spec : refactor params (#22397 )

* spec : refactor params * cont : fix * cont : rename "sparam" to "sampling" * cont : add spec params category * cont : add info about removed arguments * cont : skip param length check for spec params * cont : adapt server tests
server: use pos_next instead of n_tokens for m-rope (#22439 )
2026-04-28 09:07:33 +03:00 · 2026-04-28 08:41:00 +03:00 · 2026-04-27 15:50:59 -07:00 · 2026-04-27 23:55:00 +02:00 · 2026-04-27 08:25:45 -07:00 · 2026-04-27 16:36:59 +02:00
223 changed files with 20652 additions and 7032 deletions
@@ -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

@@ -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)
@@ -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 -->

@@ -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
@@ -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
@@ -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
@@ -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"
@@ -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
@@ -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
@@ -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)
@@ -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
@@ -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)
@@ -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
@@ -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
@@ -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);
@@ -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"];
            }
        }
    }
@@ -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();
 }
-
@@ -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);
@@ -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(),
@@ -1520,7 +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
@@ -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;

@@ -296,34 +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
-
-    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)
@@ -331,25 +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 {
@@ -420,19 +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;
@@ -580,8 +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

@@ -678,7 +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);

 //
@@ -746,6 +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() &&
@@ -841,7 +866,7 @@ 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);
@@ -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);
@@ -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);
 };
@@ -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) {
@@ -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");
+}
@@ -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);
@@ -1,4 +1,3 @@
-#include "log.h"
 #include "value.h"
 #include "runtime.h"
 #include "caps.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");
    }
 };
@@ -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;
 }
@@ -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); }
@@ -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()) {
@@ -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);
    }
 }

@@ -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 {
@@ -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) {
@@ -359,7 +372,7 @@ 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_ckpt && prompt_dft.size() > 0;
        if (reuse_n == 0 || (use_ckpt && reuse_i > 0)) {
@@ -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;
                    }
                }
@@ -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);
 }
@@ -918,8 +986,8 @@ 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;
@@ -929,7 +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);
@@ -952,16 +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);
                }
            }

@@ -991,7 +1060,7 @@ common_speculative * common_speculative_init(
                impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
                    /* .ctx_tgt      = */ ctx_tgt,
                    /* .ctx_dft      = */ ctx_dft,
-                    /* .replacements = */ params.replacements,
+                    /* .replacements = */ params.draft.replacements,
                    /* .use_ckpt     = */ use_ckpt
                ));
                break;
@@ -1001,18 +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;
@@ -1021,18 +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;
            }
@@ -1090,6 +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(),
@@ -1099,7 +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.
        }
    }

@@ -1127,6 +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;
@@ -33,6 +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);

@@ -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 {}

@@ -11855,7 +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

@@ -11994,28 +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]
@@ -12023,11 +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
@@ -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
@@ -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
@@ -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:
@@ -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 | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
@@ -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);
    }
@@ -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;

@@ -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);

@@ -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;
 }
@@ -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
        }

@@ -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");

@@ -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
@@ -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();
@@ -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;
@@ -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
+
+
@@ -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

@@ -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
+
@@ -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
+
@@ -4,8 +4,8 @@ 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/")
@@ -213,7 +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)
@@ -473,7 +473,7 @@ 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)
+        target_link_libraries(ggml-base PRIVATE ${MATH_LIBRARY})
    endif()
 endif()

@@ -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;
    }

@@ -1442,17 +1498,20 @@ 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;
@@ -1464,6 +1523,7 @@ struct ggml_backend_meta_context {

    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);
@@ -1475,7 +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 += ")";

@@ -1505,10 +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) {
@@ -1661,6 +1717,36 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,

                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;
                }
@@ -1675,10 +1761,12 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                        n_used = ggml_node_get_use_count(cgraph, id);
                    }
                }
-                if (id + 1 >= cgraph->n_nodes) {
-                    return idr;
-                }
-                {
+                // 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) {
@@ -1686,6 +1774,8 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                        id++;
                        idr = id;
                        n_used = ggml_node_get_use_count(cgraph, id);
+                    } else {
+                        break;
                    }
                }

@@ -1754,16 +1844,17 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
        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));
+                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));
+                }
            }
            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 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();
@@ -1812,11 +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));
@@ -1828,75 +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;
                }
@@ -1906,7 +2027,20 @@ 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;
    };

@@ -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) {
@@ -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
@@ -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;
 }


@@ -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);
    }

@@ -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;
@@ -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));
@@ -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)
@@ -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);
@@ -3478,10 +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) {
@@ -3495,8 +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.
@@ -3517,8 +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;

@@ -3561,40 +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;
@@ -3612,11 +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;
            }
@@ -3641,32 +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;
@@ -3708,7 +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>
@@ -3717,46 +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;
    }
@@ -3765,11 +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--;
@@ -3779,20 +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--;
@@ -3803,14 +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;
@@ -3818,6 +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) {
@@ -3827,16 +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;
    }
@@ -3856,6 +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) {
@@ -3865,16 +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];
    }
 }

@@ -3922,29 +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);
@@ -3952,40 +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);
    }
 }

@@ -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);
+    }
+}
@@ -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));
 }
@@ -58,6 +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
@@ -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
@@ -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;
@@ -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
@@ -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;
+}
@@ -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;
+}
@@ -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 */
@@ -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)
@@ -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 */
@@ -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 */
@@ -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 */
@@ -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 */
@@ -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);
@@ -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 = {
@@ -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;
+}
@@ -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]
@@ -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;
 }
@@ -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);

@@ -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) {
@@ -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
@@ -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);

@@ -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) {
@@ -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;
@@ -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
@@ -545,6 +545,9 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_convert_block_q5_K_noshuffle;
    cl_kernel kernel_restore_block_q5_K_noshuffle;
    cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
+    cl_kernel kernel_convert_block_iq4_nl, kernel_restore_block_iq4_nl;
+    cl_kernel kernel_convert_block_iq4_nl_noshuffle;
+    cl_kernel kernel_restore_block_iq4_nl_noshuffle;
    cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
    cl_kernel kernel_mul_mv_q4_1_f32;
    cl_kernel kernel_mul_mv_q4_1_f32_flat;
@@ -556,6 +559,8 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_mul_mv_q6_K_f32_flat;
    cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
    cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
+    cl_kernel kernel_mul_mv_iq4_nl_f32;
+    cl_kernel kernel_mul_mv_iq4_nl_f32_flat;
    cl_kernel kernel_solve_tri_f32;
    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
    cl_kernel kernel_argsort_f32_i32;
@@ -594,6 +599,7 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_mul_mm_q4_k_f32_l4_lm;
    cl_kernel kernel_mul_mm_q5_k_f32_l4_lm;
    cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
+    cl_kernel kernel_mul_mm_iq4_nl_f32_l4_lm;

    std::vector<ProfilingInfo> profiling_info;

@@ -734,6 +740,8 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_gemm_noshuffle_q6_K_f32;
    cl_kernel kernel_gemv_noshuffle_q5_k_f32;
    cl_kernel kernel_gemm_noshuffle_q5_k_f32;
+    cl_kernel kernel_gemv_noshuffle_iq4_nl_f32;
+    cl_kernel kernel_gemm_noshuffle_iq4_nl_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS

    void free() {
@@ -954,6 +962,10 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
        CL_CHECK((backend_ctx->kernel_convert_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K_noshuffle", &err), err));
        CL_CHECK((backend_ctx->kernel_restore_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_iq4_nl = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_iq4_nl", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_iq4_nl = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_iq4_nl", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_iq4_nl_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_iq4_nl_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_iq4_nl_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_iq4_nl_noshuffle", &err), err));
        GGML_LOG_CONT(".");
    }

@@ -1359,6 +1371,40 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mv_iq4_nl_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_iq4_nl_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_iq4_nl_f32.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_iq4_nl_f32 = clCreateKernel(prog, "kernel_mul_mv_iq4_nl_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_iq4_nl_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_iq4_nl_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_iq4_nl_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_iq4_nl_f32_flat = clCreateKernel(prog, "kernel_mul_mv_iq4_nl_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mv_mxfp4_f32
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1567,6 +1613,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // mul_mm_iq4_nl_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_iq4_nl_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_iq4_nl_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_iq4_nl_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_iq4_nl_f32_l4_lm", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mm_q4_k_f32_l4_lm
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -2647,6 +2710,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+    // gemm_noshuffle_iq4_nl_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_iq4_nl_f32.cl.h"
+       };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_iq4_nl_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_iq4_nl_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_iq4_nl_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_iq4_nl_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAST ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_iq4_nl_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_iq4_nl_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_iq4_nl_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_iq4_nl_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
    // mul_mm_q8_0_f32_8x4
    {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3597,6 +3699,30 @@ struct ggml_tensor_extra_cl_q8_0 {
    }
 };

+struct ggml_tensor_extra_cl_iq4_nl {
+    cl_mem q = nullptr;
+    cl_mem q_img = nullptr;
+
+    cl_mem d = nullptr;
+    cl_mem d_img = nullptr;
+
+    size_t size_q = 0;
+    size_t size_d = 0;
+
+    ~ggml_tensor_extra_cl_iq4_nl() {
+        reset();
+    }
+
+    void reset() {
+        if (q != nullptr) { CL_CHECK(clReleaseMemObject(q)); q = nullptr; }
+        if (d != nullptr) { CL_CHECK(clReleaseMemObject(d)); d = nullptr; }
+        q_img = nullptr;
+        d_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+    }
+};
+
 struct ggml_tensor_extra_cl_q4_K {
    // Quantized values
    cl_mem q = nullptr;
@@ -4097,6 +4223,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                return op->src[1]->type == GGML_TYPE_F32;
            } else if (op->src[0]->type == GGML_TYPE_Q4_0  || op->src[0]->type == GGML_TYPE_Q4_1 ||
                       op->src[0]->type == GGML_TYPE_MXFP4 ||
+                       op->src[0]->type == GGML_TYPE_IQ4_NL ||
                       op->src[0]->type == GGML_TYPE_Q4_K  ||
                       op->src[0]->type == GGML_TYPE_Q5_K  ||
                       op->src[0]->type == GGML_TYPE_Q6_K) {
@@ -4295,6 +4422,12 @@ struct ggml_backend_opencl_buffer_context {
        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
            delete e;
        }
+        for (ggml_tensor_extra_cl_iq4_nl * e : temp_tensor_extras_iq4_nl) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_iq4_nl * e : temp_tensor_extras_iq4_nl_in_use) {
+            delete e;
+        }
        for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K) {
            delete e;
        }
@@ -4390,6 +4523,21 @@ struct ggml_backend_opencl_buffer_context {
        return extra;
    }

+    ggml_tensor_extra_cl_iq4_nl * ggml_opencl_alloc_temp_tensor_extra_iq4_nl() {
+        ggml_tensor_extra_cl_iq4_nl * extra;
+        if (temp_tensor_extras_iq4_nl.empty()) {
+            extra = new ggml_tensor_extra_cl_iq4_nl();
+        } else {
+            extra = temp_tensor_extras_iq4_nl.back();
+            temp_tensor_extras_iq4_nl.pop_back();
+        }
+
+        temp_tensor_extras_iq4_nl_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
    ggml_tensor_extra_cl_q4_K * ggml_opencl_alloc_temp_tensor_extra_q4_K() {
        ggml_tensor_extra_cl_q4_K * extra;
        if (temp_tensor_extras_q4_K.empty()) {
@@ -4461,6 +4609,11 @@ struct ggml_backend_opencl_buffer_context {
        }
        temp_tensor_extras_q8_0_in_use.clear();

+        for (ggml_tensor_extra_cl_iq4_nl * e : temp_tensor_extras_iq4_nl_in_use) {
+            temp_tensor_extras_iq4_nl.push_back(e);
+        }
+        temp_tensor_extras_iq4_nl_in_use.clear();
+
        for (ggml_tensor_extra_cl_q4_K * e : temp_tensor_extras_q4_K_in_use) {
            temp_tensor_extras_q4_K.push_back(e);
        }
@@ -4492,6 +4645,8 @@ struct ggml_backend_opencl_buffer_context {
    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
+    std::vector<ggml_tensor_extra_cl_iq4_nl *> temp_tensor_extras_iq4_nl;
+    std::vector<ggml_tensor_extra_cl_iq4_nl *> temp_tensor_extras_iq4_nl_in_use;
    std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K;
    std::vector<ggml_tensor_extra_cl_q4_K *> temp_tensor_extras_q4_K_in_use;
    std::vector<ggml_tensor_extra_cl_q5_K *> temp_tensor_extras_q5_K;
@@ -5123,6 +5278,87 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,

        return;
    }
+    if (tensor->type == GGML_TYPE_IQ4_NL) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tensors in OpenCL backend should have been allocated and initialized");
+
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_iq4_nl * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_iq4_nl();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(ggml_blck_size(tensor->type)/2);
+        GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+    #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        cl_kernel kernel = backend_ctx->kernel_convert_block_iq4_nl;
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_iq4_nl_noshuffle;
+        }
+    #else
+        cl_kernel kernel = backend_ctx->kernel_convert_block_iq4_nl;
+    #endif
+        cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+        cl_uchar mask_0F = 0x0F;
+        cl_uchar mask_F0 = 0xF0;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_uchar), &mask_0F));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_F0));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &n_blk));
+
+        size_t global_work_size[] = {(size_t)CEIL_DIV(n_blk, 64)*64, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose q as ushort
+            transpose_2d_as_16b(backend_ctx, extra->q, extra->q, size_q, K/4, M);
+            // Transpose d as ushort
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/32, M);
+        }
+#endif
+        return;
+    }
    if (tensor->type == GGML_TYPE_Q4_K) {
        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
@@ -5775,6 +6011,78 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
        CL_CHECK(clReleaseMemObject(data_device));
        return;
    }
+    if (tensor->type == GGML_TYPE_IQ4_NL) {
+        ggml_tensor_extra_cl_iq4_nl * extra = (ggml_tensor_extra_cl_iq4_nl *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            static ggml_cl_buffer buf_trans_q;
+            static ggml_cl_buffer buf_trans_d;
+            static ggml_cl_buffer buf_unpacked;
+
+            cl_int M = tensor->ne[1];
+            cl_int K = tensor->ne[0];
+            GGML_ASSERT(K % 32 == 0);
+
+            size_t size_q = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*(ggml_blck_size(tensor->type)/2);
+            size_t size_d = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+            buf_trans_q.allocate(backend_ctx->context, size_q);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_unpacked.allocate(backend_ctx->context, ggml_nbytes(tensor));
+
+            // transpose q, d back
+            transpose_2d_as_16b(backend_ctx, extra->q, buf_trans_q.buffer, size_q, M, K/4);
+            transpose_2d_as_16b(backend_ctx, extra->d, buf_trans_d.buffer, size_d, M, K/32);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_iq4_nl_noshuffle;
+            cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_q.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &buf_unpacked.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_F0));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &n_blk));
+
+            size_t global_work_size[] = {(size_t)n_blk, 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, buf_unpacked.buffer, CL_TRUE, offset, size, data, 0, NULL, NULL));
+            return;
+        }
+#endif
+        cl_kernel kernel = backend_ctx->kernel_restore_block_iq4_nl;
+        cl_ulong n_blk = ggml_nelements(tensor)/ggml_blck_size(tensor->type);
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &n_blk));
+
+        size_t global_work_size[] = {(size_t)n_blk, 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
    if (tensor->type == GGML_TYPE_Q4_K) {
        ggml_tensor_extra_cl_q4_K * extra = (ggml_tensor_extra_cl_q4_K *)tensor->extra;

@@ -9840,6 +10148,178 @@ static void ggml_cl_mul_mat_q4_1_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }

+static void ggml_cl_mul_mat_iq4_nl_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl_iq4_nl * extra0_iq4_nl = (ggml_tensor_extra_cl_iq4_nl *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 % 32 == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
+    cl_buffer_region    region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (ne1 == 1) {
+        cl_mem q_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for q
+        img_fmt = { CL_R, CL_UNSIGNED_INT32};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer = extra0_iq4_nl->q;
+        CL_CHECK((q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_iq4_nl_f32;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &q_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_iq4_nl->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &ne01));
+
+        size_t local_work_size[3] = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(q_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // transpose activations
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_iq4_nl_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0_iq4_nl->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra0_iq4_nl->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_int),   &ne1));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3] = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    GGML_ASSERT(src0);
@@ -10634,6 +11114,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
    ggml_tensor_extra_cl_q4_1 * extra0_q4_1 = (ggml_tensor_extra_cl_q4_1 *)src0->extra;
    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+    ggml_tensor_extra_cl_iq4_nl * extra0_iq4_nl = (ggml_tensor_extra_cl_iq4_nl *)src0->extra;
    ggml_tensor_extra_cl_q4_K * extra0_q4_K = (ggml_tensor_extra_cl_q4_K *)src0->extra;
    ggml_tensor_extra_cl_q5_K * extra0_q5_K = (ggml_tensor_extra_cl_q5_K *)src0->extra;
    ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
@@ -10738,6 +11219,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            return;
    }

+    // iq4_nl x fp32
+    if (src0t == GGML_TYPE_IQ4_NL && src1t == GGML_TYPE_F32) {
+        ggml_cl_mul_mat_iq4_nl_f32_adreno(backend, src0, src1, dst);
+        return;
+    }
+
    // q8_0 x fp32
    if (src0t == GGML_TYPE_Q8_0 && src1t == GGML_TYPE_F32 &&
        enable_adreno_trans_weight(backend_ctx, src0)) {
@@ -11302,6 +11789,48 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
                return;
            }
+            case GGML_TYPE_IQ4_NL: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_iq4_nl_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_iq4_nl->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_iq4_nl->d));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
            case GGML_TYPE_Q4_K: {
                if (ne11 < 32) {
                    break;
@@ -11829,6 +12358,70 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
+        case GGML_TYPE_IQ4_NL: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_iq4_nl_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_iq4_nl->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_iq4_nl->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_iq4_nl_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
 #endif // GGML_OPENCL_SOA_Q
            break;
        }
@@ -12131,6 +12724,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
        src0t == GGML_TYPE_Q4_1 ||
        src0t == GGML_TYPE_Q8_0 ||
+        src0t == GGML_TYPE_IQ4_NL ||
        src0t == GGML_TYPE_Q2_K) {
        // Each SIMD group produces N_DST values in the result. Assuming each
        // workgroup has N_SIMDGROUP SIMD groups, then each workgroup will
@@ -87,6 +87,17 @@ struct block_q6_K {
    half d;                  // super-block scale
 };

+//------------------------------------------------------------------------------
+// block_iq4_nl
+//------------------------------------------------------------------------------
+#define QK4_NL 32
+
+struct block_iq4_nl
+{
+    half d;
+    uint8_t qs[QK4_NL / 2];
+};
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q4_0
 // Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
@@ -895,3 +906,99 @@ kernel void kernel_restore_block_q6_K_noshuffle(
        b->scales[i] = s[i];
    }
 }
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_iq4_nl
+// Convert the block_iq4_nl format to 2 separate arrays (AOS -> SOA).
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_iq4_nl(
+    global struct block_iq4_nl * src0,
+    global uchar * dst_q,
+    global half  * dst_d,
+    uchar          mask_0F,
+    uchar          mask_F0,
+    ulong          n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_iq4_nl * b = (global struct block_iq4_nl *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_NL/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK4_NL/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_iq4_nl(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_iq4_nl * dst,
+    ulong          n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_iq4_nl * b = (global struct block_iq4_nl *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_NL/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+
+    for (int i = 0; i < QK4_NL/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+kernel void kernel_convert_block_iq4_nl_noshuffle(
+    global struct block_iq4_nl * src0,
+    global uchar * dst_q,
+    global half  * dst_d,
+    uchar          mask_0F,
+    uchar          mask_F0,
+    ulong          n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_iq4_nl * b = (global struct block_iq4_nl *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_NL/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+    for (int i = 0; i < QK4_NL/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0       ] = convert_uchar(x0 & mask_0F) | convert_uchar((x1 & mask_0F) << 4);
+        q[i + QK4_NL/4] = convert_uchar((x0 & mask_F0) >> 4) | convert_uchar(x1 & mask_F0);
+    }
+}
+
+kernel void kernel_restore_block_iq4_nl_noshuffle(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_iq4_nl * dst,
+    uchar mask_0F,
+    uchar mask_F0,
+    ulong n_blk
+) {
+    if (get_global_id(0) >= n_blk) {
+        return;
+    }
+    global struct block_iq4_nl * b = (global struct block_iq4_nl *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_NL/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_NL/4; ++i) {
+        uchar x0 = q[i + 0       ];
+        uchar x1 = q[i + QK4_NL/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
@@ -0,0 +1,150 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+constant half kvalues_iq4nl[16] = {
+    (half)-127.f, (half)-104.f, (half)-83.f, (half)-65.f,
+    (half) -49.f, (half) -35.f, (half)-22.f, (half)-10.f,
+    (half)   1.f, (half)  13.f, (half) 25.f, (half) 38.f,
+    (half)  53.f, (half)  69.f, (half) 89.f, (half)113.f
+};
+
+// Packed LUT: 2 FP16 values per uint, 8 unique constant loads instead of 16
+constant uint iq4nl_packed[8] = {
+    0xD680D7F0u,  // idx 0,1: -127, -104
+    0xD410D530u,  // idx 2,3: -83, -65
+    0xD060D220u,  // idx 4,5: -49, -35
+    0xC900CD80u,  // idx 6,7: -22, -10
+    0x4A803C00u,  // idx 8,9: 1, 13
+    0x50C04E40u,  // idx 10,11: 25, 38
+    0x545052A0u,  // idx 12,13: 53, 69
+    0x57105590u   // idx 14,15: 89, 113
+};
+
+// Packed dequant: 1 uint constant load (8-way divergence) + shift + as_half
+#define IQ4_NL_DEQUANT(nibble) as_half((ushort)(iq4nl_packed[(nibble) >> 1] >> (((nibble) & 1u) << 4)))
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_gemm_noshuffle_iq4_nl_f32(
+        global const ushort * src0_q,
+        global const half  * src0_d,
+        read_only image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int m,
+        int n,
+        int k,
+        int n_no_padding
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    global const ushort * weight_ptr = src0_q + gx_2;
+    global const half * scale_ptr = src0_d + gx_2;
+
+    for (int i = 0; i < k; i += 4) {
+        B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1);
+
+        ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m));
+
+        half4 scale = vload4(0, scale_ptr + (i/32)*(m));
+
+        // j=0
+        dequantized_weights.s0 = IQ4_NL_DEQUANT(bits4.s0 & 0x000Fu) * scale.s0;
+        dequantized_weights.s1 = IQ4_NL_DEQUANT(bits4.s1 & 0x000Fu) * scale.s1;
+        dequantized_weights.s2 = IQ4_NL_DEQUANT(bits4.s2 & 0x000Fu) * scale.s2;
+        dequantized_weights.s3 = IQ4_NL_DEQUANT(bits4.s3 & 0x000Fu) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1);
+        dequantized_weights.s0 = IQ4_NL_DEQUANT((bits4.s0 >> 4) & 0x000Fu) * scale.s0;
+        dequantized_weights.s1 = IQ4_NL_DEQUANT((bits4.s1 >> 4) & 0x000Fu) * scale.s1;
+        dequantized_weights.s2 = IQ4_NL_DEQUANT((bits4.s2 >> 4) & 0x000Fu) * scale.s2;
+        dequantized_weights.s3 = IQ4_NL_DEQUANT((bits4.s3 >> 4) & 0x000Fu) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1);
+        dequantized_weights.s0 = IQ4_NL_DEQUANT((bits4.s0 >> 8) & 0x000Fu) * scale.s0;
+        dequantized_weights.s1 = IQ4_NL_DEQUANT((bits4.s1 >> 8) & 0x000Fu) * scale.s1;
+        dequantized_weights.s2 = IQ4_NL_DEQUANT((bits4.s2 >> 8) & 0x000Fu) * scale.s2;
+        dequantized_weights.s3 = IQ4_NL_DEQUANT((bits4.s3 >> 8) & 0x000Fu) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1);
+        dequantized_weights.s0 = IQ4_NL_DEQUANT((bits4.s0 >> 12) & 0x000Fu) * scale.s0;
+        dequantized_weights.s1 = IQ4_NL_DEQUANT((bits4.s1 >> 12) & 0x000Fu) * scale.s1;
+        dequantized_weights.s2 = IQ4_NL_DEQUANT((bits4.s2 >> 12) & 0x000Fu) * scale.s2;
+        dequantized_weights.s3 = IQ4_NL_DEQUANT((bits4.s3 >> 12) & 0x000Fu) * scale.s3;
+        c0 += B * dequantized_weights.s0;
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}
@@ -0,0 +1,302 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK4_NL 32
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+constant half kvalues_iq4nl[16] = {
+    (half)-127.f, (half)-104.f, (half)-83.f, (half)-65.f,
+    (half) -49.f, (half) -35.f, (half)-22.f, (half)-10.f,
+    (half)   1.f, (half)  13.f, (half) 25.f, (half) 38.f,
+    (half)  53.f, (half)  69.f, (half) 89.f, (half)113.f
+};
+
+// Packed LUT: 2 FP16 values per uint, 8 unique constant loads instead of 16
+constant uint iq4nl_packed[8] = {
+    0xD680D7F0u,  // idx 0,1: -127, -104
+    0xD410D530u,  // idx 2,3: -83, -65
+    0xD060D220u,  // idx 4,5: -49, -35
+    0xC900CD80u,  // idx 6,7: -22, -10
+    0x4A803C00u,  // idx 8,9: 1, 13
+    0x50C04E40u,  // idx 10,11: 25, 38
+    0x545052A0u,  // idx 12,13: 53, 69
+    0x57105590u   // idx 14,15: 89, 113
+};
+
+// Packed dequant: 1 uint constant load (8-way divergence) + shift + as_half
+#define IQ4_NL_DEQUANT(nibble) as_half((ushort)(iq4nl_packed[(nibble) >> 1] >> (((nibble) & 1u) << 4)))
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s0 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s1 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s2 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s3 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s4 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s5 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s6 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s7 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s0 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s1 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s2 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s3 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s4 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s5 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s6 & 0x000F)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s7 & 0x000F)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x00F0) >> 4)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x00F0) >> 4)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x0F00) >> 8)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x0F00) >> 8)) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0xF000) >> 12)) * scale.s0 * shared_y; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0xF000) >> 12)) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s0 & 0x000F))         * scale.s0 * shared_y.s0; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x00F0) >> 4))  * scale.s0 * shared_y.s1; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x0F00) >> 8))  * scale.s0 * shared_y.s2; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0xF000) >> 12)) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s2 & 0x000F))         * scale.s0 * shared_y.s4; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x00F0) >> 4))  * scale.s0 * shared_y.s5; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x0F00) >> 8))  * scale.s0 * shared_y.s6; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0xF000) >> 12)) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s1 & 0x000F))         * scale.s1 * shared_y.s0; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x00F0) >> 4))  * scale.s1 * shared_y.s1; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x0F00) >> 8))  * scale.s1 * shared_y.s2; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0xF000) >> 12)) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s3 & 0x000F))         * scale.s1 * shared_y.s4; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x00F0) >> 4))  * scale.s1 * shared_y.s5; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x0F00) >> 8))  * scale.s1 * shared_y.s6; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0xF000) >> 12)) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s4 & 0x000F))         * scale.s0 * shared_y.s0; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x00F0) >> 4))  * scale.s0 * shared_y.s1; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x0F00) >> 8))  * scale.s0 * shared_y.s2; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0xF000) >> 12)) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s6 & 0x000F))         * scale.s0 * shared_y.s4; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x00F0) >> 4))  * scale.s0 * shared_y.s5; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x0F00) >> 8))  * scale.s0 * shared_y.s6; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0xF000) >> 12)) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s5 & 0x000F))         * scale.s1 * shared_y.s0; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x00F0) >> 4))  * scale.s1 * shared_y.s1; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x0F00) >> 8))  * scale.s1 * shared_y.s2; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0xF000) >> 12)) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s7 & 0x000F))         * scale.s1 * shared_y.s4; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x00F0) >> 4))  * scale.s1 * shared_y.s5; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x0F00) >> 8))  * scale.s1 * shared_y.s6; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0xF000) >> 12)) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s0 & 0x000F))         * scale.s0 * shared_y.s0; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x00F0) >> 4))  * scale.s0 * shared_y.s1; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0x0F00) >> 8))  * scale.s0 * shared_y.s2; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s0 & 0xF000) >> 12)) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s2 & 0x000F))         * scale.s0 * shared_y.s4; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x00F0) >> 4))  * scale.s0 * shared_y.s5; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0x0F00) >> 8))  * scale.s0 * shared_y.s6; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s2 & 0xF000) >> 12)) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s1 & 0x000F))         * scale.s1 * shared_y.s0; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x00F0) >> 4))  * scale.s1 * shared_y.s1; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0x0F00) >> 8))  * scale.s1 * shared_y.s2; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s1 & 0xF000) >> 12)) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s3 & 0x000F))         * scale.s1 * shared_y.s4; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x00F0) >> 4))  * scale.s1 * shared_y.s5; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0x0F00) >> 8))  * scale.s1 * shared_y.s6; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s3 & 0xF000) >> 12)) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s4 & 0x000F))         * scale.s0 * shared_y.s0; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x00F0) >> 4))  * scale.s0 * shared_y.s1; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0x0F00) >> 8))  * scale.s0 * shared_y.s2; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s4 & 0xF000) >> 12)) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += IQ4_NL_DEQUANT((bits4.s6 & 0x000F))         * scale.s0 * shared_y.s4; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x00F0) >> 4))  * scale.s0 * shared_y.s5; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0x0F00) >> 8))  * scale.s0 * shared_y.s6; \
+    total_sums.s0 += IQ4_NL_DEQUANT(((bits4.s6 & 0xF000) >> 12)) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s5 & 0x000F))         * scale.s1 * shared_y.s0; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x00F0) >> 4))  * scale.s1 * shared_y.s1; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0x0F00) >> 8))  * scale.s1 * shared_y.s2; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s5 & 0xF000) >> 12)) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += IQ4_NL_DEQUANT((bits4.s7 & 0x000F))         * scale.s1 * shared_y.s4; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x00F0) >> 4))  * scale.s1 * shared_y.s5; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0x0F00) >> 8))  * scale.s1 * shared_y.s6; \
+    total_sums.s1 += IQ4_NL_DEQUANT(((bits4.s7 & 0xF000) >> 12)) * scale.s1 * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_iq4_nl_f32(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+
+    private uint4     regA;
+    private half2     regS;
+    private float8    regB;
+
+    private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_NL); k += NSUBGROUPS) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}
@@ -0,0 +1,171 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 8
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+constant float kvalues_iq4nl[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f,
+      1.f,   13.f,  25.f,  38.f,  53.f,  69.f,  89.f, 113.f
+};
+
+kernel void kernel_mul_mm_iq4_nl_f32_l4_lm(
+    global uchar4 * src0_q,
+    global half   * src0_d,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                int ib  = idx / 4;
+                int iqs = idx % 4;
+
+                float d = (float)src0_d[ib];
+                global uchar4 * qs = src0_q + ib*4 + iqs;
+                uchar4 q = *qs;
+                // IQ4_NL: use lookup table instead of linear (nibble - 8)
+                float4 v1 = (float4)(kvalues_iq4nl[(q.s0   )&0x0F], kvalues_iq4nl[(q.s1   )&0x0F],
+                                     kvalues_iq4nl[(q.s2   )&0x0F], kvalues_iq4nl[(q.s3   )&0x0F])*d;
+                float4 v2 = (float4)(kvalues_iq4nl[(q.s0>>4)&0x0F], kvalues_iq4nl[(q.s1>>4)&0x0F],
+                                     kvalues_iq4nl[(q.s2>>4)&0x0F], kvalues_iq4nl[(q.s3>>4)&0x0F])*d;
+
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = v1.s3;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = v2.s0;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = v2.s1;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = v2.s2;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = v2.s3;
+            } else {
+                buf_a[(loadr_a * 4 +  0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 +  3) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 16) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 17) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 18) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * 4 + 19) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
@@ -0,0 +1,164 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_NL 32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+constant float kvalues_iq4nl[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f,
+      1.f,   13.f,  25.f,  38.f,  53.f,  69.f,  89.f, 113.f
+};
+
+//------------------------------------------------------------------------------
+// block_iq4_nl
+//------------------------------------------------------------------------------
+struct block_iq4_nl
+{
+    half d;
+    uint8_t qs[QK4_NL / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32
+//------------------------------------------------------------------------------
+// Compute inner product between half a block of iq4_nl and 16 floats (yl).
+// il indicates where the quants begin (0 or 8).
+inline float block_iq4_nl_dot_y(
+        global struct block_iq4_nl * qb_curr,
+        private float * yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float acc = 0.f;
+    global uchar * qs = qb_curr->qs + il;
+    for (int i = 0; i < 8; ++i) {
+        acc += yl[i]   * kvalues_iq4nl[qs[i] & 0x0F];
+        acc += yl[i+8] * kvalues_iq4nl[qs[i] >> 4];
+    }
+    return d * acc;
+}
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each subgroup group works on 4 rows
+#define N_SUBGROUP 1 // number of subgroups in a thread group
+#define N_SUBGROUP_SIZE 16 // assuming subgroup size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SUBGROUP 1
+#define N_SUBGROUP_SIZE 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+
+    const ulong nb = ne00/QK4_NL;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SUBGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_iq4_nl * x = (global struct block_iq4_nl *) src0 + offset0;
+    global float               * y = (global float               *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];       // src1 vector cache
+    float sumf[N_DST]={0.f};
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_NL + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SUBGROUP_SIZE/2) {
+        for (int i = 0; i < 8; ++i) {
+            yl[i]   = yb[i];
+            yl[i+8] = yb[i+16];
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            sumf[row] += block_iq4_nl_dot_y(x+ib+row*nb, yl, il);
+        }
+
+        yb += QK4_NL * (N_SUBGROUP_SIZE/2);
+    }
+
+    float tot[N_DST] = {
+        sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]),
+        sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])};
+    for (int row = 0; row < N_DST; ++row) {
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row];
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq4_nl_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
@@ -0,0 +1,202 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_NL 32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+constant float kvalues_iq4nl[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f,
+      1.f,   13.f,  25.f,  38.f,  53.f,  69.f,  89.f, 113.f
+};
+
+//------------------------------------------------------------------------------
+// block_iq4_nl
+//------------------------------------------------------------------------------
+struct block_iq4_nl
+{
+    half d;
+    uint8_t qs[QK4_NL / 2];
+};
+
+// Compute dot product between half a block of iq4_nl quants and activations.
+// x points to the quant bytes, dh points to the scale.
+// yl has 16 activation values: [0..7] for low nibbles, [8..15] for high nibbles.
+// il indicates offset into the quant bytes (0 or 8).
+inline float block_iq4_nl_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        private float * yl,
+        int il
+) {
+    float d = *dh;
+    global uchar * qs = x + il;
+    float acc = 0.f;
+    for (int i = 0; i < 8; ++i) {
+        acc += yl[i]   * kvalues_iq4nl[qs[i] & 0x0F];
+        acc += yl[i+8] * kvalues_iq4nl[qs[i] >> 4];
+    }
+    return d * acc;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each subgroup works on 8 rows
+#define N_SUBGROUP 1 // number of subgroups in a thread group
+#define N_SUBGROUP_SIZE 16 // assuming subgroup size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SUBGROUP 1
+#define N_SUBGROUP_SIZE 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_NL;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SUBGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_NL/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_NL/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_NL + il;
+
+    for (int ib = ix; ib < nb; ib += N_SUBGROUP_SIZE/2) {
+        for (int i = 0; i < 8; ++i) {
+            yl[i]   = yb[i];
+            yl[i+8] = yb[i+16];
+        }
+
+        sumf.s0 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 0*nb*QK4_NL/2, d + ib + 0*nb, yl, il);
+        sumf.s1 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 1*nb*QK4_NL/2, d + ib + 1*nb, yl, il);
+        sumf.s2 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 2*nb*QK4_NL/2, d + ib + 2*nb, yl, il);
+        sumf.s3 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 3*nb*QK4_NL/2, d + ib + 3*nb, yl, il);
+
+        sumf.s4 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 4*nb*QK4_NL/2, d + ib + 4*nb, yl, il);
+        sumf.s5 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 5*nb*QK4_NL/2, d + ib + 5*nb, yl, il);
+        sumf.s6 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 6*nb*QK4_NL/2, d + ib + 6*nb, yl, il);
+        sumf.s7 += block_iq4_nl_dot_y_flat(x + ib*QK4_NL/2 + 7*nb*QK4_NL/2, d + ib + 7*nb, yl, il);
+
+        yb += QK4_NL * (N_SUBGROUP_SIZE/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq4_nl_f32_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
@@ -19,7 +19,6 @@
 #include <iomanip>
 #include <map>
 #include <memory>
-#include <mutex>
 #include <openvino/core/dimension.hpp>
 #include <openvino/core/except.hpp>
 #include <openvino/core/node.hpp>
@@ -207,8 +206,22 @@ int GgmlOvDecoder::compute_op_case(const ggml_tensor * node) const {
        break;
    }
    case GGML_OP_ROPE: {
+        const int mode = node->op_params[2];
+        switch (mode) {
+       case GGML_ROPE_TYPE_NEOX: {
+            op_case = 0x00010000;
+            break;
+        }
+       case GGML_ROPE_TYPE_IMROPE: {
+            op_case = 0x00020000;
+            break;
+        }
+        default:
+            op_case = 0x00000000;
+            break;
+        }
        if (node->src[0]->op == GGML_OP_VIEW) {
-            op_case = 2;
+            op_case = (op_case | 0x00000002);
        }
        break;
    }
@@ -573,9 +586,6 @@ std::map<std::string, std::string> GgmlOvDecoder::get_kv_param_res_names() const
 }

 std::map<std::string, std::shared_ptr<ov::Node>> GgmlOvDecoder::create_weight_nodes(ggml_cgraph * cgraph, bool naive) {
-    static std::mutex weights_mutex;
-    std::lock_guard<std::mutex> lock(weights_mutex);
-
    std::map<std::string, std::shared_ptr<ov::Node>> model_weights;
    auto * nodes = cgraph->nodes;
    auto n_nodes = cgraph->n_nodes;
@@ -6,6 +6,7 @@
 #include <cstring>
 #include <openvino/runtime/intel_gpu/ocl/ocl.hpp>
 #include <openvino/runtime/intel_npu/level_zero/level_zero.hpp>
+#include <openvino/runtime/properties.hpp>
 #include <optional>

 ov::Core & ov_singleton_core() {
@@ -42,11 +43,13 @@ void ggml_openvino_device_config::init() {
            {"NPUW_DQ",                           "YES"   },
            {"NPUW_DQ_FULL",                      "NO"    },
        };
-        if (cache_dir) {
+        if (cache_dir && strlen(cache_dir) > 0) {
            compile_config["NPUW_CACHE_DIR"] = cache_dir;
+            compile_config.insert(ov::cache_mode(ov::CacheMode::OPTIMIZE_SIZE));
        }
-    } else if (cache_dir) {
-        ov_singleton_core().set_property(ov::cache_dir(cache_dir));
+    } else if (cache_dir && strlen(cache_dir) > 0) {
+        compile_config.insert(ov::cache_dir(cache_dir));
+        compile_config.insert(ov::cache_mode(ov::CacheMode::OPTIMIZE_SIZE));
    }

    // Initialize remote context with queue sharing for GPU
@@ -259,10 +262,12 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten
            layout.weights_size = layout.is_u4 ? (n_elements / 2) : n_elements;
            int64_t n_blocks = n_elements / layout.weights_per_block;
            layout.scales_size = n_blocks * sizeof(uint16_t);
-            // For symmetric quantization, we only need one zp value (not one per block)
-            // Zero points are stored in U4 or U8 format matching the weight type
-            size_t n_zp_elements = layout.is_symmetric ? 1 : n_blocks;
-            layout.zp_size = layout.is_u4 ? ((n_zp_elements + 1) / 2) : n_zp_elements;
+            // For symmetric quantization, no zp needed (weights stored as signed)
+            if (layout.is_symmetric) {
+                layout.zp_size = 0;
+            } else {
+                layout.zp_size = layout.is_u4 ? ((n_blocks + 1) / 2) : n_blocks;
+            }

            layout.weights_offset = 0;
            layout.scales_offset = ((layout.weights_size + alignment - 1) / alignment) * alignment;
@@ -313,10 +318,12 @@ ggml_openvino_extracted_layout ggml_openvino_get_extracted_layout(const ggml_ten
    // Scales: F16 per block
    int64_t n_blocks = n_elements / layout.weights_per_block;
    layout.scales_size = n_blocks * sizeof(uint16_t);  // F16 = 2 bytes
-    // Zero points: U4 or U8 matching weight type
-    // For symmetric quantization, we only need one zp value (not one per block)
-    size_t n_zp_elements = layout.is_symmetric ? 1 : n_blocks;
-    layout.zp_size = layout.is_u4 ? ((n_zp_elements + 1) / 2) : n_zp_elements;
+    // For symmetric quantization, no zp needed (weights stored as signed)
+    if (layout.is_symmetric) {
+        layout.zp_size = 0;
+    } else {
+        layout.zp_size = layout.is_u4 ? ((n_blocks + 1) / 2) : n_blocks;
+    }

    // Layout in buffer: [weights | scales | zp] with alignment
    layout.weights_offset = 0;
--- a/Show More
+++ b/Show More