {
         // Add new op
-        htp_op_desc &o = ops[n_ops++];
+
+        unsigned int n = n_ops++;
         GGML_ASSERT(n_ops <= n_ops_max);
 
+        ops[n] = t;
+
+        htp_op_desc &o = h_ops[n];
         memcpy(&o.params, &t->op_params, sizeof(t->op_params));
         o.opcode = opcode;
         o.flags  = 0;
 
-        if (!(opt_opmask & HTP_OPMASK_COMPUTE)) {
+        if (!(opt_opstage & HTP_OPSTAGE_COMPUTE)) {
             o.flags |= HTP_OPFLAGS_SKIP_COMPUTE;
         }
 
-        ggml_hexagon_dump_op_exec(name, t, o.flags);
+        ggml_hexagon_dump_op_exec(sess->c_name(), t, o.flags);
 
         for (unsigned int i=0; i < HTP_OP_MAX_INPUTS; i++) {
             o.src[i] = t->src[i] ? add_tensor(t->src[i]) : 0xffff;
         }
         o.dst = add_tensor(t);
     }
+};
 
-    size_t flush(uint8_t * mem_addr, size_t mem_size) {
-        static_assert(sizeof(htp_buf_desc) % 8 == 0, "sizeof(htp_buf_desc) must be multiple of 8");
-        static_assert(sizeof(htp_tensor)   % 8 == 0, "sizeof(htp_tensor) must be multiple of 8");
-        static_assert(sizeof(htp_op_desc)  % 8 == 0, "sizeof(htp_op_desc) must be multiple of 8");
+struct ggml_hexagon_opqueue {
+    // Shared buffer for storing batches
+    ggml_hexagon_shared_buffer *shm_buf;
+    size_t                      shm_blk_size;
 
-        const size_t b_size = sizeof(htp_buf_desc) * n_bufs;
-        const size_t t_size = sizeof(htp_tensor)   * n_tens;
-        const size_t o_size = sizeof(htp_op_desc)  * n_ops;
+    using opvec = std::vector<const ggml_tensor*>;
 
-        const size_t m_size = b_size + t_size + o_size;
-        GGML_ASSERT(m_size <= mem_size);
+    std::queue<unsigned int>    done;       // completed batch ids
+    std::vector<opvec>          op_cache;   // per batch op cache
+    std::vector<uint64_t>       start_usec; // per batch start time
 
-        uint8_t * b_ptr = (uint8_t *) mem_addr;
-        uint8_t * t_ptr = (uint8_t *) b_ptr + b_size;
-        uint8_t * o_ptr = (uint8_t *) t_ptr + t_size;
+    ggml_hexagon_opqueue(ggml_hexagon_session *sess, size_t batch_size, size_t depth) {
+        size_t n_bufs    = HTP_OP_MAX_BUFS;
+        size_t n_ops     = batch_size;
+        size_t n_tensors = n_ops + n_ops * HTP_OP_MAX_INPUTS;
 
-        memcpy(b_ptr, (void *) buffers.data(), b_size);
-        memcpy(t_ptr, (void *) tensors.data(), t_size);
-        memcpy(o_ptr, (void *) ops.data(),     o_size);
+        shm_blk_size = sizeof(htp_buf_desc)  * n_bufs    +
+                       sizeof(htp_tensor)    * n_tensors +
+                       sizeof(htp_op_desc)   * n_ops     +
+                       sizeof(htp_prof_desc) * n_ops;
 
-        HEX_VERBOSE("ggml-hex: %s flush-opbatch : n-bufs %u n-tensors %u n-ops %u vmem %zu : b-size %zu t-size %zu o-size %zu\n",
-                name, n_bufs, n_tens, n_ops, b_vmem, b_size, t_size, o_size);
+        shm_buf = new ggml_hexagon_shared_buffer(sess, shm_blk_size * depth, true /* pinned */);
+
+        op_cache.resize(depth);
+        start_usec.resize(depth, 0);
+
+        // init done queue
+        for (unsigned int i = 0; i < depth; i++) { done.push(i); }
+
+        if (opt_verbose) {
+            GGML_LOG_INFO("ggml-hex: %s allocated op-queue : batch-size %zu depth %zu shm-size %zu shm-block-size %zu\n",
+                    sess->c_name(), batch_size, depth, shm_buf->size, shm_blk_size);
+        }
+    }
+
+    ~ggml_hexagon_opqueue() {
+        delete shm_buf;
+    }
+
+    // push new batch
+    bool push(htp_opbatch_req& req, dspqueue_buffer& dbuf, ggml_hexagon_opbatch* op_batch) {
+        static_assert(sizeof(htp_opbatch_req) % 8 == 0, "sizeof(htp_opbatch_req) must be multiple of 8");
+        static_assert(sizeof(htp_opbatch_rsp) % 8 == 0, "sizeof(htp_opbatch_rsp) must be multiple of 8");
+        static_assert(sizeof(htp_buf_desc)    % 8 == 0, "sizeof(htp_buf_desc) must be multiple of 8");
+        static_assert(sizeof(htp_tensor)      % 8 == 0, "sizeof(htp_tensor) must be multiple of 8");
+        static_assert(sizeof(htp_op_desc)     % 8 == 0, "sizeof(htp_op_desc) must be multiple of 8");
+        static_assert(sizeof(htp_prof_desc)   % 8 == 0, "sizeof(htp_prof_desc) must be multiple of 8");
+
+        if (done.empty()) { return false; }
+
+        req.id        = done.front(); done.pop(); // batch id
+        req.n_bufs    = op_batch->n_bufs;
+        req.n_tensors = op_batch->n_tens;
+        req.n_ops     = op_batch->n_ops;
+
+        op_cache[req.id]   = op_batch->ops;
+        start_usec[req.id] = ggml_time_us();
+
+        const size_t b_size = sizeof(htp_buf_desc)  * req.n_bufs;
+        const size_t t_size = sizeof(htp_tensor)    * req.n_tensors;
+        const size_t o_size = sizeof(htp_op_desc)   * req.n_ops;
+        const size_t p_size = sizeof(htp_prof_desc) * req.n_ops;
+
+        dbuf.ptr      = shm_buf->base + (req.id * shm_blk_size);
+        dbuf.fd       = shm_buf->fd;
+        dbuf.flags    = DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT;
+        dbuf.offset   = (uint8_t*) dbuf.ptr - (uint8_t*) shm_buf->base;
+        dbuf.size     = b_size + t_size + o_size + p_size;
+
+        GGML_ASSERT(dbuf.size <= shm_blk_size);
+
+        uint8_t * m_ptr = (uint8_t*) dbuf.ptr;
+        uint8_t * b_ptr = m_ptr; m_ptr += b_size;
+        uint8_t * t_ptr = m_ptr; m_ptr += t_size;
+        uint8_t * o_ptr = m_ptr;
+
+        memcpy(b_ptr, (void *) op_batch->h_bufs.data(), b_size);
+        memcpy(t_ptr, (void *) op_batch->h_tens.data(), t_size);
+        memcpy(o_ptr, (void *) op_batch->h_ops.data(),  o_size);
+
+        HEX_VERBOSE("ggml-hex: %s op-queue push batch #%u : n-bufs %u n-tensors %u n-ops %u vmem %zu : b-size %zu t-size %zu o-size %zu m-size %zu\n",
+                shm_buf->sess->c_name(), req.id, req.n_bufs, req.n_tensors, req.n_ops, op_batch->b_vmem,
+                b_size, t_size, o_size, (size_t) dbuf.size);
+
+        op_batch->reset();
 
         if (opt_verbose > 1) {
             htp_buf_desc *b = (htp_buf_desc*) b_ptr;
-            for (unsigned int i=0; i < n_bufs; i++) {
-                GGML_LOG_DEBUG("ggml-hex: %s htp-buf #%u : fd %d base %p size %zu\n", name, i,
+            for (unsigned int i=0; i < req.n_bufs; i++) {
+                GGML_LOG_DEBUG("ggml-hex: %s htp-buf #%u : fd %d base %p size %zu\n", shm_buf->sess->c_name(), i,
                             b[i].fd, (void *) b[i].base, (size_t) b[i].size);
             }
             htp_tensor *t = (htp_tensor*) t_ptr;
-            for (unsigned int i=0; i < n_tens; i++) {
+            for (unsigned int i=0; i < req.n_tensors; i++) {
                 GGML_LOG_DEBUG("ggml-hex: %s htp-tensor #%u : bi %u offset %u size %u : %zu:%zu:%zu:%zu\n",
-                            name, i, t[i].bi, t[i].data, t[i].size,
+                            shm_buf->sess->c_name(), i, t[i].bi, t[i].data, t[i].size,
                             (size_t) t[i].ne[0], (size_t) t[i].ne[1], (size_t) t[i].ne[2], (size_t) t[i].ne[3]);
             }
         }
 
-        reset();
+        return true;
+    }
 
-        return m_size;
+    void pop(htp_opbatch_rsp rsp, dspqueue_buffer dbuf) {
+        GGML_ASSERT(rsp.id < op_cache.size());
+
+        done.push(rsp.id);
+
+        const size_t b_size = sizeof(htp_buf_desc)  * rsp.n_bufs;
+        const size_t t_size = sizeof(htp_tensor)    * rsp.n_tensors;
+        const size_t o_size = sizeof(htp_op_desc)   * rsp.n_ops;
+        const size_t p_size = sizeof(htp_prof_desc) * rsp.n_ops;
+
+        const size_t m_size = b_size + t_size + o_size + p_size;
+        GGML_ASSERT(m_size <= shm_blk_size);
+
+        HEX_VERBOSE("ggml-hex: %s op-queue pop batch #%u : n-bufs %u n-tensors %u n-ops %u : m-size %zu b-size %zu t-size %zu o-size %zu\n",
+                shm_buf->sess->c_name(), rsp.id, rsp.n_bufs, rsp.n_tensors, rsp.n_ops,
+                (size_t) dbuf.size, b_size, t_size, o_size);
+
+        uint8_t * m_ptr = (uint8_t*) dbuf.ptr;
+        uint8_t * p_ptr = m_ptr + (b_size + t_size + o_size);
+
+        if (opt_profile && rsp.n_ops > 0) {
+            auto & ops = op_cache[rsp.id];
+
+            uint64_t batch_usec = ggml_time_us() - start_usec[rsp.id];
+            uint32_t htp_usec   = 0;
+
+            GGML_ASSERT(rsp.n_ops <= ops.size());
+
+            const htp_prof_desc * pd = (const htp_prof_desc *) p_ptr;
+            for (uint32_t i = 0; i < rsp.n_ops; i++) {
+                htp_usec += pd[i].usecs;
+                ggml_hexagon_dump_op_prof(shm_buf->sess->name, ops[i], pd[i].usecs, pd[i].cycles, pd[i].pmu);
+            }
+
+            GGML_LOG_DEBUG("ggml-hex: %s profile-batch n-ops %u batch-dur-usec %lld htp-ops-usec %u\n",
+                           shm_buf->sess->c_name(), rsp.n_ops, (long long) batch_usec, htp_usec);
+        }
     }
 };
 
@@ -1824,17 +1898,12 @@ void ggml_hexagon_session::flush_pending(bool all) {
             GGML_ABORT("ggml-hex: %s dspcall : bad response : size %u dspbufs %u\n", this->c_name(), rsp_size, n_dbufs);
         }
 
-        op_shm->release((uint8_t*) dbuf.ptr);
-
         if (rsp.status != HTP_STATUS_OK) {
             GGML_LOG_ERROR("ggml-hex: %s dspcall : dsp-rsp: %s\n", this->c_name(), status_to_str(rsp.status));
             // TODO: handle errors
         }
 
-        // FIXME: profile will be per opreq
-        // this->prof_usecs  = rsp.prof_usecs;
-        // this->prof_cycles = rsp.prof_cycles;
-        // this->prof_pkts   = rsp.prof_pkts;
+        op_queue->pop(rsp, dbuf);
 
         this->op_pending--;  // atomic dec
 
@@ -1845,28 +1914,17 @@ void ggml_hexagon_session::flush_pending(bool all) {
 void ggml_hexagon_session::flush_batch() {
     if (op_batch->empty()) { return; }
 
-    htp_opbatch_req req;
-    req.n_bufs    = op_batch->n_bufs;
-    req.n_tensors = op_batch->n_tens;
-    req.n_ops     = op_batch->n_ops;
+    htp_opbatch_req req {};
+    dspqueue_buffer dbuf{};
 
-    dspqueue_buffer dbuf;
-    dbuf.fd     = op_shm->fd();
-    dbuf.flags  = DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT;
-    dbuf.ptr    = op_shm->allocate();
-    if (!dbuf.ptr) {
+    if (!op_queue->push(req, dbuf, op_batch)) {
         flush_pending(false);
-        dbuf.ptr = op_shm->allocate();
+        op_queue->push(req, dbuf, op_batch);
     }
 
-    dbuf.offset = (uint8_t*) dbuf.ptr - (uint8_t*) op_shm->base();
-    dbuf.size   = op_batch->flush((uint8_t*) dbuf.ptr, op_shm->block_size);
-
     // Bump pending flag (cleared in the session::flush once we get the response)
     this->op_pending++;  // atomic inc
 
-    HEX_VERBOSE("ggml-hex: %s: queue-opbatch : %p size %u\n", this->c_name(), dbuf.ptr, dbuf.size);
-
     int err = dspqueue_write(this->queue, 0, 1, &dbuf, sizeof(req), (const uint8_t*) &req, DSPQUEUE_TIMEOUT);
     if (err != 0) {
         GGML_ABORT("ggml-hex: %s dspqueue_write failed: 0x%08x\n", this->c_name(), (unsigned) err);
@@ -2016,25 +2074,33 @@ void ggml_hexagon_session::allocate(int dev_id) noexcept(false) {
     }
 
     if (opt_etm) {
-        err = htp_iface_enable_etm(this->handle);
+        err = htp_iface_etm(this->handle, 1);
         if (err != 0) {
             GGML_LOG_ERROR("ggml-hex: failed to enable ETM tracing: 0x%08x\n", (unsigned) err);
         }
     }
 
-    // Start the DSP-side service. We need to pass the queue ID to the
-    // DSP in a FastRPC call; the DSP side will import the queue and start
-    // listening for packets in a callback.
+    if (opt_profile) {
+        htp_iface_pmu_conf pmu_conf{};
+        std::copy(opt_pmu_evt.begin(), opt_pmu_evt.end(), pmu_conf.events);
+
+        err = htp_iface_profiler(this->handle, opt_profile, &pmu_conf);
+        if (err != 0) {
+            GGML_LOG_ERROR("ggml-hex: failed to enable profiling: 0x%08x\n", (unsigned) err);
+        }
+    }
+
+    // Allocate buffers and state for op batching
+    this->op_batch = new ggml_hexagon_opbatch(this, opt_opbatch);
+    this->op_queue = new ggml_hexagon_opqueue(this, opt_opbatch, opt_opqueue);
+
+    // Start processing op batch requests
     err = htp_iface_start(this->handle, dev_id, this->queue_id, opt_nhvx, opt_use_hmx);
     if (err != 0) {
         GGML_LOG_ERROR("ggml-hex: failed to start session: 0x%08x\n", (unsigned) err);
         throw std::runtime_error("ggml-hex: iface start failed (see log for details)");
     }
     this->valid_iface = true;
-
-    // Allocate buffers and state for op batching
-    this->op_batch = new ggml_hexagon_opbatch(this, opt_opbatch);
-    this->op_shm   = new ggml_hexagon_opshm(this, opt_opbatch, opt_opqueue);
 }
 
 void ggml_hexagon_session::release() noexcept(true) {
@@ -2043,7 +2109,7 @@ void ggml_hexagon_session::release() noexcept(true) {
     int err;
 
     delete this->op_batch;
-    delete this->op_shm;
+    delete this->op_queue;
 
     // Stop the DSP-side service and close the queue
     if (this->valid_iface) {
@@ -2054,12 +2120,20 @@ void ggml_hexagon_session::release() noexcept(true) {
     }
 
     if (opt_etm) {
-        err = htp_iface_disable_etm(this->handle);
+        err = htp_iface_etm(this->handle, 0);
         if (err != 0) {
             GGML_LOG_ERROR("ggml-hex: warn : failed to disable ETM tracing: 0x%08x\n", (unsigned) err);
         }
     }
 
+    if (opt_profile) {
+        htp_iface_pmu_conf pmu_conf{};
+        err = htp_iface_profiler(this->handle, 0, &pmu_conf);
+        if (err != 0) {
+            GGML_LOG_ERROR("ggml-hex: warn : failed to disable profiling: 0x%08x\n", (unsigned) err);
+        }
+    }
+
     if (this->valid_queue) {
         err = dspqueue_close(queue);
         if (err != 0) {
@@ -2077,7 +2151,7 @@ ggml_hexagon_session::ggml_hexagon_session(int dev_id, ggml_backend_dev_t dev) n
     repack_buffer_type.device = dev;
 
     op_batch = nullptr;
-    op_shm   = nullptr;
+    op_queue = nullptr;
 
     try {
         allocate(dev_id);
@@ -2596,6 +2670,62 @@ static bool ggml_hexagon_supported_cumsum(const struct ggml_hexagon_session * se
     return true;
 }
 
+static bool ggml_hexagon_supported_diag(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * dst  = op;
+
+    // diag only supports F32 currently
+    if (src0->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    // Input must have ne[1] == 1 (vector input)
+    if (src0->ne[1] != 1) {
+        return false;
+    }
+
+    // Output must be square in first two dimensions
+    if (dst->ne[0] != dst->ne[1] || dst->ne[0] != src0->ne[0]) {
+        return false;
+    }
+
+    GGML_UNUSED(sess);
+    return true;
+}
+
+static bool ggml_hexagon_supported_solve_tri(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0]; // A
+    const struct ggml_tensor * src1 = op->src[1]; // B
+    const struct ggml_tensor * dst  = op;         // X
+
+    if (!src0 || !src1) {
+        return false;
+    }
+
+    if (src0->type != GGML_TYPE_F32 || src1->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    if (src0->ne[0] != src0->ne[1]) {
+        return false;
+    }
+
+    if (src0->ne[1] != src1->ne[1]) {
+        return false;
+    }
+
+    if (src0->ne[2] != src1->ne[2] || src0->ne[3] != src1->ne[3]) {
+        return false;
+    }
+
+    if (dst->ne[0] != src1->ne[0] || dst->ne[1] != src1->ne[1] || dst->ne[2] != src1->ne[2] || dst->ne[3] != src1->ne[3]) {
+        return false;
+    }
+
+    GGML_UNUSED(sess);
+    return true;
+}
+
 static const char * ggml_backend_hexagon_name(ggml_backend_t backend) {
     auto sess = static_cast<ggml_hexagon_session *>(backend->context);
     return sess->c_name();
@@ -2632,7 +2762,9 @@ static htp_op_code op_remap_to_htp(const ggml_tensor * t) {
         case GGML_OP_ROPE:           return HTP_OP_ROPE;
         case GGML_OP_REPEAT:         return HTP_OP_REPEAT;
         case GGML_OP_CUMSUM:         return HTP_OP_CUMSUM;
-
+        case GGML_OP_FILL:           return HTP_OP_FILL;
+        case GGML_OP_DIAG:           return HTP_OP_DIAG;
+        case GGML_OP_SOLVE_TRI:      return HTP_OP_SOLVE_TRI;
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(t)) {
                 case GGML_UNARY_OP_SILU:     return HTP_OP_UNARY_SILU;
@@ -2673,7 +2805,7 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg
 
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * n = graph->nodes[i];
-        if (op_is_compute(n)) {
+        if (op_is_compute(n) && (opt_opstage & HTP_OPSTAGE_QUEUE)) {
             sess->enqueue_op(op_remap_to_htp(n), n);
         }
     }
@@ -3029,6 +3161,17 @@ static bool ggml_hexagon_supported_repeat(const struct ggml_hexagon_session * se
     return true;
 }
 
+static bool ggml_hexagon_supported_fill(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
+    const struct ggml_tensor * dst = op;
+
+    if (dst->type != GGML_TYPE_F32 && dst->type != GGML_TYPE_F16) {
+        return false;
+    }
+
+    GGML_UNUSED(sess);
+    return true;
+}
+
 static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
     auto sess = static_cast<ggml_hexagon_session *>(dev->context);
 
@@ -3159,6 +3302,18 @@ static bool ggml_backend_hexagon_device_supports_op(ggml_backend_dev_t dev, cons
             supp = ggml_hexagon_supported_cumsum(sess, op);
             break;
 
+        case GGML_OP_FILL:
+            supp = ggml_hexagon_supported_fill(sess, op);
+            break;
+
+        case GGML_OP_DIAG:
+            supp = ggml_hexagon_supported_diag(sess, op);
+            break;
+
+        case GGML_OP_SOLVE_TRI:
+            supp = ggml_hexagon_supported_solve_tri(sess, op);
+            break;
+
         default:
             break;
     }
@@ -3294,6 +3449,26 @@ static void * ggml_backend_hexagon_get_proc_address(ggml_backend_reg_t reg, cons
     return NULL;
 }
 
+template<typename T> std::vector<T> str_to_vec(const char* str) {
+    std::stringstream ss(str);
+    std::vector<T> v;
+    std::string    t;
+
+    while (std::getline(ss, t, ',')) {
+        v.push_back(std::stoul(t, nullptr, 0));
+    }
+
+    return v;
+}
+
+template<typename T, int BASE=10> std::string vec_to_str(std::vector<T> v) {
+    std::stringstream ss;
+    ss << std::setbase(BASE) << std::showbase;
+    for (auto i : v) { ss << i << ','; }
+    auto str = ss.str(); str.pop_back(); // drop last comma
+    return str;
+}
+
 static void ggml_hexagon_init(ggml_backend_reg * reg) {
     // Basic sanity checks to make sure definitions match
     static_assert((unsigned int) HTP_TYPE_Q4_0 == (unsigned int) GGML_TYPE_Q4_0,
@@ -3307,8 +3482,7 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
 
     const char * str_verbose = getenv("GGML_HEXAGON_VERBOSE");
     const char * str_hostbuf = getenv("GGML_HEXAGON_HOSTBUF");
-    const char * str_opmask  = getenv("GGML_HEXAGON_OPMASK");
-    const char * str_opsync  = getenv("GGML_HEXAGON_OPSYNC");
+    const char * str_opstage = getenv("GGML_HEXAGON_OPSTAGE");
     const char * str_opbatch = getenv("GGML_HEXAGON_OPBATCH");
     const char * str_opqueue = getenv("GGML_HEXAGON_OPQUEUE");
     const char * str_opfilter= getenv("GGML_HEXAGON_OPFILTER");
@@ -3321,19 +3495,30 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
 
     auto RE_ICASE = std::regex_constants::icase;
 
-    opt_opfilter     = str_opfilter     ? new std::regex(str_opfilter, RE_ICASE) : NULL;
-    opt_verbose      = str_verbose ? atoi(str_verbose) : 0;
-    opt_hostbuf      = str_hostbuf ? atoi(str_hostbuf) : opt_hostbuf;
-    opt_opmask       = str_opmask  ? strtoul(str_opmask, NULL, 0)  : opt_opmask;
-    opt_opsync       = str_opsync  ? atoi(str_opsync)              : opt_opsync;
-    opt_opbatch      = str_opbatch ? strtoul(str_opbatch, NULL, 0) : opt_opbatch;
-    opt_opqueue      = str_opqueue ? strtoul(str_opqueue, NULL, 0) : opt_opqueue;
-    opt_profile      = str_profile ? atoi(str_profile) : 0;
-    opt_etm          = str_etm     ? atoi(str_etm)     : 0;
-    opt_nhvx         = str_nhvx    ? strtoul(str_nhvx, NULL, 0) : opt_nhvx;
-    opt_use_hmx      = str_use_hmx ? atoi(str_use_hmx) : opt_use_hmx;
-    opt_ndev         = str_ndev    ? strtoul(str_ndev, NULL, 0) : opt_ndev;
-    opt_hostbuf      = str_hostbuf ? atoi(str_hostbuf) : opt_hostbuf;
+    opt_opfilter     = str_opfilter ? new std::regex(str_opfilter, RE_ICASE) : NULL;
+    opt_verbose      = str_verbose  ? atoi(str_verbose)             : 0;
+    opt_hostbuf      = str_hostbuf  ? atoi(str_hostbuf)             : opt_hostbuf;
+    opt_opstage      = str_opstage  ? strtoul(str_opstage, NULL, 0) : opt_opstage;
+    opt_opbatch      = str_opbatch  ? strtoul(str_opbatch, NULL, 0) : opt_opbatch;
+    opt_opqueue      = str_opqueue  ? strtoul(str_opqueue, NULL, 0) : opt_opqueue;
+    opt_etm          = str_etm      ? atoi(str_etm)                 : 0;
+    opt_nhvx         = str_nhvx     ? strtoul(str_nhvx, NULL, 0)    : opt_nhvx;
+    opt_use_hmx      = str_use_hmx  ? atoi(str_use_hmx)             : opt_use_hmx;
+    opt_ndev         = str_ndev     ? strtoul(str_ndev, NULL, 0)    : opt_ndev;
+    opt_hostbuf      = str_hostbuf  ? atoi(str_hostbuf)             : opt_hostbuf;
+
+    if (str_profile) {
+        opt_pmu_evt = [&]() -> std::vector<uint32_t> {
+            auto v  = str_to_vec<uint32_t>(str_profile);
+            switch (v.size()) {
+                case 1:  opt_profile = v[0]; return opt_pmu_evt; // mode with default pmu events
+                case 8:  opt_profile = 2;    return v;           // mode with custom  pmu events
+                default: opt_profile = 0;    return {};          // garbage input
+            }}();
+        if (opt_profile == 1) opt_pmu_evt = {};
+        GGML_LOG_INFO("ggml-hex: Profiling mode %u : pmu-evt [ %s ]\n", opt_profile,
+                vec_to_str<uint32_t, 16>(opt_pmu_evt).c_str());
+    }
 
     if (opt_ndev > GGML_HEXAGON_MAX_SESSIONS) {
         opt_ndev = GGML_HEXAGON_MAX_SESSIONS;

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -34,6 +34,9 @@ add_library(${HTP_LIB} SHARED
     argsort-ops.c
     ssm-conv.c
     cumsum-ops.c
+    fill-ops.c
+    diag-ops.c
+    solve-tri-ops.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

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

@@ -0,0 +1,216 @@
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "hvx-types.h"
+#include "hex-utils.h"
+#include "hvx-copy.h"
+#include "hex-dma.h"
+
+#define htp_diag_tensors_preamble                           \
+    const struct htp_tensor * restrict src0 = octx->src[0]; \
+    const struct htp_tensor * restrict dst  = octx->dst;    \
+                                                     \
+    const uint32_t ne02 = src0->ne[2];               \
+                                                     \
+    const uint32_t ne0 = dst->ne[0];                 \
+    const uint32_t ne1 = dst->ne[1];                 \
+                                                     \
+    const uint32_t nb02 = src0->nb[2];               \
+    const uint32_t nb03 = src0->nb[3];               \
+                                                     \
+    const uint32_t nb1 = dst->nb[1];                 \
+    const uint32_t nb2 = dst->nb[2];                 \
+    const uint32_t nb3 = dst->nb[3];
+
+struct htp_diag_context {
+    struct htp_ops_context * octx;
+    size_t          src_batch_size;
+    size_t          dst_row_size;
+    size_t          src_batch_size_aligned;
+    size_t          dst_row_size_aligned;
+    uint32_t        batches_per_thread;
+    uint32_t        total_batches;
+};
+
+#define htp_diag_preamble                                              \
+    struct htp_diag_context * dctx = (struct htp_diag_context *) data; \
+    struct htp_ops_context *  octx = dctx->octx;                       \
+    htp_diag_tensors_preamble;
+
+static inline void hvx_diag_row_f32(const float * restrict src, float * restrict dst,
+                                    uint32_t row_idx, uint32_t n) {
+    hvx_splat_f32_a((uint8_t *) dst, 0.0f, n);
+    dst[row_idx] = src[row_idx];
+}
+
+// ---------------------------------------------------------------------------
+// Per thread worker: DMA src fetch, compute in VTCM, DMA dst writeback
+// ---------------------------------------------------------------------------
+
+static void diag_thread_f32_dma(unsigned int nth, unsigned int ith, void * data) {
+    htp_diag_preamble;
+    dma_queue * dma_queue = octx->ctx->dma[ith];
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint32_t ib0 = dctx->batches_per_thread * ith;
+    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->total_batches);
+
+    if (ib0 >= ib1) {
+        return;
+    }
+
+    const size_t src_batch_size         = dctx->src_batch_size;
+    const size_t dst_row_size           = dctx->dst_row_size;
+    const size_t src_batch_size_aligned = dctx->src_batch_size_aligned;
+    const size_t dst_row_size_aligned   = dctx->dst_row_size_aligned;
+
+    const uint8_t * src_data = (const uint8_t *) src0->data;
+    uint8_t *       dst_data = (uint8_t *) dst->data;
+
+    // 1 src buffer + 1 dst row buffer per thread in VTCM
+    uint8_t * src_spad = octx->src0_spad.data + (ith * src_batch_size_aligned);
+    uint8_t * dst_spad = octx->dst_spad.data  + (ith * dst_row_size_aligned);
+
+    for (uint32_t ib = ib0; ib < ib1; ib++) {
+        const uint32_t i3 = ib / ne02;
+        const uint32_t i2 = ib % ne02;
+
+        const uint8_t * src_batch = src_data + i3 * nb03 + i2 * nb02;
+
+        // Fetch source vector into VTCM
+        dma_queue_push_ddr_to_vtcm(dma_queue,
+                                   dma_make_ptr(src_spad, src_batch),
+                                   src_batch_size_aligned, src_batch_size, 1);
+        dma_queue_flush(dma_queue);
+
+        const float * src_spad_f32 = (const float *) src_spad;
+        float       * dst_spad_f32 = (float *) dst_spad;
+
+        for (uint32_t i1 = 0; i1 < ne1; i1++) {
+            // Compute row in VTCM
+            hvx_diag_row_f32(src_spad_f32, dst_spad_f32, i1, ne0);
+
+            // Write completed row back to DDR
+            uint8_t * dst_row = dst_data + i3 * nb3 + i2 * nb2 + i1 * nb1;
+            dma_queue_push_vtcm_to_ddr(dma_queue,
+                                       dma_make_ptr(dst_row, dst_spad),
+                                       dst_row_size, dst_row_size_aligned, 1);
+            dma_queue_flush(dma_queue);
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "diag-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+// ---------------------------------------------------------------------------
+// Per thread worker: Direct HVX (no DMA)
+// ---------------------------------------------------------------------------
+
+static void diag_thread_f32(unsigned int nth, unsigned int ith, void * data) {
+    htp_diag_preamble;
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    const uint8_t * src_data = (const uint8_t *) src0->data;
+    uint8_t *       dst_data = (uint8_t *) dst->data;
+
+    const uint32_t ib0 = dctx->batches_per_thread * ith;
+    const uint32_t ib1 = MIN(ib0 + dctx->batches_per_thread, dctx->total_batches);
+
+    for (uint32_t ib = ib0; ib < ib1; ib++) {
+        const uint32_t i3 = ib / ne02;
+        const uint32_t i2 = ib % ne02;
+
+        const float * restrict src_batch = (const float *)(src_data + i3 * nb03 + i2 * nb02);
+
+        for (uint32_t i1 = 0; i1 < ne1; i1++) {
+            float * restrict dst_row = (float *)(dst_data + i3 * nb3 + i2 * nb2 + i1 * nb1);
+            hvx_diag_row_f32(src_batch, dst_row, i1, ne0);
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "diag-f32 %d/%d: %ux%ux%ux%u (%u:%u) -> %ux%ux%ux%u usec %u\n",
+         ith, nth, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3], ib0, ib1,
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+int op_diag_f32(struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * dst  = octx->dst;
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t total_batches = src0->ne[2] * src0->ne[3];
+    const uint32_t n_threads     = MIN(octx->n_threads, total_batches);
+
+    const size_t src_batch_size         = src0->ne[0] * sizeof(float);
+    const size_t dst_row_size           = dst->ne[0] * sizeof(float);
+    const size_t src_batch_size_aligned = hex_round_up(src_batch_size, VLEN);
+    const size_t dst_row_size_aligned   = hex_round_up(dst_row_size, VLEN);
+
+    // 1 src buffer + 1 dst row buffer per thread
+    const size_t spad_per_thread = src_batch_size_aligned + dst_row_size_aligned;
+
+    octx->src0_spad.size_per_thread = src_batch_size_aligned;
+    octx->dst_spad.size_per_thread  = dst_row_size_aligned;
+
+    octx->src0_spad.size = n_threads * octx->src0_spad.size_per_thread;
+    octx->dst_spad.size  = n_threads * octx->dst_spad.size_per_thread;
+
+    octx->src0_spad.data = octx->ctx->vtcm_base;                        octx->src0_spad.src = NULL;
+    octx->dst_spad.data  = octx->src0_spad.data + octx->src0_spad.size; octx->dst_spad.src  = NULL;
+
+    struct htp_diag_context dctx = {
+        .octx                   = octx,
+        .src_batch_size         = src_batch_size,
+        .dst_row_size           = dst_row_size,
+        .src_batch_size_aligned = src_batch_size_aligned,
+        .dst_row_size_aligned   = dst_row_size_aligned,
+        .batches_per_thread     = (total_batches + n_threads - 1) / n_threads,
+        .total_batches          = total_batches,
+    };
+
+    if (octx->ctx->vtcm_size < spad_per_thread * n_threads) {
+        worker_pool_run_func(octx->ctx->worker_pool, diag_thread_f32, &dctx, n_threads);
+    } else {
+        worker_pool_run_func(octx->ctx->worker_pool, diag_thread_f32_dma, &dctx, n_threads);
+    }
+
+    return HTP_STATUS_OK;
+}
+
+int op_diag(struct htp_ops_context * octx) {
+    const struct htp_tensor * dst = octx->dst;
+
+    int err = HTP_STATUS_OK;
+
+    switch (dst->type) {
+        case HTP_TYPE_F32:
+            err = op_diag_f32(octx);
+            break;
+        default:
+            err = HTP_STATUS_NO_SUPPORT;
+            break;
+    }
+
+    return err;
+}

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

@@ -0,0 +1,123 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <string.h>
+
+#include "hvx-copy.h"
+#include "hvx-utils.h"
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-ops.h"
+
+// ggml op_params layout for FILL:
+//   op_params[0] (as float) - the scalar fill value
+
+#define fill_preamble \
+    const struct htp_tensor * dst = octx->dst; \
+    \
+    const uint32_t ne0 = dst->ne[0]; \
+    const uint32_t ne1 = dst->ne[1]; \
+    const uint32_t ne2 = dst->ne[2]; \
+    const uint32_t ne3 = dst->ne[3]; \
+    \
+    const uint32_t nb1 = dst->nb[1]; \
+    const uint32_t nb2 = dst->nb[2]; \
+    const uint32_t nb3 = dst->nb[3]; \
+    \
+    const uint32_t nr = ne1 * ne2 * ne3;
+
+struct htp_fill_context {
+    struct htp_ops_context * octx;
+    uint32_t nrows_per_thread;
+    uint32_t total_rows;  // ne1 * ne2 * ne3
+    bool     opt_path;
+    HVX_Vector splat_vec;
+    uint32_t   elem_size;
+};
+
+static void fill_thread(unsigned int nth, unsigned int ith, void * data) {
+    const struct htp_fill_context * fctx = (const struct htp_fill_context *) data;
+    struct htp_ops_context        * octx = fctx->octx;
+    fill_preamble;
+
+    // Parallelise over the flat row index spanning ne1*ne2*ne3
+    const uint32_t ir0 = fctx->nrows_per_thread * ith;
+    const uint32_t ir1 = MIN(ir0 + fctx->nrows_per_thread, fctx->total_rows);
+
+    uint64_t t1 = HAP_perf_get_qtimer_count();
+
+    if (fctx->opt_path) {
+        // Opt path: tensor is fully contiguous, treat as flat array
+        const uint32_t elem_start = ir0 * ne0;
+        const uint32_t elem_end = ir1 * ne0;
+        uint8_t * dst_ptr = (uint8_t *) dst->data + elem_start * fctx->elem_size;
+        hvx_splat_u(dst_ptr, fctx->splat_vec, elem_end - elem_start, fctx->elem_size);
+    } else {
+        // Non-contiguous path: must respect strides
+        for (uint32_t ir = ir0; ir < ir1; ++ir) {
+            const uint32_t i1 = ir % ne1;
+            const uint32_t i2 = (ir / ne1) % ne2;
+            const uint32_t i3 = ir / (ne1 * ne2);
+            uint8_t * dst_ptr = (uint8_t *) dst->data + i1*nb1 + i2*nb2 + i3*nb3;
+            hvx_splat_u(dst_ptr, fctx->splat_vec, ne0, fctx->elem_size);
+        }
+    }
+
+    uint64_t t2 = HAP_perf_get_qtimer_count();
+    FARF(HIGH, "fill %u/%u: rows %u:%u usec %u\n",
+         ith, nth, ir0, ir1, (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+int op_fill(struct htp_ops_context * octx) {
+    fill_preamble;
+
+    if (dst->type != HTP_TYPE_F32 && dst->type != HTP_TYPE_F16) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    // nr = ne1*ne2*ne3 (flat row count across all outer dims); parallelise over it.
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
+    // Optimize if fully contiguous: skip stride arithmetic, treat as flat array
+    const bool opt_path = (nb2 == nb1 * ne1) && (nb3 == nb2 * ne2);
+
+    FARF(HIGH, "fill: (%ux%ux%ux%u) type=%u opt=%d\n",
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], dst->type, (int) opt_path);
+
+    float val_f32 = 0.f;
+    memcpy(&val_f32, &octx->op_params[0], sizeof(float));
+
+    struct htp_fill_context fctx = {
+        .octx             = octx,
+        .nrows_per_thread = (nr + n_threads - 1) / n_threads,
+        .total_rows       = nr,
+        .opt_path         = opt_path,
+    };
+
+    switch (dst->type) {
+    case HTP_TYPE_F32:
+        fctx.splat_vec = hvx_vec_splat_f32(val_f32);
+        fctx.elem_size = sizeof(float);
+        break;
+    case HTP_TYPE_F16:
+        fctx.splat_vec = hvx_vec_splat_f16((_Float16) val_f32);
+        fctx.elem_size = sizeof(_Float16);
+        break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    worker_pool_run_func(octx->ctx->worker_pool, fill_thread, &fctx, n_threads);
+
+    return HTP_STATUS_OK;
+}

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

@@ -4,6 +4,7 @@
 #include <stdbool.h>
 #include <stdint.h>
 #include <qurt_memory.h>
+#include <qurt.h>
 
 #include "hexagon_types.h"
 #include "hexagon_protos.h"
@@ -100,4 +101,31 @@ static inline void hex_pause() {
     asm volatile(" pause(#255)\n");
 }
 
+#ifndef HEX_NUM_PMU_COUNTERS
+#define HEX_NUM_PMU_COUNTERS 8
+#endif
+
+static inline void hex_get_pmu(uint32_t counters[]) {
+#if __HVX_ARCH__ >= 79
+    asm volatile("%0 = upmucnt0" : "=r"(counters[0]));
+    asm volatile("%0 = upmucnt1" : "=r"(counters[1]));
+    asm volatile("%0 = upmucnt2" : "=r"(counters[2]));
+    asm volatile("%0 = upmucnt3" : "=r"(counters[3]));
+    asm volatile("%0 = upmucnt4" : "=r"(counters[4]));
+    asm volatile("%0 = upmucnt5" : "=r"(counters[5]));
+    asm volatile("%0 = upmucnt6" : "=r"(counters[6]));
+    asm volatile("%0 = upmucnt7" : "=r"(counters[7]));
+#else
+    counters[0] = qurt_pmu_get(QURT_PMUCNT0);
+    counters[1] = qurt_pmu_get(QURT_PMUCNT1);
+    counters[2] = qurt_pmu_get(QURT_PMUCNT2);
+    counters[3] = qurt_pmu_get(QURT_PMUCNT3);
+    counters[4] = qurt_pmu_get(QURT_PMUCNT4);
+    counters[5] = qurt_pmu_get(QURT_PMUCNT5);
+    counters[6] = qurt_pmu_get(QURT_PMUCNT6);
+    counters[7] = qurt_pmu_get(QURT_PMUCNT7);
+    // qurt_pmu_get_pmucnt(counters);
+#endif
+}
+
 #endif /* HEX_UTILS_H */

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

@@ -1683,7 +1683,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
     __fp16  *vtcm_scales     = (__fp16 *) vtcm_seq_alloc(&vtcm_ptr, 256);
     assert((size_t)(vtcm_ptr - (uint8_t *)ctx->vtcm_base) <= vtcm_budget);
 
-    FARF(HIGH, "hmx-mm: m=%d k=%d n=%d wtype=%d block M=%zu N=%zu K=%zu vtcm=%zu/%zu", __func__, m, k, n, weight_type,
+    FARF(HIGH, "hmx-mm: m=%d k=%d n=%d wtype=%d block M=%zu N=%zu K=%zu vtcm=%zu/%zu", m, k, n, weight_type,
          M_BLOCK_SIZE, N_BLOCK_SIZE, K_BLOCK_SIZE, (size_t) (vtcm_ptr - (uint8_t *) ctx->vtcm_base), vtcm_budget);
 
     // initialize eye tile (32x32 identity matrix)

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

@@ -10,6 +10,7 @@
 #include <dspqueue.h>
 #include <stdatomic.h>
 #include <stdint.h>
+#include <stdbool.h>
 
 #define HTP_MAX_NTHREADS 10
 #define HTP_MAX_MMAPS    16
@@ -66,7 +67,9 @@ struct htp_context {
     int                    thread_id;
     int                    thread_prio;
 
-    int                    hmx_enabled;
+    bool                   hmx_enabled;
+    bool                   etm;
+    uint32_t               profiler;
 
     uint8_t *              vtcm_base;
     size_t                 vtcm_size;
@@ -98,5 +101,8 @@ int op_repeat(struct htp_ops_context * octx);
 int op_argsort(struct htp_ops_context * octx);
 int op_ssm_conv(struct htp_ops_context * octx);
 int op_cumsum(struct htp_ops_context * octx);
+int op_fill(struct htp_ops_context * octx);
+int op_diag(struct htp_ops_context * octx);
+int op_solve_tri(struct htp_ops_context * octx);
 
 #endif /* HTP_CTX_H */

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

@@ -42,9 +42,9 @@ enum htp_data_type {
 
 // Mask to enable various stages of the Ops.
 // Used for debugging and profiling.
-enum htp_op_mask {
-    HTP_OPMASK_QUEUE    = (1 << 0),  // Enable Queueing (ie calls into the DSP)
-    HTP_OPMASK_COMPUTE  = (1 << 1),  // Enable Compute
+enum htp_op_stage {
+    HTP_OPSTAGE_QUEUE    = (1 << 0),  // Enable Queueing (ie calls into NPU)
+    HTP_OPSTAGE_COMPUTE  = (1 << 1),  // Enable Compute
 };
 
 // Do not reorder first 4 (used as an index)
@@ -80,7 +80,9 @@ enum htp_op_code {
     HTP_OP_SSM_CONV,
     HTP_OP_REPEAT,
     HTP_OP_CUMSUM,
-
+    HTP_OP_FILL,
+    HTP_OP_DIAG,
+    HTP_OP_SOLVE_TRI,
     HTP_OP_INVALID
 };
 
@@ -135,27 +137,45 @@ struct htp_op_desc {
     int32_t  params[HTP_OP_MAX_PARAMS]; // Params for the op, e.g. epsilon of RMS norm
     uint16_t src[HTP_OP_MAX_INPUTS];    // Input tensors indices
     uint16_t dst;                       // Output tensor index
+};
 
-    // the rest is filled in-place by the NPU
-    uint32_t prof_usecs;                // Number of usec per request
-    uint32_t prof_cycles;               // Number of cycles per request
-    uint32_t prof_pkts;                 // Number of instruction packets per request
-    uint32_t unused;
+enum htp_profiler_mode {
+    HTP_PROF_DISABLED = 0,
+    HTP_PROF_BASIC    = 1,
+    HTP_PROF_PMU      = 2,
+};
+
+#define HTP_PROF_PMU_NCNT 8
+
+// Profile descriptor
+struct htp_prof_desc {
+    uint32_t opcode;                 // GGML/HTP Op
+    uint32_t usecs;                  // Number of usec
+    uint32_t cycles;                 // Number of cycles
+    uint32_t pad;                    // Unused
+    uint32_t pmu[HTP_PROF_PMU_NCNT]; // PMU counters
 };
 
 struct htp_opbatch_req {
+    uint32_t id;          // Batch id
     uint32_t n_bufs;      // Number of buffers
     uint32_t n_tensors;   // Number of tensors
     uint32_t n_ops;       // Number of ops
     uint32_t flags;       // unused
+    uint32_t pad;         // unused
     // struct htp_buf_desc  bufs[];    -- dspqueue buf 0
     // struct htp_tensor    tensors[]; -- dspqueue buf 0
     // struct htp_op_desc   ops[];     -- dspqueue buf 0
 };
 
 struct htp_opbatch_rsp {
+    uint32_t id;         // Batch id
     uint32_t status;     // HTP_STATUS_...
-    // struct htp_op_req ops[];     -- dspqueue buf 0
+    uint32_t n_bufs;     // Number of buffers
+    uint32_t n_tensors;  // Number of tensors
+    uint32_t n_ops;      // Number of op profile descriptors
+    uint32_t pad;        // unused
+    // struct htp_prof_desc profs[];  -- dspqueue buf 0
 };
 
 #endif /* HTP_OPS_H */

ggml/src/ggml-hexagon/htp/htp_iface.idl

@@ -6,13 +6,17 @@
 #include "AEEStdDef.idl"
 #include "remote.idl"
 
+struct htp_iface_pmu_conf {
+    uint32 events[8];
+};
+
 interface htp_iface : remote_handle64 {
     AEEResult start(in uint32 sess_id, in uint64 dsp_queue_id, in uint32 n_hvx, in uint32 use_hmx);
     AEEResult stop();
     AEEResult mmap(in uint32 fd, in uint32 size, in uint32 pinned);
     AEEResult munmap(in uint32 fd);
-    AEEResult enable_etm();
-    AEEResult disable_etm();
+    AEEResult profiler(in uint32 mode, in htp_iface_pmu_conf pmu);
+    AEEResult etm(in uint32 enable);
 };
 
 #endif /* HTP_IDL */

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

@@ -256,6 +256,18 @@ static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
     return Q6_Vhf_equals_Wqf32(Q6_Wqf32_vmpy_VhfVhf(a, b));
 }
 
+static inline HVX_Vector hvx_vec_add_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_VsfVsf(a, b));
+}
+
+static inline HVX_Vector hvx_vec_sub_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(a, b));
+}
+
+static inline HVX_Vector hvx_vec_mul_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(a, b));
+}
+
 #else
 
 static inline HVX_Vector hvx_vec_add_f16_f16(HVX_Vector a, HVX_Vector b)
@@ -273,6 +285,18 @@ static inline HVX_Vector hvx_vec_mul_f16_f16(HVX_Vector a, HVX_Vector b)
     return Q6_Vhf_vmpy_VhfVhf(a, b);
 }
 
+static inline HVX_Vector hvx_vec_add_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_vadd_VsfVsf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_sub_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_vsub_VsfVsf(a, b);
+}
+
+static inline HVX_Vector hvx_vec_mul_f32_f32(HVX_Vector a, HVX_Vector b) {
+    return Q6_Vsf_vmpy_VsfVsf(a, b);
+}
+
 #endif // __HVX_ARCH__ < 79
 
 #endif /* HVX_BASE_H */

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

@@ -27,6 +27,7 @@
 #include "htp-ctx.h"
 #include "htp-ops.h"
 #include "htp-ops.h"
+#include "htp_iface.h"
 #include "worker-pool.h"
 
 AEEResult htp_iface_open(const char * uri, remote_handle64 * handle) {
@@ -100,6 +101,74 @@ AEEResult htp_iface_open(const char * uri, remote_handle64 * handle) {
         }
     }
 
+#if __HVX_ARCH__ >= 75
+    {
+        // Set HMX clock
+        HAP_power_request_t request;
+        memset(&request, 0, sizeof(HAP_power_request_t));
+        request.type = HAP_power_set_HMX_v2;
+        request.hmx_v2.set_clock = TRUE;
+        request.hmx_v2.target_corner = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.min_corner = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.max_corner = HAP_DCVS_EXP_VCORNER_MAX;
+        request.hmx_v2.perf_mode = HAP_CLK_PERF_HIGH;
+        FARF(ALWAYS, "Setting HMX clock\n");
+        err = HAP_power_set((void *) &ctx, &request);
+        if (err != AEE_SUCCESS) {
+            FARF(ERROR, "Error setting HMX clock.");
+            return err;
+        }
+    }
+#endif
+
+    return AEE_SUCCESS;
+}
+
+AEEResult htp_iface_etm(remote_handle64 handle, uint32_t enable) {
+    int err = enable ? HAP_user_etm_enable() : HAP_user_etm_disable();
+    if (err) {
+        if (err == AEE_EVERSIONNOTSUPPORT) {
+            FARF(ERROR, "API HAP_user_etm_enable/disable is not supported\n");
+        } else {
+            FARF(ERROR, "Error executing HAP_user_etm_enable/disable with error code : 0x%x\n", err);
+        }
+    }
+    return err;
+}
+
+AEEResult htp_iface_profiler(remote_handle64 handle, uint32_t mode, const htp_iface_pmu_conf* pmu_conf) {
+    struct htp_context * ctx = (struct htp_context *) handle;
+    if (!ctx) {
+        return AEE_EBADPARM;
+    }
+
+    if (mode == HTP_PROF_PMU) {
+        const uint32_t* events = pmu_conf->events;
+
+        // Pack 4 event IDs (low 8 bits) into each 32-bit config register
+        uint32_t evtcfg = 0, evtcfg1 = 0, cfg = 0, i = 0;
+        for (; i < HEX_NUM_PMU_COUNTERS/2; i++) {
+            evtcfg  |= ((events[i + 0] & 0xFF) << (i * 8));
+            evtcfg1 |= ((events[i + 4] & 0xFF) << (i * 8));
+        }
+
+        // For events >255 pack high 2 bits of all 8 event IDs into cfg register
+        // 2 bits per counter: bits [1:0] for counter 0, [3:2] for counter 1, etc.
+        for (i = 0; i < HEX_NUM_PMU_COUNTERS; i++) {
+            cfg |= (((events[i] >> 8) & 3) << (i * 2));
+        }
+
+        FARF(ALWAYS, "Configuring PMU registers: evtcfg = 0x%x, evtcfg1 = 0x%x, pmucfg = 0x%x", evtcfg, evtcfg1, cfg);
+
+        // Configure PMU registers
+        qurt_pmu_set(QURT_PMUCFG,     cfg);
+        qurt_pmu_set(QURT_PMUEVTCFG,  evtcfg);
+        qurt_pmu_set(QURT_PMUEVTCFG1, evtcfg1);
+        qurt_pmu_enable(1);
+    }
+
+    ctx->profiler = mode;
+
     return AEE_SUCCESS;
 }
 
@@ -129,35 +198,19 @@ AEEResult htp_iface_close(remote_handle64 handle) {
         }
     }
 
+    if (ctx->profiler) {
+        qurt_pmu_enable(1);
+    }
+
+    if (ctx->etm) {
+        HAP_user_etm_disable();
+    }
+
     free(ctx);
     return AEE_SUCCESS;
 }
 
-AEEResult htp_iface_enable_etm(remote_handle64 handle) {
-    int err = HAP_user_etm_enable();
-    if (err) {
-        if (err == AEE_EVERSIONNOTSUPPORT) {
-            FARF(ERROR, "API HAP_user_etm_enable is not supported\n");
-        } else {
-            FARF(ERROR, "Error executing HAP_user_etm_enable with error code : 0x%x\n", err);
-        }
-    }
-    return err;
-}
-
-AEEResult htp_iface_disable_etm(remote_handle64 handle) {
-    int err = HAP_user_etm_disable();
-    if (err) {
-        if (err == AEE_EVERSIONNOTSUPPORT) {
-            FARF(ERROR, "API HAP_user_etm_disable is not supported\n");
-        } else {
-            FARF(ERROR, "Error executing HAP_user_etm_disable with error code : 0x%x\n", err);
-        }
-    }
-    return err;
-}
-
-AEEResult htp_iface_mmap(remote_handle64 handle, int fd, uint32_t size, uint32_t pinned) {
+AEEResult htp_iface_mmap(remote_handle64 handle, uint32 fd, uint32 size, uint32 pinned) {
     struct htp_context * ctx = (struct htp_context *) handle;
     if (!ctx) {
         return AEE_EBADPARM;
@@ -204,7 +257,7 @@ AEEResult htp_iface_mmap(remote_handle64 handle, int fd, uint32_t size, uint32_t
     return AEE_ENOMEMORY;
 }
 
-AEEResult htp_iface_munmap(remote_handle64 handle, int fd) {
+AEEResult htp_iface_munmap(remote_handle64 handle, uint32 fd) {
     struct htp_context * ctx = (struct htp_context *) handle;
     if (!ctx) {
         return AEE_EBADPARM;
@@ -434,19 +487,39 @@ static void htp_error_callback(dspqueue_t queue, int error, void * context) {
 struct profile_data {
     uint64_t usecs;
     uint64_t cycles;
-    uint64_t pkts;
+    uint32_t pmu_counters[HEX_NUM_PMU_COUNTERS];
 };
 
-static inline void profile_start(struct profile_data * d) {
-    d->usecs  = HAP_perf_get_qtimer_count();
-    d->cycles = hex_get_cycles();
-    d->pkts   = hex_get_pktcnt();
+static inline void profile_start(uint32_t mode, struct profile_data * d) {
+    switch (mode) {
+        case HTP_PROF_PMU:
+            hex_get_pmu(d->pmu_counters);
+            // fallthrough
+        case HTP_PROF_BASIC:
+            d->usecs  = HAP_perf_get_qtimer_count();
+            d->cycles = hex_get_cycles();
+            break;
+        default:
+            break;
+    }
 }
 
-static inline void profile_stop(struct profile_data * d) {
-    d->usecs  = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - d->usecs);
-    d->cycles = hex_get_cycles() - d->cycles;
-    d->pkts   = hex_get_pktcnt() - d->pkts;
+static inline void profile_stop(uint32_t mode, struct profile_data * d) {
+    uint32_t pmu_counters[HEX_NUM_PMU_COUNTERS];
+    switch (mode) {
+        case HTP_PROF_PMU:
+            hex_get_pmu(pmu_counters);
+            for (int i = 0; i < HEX_NUM_PMU_COUNTERS; i++) {
+                d->pmu_counters[i] = pmu_counters[i] - d->pmu_counters[i];
+            }
+            // fallthrough
+        case HTP_PROF_BASIC:
+            d->usecs  = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - d->usecs);
+            d->cycles = hex_get_cycles() - d->cycles;
+            break;
+        default:
+            break;
+    }
 }
 
 static int execute_op(struct htp_ops_context * octx) {
@@ -514,6 +587,15 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_CUMSUM:
             return op_cumsum(octx);
 
+        case HTP_OP_FILL:
+            return op_fill(octx);
+
+        case HTP_OP_DIAG:
+            return op_diag(octx);
+
+        case HTP_OP_SOLVE_TRI:
+            return op_solve_tri(octx);
+
         case HTP_OP_INVALID:
             break;
 
@@ -720,30 +802,33 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
             continue;
         }
 
+        // Reset poll count for valid requests
+        poll_count = DSPQUEUE_POLL_COUNT;
+
         const uint32_t n_bufs = req.n_bufs;
         const uint32_t n_tens = req.n_tensors;
         const uint32_t n_ops  = req.n_ops;
 
-        const uint32_t b_size = sizeof(struct htp_buf_desc) * n_bufs;
-        const uint32_t t_size = sizeof(struct htp_tensor)   * n_tens;
-        const uint32_t o_size = sizeof(struct htp_op_desc)  * n_ops;
+        const uint32_t b_size = sizeof(struct htp_buf_desc)  * n_bufs;
+        const uint32_t t_size = sizeof(struct htp_tensor)    * n_tens;
+        const uint32_t o_size = sizeof(struct htp_op_desc)   * n_ops;
+        const uint32_t p_size = sizeof(struct htp_prof_desc) * n_ops;
 
-        if (dbuf.size < b_size + t_size + o_size) {
+        if (dbuf.size < b_size + t_size + o_size + p_size) {
             FARF(ERROR, "invalid opbatch memory block size %u", dbuf.size);
             break;
         }
 
-        // Reset poll count for valid requests
-        poll_count = DSPQUEUE_POLL_COUNT;
-
-        uint8_t * m_ptr = dbuf.ptr;
-        struct htp_buf_desc* bufs = (struct htp_buf_desc*) m_ptr; m_ptr += b_size;
-        struct htp_tensor*   tens = (struct htp_tensor*)   m_ptr; m_ptr += t_size;
-        struct htp_op_desc*   ops = (struct htp_op_desc*)  m_ptr;
-
-        FARF(HIGH, "processing opbatch: n-bufs %u n-tensors %u n-ops %u : m-size %u b-size %u t-size %u o-size %u",
+        FARF(HIGH, "processing opbatch #%u: n-bufs %u n-tensors %u n-ops %u : m-size %u b-size %u t-size %u o-size %u", req.id,
                 n_bufs, n_tens, n_ops, dbuf.size, b_size, t_size, o_size);
 
+        // Setup descriptor pointers
+        uint8_t * m_ptr = dbuf.ptr;
+        struct htp_buf_desc* bufs = (struct htp_buf_desc*)  m_ptr; m_ptr += b_size;
+        struct htp_tensor*   tens = (struct htp_tensor*)    m_ptr; m_ptr += t_size;
+        struct htp_op_desc*   ops = (struct htp_op_desc*)   m_ptr; m_ptr += o_size;
+        struct htp_prof_desc* pds = (struct htp_prof_desc*) m_ptr;
+
         prep_op_bufs(ctx, bufs, n_bufs);
         prep_tensors(ctx, bufs, tens, n_tens);
 
@@ -754,22 +839,34 @@ static void htp_packet_callback(dspqueue_t queue, int error, void * context) {
 
         for (uint32_t i=0; i < n_ops; i++) {
             struct profile_data prof;
-            profile_start(&prof);
+
+            profile_start(ctx->profiler, &prof);
 
             proc_op_req(octx, tens, i, &ops[i]);
 
-            profile_stop(&prof);
-            ops[i].prof_usecs  = prof.usecs;
-            ops[i].prof_cycles = prof.cycles;
-            ops[i].prof_pkts   = prof.pkts;
+            profile_stop(ctx->profiler, &prof);
+
+            if (ctx->profiler) {
+                pds[i].opcode = ops[i].opcode;
+                pds[i].usecs  = prof.usecs;
+                pds[i].cycles = prof.cycles;
+                for (int j = 0; j < HEX_NUM_PMU_COUNTERS; j++) {
+                    pds[i].pmu[j] = prof.pmu_counters[j];
+                }
+            }
         }
 
         // dspqueue_write_early_wakeup_noblock(ctx->queue, 10, 0);
 
         struct htp_opbatch_rsp rsp;
-        rsp.status = HTP_STATUS_OK; // FIXME
+        rsp.id        = req.id;
+        rsp.status    = HTP_STATUS_OK;
+        rsp.n_bufs    = n_bufs;
+        rsp.n_tensors = n_tens;
+        rsp.n_ops     = n_ops;
 
         dbuf.flags = DSPQUEUE_BUFFER_FLAG_FLUSH_SENDER | DSPQUEUE_BUFFER_FLAG_INVALIDATE_RECIPIENT;
+
         err = dspqueue_write(queue, 0, 1, &dbuf, sizeof(rsp), (const uint8_t *) &rsp, DSPQUEUE_TIMEOUT_NONE);
         if (err != 0) {
             FARF(ERROR, "dspqueue_write failed: 0x%08x", (unsigned) err);

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

@@ -3017,6 +3017,10 @@ int op_matmul(struct htp_ops_context * octx) {
     const int act_stride = (int)(src1->nb[1] / sizeof(float));
     const int wgt_stride = (int)(src0->nb[1] / sizeof(__fp16));
 
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
     if (src0->type == HTP_TYPE_F16) {
         if (is_batched) {
             hmx_matmul_w16a32_batched_params_t batch_params = {

ggml/src/ggml-hexagon/htp/solve-tri-ops.c

@@ -0,0 +1,267 @@
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "hvx-types.h"
+#include "hvx-utils.h"
+
+struct htp_solve_tri_context {
+    struct htp_ops_context * octx;
+    uint32_t                 jobs_per_thread;
+    uint32_t                 total_jobs;
+    uint32_t                 k_chunks;
+    uint32_t                 col_block;
+};
+
+static inline void solve_tri_row_scalar(const float * A_row,
+                                        const float * B_row,
+                                        float *       X,
+                                        uint32_t      row,
+                                        uint32_t      k,
+                                        uint32_t      col0,
+                                        uint32_t      coln,
+                                        float         inv_diag) {
+    for (uint32_t col = col0; col < col0 + coln; ++col) {
+        float sum = 0.0f;
+        for (uint32_t t = 0; t < row; ++t) {
+            sum += A_row[t] * X[t * k + col];
+        }
+        X[row * k + col] = (B_row[col] - sum) * inv_diag;
+    }
+}
+
+static inline HVX_Vector hvx_load_partial_f32(const float * src, uint32_t n) {
+    HVX_Vector v = *((const HVX_UVector *) src);
+    HVX_VectorPred mask = Q6_Q_vsetq2_R(n * sizeof(float));
+    return Q6_V_vmux_QVV(mask, v, Q6_V_vzero());
+}
+
+static inline void solve_tri_row_hvx(const float * A_row,
+                                     const float * B_row,
+                                     float *       X,
+                                     uint32_t      row,
+                                     uint32_t      k,
+                                     uint32_t      col0,
+                                     uint32_t      coln,
+                                     float         inv_diag) {
+    const bool full = (coln == VLEN_FP32);
+
+    HVX_Vector sum_v = Q6_V_vzero();
+    for (uint32_t t = 0; t < row; ++t) {
+        const float   a         = A_row[t];
+        const float * x_row_col = X + t * k + col0;
+
+        HVX_Vector x_v = full ? *((const HVX_UVector *) x_row_col) : hvx_load_partial_f32(x_row_col, coln);
+        HVX_Vector a_v = hvx_vec_splat_f32(a);
+        sum_v          = hvx_vec_add_f32_f32(sum_v, hvx_vec_mul_f32_f32(x_v, a_v));
+    }
+
+    const float * b_row_col = B_row + col0;
+    float *       x_out_col = X + row * k + col0;
+
+    HVX_Vector b_v        = full ? *((const HVX_UVector *) b_row_col) : hvx_load_partial_f32(b_row_col, coln);
+    HVX_Vector inv_diag_v = hvx_vec_splat_f32(inv_diag);
+
+    HVX_Vector out_v = hvx_vec_mul_f32_f32(hvx_vec_sub_f32_f32(b_v, sum_v), inv_diag_v);
+    hvx_vec_store_u((void *) x_out_col, coln * sizeof(float), out_v);
+}
+
+// Batch-level thread: each job is one full batch.
+static void solve_tri_batch_thread_f32(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_solve_tri_context * sctx = (struct htp_solve_tri_context *) data;
+    struct htp_ops_context *       octx = sctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];  // A
+    const struct htp_tensor * src1 = octx->src[1];  // B
+    const struct htp_tensor * dst  = octx->dst;     // X
+
+    const uint32_t n = src0->ne[0];
+    const uint32_t k = src1->ne[0];
+
+    const uint32_t ne02 = src0->ne[2];
+
+    const uint32_t col_block = VLEN_FP32;
+    const uint32_t k_full    = (k / col_block) * col_block;
+
+    const uint32_t start_batch = sctx->jobs_per_thread * ith;
+    const uint32_t end_batch   = MIN(start_batch + sctx->jobs_per_thread, sctx->total_jobs);
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    for (uint32_t batch = start_batch; batch < end_batch; ++batch) {
+        const uint32_t i03 = batch / ne02;
+        const uint32_t i02 = batch - i03 * ne02;
+
+        const float * A_batch =
+            (const float *) ((const uint8_t *) (uintptr_t) src0->data + i02 * src0->nb[2] + i03 * src0->nb[3]);
+        const float * B_batch =
+            (const float *) ((const uint8_t *) (uintptr_t) src1->data + i02 * src1->nb[2] + i03 * src1->nb[3]);
+        float * X_batch = (float *) ((uint8_t *) (uintptr_t) dst->data + i02 * dst->nb[2] + i03 * dst->nb[3]);
+
+        for (uint32_t row = 0; row < n; ++row) {
+            const float   diag     = A_batch[row * n + row];
+            const float   inv_diag = 1.0f / diag;
+            const float * A_row    = A_batch + row * n;
+            const float * B_row    = B_batch + row * k;
+
+            uint32_t col0 = 0;
+            for (; col0 < k_full; col0 += col_block) {
+                solve_tri_row_hvx(A_row, B_row, X_batch, row, k, col0, col_block, inv_diag);
+            }
+
+            if (col0 < k) {
+                const uint32_t coln = k - col0;
+                if (coln >= 8) {
+                    solve_tri_row_hvx(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
+                } else {
+                    solve_tri_row_scalar(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
+                }
+            }
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "solve-tri-batch %d/%d: A=(%ux%u) B=(%ux%u) batch %u:%u usec %u\n",
+         ith, nth, n, n, k, n, start_batch, end_batch,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+// Chunk-level thread: each job is one (batch, col_chunk) pair.
+static void solve_tri_chunk_thread_f32(unsigned int nth, unsigned int ith, void * data) {
+    struct htp_solve_tri_context * sctx = (struct htp_solve_tri_context *) data;
+    struct htp_ops_context *       octx = sctx->octx;
+
+    const struct htp_tensor * src0 = octx->src[0];  // A
+    const struct htp_tensor * src1 = octx->src[1];  // B
+    const struct htp_tensor * dst  = octx->dst;     // X
+
+    const uint32_t n = src0->ne[0];
+    const uint32_t k = src1->ne[0];
+
+    const uint32_t ne02 = src0->ne[2];
+
+    const uint32_t start_job = sctx->jobs_per_thread * ith;
+    const uint32_t end_job   = MIN(start_job + sctx->jobs_per_thread, sctx->total_jobs);
+
+    uint64_t t1, t2;
+    t1 = HAP_perf_get_qtimer_count();
+
+    for (uint32_t job = start_job; job < end_job; ++job) {
+        const uint32_t batch = job / sctx->k_chunks;
+        const uint32_t chunk = job - batch * sctx->k_chunks;
+
+        const uint32_t i03 = batch / ne02;
+        const uint32_t i02 = batch - i03 * ne02;
+
+        const uint32_t col0 = chunk * sctx->col_block;
+        const uint32_t coln = MIN(sctx->col_block, k - col0);
+
+        const float * A_batch =
+            (const float *) ((const uint8_t *) (uintptr_t) src0->data + i02 * src0->nb[2] + i03 * src0->nb[3]);
+        const float * B_batch =
+            (const float *) ((const uint8_t *) (uintptr_t) src1->data + i02 * src1->nb[2] + i03 * src1->nb[3]);
+        float * X_batch = (float *) ((uint8_t *) (uintptr_t) dst->data + i02 * dst->nb[2] + i03 * dst->nb[3]);
+
+        const bool use_hvx = (coln >= 8);
+
+        for (uint32_t row = 0; row < n; ++row) {
+            const float diag     = A_batch[row * n + row];
+            const float inv_diag = 1.0f / diag;
+
+            const float * A_row = A_batch + row * n;
+            const float * B_row = B_batch + row * k;
+
+            if (use_hvx) {
+                solve_tri_row_hvx(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
+            } else {
+                solve_tri_row_scalar(A_row, B_row, X_batch, row, k, col0, coln, inv_diag);
+            }
+        }
+    }
+
+    t2 = HAP_perf_get_qtimer_count();
+
+    FARF(HIGH, "solve-tri-chunk %d/%d: A=(%ux%u) B=(%ux%u) job %u:%u usec %u\n",
+         ith, nth, n, n, k, n, start_job, end_job,
+         (unsigned) HAP_perf_qtimer_count_to_us(t2 - t1));
+}
+
+int op_solve_tri(struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = octx->src[0];  // A
+    const struct htp_tensor * src1 = octx->src[1];  // B
+    const struct htp_tensor * dst  = octx->dst;     // X
+
+    if (src0->type != HTP_TYPE_F32 || src1->type != HTP_TYPE_F32 || dst->type != HTP_TYPE_F32) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    // left=true, lower=true, uni=false only
+    if (src0->ne[0] != src0->ne[1]) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+    if (src0->ne[1] != src1->ne[1]) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+    if (src0->ne[2] != src1->ne[2] || src0->ne[3] != src1->ne[3]) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+    if (dst->ne[0] != src1->ne[0] || dst->ne[1] != src1->ne[1] || dst->ne[2] != src1->ne[2] ||
+        dst->ne[3] != src1->ne[3]) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t k = src1->ne[0];
+
+    const uint32_t col_block     = VLEN_FP32;
+    const uint32_t k_chunks      = (k + col_block - 1) / col_block;
+    const uint32_t total_batches = src0->ne[2] * src0->ne[3];
+    const bool     batched       = total_batches >= (uint32_t) octx->n_threads;
+
+    FARF(HIGH, "solve-tri: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : batched %d\n",
+         src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], batched);
+
+    if (batched) {
+        // Batch-level parallelism
+        const uint32_t n_threads = MIN((uint32_t) octx->n_threads, total_batches);
+
+        struct htp_solve_tri_context sctx = {
+            .octx            = octx,
+            .jobs_per_thread = (total_batches + n_threads - 1) / n_threads,
+            .total_jobs      = total_batches,
+            .k_chunks        = k_chunks,
+            .col_block       = col_block,
+        };
+
+        worker_pool_run_func(octx->ctx->worker_pool, solve_tri_batch_thread_f32, &sctx, n_threads);
+    } else {
+        // Chunk-level parallelism
+        const uint32_t total_jobs = total_batches * k_chunks;
+        const uint32_t n_threads  = MIN((uint32_t) octx->n_threads, MAX(total_jobs, 1));
+
+        struct htp_solve_tri_context sctx = {
+            .octx            = octx,
+            .jobs_per_thread = (total_jobs + n_threads - 1) / n_threads,
+            .total_jobs      = total_jobs,
+            .k_chunks        = k_chunks,
+            .col_block       = col_block,
+        };
+
+        worker_pool_run_func(octx->ctx->worker_pool, solve_tri_chunk_thread_f32, &sctx, n_threads);
+    }
+
+    return HTP_STATUS_OK;
+}

ggml/src/ggml-hexagon/libggml-htp.inf

@@ -8,7 +8,7 @@ CatalogFile = libggml-htp.cat
 PnpLockDown = 1
 
 [DestinationDirs]
-Drivers_Dir = 6
+Drivers_Dir = 13
 
 [SourceDisksNames]
 1 = %DiskId%
@@ -18,6 +18,7 @@ libggml-htp-v68.so = 1
 libggml-htp-v69.so = 1
 libggml-htp-v73.so = 1
 libggml-htp-v75.so = 1
+libggml-htp-v79.so = 1
 libggml-htp-v81.so = 1
 
 [ControlFlags]
@@ -31,6 +32,7 @@ libggml-htp-v68.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v69.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v73.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v75.so,,,0x10 ;COPYFLG_NO_OVERWRITE
+libggml-htp-v79.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 libggml-htp-v81.so,,,0x10 ;COPYFLG_NO_OVERWRITE
 
 [Strings]

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

@@ -677,7 +677,15 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm(ggml_meta
     const ggml_type tsrc1 = op->src[1]->type;
 
     const bool bc_inp = op->src[0]->ne[0] % 32 != 0;
-    const bool bc_out = op->ne[0] % 64 != 0 || op->ne[1] % 32 != 0;
+
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+
+    const bool has_tensor = ggml_metal_device_get_props(ggml_metal_library_get_device(lib))->has_tensor;
+
+    const bool bc_out = has_tensor
+        ? (op->ne[0] % NRA != 0 || op->ne[1] % NRB != 0)
+        : (op->ne[0] % 64  != 0 || op->ne[1] % 32  != 0);
 
     snprintf(base, 256, "kernel_mul_mm_%s_%s", ggml_type_name(tsrc0), ggml_type_name(tsrc1));
     snprintf(name, 256, "%s_bci=%d_bco=%d", base, bc_inp, bc_out);
@@ -694,8 +702,20 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm(ggml_meta
         ggml_metal_cv_free(cv);
     }
 
-    // when the output size is not multiple of 64x32, we need extra smem to prevent out-of-bounds writes
-    res.smem = bc_out ? 8192 : 4096 + 2048;
+    if (has_tensor) {
+        res.nr0 = NRA;
+        res.nr1 = NRB;
+
+        const size_t smem_a = NRA * N_MM_NK_TOTAL * sizeof(ggml_fp16_t);
+        res.smem = smem_a;
+    } else {
+        res.nr0 = 64;
+        res.nr1 = 32;
+
+        res.smem = bc_out ? 8192 : (4096 + 2048);
+    }
+
+    res.nsg = N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
 
     return res;
 }

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

@@ -102,6 +102,8 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
 
 void ggml_metal_library_free(ggml_metal_library_t lib);
 
+ggml_metal_device_t ggml_metal_library_get_device(ggml_metal_library_t lib);
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline    (ggml_metal_library_t lib, const char * name);
 struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv);
 

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

@@ -95,8 +95,8 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
 
 struct ggml_metal_library {
     id<MTLLibrary> obj;
-    id<MTLDevice> device;
 
+    ggml_metal_device_t dev;
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
 
     NSLock * lock;
@@ -251,7 +251,7 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
     ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
 
     res->obj       = library;
-    res->device    = device;
+    res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
@@ -318,7 +318,7 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
     }
 
     res->obj       = library;
-    res->device    = device;
+    res->dev       = dev;
     res->pipelines = ggml_metal_pipelines_init();
     res->lock      = [NSLock new];
 
@@ -341,6 +341,10 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
     free(lib);
 }
 
+ggml_metal_device_t ggml_metal_library_get_device(ggml_metal_library_t lib) {
+    return lib->dev;
+}
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_metal_library_t lib, const char * name) {
     [lib->lock lock];
 
@@ -405,7 +409,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_
             return res;
         }
 
-        id<MTLComputePipelineState> obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
+        id<MTLDevice> device = ggml_metal_device_get_obj(lib->dev);
+        id<MTLComputePipelineState> obj = [device newComputePipelineStateWithFunction:mtl_function error:&error];
 
         [mtl_function release];
 
@@ -699,7 +704,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
                     "    auto sB = tB.slice(0, 0); \n"
                     "    mm.run(sB, sA, cT); \n"
                     " \n"
-                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(16, 16)); \n"
                     " \n"
                     "    cT.store(tC); \n"
                     "}";
@@ -749,7 +754,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
                     "    auto sB = tB.slice(0, 0); \n"
                     "    mm.run(sB, sA, cT); \n"
                     " \n"
-                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(4, 4)); \n"
+                    "    auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(16, 16)); \n"
                     " \n"
                     "    cT.store(tC); \n"
                     "}";
@@ -814,7 +819,7 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
             }
 
             // print MTL GPU family:
-            GGML_LOG_INFO("%s: GPU name:   %s\n", __func__, dev->props.name);
+            GGML_LOG_INFO("%s: GPU name:   %s (%s)\n", __func__, dev->props.name, dev->props.desc);
 
             // determine max supported GPU family
             // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
@@ -931,13 +936,13 @@ void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev) {
 }
 
 struct ggml_metal_event {
-    void * obj; // id<MTLEvent>
+    void * obj; // id<MTLSharedEvent>
 
     atomic_int value;
 };
 
 void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
-    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+    id<MTLSharedEvent> event = (id<MTLSharedEvent>)ev->obj;
 
     id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
 
@@ -945,7 +950,7 @@ void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t
 }
 
 void ggml_metal_event_encode_wait(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf_raw) {
-    id<MTLEvent> event = (id<MTLEvent>)ev->obj;
+    id<MTLSharedEvent> event = (id<MTLSharedEvent>)ev->obj;
 
     id<MTLCommandBuffer> cmd_buf = (id<MTLCommandBuffer>) cmd_buf_raw;
 
@@ -953,7 +958,7 @@ void ggml_metal_event_encode_wait(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cm
 }
 
 ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
-    id<MTLEvent> event = [dev->mtl_device newEvent];
+    id<MTLSharedEvent> event = [dev->mtl_device newSharedEvent];
 
     ggml_metal_event_t ev = calloc(1, sizeof(struct ggml_metal_event));
 
@@ -964,7 +969,7 @@ ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
 }
 
 void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
-    id<MTLEvent> event = ev->obj;
+    id<MTLSharedEvent> event = ev->obj;
     [event release];
 
     free(ev);
@@ -973,14 +978,13 @@ void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev
 }
 
 void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
-    @autoreleasepool {
-        id<MTLEvent> event = ev->obj;
-
-        id<MTLCommandBuffer> cmd_buf = [dev->mtl_queue commandBuffer];
-        [cmd_buf encodeWaitForEvent:event value:atomic_load_explicit(&ev->value, memory_order_relaxed)];
-        [cmd_buf commit];
-        [cmd_buf waitUntilCompleted];
+    id<MTLSharedEvent> event = ev->obj;
+    const bool res = [event waitUntilSignaledValue:atomic_load_explicit(&ev->value, memory_order_relaxed) timeoutMS:60000];
+    if (!res) {
+        GGML_ABORT("%s: failed to wait for event\n", __func__);
     }
+
+    GGML_UNUSED(dev);
 }
 
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {

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

@@ -1,6 +1,19 @@
 #ifndef GGML_METAL_IMPL
 #define GGML_METAL_IMPL
 
+// kernel parameters for mat-mat threadgroups
+//
+// TODO: become function constants
+
+#define SZ_SIMDGROUP 16
+#define N_MM_NK 2
+#define N_MM_NK_TOTAL (SZ_SIMDGROUP * N_MM_NK)
+
+#define N_MM_BLOCK_X 4
+#define N_MM_BLOCK_Y 2
+#define N_MM_SIMD_GROUP_X 2
+#define N_MM_SIMD_GROUP_Y 2
+
 // kernel parameters for mat-vec threadgroups
 //
 // N_R0: number of src0 rows to process per simdgroup

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

@@ -2195,7 +2195,12 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
         const size_t smem = pipeline.smem;
 
         ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
-        ggml_metal_encoder_dispatch_threadgroups(enc, ((ne11 + 31)/32), ((ne01 + 63)/64), ne12*ne13, 128, 1, 1);
+
+        const int nr0 = pipeline.nr0;
+        const int nr1 = pipeline.nr1;
+        const int nsg = pipeline.nsg;
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, ((ne11 + nr1 - 1) / nr1), ((ne01 + nr0 - 1) / nr0), ne12 * ne13, 32, nsg, 1);
     } else {
         auto pipeline = ggml_metal_library_get_pipeline_mul_mv(lib, op);
 

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

@@ -918,6 +918,10 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) {
         static std::vector<ggml_backend_device_ptr> devs;
 
         if (!initialized) {
+            // workaround macOS limitation (kIOGPUCommandBufferCallbackErrorImpactingInteractivity) until proper fix becomes possible
+            // ref: https://github.com/ggml-org/llama.cpp/issues/20141#issuecomment-4272947703
+            setenv("AGX_RELAX_CDM_CTXSTORE_TIMEOUT", "1", true);
+
             static ggml_backend_metal_reg_ptr reg_ctx(ggml_backend_metal_reg_init());
 
             for (int i = 0; i < g_devices; ++i) {

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

@@ -9306,7 +9306,137 @@ constant bool FC_mul_mm_bc_inp [[function_constant(FC_MUL_MM + 0)]];
 constant bool FC_mul_mm_bc_out [[function_constant(FC_MUL_MM + 1)]];
 
 // each block_q contains 16*nl weights
-template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+#ifdef GGML_METAL_HAS_TENSOR
+template<
+    typename SA, typename SA_4x4, typename SA_8x8,
+    typename SB, typename SB_2x4, typename SB_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread SA_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * srcA,
+        device const char * srcB,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig [[threadgroup_position_in_grid]],
+        ushort tiitg [[thread_index_in_threadgroup]],
+        ushort sgitg [[simdgroup_index_in_threadgroup]]) {
+    (void) sgitg;
+
+    // Matrix dimensions: A(M,K) x B(K,N) -> C(M,N)
+    const int K = args.ne00;
+    const int M = args.ne0;
+    const int N = args.ne1;
+
+    // Batch dimension handling
+    const int im = tgpig.z;
+    const int i12 = im % args.ne12;
+    const int i13 = im / args.ne12;
+
+    // Batch offsets for srcA and srcB
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+
+    // Tile dimensions
+    constexpr int NRB = SZ_SIMDGROUP * N_MM_BLOCK_X * N_MM_SIMD_GROUP_X;
+    constexpr int NRA = SZ_SIMDGROUP * N_MM_BLOCK_Y * N_MM_SIMD_GROUP_Y;
+
+    // Tile offsets in output matrix
+    const int ra = tgpig.y * NRA;
+    const int rb = tgpig.x * NRB;
+
+    // Threadgroup memory for dequantized A tile only
+    threadgroup SA * sa = (threadgroup SA *)(shmem);
+
+    // Work-item count for A loading
+    constexpr int A_WORK_ITEMS = NRA * N_MM_NK;
+    constexpr int NUM_THREADS = N_SIMDWIDTH * N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y;
+
+    // tA wraps threadgroup memory
+    auto tA = tensor(sa, dextents<int32_t, 2>(N_MM_NK_TOTAL, NRA));
+
+    // tB wraps device memory directly
+    device T1 * ptrB = (device T1 *)(srcB + args.nb12*i12 + args.nb13*i13);
+    const int strideB = args.nb11 / sizeof(T1);
+    auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
+
+    // Configure matmul operation
+    mpp::tensor_ops::matmul2d<
+        mpp::tensor_ops::matmul2d_descriptor(
+            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
+        execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
+
+    auto cT = mm.get_destination_cooperative_tensor<decltype(tB), decltype(tA), float>();
+
+    // Accumulate partial results over K dimension
+    for (int loop_k = 0; loop_k < K; loop_k += N_MM_NK_TOTAL) {
+        // === PHASE 1: Dequantization of A into threadgroup memory ===
+        for (int work = tiitg; work < A_WORK_ITEMS; work += NUM_THREADS) {
+            const int row = work / N_MM_NK;
+            const int k_chunk = work % N_MM_NK;
+            const int k_pos = loop_k + k_chunk * 16;
+            const short k_base = k_chunk * 16;
+
+            // Bounds check: skip device read if row is out of matrix bounds
+            if (ra + row < M) {
+                if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
+                    // Element-wise reads when K is not aligned (nb01 not aligned for half4x4/float4x4).
+                    // MSL spec Table 2.5: half4x4 requires 8-byte alignment. When K is odd,
+                    // nb01 = K*2 is not 8-byte aligned, so odd-row pointers are misaligned.
+                    // Mirrors the legacy kernel's existing guard.
+                    device const T0 * row_ptr = (device const T0 *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? (SA) row_ptr[k_pos + i] : (SA)0;
+                    }
+                } else {
+                    const int block_idx = k_pos / (16 * nl);
+                    const short il = (k_pos / 16) % nl;
+
+                    device const block_q * row_ptr = (device const block_q *)(srcA + args.nb01 * (ra + row) + offset0);
+
+                    SA_4x4 temp_a;
+                    dequantize_func(row_ptr + block_idx, il, temp_a);
+
+                    FOR_UNROLL (short i = 0; i < 16; i++) {
+                        // Zero-pad A for K positions beyond valid range (handles partial K iterations)
+                        sa[row * N_MM_NK_TOTAL + (k_base + i)] = (k_pos + i < K) ? temp_a[i/4][i%4] : (SA)0;
+                    }
+                }
+            } else {
+                // Zero-pad rows beyond matrix bounds
+                FOR_UNROLL (short i = 0; i < 16; i++) {
+                    sa[row * N_MM_NK_TOTAL + (k_base + i)] = (SA)0;
+                }
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // === PHASE 2: Tensor matmul ===
+        auto mA = tA.slice(0, 0);
+        auto mB = tB.slice(loop_k, rb);
+
+        mm.run(mB, mA, cT);
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    // Store result tile to output matrix (with batch offset)
+    // cT.store handles bounds checking via tD's extents (M, N)
+    device float * dstBatch = (device float *)dst + im * N * M;
+
+    auto tD = tensor(dstBatch, dextents<int32_t, 2>(M, N), array<int, 2>({1, M}));
+    cT.store(tD.slice(ra, rb));
+}
+
+#else
+
+template<
+    typename S0, typename S0_4x4, typename S0_8x8,
+    typename S1, typename S1_2x4, typename S1_8x8,
+    typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &),
+    typename T0, typename T0_4x4, typename T1, typename T1_2x4>
 kernel void kernel_mul_mm(
         constant ggml_metal_kargs_mul_mm & args,
         device const char * src0,
@@ -9320,10 +9450,6 @@ kernel void kernel_mul_mm(
     threadgroup S0 * sa = (threadgroup S0 *)(shmem);
     threadgroup S1 * sb = (threadgroup S1 *)(shmem + 4096);
 
-#ifdef GGML_METAL_HAS_TENSOR
-    threadgroup float * sc = (threadgroup float *)(shmem);
-#endif
-
     constexpr int NR0 = 64;
     constexpr int NR1 = 32;
 
@@ -9363,7 +9489,6 @@ kernel void kernel_mul_mm(
         + args.nb11*(r1 + lr1)
         + args.nb10*iy);
 
-#ifndef GGML_METAL_HAS_TENSOR
     S0_8x8 ma[4];
     S1_8x8 mb[2];
 
@@ -9372,19 +9497,8 @@ kernel void kernel_mul_mm(
     for (short i = 0; i < 8; i++){
         mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
-#else
-    auto tA = tensor<threadgroup S0, dextents<int32_t, 2>, tensor_inline>(sa, dextents<int32_t, 2>(NK,  NR0));
-    auto tB = tensor<threadgroup S1, dextents<int32_t, 2>, tensor_inline>(sb, dextents<int32_t, 2>(NR1, NK ));
-
-    mpp::tensor_ops::matmul2d<
-        mpp::tensor_ops::matmul2d_descriptor(NR1, NR0, NK, false, true, false, mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
-        execution_simdgroups<4>> mm;
-
-    auto cT = mm.get_destination_cooperative_tensor<decltype(tA), decltype(tB), float>();
-#endif
 
     for (int loop_k = 0; loop_k < args.ne00; loop_k += NK) {
-#ifndef GGML_METAL_HAS_TENSOR
         // load data and store to threadgroup memory
         if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
             threadgroup_barrier(mem_flags::mem_threadgroup);
@@ -9454,66 +9568,6 @@ kernel void kernel_mul_mm(
 
             *(threadgroup S1_2x4 *)(sb + 64*ib + 8*ly) = (S1_2x4)(*((device T1_2x4 *) y));
         }
-#else
-        // load data and store to threadgroup memory
-        if (is_same<T0_4x4, block_q>::value && FC_mul_mm_bc_inp) {
-            threadgroup_barrier(mem_flags::mem_threadgroup);
-
-            // no need for dequantization
-            for (short i = 0; i < 16; i++) {
-                const short sx = 2*il0 + i/8;
-                const short sy = (tiitg/NL0)/8;
-
-                const short lx = i%8;
-                const short ly = (tiitg/NL0)%8;
-                //const short lx = (tiitg/NL0)%8;
-                //const short ly = i%8;
-
-                *(sa + NK*(8*sy + ly) + 8*sx + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
-            }
-        } else {
-            S0_4x4 temp_a;
-            dequantize_func(x, il, temp_a);
-
-            threadgroup_barrier(mem_flags::mem_threadgroup);
-
-            FOR_UNROLL (short i = 0; i < 16; i++) {
-                const short sx = 2*il0 + i/8;
-                const short sy = (tiitg/NL0)/8;
-
-                const short lx = i%8;
-                const short ly = (tiitg/NL0)%8;
-                //const short lx = (tiitg/NL0)%8;
-                //const short ly = i%8;
-
-                *(sa + NK*(8*sy + ly) + 8*sx + lx) = temp_a[i/4][i%4];
-            }
-        }
-
-        if (FC_mul_mm_bc_inp) {
-            for (short i = 0; i < 8; ++i) {
-                const short sx = (tiitg%NL1);
-                const short sy = (tiitg/NL1)/8;
-
-                const short lx = i;
-                const short ly = (tiitg/NL1)%8;
-                //const short lx = (tiitg/NL1)%8;
-                //const short ly = i;
-
-                *(sb + NK*(8*sy + ly) + 8*sx + lx) = loop_k + iy + i < args.ne00 ? (S1) *((device T1 *) y + i) : 0;
-            }
-        } else {
-            const short sx = (tiitg%NL1);
-            const short sy = (tiitg/NL1)/8;
-
-            //const short lx = i;
-            const short ly = (tiitg/NL1)%8;
-            //const short lx = (tiitg/NL1)%8;
-            //const short ly = i;
-
-            *(threadgroup S1_2x4 *)(sb + NK*(8*sy + ly) + 8*sx) = (S1_2x4)(*((device T1_2x4 *) y));
-        }
-#endif
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
         x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
@@ -9522,7 +9576,6 @@ kernel void kernel_mul_mm(
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-#ifndef GGML_METAL_HAS_TENSOR
         // load matrices from threadgroup memory and conduct outer products
         threadgroup const S0 * lsma = (sa + 4*64*(sgitg%2));
         threadgroup const S1 * lsmb = (sb + 2*64*(sgitg/2));
@@ -9549,24 +9602,10 @@ kernel void kernel_mul_mm(
             lsma += 8*64;
             lsmb += 4*64;
         }
-#else
-        auto sA = tA.slice(0, 0);
-        auto sB = tB.slice(0, 0);
-
-        mm.run(sB, sA, cT);
-#endif
     }
 
     if (!FC_mul_mm_bc_out || (r0 + NR0 <= args.ne0 && r1 + NR1 <= args.ne1)) {
         // if no bounds checks on the output are needed, we can directly write to device memory
-#ifdef GGML_METAL_HAS_TENSOR
-        device float * C = (device float *) dst +
-            r0 + \
-            r1 * args.ne0 + im*args.ne1*args.ne0;
-
-        auto tC = tensor<device float, dextents<int32_t, 2>, tensor_inline>(C, dextents<int32_t, 2>(args.ne0, NR1));
-        cT.store(tC);
-#else
         device float * C = (device float *) dst +
             (r0 + 32*(sgitg &  1)) + \
             (r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
@@ -9574,21 +9613,15 @@ kernel void kernel_mul_mm(
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], C + 8*(i%4) + 8*args.ne0*(i/4), args.ne0, 0, false);
         }
-#endif
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         threadgroup float * temp_str = ((threadgroup float *) shmem) + 32*(sgitg&1) + (16*(sgitg >> 1))*NR0;
 
-#ifdef GGML_METAL_HAS_TENSOR
-        auto tC = tensor<threadgroup float, dextents<int32_t, 2>, tensor_inline>(sc, dextents<int32_t, 2>(NR0, NR1));
-        cT.store(tC);
-#else
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*NR0*(i/4), NR0, 0, false);
         }
-#endif
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
@@ -9614,6 +9647,8 @@ kernel void kernel_mul_mm(
     }
 }
 
+#endif // GGML_METAL_HAS_TENSOR
+
 template<short ne20> // n_expert_used
 kernel void kernel_mul_mm_id_map0(
         constant ggml_metal_kargs_mul_mm_id_map0 & args,
@@ -9789,7 +9824,7 @@ kernel void kernel_mul_mm_id(
 
                 const short ib = 8*sx + sy;
 
-                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? *((device T0 *) x + i) : 0;
+                *(sa + 64*ib + 8*ly + lx) = loop_k + 16*il + i < args.ne00 ? (S0) *((device T0 *) x + i) : (S0) 0;
             }
         } else {
             S0_4x4 temp_a;

ggml/src/ggml-opencl/CMakeLists.txt

@@ -96,6 +96,8 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q6_k_f32_flat
     mul_mv_q8_0_f32
     mul_mv_q8_0_f32_flat
+    mul_mv_iq4_nl_f32
+    mul_mv_iq4_nl_f32_flat
     mul_mv_mxfp4_f32
     mul_mv_mxfp4_f32_flat
     mul_mv_id_q4_0_f32_8x_flat
@@ -110,12 +112,15 @@ set(GGML_OPENCL_KERNELS
     mul_mm_q4_0_f32_l4_lm
     mul_mm_q4_1_f32_l4_lm
     mul_mm_q8_0_f32_l4_lm
+    mul_mm_iq4_nl_f32_l4_lm
     mul_mm_q4_k_f32_l4_lm
     mul_mm_q5_k_f32_l4_lm
     mul_mm_q6_k_f32_l4_lm
     mul_mm_q8_0_f32_8x4
     gemv_noshuffle_q4_1_f32
     gemm_noshuffle_q4_1_f32
+    gemv_noshuffle_iq4_nl_f32
+    gemm_noshuffle_iq4_nl_f32
     gemv_noshuffle_general_q8_0_f32
     gemv_noshuffle_q4_k_f32
     gemm_noshuffle_q4_k_f32