Merge branch 'master' into xsn/common_json

.devops/openvino.Dockerfile

@@ -1,12 +1,12 @@
-ARG OPENVINO_VERSION_MAJOR=2026.3.1
-ARG OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04
 
 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.40.13
-ARG IGC_VERSION_FULL=2_2.40.13+22418
-ARG COMPUTE_RUNTIME_VERSION=26.31.39395.13
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.31.39395.13-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0
 
 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
@@ -90,9 +90,6 @@ RUN bash -c "source ${OpenVINO_DIR}/setupvars.sh && \
     cmake -B build/ReleaseOV -G Ninja \
         -DCMAKE_BUILD_TYPE=Release \
         -DLLAMA_BUILD_TESTS=OFF \
-        -DGGML_NATIVE=OFF \
-        -DGGML_BACKEND_DL=ON \
-        -DGGML_CPU_ALL_VARIANTS=ON \
         -DGGML_OPENVINO=ON && \
     cmake --build build/ReleaseOV --parallel "
 

.github/actions/ccache-buckets/action.yml

@@ -1,95 +0,0 @@
-name: "ccache-buckets"
-description: "Save/restore latest GitHub Actions ccache matching a key prefix to/from HF buckets"
-inputs:
-  key:
-    description: "Cache key prefix to match and load"
-    required: true
-  folder:
-    description: "Bucket folder containing ccache files"
-    required: true
-  evict-old-files:
-    description: "Corresponds to the ccache --evict-older-than AGE option, where AGE is the number of seconds or days followed by the 's' or 'd' suffix respectively."
-    default: ''
-  save:
-    description: "Save ccache"
-    required: false
-    default: false
-    type: boolean
-  hf_bucket:
-    description: 'Hugging Face buckets path'
-    required: true
-
-runs:
-  using: "composite"
-  steps:
-    - name: Install Hugging Face Hub CLI
-      shell: bash
-      run: |
-        python3 -m venv .venv-hf
-        .venv-hf/bin/pip install -U huggingface_hub==1.28.0
-
-    - name: Restore ccache from buckets
-      if: ${{ inputs.save != 'true' }}
-      shell: bash
-      run: |
-        set +e -uo pipefail
-        source .venv-hf/bin/activate
-        CCACHE_DIR=$(ccache -k cache_dir)
-        if [[ -d "$CCACHE_DIR" ]]; then
-            CACHE_PATH=$(hf buckets list "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}" --json | jq -r '[.[] | select(.type == "file") | select(.path | startswith("${{ inputs.folder }}/${{ inputs.key }}") and endswith(".tar.gz"))] | sort_by(.path) | last | .path // ""')
-            if [[ -n "$CACHE_PATH" ]]; then
-                echo "Restoring ccache from '$CACHE_PATH'."
-                hf buckets cp "hf://buckets/${{ inputs.hf_bucket }}/$CACHE_PATH" ccache_bucket.tar.gz
-                mkdir -p ccache_bucket
-                if tar -xzf ccache_bucket.tar.gz -C ccache_bucket; then
-                    rm -rf "$CCACHE_DIR"
-                    mv ccache_bucket "$CCACHE_DIR"
-                    ccache -z
-                fi
-                rm ccache_bucket.tar.gz
-            else
-                echo "No ccache found."
-            fi
-        else
-            echo "'$CCACHE_DIR' not found."
-        fi
-
-    - name: Save ccache to buckets
-      if: ${{ inputs.save == 'true' }}
-      shell: bash
-      run: |
-        if [[ -n "$HF_TOKEN" ]]; then
-            set +e -uo pipefail
-            source .venv-hf/bin/activate
-            CCACHE_DIR=$(ccache -k cache_dir)
-            if [[ -d "$CCACHE_DIR" ]]; then
-                ccache -s
-                if [[ -n "${{ inputs.evict-old-files }}" ]]; then
-                    ccache --evict-older-than "${{ inputs.evict-old-files }}"
-                fi
-                DATESTAMP=$(date -u +'%Y-%m-%dT%H:%M:%SZ')
-                CACHEFILE="${{ inputs.key }}-$DATESTAMP.tar.gz"
-                if tar -czf ccache_bucket.tar.gz -C "$CCACHE_DIR" .; then
-                    hf buckets cp ccache_bucket.tar.gz "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}/$CACHEFILE"
-                fi
-                rm ccache_bucket.tar.gz
-            else
-                echo "'$CCACHE_DIR' not found."
-            fi
-        fi
-
-    - name: Remove old ccache files from buckets
-      if: ${{ inputs.save == 'true' }}
-      shell: bash
-      run: |
-        if [[ -n "$HF_TOKEN" ]]; then
-            set +e -uo pipefail
-            source .venv-hf/bin/activate
-            CACHE_FILES=$(hf buckets list "hf://buckets/${{ inputs.hf_bucket }}/${{ inputs.folder }}" --json | jq -r '[.[] | select(.type == "file") | select((.uploaded_at | .[:19]+"Z" | fromdateiso8601) < (now - 5 * 60)) | select(.path | startswith("${{ inputs.folder }}/${{ inputs.key }}") and endswith(".tar.gz"))] | sort_by(.path)[:-1] | .[] | [.path // ""] | @tsv')
-            if [[ -n "$CACHE_FILES" ]]; then
-                echo "Removing old ccache files..."
-                while IFS=$'\t' read -r CACHE_PATH; do
-                    hf buckets rm "hf://buckets/${{ inputs.hf_bucket }}/$CACHE_PATH" -y
-                done <<< "$CACHE_FILES"
-            fi
-        fi

.github/actions/ccache-clear/action.yml

@@ -21,30 +21,68 @@ inputs:
 runs:
   using: "composite"
   steps:
-    - name: Install GitHub CLI if missing
-      shell: bash
-      run: |
-        # e.g. in container jobs, where it is not preinstalled
-        if ! command -v gh >/dev/null 2>&1; then
-          echo "GitHub CLI not found, installing..."
-          if ! command -v curl >/dev/null 2>&1; then
-            apt-get update >/dev/null 2>&1 || true
-            apt-get install -y curl >/dev/null 2>&1 || true
-          fi
-          mkdir -p -m 755 /etc/apt/keyrings
-          curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg | tee /etc/apt/keyrings/githubcli-archive-keyring.gpg >/dev/null
-          chmod go+r /etc/apt/keyrings/githubcli-archive-keyring.gpg
-          echo "deb [arch=$(dpkg --print-architecture) signed-by=/etc/apt/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" > /etc/apt/sources.list.d/github-cli.list
-          apt-get update >/dev/null 2>&1 || true
-          apt-get install -y gh || { echo "Failed to install GitHub CLI (gh)" >&2; exit 1; }
-        fi
-        command -v gh >/dev/null 2>&1 || { echo "GitHub CLI (gh) is required but could not be installed" >&2; exit 1; }
-
     - name: Clear caches
       shell: bash
+      env:
+        CLEAR_KEY: ${{ inputs.key }}
+        CLEAR_OLDER: ${{ inputs.older }}
+        CLEAR_MIN: ${{ inputs.min }}
+        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        bash scripts/ccache-clear.sh \
-          --key "${{ inputs.key }}" \
-          --older "${{ inputs.older }}" \
-          --min "${{ inputs.min }}" \
-          ${{ inputs.dry-run == 'true' && '--dry-run' || '' }}
+        # Convert a duration (e.g. 90m, 1h, 1d, plain seconds) to seconds
+        to_seconds() {
+          local val="$1"
+          [[ "$val" =~ ^[0-9]+$ ]] && { echo "$val"; return 0; }
+          local num="${val%?}" unit="${val: -1}" mult
+          [[ "$num" =~ ^[0-9]+$ ]] || return 1
+          case "$unit" in
+            s) mult=1 ;;
+            m) mult=60 ;;
+            h) mult=3600 ;;
+            d) mult=86400 ;;
+            *) return 1 ;;
+          esac
+          echo $((num * mult))
+        }
+
+        [[ "$CLEAR_MIN" =~ ^[0-9]+$ ]] || { echo "Invalid min value: $CLEAR_MIN" >&2; exit 1; }
+        [[ "$CLEAR_DRY_RUN" =~ ^(true|false)$ ]] || { echo "Invalid dry-run value: $CLEAR_DRY_RUN" >&2; exit 1; }
+
+        CACHES=$(gh cache list --key "ccache-$CLEAR_KEY" --json id,key,createdAt --jq '.[] | [.createdAt, .id, .key] | @tsv' 2>/dev/null | LC_ALL=C sort)
+        if [ -z "$CACHES" ]; then
+          echo "No caches found with key prefix: $CLEAR_KEY"
+          exit 0
+        fi
+
+        TOTAL=$(( $(wc -l <<< "$CACHES") ))
+
+        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
+        done <<< "$CACHES"
+
+        CUTOFF=""
+        if [ -n "$CLEAR_OLDER" ]; then
+          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
+          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
+        fi
+
+        # Caches are sorted oldest first
+        DELETED=0
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
+            echo "Rest are not older than $CLEAR_OLDER, stopping"
+            break
+          fi
+          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
+            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
+            break
+          fi
+          if [ "$CLEAR_DRY_RUN" = "true" ]; then
+            echo "Would delete cache: $ID ($KEY)"
+          else
+            echo "Deleting cache: $ID ($KEY)"
+            gh cache delete "$ID"
+          fi
+          DELETED=$((DELETED + 1))
+        done <<< "$CACHES"

.github/workflows/build-android.yml

@@ -110,7 +110,7 @@ jobs:
       #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
       #
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: android-ubuntu-arm64
       #    evict-old-files: 1d

.github/workflows/build-apple.yml

@@ -22,8 +22,7 @@ on:
     types: [opened, synchronize, reopened]
     paths: [
       '.github/workflows/build-apple.yml',
-      'ggml/src/ggml-metal/**',
-      'ggml/src/ggml-rpc/**'
+      'ggml/src/ggml-metal/**'
     ]
 
 concurrency:
@@ -47,7 +46,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: apple-arm64
           evict-old-files: 1d
@@ -74,16 +73,6 @@ jobs:
           cd build
           ctest -L main -E "test-llama-archs" --verbose --timeout 900
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: apple-arm64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   macos-latest-x64:
     runs-on: macos-15-intel
 
@@ -93,7 +82,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: apple-x64
           evict-old-files: 1d
@@ -120,16 +109,6 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: apple-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   macos-latest-ios-xcode:
     runs-on: macos-latest
 
@@ -184,6 +163,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-tvos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -209,6 +196,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-visionos
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Build
         id: cmake_build
         run: |
@@ -239,6 +234,14 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      # TODO: this likely does not do anything - if yes, remove it
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: apple-swift
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
       - name: Download xcframework artifact
         uses: actions/download-artifact@v7
         with:

.github/workflows/build-cache.yml

@@ -41,8 +41,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -69,8 +69,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/build-cpu.yml

@@ -62,7 +62,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-${{ matrix.os }}
           evict-old-files: 1d
@@ -125,7 +125,7 @@ jobs:
           GH_TOKEN: ${{ github.token }}
         with:
           key: cpu-${{ matrix.os }}
-          older: 5m
+          older: 1h
           min: 1
           dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
 
@@ -156,7 +156,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-windows-2025-${{ matrix.build }}
           variant: ccache
@@ -215,13 +215,3 @@ jobs:
       #     cd build
       #     $env:LLAMA_SKIP_TESTS_SLOW_ON_EMULATOR = 1
       #     & $sde -future -- ctest -L main -C Release --verbose --timeout 900
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-windows-2025-${{ matrix.build }}
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-cuda-ubuntu.yml

@@ -50,22 +50,14 @@ jobs:
           DEBIAN_FRONTEND: noninteractive
         run: |
           apt update
-          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev jq python3 python3-venv python3-pip
+          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-24.04-cuda
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-24.04-cuda
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with CMake
         # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
@@ -80,18 +72,6 @@ jobs:
             -DGGML_CUDA_CUB_3DOT2=ON
           cmake --build build
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-24.04-cuda
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   hip:
     runs-on: ubuntu-22.04
     container: rocm/dev-ubuntu-22.04:6.1.2
@@ -105,22 +85,14 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev jq python3-venv
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-22.04-hip
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-hip
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with native CMake HIP support
         id: cmake_build
@@ -131,18 +103,6 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-hip
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true
-
   musa:
     runs-on: ubuntu-22.04
     container: mthreads/musa:rc4.3.0-devel-ubuntu22.04-amd64
@@ -156,22 +116,14 @@ jobs:
         id: depends
         run: |
           apt-get update
-          apt-get install -y build-essential git cmake libssl-dev jq
+          apt-get install -y build-essential git cmake libssl-dev
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cuda-ubuntu-22.04-musa
-          save: false
-
-      - name: ccache-buckets-restore
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-musa
-          folder: llama.cpp
-          hf_bucket: ggml-org/cache
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build with native CMake MUSA support
         id: cmake_build
@@ -179,15 +131,3 @@ jobs:
           cmake -B build -S . \
             -DGGML_MUSA=ON
           time cmake --build build --config Release -j $(nproc)
-
-      - name: ccache-buckets-save
-        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
-        uses: ./.github/actions/ccache-buckets
-        env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
-        with:
-          key: cuda-ubuntu-22.04-musa
-          folder: llama.cpp
-          evict-old-files: 1d
-          hf_bucket: ggml-org/cache
-          save: true

.github/workflows/build-cuda-windows.yml

@@ -47,7 +47,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
@@ -152,7 +152,7 @@ jobs:
           & "${env:HIP_PATH}\lib\llvm\bin\clang.exe" --version
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           # TODO: this build does not match the build in release.yml, so we use a different cache key
           #       ideally, the builds should match, similar to the CUDA build above so that we would be able

.github/workflows/build-msys.yml

@@ -35,7 +35,7 @@ jobs:
         uses: actions/checkout@v6
 
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.16
       #  with:
       #    key: msys-windows-2025-x64
       #    variant: ccache

.github/workflows/build-opencl.yml

@@ -44,7 +44,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: opencl-windows-2025-x64
           variant: ccache
@@ -80,13 +80,3 @@ jobs:
         run: |
           cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: opencl-windows-2025-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-openvino.yml

@@ -32,8 +32,6 @@ env:
   LLAMA_ARG_LOG_COLORS: 1
   LLAMA_ARG_LOG_PREFIX: 1
   LLAMA_ARG_LOG_TIMESTAMPS: 1
-  # TODO: fix and re-enable the `test-llama-archs` and `test-recurrent-state-rollback`
-  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-rollback"
 
 jobs:
   ubuntu-24-openvino:
@@ -41,8 +39,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -80,24 +78,26 @@ jobs:
 
       - name: Test (CPU)
         id: cmake_test_cpu
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000
 
       - name: Test (GPU)
         id: cmake_test_gpu
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           cd ${{ github.workspace }}
           export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000
 
   openvino-windows-2022:
     runs-on: windows-2022
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone
@@ -105,7 +105,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: openvino-windows-2022
           variant: ccache
@@ -159,20 +159,11 @@ jobs:
       - name: Test (CPU)
         id: cmake_test_cpu
         shell: cmd
+        # TODO: fix and re-enable the `test-llama-archs` test below
         run: |
           REM Find extracted OpenVINO folder dynamically
           for /d %%i in (openvino_toolkit\*) do set OPENVINO_ROOT=%%i
           call "%OPENVINO_ROOT%\setupvars.bat"
 
           cd build
-          ctest --test-dir ReleaseOV -L main -E "${{ env.CTEST_EXCLUDE }}" -C Release --verbose --timeout 3000
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: openvino-windows-2022
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000

.github/workflows/build-riscv.yml

@@ -67,7 +67,7 @@ jobs:
 
       # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
       #  with:
       #    key: riscv-ubuntu-native
       #    evict-old-files: 1d
@@ -137,7 +137,7 @@ jobs:
 
       # note: sparing some ccache since these jobs run on dedicated runners that are not part of the organitzation
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
       #  with:
       #    key: riscv-ubuntu-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
       #    evict-old-files: 1d

.github/workflows/build-sanitize.yml

@@ -55,7 +55,7 @@ jobs:
         uses: actions/checkout@v6
 
       # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.24
+      #   uses: ggml-org/ccache-action@v1.2.21
       #   if: ${{ matrix.sanitizer != 'UNDEFINED' }}
       #   with:
       #     key: ctest-${{ matrix.sanitizer }}-ubuntu-24.04

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

@@ -288,8 +288,8 @@ jobs:
 
     env:
       # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Clone

.github/workflows/build-sycl.yml

@@ -75,7 +75,7 @@ jobs:
           sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: sycl-ubuntu-24-${{ matrix.build }}
           evict-old-files: 1d
@@ -96,16 +96,6 @@ jobs:
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: sycl-ubuntu-24-${{ matrix.build }}
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   windows-latest-sycl:
     runs-on: windows-2022
 
@@ -137,7 +127,7 @@ jobs:
           "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: sycl-windows-latest
           variant: ccache
@@ -149,13 +139,3 @@ jobs:
       - name: Build
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: sycl-windows-latest
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-vulkan.yml

@@ -53,9 +53,9 @@ jobs:
           echo "CXX=g++-14" >> "$GITHUB_ENV"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: vulkan-ubuntu-24.04-arm
+          key: vulkan-ubuntu-24.04-arm-new
           variant: ccache
           evict-old-files: 1d
           save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -73,16 +73,6 @@ jobs:
         run: |
           time cmake --build build -j $(nproc)
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: vulkan-ubuntu-24.04-arm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   ubuntu-llvmpipe:
     runs-on: ubuntu-24.04
 
@@ -112,7 +102,7 @@ jobs:
           strip: 1
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: vulkan-ubuntu-24.04-llvmpipe
           evict-old-files: 1d
@@ -138,16 +128,6 @@ jobs:
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: vulkan-ubuntu-24.04-llvmpipe
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   windows:
     runs-on: windows-2025
 
@@ -160,7 +140,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: cpu-windows-2025-x64-vulkan
           variant: ccache
@@ -200,13 +180,3 @@ jobs:
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: cpu-windows-2025-x64-vulkan
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-wasm.yml

@@ -54,7 +54,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-ubuntu-24.04-arm-wasm
           evict-old-files: 1d
@@ -88,13 +88,3 @@ jobs:
             -DEMDAWNWEBGPU_DIR=emdawnwebgpu_pkg
 
           time cmake --build build-wasm --config Release --target test-backend-ops -j $(nproc)
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: webgpu-ubuntu-24.04-arm-wasm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-webgpu.yml

@@ -69,7 +69,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-macos-latest
           evict-old-files: 1d
@@ -101,16 +101,6 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: webgpu-macos-latest
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   ubuntu:
     runs-on: ubuntu-24.04
 
@@ -120,7 +110,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: webgpu-ubuntu-24.04
           evict-old-files: 1d
@@ -163,13 +153,3 @@ jobs:
           # This is using llvmpipe and runs slower than other backends
           # test-backend-ops is too slow on llvmpipe, skip it
           ctest -L main -E test-backend-ops --verbose --timeout 900
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: webgpu-ubuntu-24.04
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/copilot-setup-steps.yml

@@ -29,7 +29,7 @@ jobs:
         uses: actions/checkout@v6
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: copilot-setup-steps
           evict-old-files: 1d

.github/workflows/docker.yml

@@ -64,7 +64,7 @@ jobs:
     needs: create_tag
     uses: ./.github/workflows/ui-build.yml
     with:
-      ui_version: ${{ needs.create_tag.outputs.source_tag }}
+      hf_ui_version: ${{ needs.create_tag.outputs.source_tag }}
 
   prepare_matrices:
     name: Prepare Docker matrices
@@ -162,7 +162,7 @@ jobs:
         if: ${{ matrix.config.prebuilt_ui == true }}
         uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist
 
       - name: Set up QEMU

.github/workflows/hip-quality-check.yml

@@ -52,7 +52,7 @@ jobs:
           sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev python3
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: hip-quality-check-ubuntu-22.04
           evict-old-files: 1d
@@ -84,13 +84,3 @@ jobs:
           cd build
           make -j $(nproc) 2>&1 | tee metrics.log | grep -v 'Rpass-analysis=kernel-resource-usage\|remark:\|^$'
           python3 ../scripts/hip/gcn-cdna-vgpr-check.py metrics.log
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: hip-quality-check-ubuntu-22.04
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/make-release.yml

@@ -84,13 +84,11 @@ jobs:
 
             New version has been released.
 
-            ## Assets
-
             ${{ steps.desc.outputs.nightly }}
 
-            ## More info
+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
 
-            - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
 
             ## ${{ steps.desc.outputs.changelog_title }}
 

.github/workflows/release.yml

@@ -61,8 +61,31 @@ jobs:
             echo "should_release=false" >> $GITHUB_OUTPUT
           fi
 
+  get-version:
+    runs-on: ubuntu-slim
+    outputs:
+      ui_version: ${{ steps.version.outputs.ui_version }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+      - id: version
+        run: |
+          # Resolve UI version: BUILD_NUMBER from cmake/build-info.cmake > git hash + epoch > fallback
+          version=""
+          if grep -q "BUILD_NUMBER" cmake/build-info.cmake; then
+            build_number=$(grep "set(BUILD_NUMBER" cmake/build-info.cmake | grep -oP '\d+')
+            if [ -n "$build_number" ] && [ "$build_number" -gt 0 ]; then
+              version="b${build_number}"
+            fi
+          fi
+          if [ -z "$version" ]; then
+            version=$(git rev-parse --short HEAD)-$(date +%s)
+          fi
+          echo "ui_version=${version}" >> $GITHUB_OUTPUT
+
   macos-cpu:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -96,14 +119,15 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-${{ matrix.arch }}
 
@@ -117,6 +141,7 @@ jobs:
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
@@ -142,7 +167,7 @@ jobs:
           key: release-${{ matrix.os }}-${{ matrix.arch }}
 
   ubuntu-cpu:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     strategy:
       matrix:
@@ -166,11 +191,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Dependencies
         id: depends
@@ -187,7 +213,7 @@ jobs:
 
       - name: ccache
         if: ${{ matrix.build != 's390x' }}
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-cpu
 
@@ -201,6 +227,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -227,7 +254,7 @@ jobs:
           key: release-${{ matrix.os }}-cpu
 
   ubuntu-vulkan:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     strategy:
@@ -250,11 +277,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Dependencies
         id: depends
@@ -272,7 +300,7 @@ jobs:
           fi
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-${{ matrix.os }}-vulkan
 
@@ -286,6 +314,7 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -311,7 +340,7 @@ jobs:
           key: release-${{ matrix.os }}-vulkan
 
   android-arm64:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-latest
@@ -329,11 +358,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Set up JDK
         uses: actions/setup-java@v5
@@ -358,7 +388,7 @@ jobs:
       #   cache on:  https://github.com/ggerganov/tmp2/actions/runs/26534713799/job/78224189394
       #
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: release-android-arm64
 
@@ -377,6 +407,7 @@ jobs:
             -DLLAMA_FATAL_WARNINGS=ON \
             -DGGML_OPENMP=OFF \
             -DLLAMA_BUILD_BORINGSSL=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -402,7 +433,7 @@ jobs:
           name: llama-bin-android-arm64.tar.gz
 
   ubuntu-24-openvino:
-    needs: [check-release, ui-build]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: ubuntu-24.04
@@ -415,8 +446,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output
@@ -429,14 +460,15 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-ubuntu-24.04-openvino-release-no-preset-v1
 
@@ -476,6 +508,7 @@ jobs:
             -DGGML_OPENVINO=ON \
             -DCMAKE_INSTALL_RPATH='$ORIGIN' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build/ReleaseOV --config Release --parallel
 
@@ -519,7 +552,7 @@ jobs:
           key: release-ubuntu-24.04-openvino-release-no-preset-v1
 
   windows-openvino:
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2022
@@ -529,8 +562,8 @@ jobs:
 
     env:
       # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.3.1"
-      OPENVINO_VERSION_FULL: "2026.3.1.22476.56d9685302d"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"
 
     steps:
       - name: Set OpenVINO version output
@@ -544,14 +577,15 @@ jobs:
         with:
             fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-openvino
           variant: ccache
@@ -648,7 +682,7 @@ jobs:
 
   windows-cpu:
     name: windows-cpu / ${{ matrix.arch }}
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2025-vs2026
@@ -668,18 +702,19 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Install Ninja
         run: |
           choco install ninja
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
@@ -714,8 +749,6 @@ jobs:
         with:
           key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
-  # note: builds only the ggml-hip backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-rocm:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -736,12 +769,8 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Install Ninja
-        run: |
-          choco install ninja
-
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
@@ -793,28 +822,33 @@ jobs:
 
       - name: Build
         run: |
-          cmake -S . -B build `
-            -G "Ninja Multi-Config" `
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
             -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
             -DGGML_BACKEND_DL=ON `
             -DGGML_NATIVE=OFF `
-            -DGGML_CPU=OFF `
+            -DGGML_CPU=ON `
+            -DGGML_CPU_ALL_VARIANTS=ON `
             -DGGML_HIP=ON `
             -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
             -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
             -DCMAKE_C_FLAGS="-Wno-error=incompatible-pointer-types" `
             -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
             -DHIP_PATH="${env:HIP_PATH}" `
+            -DGGML_HIP_ROCWMMA_FATTN=ON `
             -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
-          cmake --build build --config Release --parallel ${env:NUMBER_OF_PROCESSORS} --target ggml-hip
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
 
       - name: Verify HIP backend was built
         run: |
-          $hipDll = Get-ChildItem -Path build\bin\Release -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
+          $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
           if (-not $hipDll) {
             Write-Host "##[error]ggml-hip*.dll was NOT produced. The HIP backend silently failed to build."
-            Write-Host "Contents of build\bin\Release:"
-            Get-ChildItem build\bin\Release | Format-Table -AutoSize
+            Write-Host "Contents of build\bin:"
+            Get-ChildItem build\bin | Format-Table -AutoSize
             exit 1
           }
           Write-Host "HIP backend artifact found:"
@@ -829,40 +863,10 @@ jobs:
           $rocmVersionShort = ('${{ matrix.ROCM_VERSION }}'.Split('.')[0..1] -join '.')
           echo "ROCM_VERSION_SHORT=$rocmVersionShort" >> $env:GITHUB_ENV
 
-      - name: Bundle HIP runtime DLLs (amdhip64_7.dll, rocm_kpack.dll, amd_comgr.dll)
-        run: |
-          $ErrorActionPreference = "Stop"
-          # See issue https://github.com/ggml-org/llama.cpp/issues/26929.
-          # ggml-hip.dll loads amdhip64_7.dll at run time. The Adrenalin driver
-          # ships an amdhip64_7.dll in System32, which the loader searches before PATH,
-          # so a matching DLL from PATH cannot win. Copy amdhip64 next to the
-          # binaries (exe directory is searched before System32) so the correct
-          # runtime is used. rocm_kpack.dll is amdhip64_7's direct dependency, so
-          # copy the matching version too. amd_comgr is copied as well to keep it
-          # in sync with the bundled amdhip64, avoiding a version mismatch with a
-          # amd_comgr from System32.
-          # rocblas/hipblaslt kernels resolve fine via PATH and are not copied.
-          $binPath = (rocm-sdk path --bin).Trim()
-          if (-not $binPath) { throw "rocm-sdk path --bin returned empty" }
-          write-host "ROCm bin path: $binPath"
-
-          $patterns = @("amdhip64_7.dll", "rocm_kpack.dll", "amd_comgr.dll")
-          foreach ($pattern in $patterns) {
-            $files = Get-ChildItem -Path $binPath -Filter $pattern -ErrorAction SilentlyContinue
-            if (-not $files) { throw "no match for $pattern in $binPath" }
-            foreach ($f in $files) {
-              Copy-Item $f.FullName -Destination build\bin\Release -Force
-              write-host "  copied $($f.Name)"
-            }
-          }
-
       - name: Pack artifacts
         run: |
-          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip `
-            .\build\bin\Release\ggml-hip.dll `
-            .\build\bin\Release\amdhip64_7.dll `
-            .\build\bin\Release\rocm_kpack.dll `
-            .\build\bin\Release\amd_comgr.dll
+          cp "LICENSE" "build\bin\"
+          7z a -snl llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip .\build\bin\*
 
       - name: Upload artifacts
         uses: actions/upload-artifact@v6
@@ -875,8 +879,6 @@ jobs:
         with:
           key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
 
-  # note: builds only the backend library - llama-server (with the embedded UI)
-  #       is injected from the windows-cpu zip during the release "Merge artifacts" step
   windows:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -907,6 +909,13 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Install Vulkan SDK
         id: get_vulkan
         if: ${{ matrix.backend == 'vulkan' }}
@@ -923,7 +932,7 @@ jobs:
 
       # TODO: these jobs need to use llvm toolchain in order to utilize the ccache
       #- name: ccache
-      #  uses: ggml-org/ccache-action@v1.2.24
+      #  uses: ggml-org/ccache-action@v1.2.21
       #  with:
       #    key: release-windows-2025-${{ matrix.arch }}-${{ matrix.backend }}
 
@@ -969,8 +978,6 @@ jobs:
           path: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
-  # note: builds only the ggml-cuda backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-cuda:
     name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
     needs: [check-release]
@@ -999,6 +1006,13 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Install Cuda Toolkit
         uses: ./.github/actions/windows-setup-cuda
         with:
@@ -1011,7 +1025,7 @@ jobs:
           choco install ninja
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
@@ -1070,8 +1084,6 @@ jobs:
         with:
           key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
 
-  # note: builds only the ggml-sycl backend - llama-server is injected from the
-  #       windows-cpu zip during the release "Merge artifacts" step
   windows-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1106,8 +1118,15 @@ jobs:
           Expand-Archive -Path "level-zero-win-sdk.zip" -DestinationPath "C:/level-zero-sdk" -Force
           "LEVEL_ZERO_V1_SDK_PATH=C:/level-zero-sdk" | Out-File -FilePath $env:GITHUB_ENV -Append
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-windows-2022-x64-sycl
 
@@ -1176,7 +1195,7 @@ jobs:
           key: release-windows-2022-x64-sycl
 
   ubuntu-24-sycl:
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     strategy:
@@ -1218,14 +1237,15 @@ jobs:
           wget -q "https://github.com/oneapi-src/level-zero/releases/download/v${LEVEL_ZERO_VERSION}/level-zero-devel_${LEVEL_ZERO_VERSION}%2B${LEVEL_ZERO_UBUNTU_VERSION}_amd64.deb" -O level-zero-devel.deb
           sudo apt-get install -y ./level-zero.deb ./level-zero-devel.deb
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: release-ubuntu-24.04-sycl-${{ matrix.build }}
 
@@ -1267,11 +1287,11 @@ jobs:
         with:
           key: release-ubuntu-24.04-sycl-${{ matrix.build }}
 
-  ubuntu-24-rocm:
-    needs: [check-release, ui-build]
+  ubuntu-22-rocm:
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
-    runs-on: ubuntu-24.04
+    runs-on: ubuntu-22.04
 
     permissions:
       actions: write
@@ -1290,11 +1310,12 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
         with:
-          name: llama-ui.zip
-          path: tools/ui/dist
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
 
       - name: Free up disk space
         uses: ggml-org/free-disk-space@v1.3.1
@@ -1302,9 +1323,9 @@ jobs:
           tool-cache: true
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
-          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
           max-size: "1G"
 
@@ -1367,6 +1388,7 @@ jobs:
             -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DGGML_HIP=ON \
             -DHIP_PLATFORM=amd \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }} \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -1392,10 +1414,10 @@ jobs:
       - name: ccache-clear
         uses: ./.github/actions/ccache-clear
         with:
-          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
 
   ios-xcode:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     runs-on: macos-26
 
@@ -1423,7 +1445,8 @@ jobs:
             -DLLAMA_BUILD_SERVER=OFF \
             -DCMAKE_SYSTEM_NAME=iOS \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=16.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml \
+            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
 
       - name: xcodebuild for swift package
@@ -1546,9 +1569,11 @@ jobs:
 #          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz
 
   ui-build:
-    needs: [check-release]
+    needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
     uses: ./.github/workflows/ui-build.yml
+    with:
+      hf_ui_version: ${{ needs.get-version.outputs.ui_version }}
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -1563,13 +1588,14 @@ jobs:
     runs-on: ubuntu-slim
 
     needs:
+      - get-version
       - windows
       - windows-cpu
       - windows-cuda
       - windows-sycl
       - windows-rocm
       - windows-openvino
-      - ubuntu-24-rocm
+      - ubuntu-22-rocm
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
@@ -1602,27 +1628,24 @@ jobs:
           path: ./artifact
           merge-multiple: true
 
-      - name: Merge artifacts
+      - name: Move artifacts
         id: move_artifacts
         run: |
           mkdir -p release
 
-          # the windows-cpu zip contains the full toolset (llama-server with the embedded
-          # UI, ggml-cpu) - inject it into the other windows zips so that every archive
-          # ships the same binaries, only with a different backend library on top
-          echo "Injecting windows-cpu binaries (llama-server + CPU backend) into the backend zips..."
+          echo "Adding CPU backend files to existing zips..."
           for arch in x64 arm64; do
             cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
             temp_dir=$(mktemp -d)
-            echo "Extracting windows-cpu-${arch} package..."
+            echo "Extracting CPU backend for $arch..."
             unzip "$cpu_zip" -d "$temp_dir"
 
-            echo "Merging into $arch zips..."
+            echo "Adding CPU files to $arch zips..."
             for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
               if [[ "$target_zip" == "$cpu_zip" ]]; then
                 continue
               fi
-              echo "Injecting into $(basename "$target_zip")"
+              echo "Adding CPU backend to $(basename "$target_zip")"
               realpath_target_zip=$(realpath "$target_zip")
               (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
             done
@@ -1646,7 +1669,7 @@ jobs:
         id: download_ui
         uses: actions/download-artifact@v7
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: ./ui-dist
 
       - name: Package UI

.github/workflows/server-sanitize.yml

@@ -73,6 +73,13 @@ jobs:
           fetch-depth: 0
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
+      - name: Setup Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: "24"
+          cache: "npm"
+          cache-dependency-path: "tools/ui/package-lock.json"
+
       - name: Build
         id: cmake_build
         run: |

.github/workflows/server.yml

@@ -80,7 +80,7 @@ jobs:
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: server-ubuntu-24.04-arm
           evict-old-files: 1d
@@ -128,16 +128,6 @@ jobs:
           export LLAMA_ARG_BACKEND_SAMPLING=1
           SLOW_TESTS=1 ./tests.sh
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: server-ubuntu-24.04-arm
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
-
   windows:
     runs-on: windows-2025
 
@@ -150,7 +140,7 @@ jobs:
           ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
 
       - name: ccache
-        uses: ggml-org/ccache-action@v1.2.24
+        uses: ggml-org/ccache-action@v1.2.21
         with:
           key: server-windows-2025-x64
           evict-old-files: 1d
@@ -191,13 +181,3 @@ jobs:
           cd tools/server/tests
           export SLOW_TESTS="1"
           ./tests.sh
-
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        env:
-          GH_TOKEN: ${{ github.token }}
-        with:
-          key: server-windows-2025-x64
-          older: 5m
-          min: 1
-          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

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

@@ -31,6 +31,6 @@ jobs:
       - name: Upload built UI
         uses: actions/upload-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
           retention-days: 1

.github/workflows/ui-build.yml

@@ -3,8 +3,8 @@ name: UI Build
 on:
   workflow_call:
     inputs:
-      ui_version:
-        description: 'Version string embedded in build.json (e.g. b1234); defaults to b<commit-count>'
+      hf_ui_version:
+        description: 'Version string for version.json (e.g. 12345)'
         required: false
         type: string
 
@@ -17,17 +17,6 @@ jobs:
     steps:
       - name: Checkout code
         uses: actions/checkout@v6
-        with:
-          fetch-depth: 0
-
-      - name: Resolve UI version
-        id: version
-        run: |
-          version="${{ inputs.ui_version }}"
-          if [ -z "$version" ]; then
-            version="b$(git rev-list --count HEAD)"
-          fi
-          echo "ui_version=${version}" >> $GITHUB_OUTPUT
 
       - name: Setup Node.js
         uses: actions/setup-node@v6
@@ -42,7 +31,8 @@ jobs:
 
       - name: Build application
         env:
-          LLAMA_BUILD_NUMBER: ${{ steps.version.outputs.ui_version }}
+          HF_UI_VERSION: ${{ inputs.hf_ui_version || '' }}
+          LLAMA_BUILD_NUMBER: ${{ inputs.hf_ui_version || 'b0000' }}
         run: npm run build
         working-directory: tools/ui
 
@@ -53,6 +43,6 @@ jobs:
       - name: Upload built UI
         uses: actions/upload-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
           retention-days: 1

.github/workflows/ui-publish.yml

@@ -37,7 +37,7 @@ jobs:
       - name: Download UI build artifact
         uses: actions/download-artifact@v7
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Create distribution archive

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

@@ -64,7 +64,7 @@ jobs:
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Run type checking
@@ -106,7 +106,7 @@ jobs:
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Build Storybook

.github/workflows/ui.yml

@@ -63,7 +63,7 @@ jobs:
       - name: Download built UI artifacts
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Install dependencies
@@ -126,7 +126,7 @@ jobs:
       - name: Download built UI artifacts (reuses ui-build)
         uses: actions/download-artifact@v6
         with:
-          name: llama-ui.zip
+          name: ui-build
           path: tools/ui/dist/
 
       - name: Install Playwright browsers

CMakeLists.txt

@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)
 
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 3)
+set(LLAMA_VERSION_MINOR 2)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
 
@@ -134,8 +134,8 @@ option(LLAMA_BUILD_TOOLS     "llama: build tools"
 option(LLAMA_BUILD_EXAMPLES  "llama: build examples"                                                             ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER    "llama: build server example"                                                       ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_APP       "llama: build the unified binary"                                                   ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       OFF)
-option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available" ON)
+option(LLAMA_BUILD_UI        "llama: build the embedded Web UI for server"                                       ON)
+option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
 
 option(LLAMA_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
 option(LLAMA_TESTS_INSTALL "llama: install tests" ON)

CODEOWNERS

@@ -57,7 +57,6 @@
 /ggml/src/ggml-cann/                    @ggml-org/ggml-cann
 /ggml/src/ggml-common.h                 @ggerganov
 /ggml/src/ggml-cpu/                     @ggerganov
-/ggml/src/ggml-cpu/iqp.*                @bartowski1182
 /ggml/src/ggml-cpu/spacemit/            @alex-spacemit
 /ggml/src/ggml-cuda/                    @ggml-org/ggml-cuda
 /ggml/src/ggml-cuda/vendors/hip.h       @IMbackK

CONTRIBUTING.md

@@ -74,7 +74,6 @@ For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 - If a PR does not warrant a new release, add `[no release]` in the squashed commit to spare CI resources
 - Be mindful of maintenance: most of the work going into a feature happens after the PR is merged. If the PR author is not committed to contribute long-term, someone else needs to take responsibility (you)
 - Add the ["merge ready"](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+is%3Aopen+draft%3Ano+sort%3Aupdated-desc+label%3A%22merge+ready%22+) label to a PR to indicate when a PR can be fast-merged without waiting for 2 independent reviews. [(more info)](https://github.com/ggml-org/llama.cpp/pull/26178)
-- Wait for CI results before merging
 
 Maintainers reserve the right to decline review or close pull requests for any reason, without any questions, particularly under any of the following conditions:
 - The proposed change is already mentioned in the roadmap or an existing issue, and it has been assigned to someone.

README.md

@@ -13,7 +13,7 @@
 [![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)
 
-[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
 
 </div>
 

build-xcframework.sh

@@ -18,7 +18,7 @@ LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
-GGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY:-ON}
+GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
 GGML_OPENMP=OFF
 
@@ -169,14 +169,6 @@ setup_framework_structure() {
     cp tools/mtmd/mtmd.h           ${header_path}
     cp tools/mtmd/mtmd-helper.h    ${header_path}
 
-    if [[ "$GGML_METAL_EMBED_LIBRARY" == "OFF" ]]; then
-        if [[ "$platform" == "macos" ]]; then
-            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/
-        else
-            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/
-        fi
-    fi
-
     # Create module map (common for all platforms)
     cat > ${module_path}module.modulemap << EOF
 framework module llama {
@@ -458,7 +450,6 @@ build_ios_sim() {
         -DIOS=ON \
         -DCMAKE_SYSTEM_NAME=iOS \
         -DCMAKE_OSX_SYSROOT=iphonesimulator \
-        -DGGML_METAL_TARGET_OS=ios \
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -476,7 +467,6 @@ build_ios_device() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=iOS \
         -DCMAKE_OSX_SYSROOT=iphoneos \
-        -DGGML_METAL_TARGET_OS=ios \
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -508,7 +498,6 @@ build_visionos() {
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DCMAKE_SYSTEM_NAME=visionOS \
         -DCMAKE_OSX_SYSROOT=xros \
-        -DGGML_METAL_TARGET_OS=xros \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
         -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -527,7 +516,6 @@ build_visionos_sim() {
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DCMAKE_SYSTEM_NAME=visionOS \
         -DCMAKE_OSX_SYSROOT=xrsimulator \
-        -DGGML_METAL_TARGET_OS=xros \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
         -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
         -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -546,7 +534,6 @@ build_tvos_sim() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=tvOS \
         -DCMAKE_OSX_SYSROOT=appletvsimulator \
-        -DGGML_METAL_TARGET_OS=tvos \
         -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
         -DGGML_METAL=ON \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
@@ -565,7 +552,6 @@ build_tvos_device() {
         -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
         -DCMAKE_SYSTEM_NAME=tvOS \
         -DCMAKE_OSX_SYSROOT=appletvos \
-        -DGGML_METAL_TARGET_OS=tvos \
         -DCMAKE_OSX_ARCHITECTURES="arm64" \
         -DGGML_METAL=ON \
         -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \

ci/run.sh

@@ -189,8 +189,8 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
     fi
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"
 
-    # TODO: fix and re-enable the `test-llama-archs` and `test-recurrent-state-rollback*`
-    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-rollback"
+    # TODO: fix and re-enable the `test-llama-archs` test below
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi
 
 ## helpers
@@ -300,40 +300,6 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
-# test_llama_archs_tensor_split
-
-function gg_run_test_llama_archs_tensor_split {
-    cd ${SRC}
-
-    set -e
-
-    if [ ! -z ${GG_BUILD_CUDA} ]; then
-        GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-    fi
-
-    if [ ! -z ${GG_BUILD_METAL} ]; then
-        GGML_METAL_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-        GGML_METAL_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
-    fi
-
-    set +e
-}
-
-function gg_sum_test_llama_archs_tensor_split {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs test-llama-archs with 1 to 4 devices\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}.log)"
-    gg_printf '```\n'
-}
-
 # test_scripts
 
 function gg_run_test_scripts {
@@ -732,11 +698,6 @@ function gg_check_build_requirements {
         gg_printf 'ctest not found, please install\n'
         exit 1
     fi
-
-    if ! command -v unzip &> /dev/null; then
-        gg_printf 'unzip not found, please install\n'
-        exit 1
-    fi
 }
 
 function gg_run_test_backend_ops_cpu {
@@ -790,8 +751,6 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
-test $ret -eq 0 && gg_run test_llama_archs_tensor_split
-
 if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
     test $ret -eq 0 && gg_run test_backend_ops_cpu
 fi

common/arg.cpp

@@ -1643,14 +1643,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_CTX_SIZE"));
-    add_opt(common_arg(
-        { "--kv-unified-per-slot" }, "N",
-        "context limit per parallel slot (default: unset, behavior unchanged).\n"
-        "when set without -c/--ctx-size, the shared KV pool is sized to n_parallel*N",
-        [](common_params & params, int value) {
-            params.kv_unified_per_slot = value;
-        }
-    ).set_env("LLAMA_ARG_KV_UNIFIED_PER_SLOT").set_examples({ LLAMA_EXAMPLE_SERVER }));
     add_opt(common_arg(
         {"-n", "--predict", "--n-predict"}, "N",
         string_format(
@@ -2652,27 +2644,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mtmd_batch_max_tokens = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    add_opt(common_arg(
-        {"--video-fps"}, "N",
-        string_format("target video frame rate (default: %.1f)", params.video_fps),
-        [](common_params & params, const std::string & value) {
-            params.video_fps = std::stof(value);
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FPS"));
-    add_opt(common_arg(
-        {"--video-timestamp-interval"}, "N",
-        string_format("interval in milliseconds between text timestamps (default: %" PRId64 ")", params.video_timestamp_interval_ms),
-        [](common_params & params, int value) {
-            params.video_timestamp_interval_ms = value;
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_TIMESTAMP_INTERVAL"));
-    add_opt(common_arg(
-        {"--video-ffmpeg-dir"}, "DIR",
-        "path to the directory containing ffmpeg and ffprobe (default: search in PATH)",
-        [](common_params & params, const std::string & value) {
-            params.video_ffmpeg_bin_dir = value;
-        }
-    ).set_examples(mmproj_examples).set_env("LLAMA_ARG_VIDEO_FFMPEG_DIR"));
     if (params.is_gen_docs || llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
@@ -2728,19 +2699,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else { throw std::invalid_argument("invalid value"); }
         }
     ).set_env("LLAMA_ARG_LOAD_MODE"));
-    add_opt(common_arg(
-        {"-lzm", "--lazy-mode"}, "MODE",
-        "on-demand reading of certain tensors, for example per-layer embeddings (default: auto)\n"
-        "- on: read the rows of such tensors from disk on demand instead of keeping them resident (requires mmap)\n"
-        "- auto: on, but only for tensors larger than 4 GiB\n"
-        "- off: always keep them resident",
-        [](common_params & params, const std::string & value) {
-            /**/ if (value == "on")   { params.lazy_mode = LLAMA_LAZY_MODE_ON;   }
-            else if (value == "auto") { params.lazy_mode = LLAMA_LAZY_MODE_AUTO; }
-            else if (value == "off")  { params.lazy_mode = LLAMA_LAZY_MODE_OFF;  }
-            else { throw std::invalid_argument("invalid value"); }
-        }
-    ).set_env("LLAMA_ARG_LAZY_MODE"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"
@@ -2792,20 +2750,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             if (value < 0) {
                 throw std::invalid_argument("invalid value");
             }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.tensor_buft_overrides);
+            for (int i = 0; i < value; ++i) {
+                // keep strings alive and avoid leaking memory by storing them in a static vector
+                static std::list<std::string> buft_overrides;
+                buft_overrides.push_back(llm_ffn_exps_block_regex(i));
+                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
         }
     ).set_env("LLAMA_ARG_N_CPU_MOE"));
-    add_opt(common_arg(
-        {"-ncffn", "--n-cpu-ffn"}, "N",
-        "keep the dense FFN weights of the first N layers in the CPU\n"
-        "(dense models; for MoE expert weights use --n-cpu-moe)",
-        [](common_params & params, int value) {
-            if (value < 0) {
-                throw std::invalid_argument("invalid value");
-            }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_DENSE_REGEX, params.tensor_buft_overrides);
-        }
-    ).set_env("LLAMA_ARG_N_CPU_FFN"));
     GGML_ASSERT(params.n_gpu_layers < 0); // string_format would need to be extended for a default >= 0
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
@@ -4132,7 +4084,11 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             if (value < 0) {
                 throw std::invalid_argument("invalid value");
             }
-            llm_add_n_cpu_ffn_overrides(value, LLM_FFN_EXPS_REGEX, params.speculative.draft.tensor_buft_overrides);
+            for (int i = 0; i < value; ++i) {
+                static std::list<std::string> buft_overrides_draft;
+                buft_overrides_draft.push_back(llm_ffn_exps_block_regex(i));
+                params.speculative.draft.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_CPU_MOE"));
 
@@ -4153,38 +4109,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.draft.n_min = value;
         }
     ).set_spec().set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_LOOKUP, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_DRAFT_N_MIN"));
-    add_opt(common_arg(
-        {"--spec-synth-len"}, "L",
-        "target mean synthetic acceptance length, including the target token (benchmarking only)",
-        [](common_params & params, const std::string & value) {
-            const std::string text = string_strip(value);
-            size_t pos = 0;
-            const double length = std::stod(text, &pos);
-            if (pos != text.size() || length == -1.0) {
-                throw std::invalid_argument("invalid value");
-            }
-            params.speculative.synth_len = length;
-        }
-    ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_LEN"));
-    add_opt(common_arg(
-        {"--spec-synth-rates"}, "P0,P1,...",
-        "comma-separated unconditional per-position synthetic acceptance probabilities (benchmarking only)",
-        [](common_params & params, const std::string & value) {
-            const auto values = string_split<std::string>(value, ',');
-            std::vector<double> rates;
-            rates.reserve(values.size());
-            for (const auto & raw : values) {
-                const std::string text = string_strip(raw);
-                size_t pos = 0;
-                const double rate = std::stod(text, &pos);
-                if (pos != text.size()) {
-                    throw std::invalid_argument("invalid value");
-                }
-                rates.push_back(rate);
-            }
-            params.speculative.synth_rates = std::move(rates);
-        }
-    ).set_spec().set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_SPEC_SYNTH_RATES"));
 
     add_opt(common_arg(
         {"--spec-draft-p-split", "--draft-p-split"}, "P",

common/chat.cpp

@@ -1177,8 +1177,6 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
         "</tool_call>",
     };
 
-    auto is_qwen3_coder  = !supports_reasoning;
-
     if (supports_reasoning) {
         data.thinking_start_tag = "<think>";
         // Support both </think> and <tool_call> as reasoning end sequences.
@@ -1219,15 +1217,13 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
 
     std::vector<std::string> tool_call_starts = { "<tool_call>" };
 
-    if (is_qwen3_coder) {
-        // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
-        // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
-        // constraining on <function which may occur in valid content generation, e.g. #include <functional>
-        foreach_function(inputs.tools, [&](const json & tool) {
-            const std::string name = tool.at("function").at("name");
-            tool_call_starts.push_back("<function=" + name + ">");
-        });
-    }
+    // Match complete <function=name> opener for Qwen3-Coder models that occasionally omit the
+    // starting <tool_call>. The model may hallucinate a tool name, but it is preferable over
+    // constraining on <function which may occur in valid content generation, e.g. #include <functional>
+    foreach_function(inputs.tools, [&](const json & tool) {
+        const std::string name = tool.at("function").at("name");
+        tool_call_starts.push_back("<function=" + name + ">");
+    });
 
     auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
         auto generation_prompt = p.literal(GEN_PREFIX);
@@ -1292,13 +1288,10 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_
 
             auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
 
-            auto tool_call_body = tool_choice + "</tool_call>" + p.space();
-            auto tool_call      = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
-
             // Qwen3-Coder models may occasionally omit the <tool_call> token.
-            auto tool_call_first = is_qwen3_coder ?
-                p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body) :
-                tool_call;
+            auto tool_call_body  = tool_choice + "</tool_call>" + p.space();
+            auto tool_call_first = p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body);
+            auto tool_call       = p.rule("tool-call", "<tool_call>\n" + tool_call_body);
 
             auto calls      = inputs.parallel_tool_calls ? tool_call_first + p.zero_or_more(tool_call) : tool_call_first;
             auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(calls, min_calls, 1));

common/common.cpp

@@ -402,11 +402,10 @@ void common_params_print_info(const common_params & params, bool print_devices)
 #endif
     COM_TRC("%s: build %d (%s) with %s for %s%s\n", __func__, llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 
-    const int verbosity = common_log_get_verbosity_thold();
-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, verbosity);
+    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
 
     // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
-    if (print_devices && verbosity >= LOG_LEVEL_TRACE) {
+    if (print_devices) {
         COM_TRC("%s", "device_info:\n");
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             auto * dev = ggml_backend_dev_get(i);
@@ -1295,34 +1294,11 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     if (params.fit_params) {
         COM_TRC("%s", "fitting params to device memory ...\n");
         COM_TRC("%s", "(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");
-
-        // the draft context is created from the same base params and follows the main context, fit both together
-        const bool has_draft = params.speculative.has_dft();
-        const bool spec_mtp  = std::find(params.speculative.types.begin(), params.speculative.types.end(),
-            COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
-
-        common_params params_dft = common_base_params_to_speculative(params);
-
-        auto mparams_dft = common_model_params_to_llama(params_dft);
-        auto cparams_dft = common_context_params_to_llama(params_dft);
-        if (spec_mtp) {
-            cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
-        }
-        cparams_dft.n_rs_seq = 0;
-
-        const common_fit_extra_model extra = {
-            /*.path_model   =*/ params_dft.model.path.c_str(),
-            /*.mparams      =*/ &mparams_dft,
-            /*.cparams      =*/ &cparams_dft,
-            /*.shares_model =*/ !has_draft, // an MTP context runs on the weights of the main model
-        };
-
         common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
-            has_draft || spec_mtp ? &extra : nullptr,
             params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1688,7 +1664,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.main_gpu        = params.main_gpu;
     mparams.split_mode      = params.split_mode;
     mparams.load_mode       = params.load_mode;
-    mparams.lazy_mode = params.lazy_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

common/common.h

@@ -8,7 +8,6 @@
 #include "ggml.h"
 #include "llama.h"
 
-#include <list>
 #include <set>
 #include <sstream>
 #include <string>
@@ -370,9 +369,6 @@ struct common_params_speculative_ngram_cache {
 struct common_params_speculative {
     std::vector<enum common_speculative_type> types = { COMMON_SPECULATIVE_TYPE_NONE };
 
-    double synth_len = -1.0;
-    std::vector<double> synth_rates;
-
     // used by Simple, MTP, Eagle3, etc. - all methods that require some kind of draft model
     common_params_speculative_draft draft;
 
@@ -387,10 +383,6 @@ struct common_params_speculative {
         return !draft.mparams.empty();
     }
 
-    bool has_synth() const {
-        return synth_len != -1.0 || !synth_rates.empty();
-    }
-
     uint32_t need_n_rs_seq() const {
         bool needs_rs_seq = std::any_of(types.begin(), types.end(), [&](auto t) {
             return t == COMMON_SPECULATIVE_TYPE_DRAFT_MTP || t == COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3 || t == COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH || t == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK;
@@ -483,8 +475,6 @@ struct common_params {
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_load_mode  load_mode  = LLAMA_LOAD_MODE_AUTO; // how to load the model
 
-    enum llama_lazy_mode lazy_mode = LLAMA_LAZY_MODE_AUTO; // on-demand reading of tensors marked by the arch
-
     common_cpu_params cpuparams;
     common_cpu_params cpuparams_batch;
 
@@ -599,11 +589,6 @@ struct common_params {
     int image_max_tokens = -1;
     int mtmd_batch_max_tokens = 1024;
 
-    // for video input
-    float       video_fps                   = 4.0f;
-    int64_t     video_timestamp_interval_ms = 5000;
-    std::string video_ffmpeg_bin_dir        = "";
-
     // finetune
     struct lr_opt lr;
     enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
@@ -627,7 +612,6 @@ struct common_params {
     bool    cache_prompt        = true;  // whether to enable prompt caching
     bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
-    int32_t kv_unified_per_slot = 0;     // max context per parallel slot; 0 = unset
     int32_t checkpoint_min_step = 8192;  // minimum spacing between context checkpoints
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 
@@ -1124,30 +1108,19 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 }
 
 //
-// FFN offload utils
+// MoE utils
 //
 
 const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";
 
-const char * const LLM_FFN_DENSE_REGEX = "\\.ffn_(up|down|gate)\\.";
-
-inline std::string llm_ffn_block_regex(int idx, const char * ffn_regex) {
-    return string_format("blk\\.%d%s", idx, ffn_regex);
+inline std::string llm_ffn_exps_block_regex(int idx) {
+    return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);
 }
 
 inline llama_model_tensor_buft_override llm_ffn_exps_cpu_override() {
     return { LLM_FFN_EXPS_REGEX, ggml_backend_cpu_buffer_type() };
 }
 
-inline void llm_add_n_cpu_ffn_overrides(int n, const char * ffn_regex, std::vector<llama_model_tensor_buft_override> & overrides) {
-    // keep strings alive and avoid leaking memory by storing them in a static list
-    static std::list<std::string> buft_override_strings;
-    for (int i = 0; i < n; ++i) {
-        buft_override_strings.push_back(llm_ffn_block_regex(i, ffn_regex));
-        overrides.push_back({buft_override_strings.back().c_str(), ggml_backend_cpu_buffer_type()});
-    }
-}
-
 //
 // training utils
 //

common/fit.cpp

@@ -178,7 +178,7 @@ common_device_memory_data_vec common_get_device_memory_data(
 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, const common_fit_extra_model * extra, enum ggml_log_level log_level) {
+        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");
     }
@@ -191,92 +191,10 @@ static void common_params_fit_impl(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
-    // with non-unified kv, we need to take into account n_streams
-    // for example, if memory can hold more than model's trained context size, we must extend the n_ctx to hold enough n_streams
-    const uint32_t n_streams  = cparams->kv_unified ? 1 : std::max<uint32_t>(1, cparams->n_seq_max);
-    const bool     n_ctx_auto = cparams->n_ctx == 0;
-
-    dmds_t   dmds_extra;       // memory of the extra model, laid out on the devices of the main model
-    uint32_t n_ctx_extra = 0;  // context that memory was measured at
-
-    // the extra model competes for the same memory as the main model, add it to every measurement
-    // its memory is measured again whenever the context it follows changes
-    auto add_extra_memory = [&](dmds_t & dmds) {
-        if (extra == nullptr) {
-            return;
-        }
-
-        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
-            std::vector<ggml_backend_dev_t> devs_extra;
-            uint32_t ngl_extra = 0;
-            uint32_t nct_extra = 0;
-            uint32_t nex_extra = 0;
-
-            extra->cparams->n_ctx = cparams->n_ctx;
-
-            LOG_TRC("%s: getting device memory data for the extra model at a context size of %" PRIu32 ":\n",
-                __func__, cparams->n_ctx);
-
-            dmds_t measured;
-            try {
-                measured = common_get_device_memory_data_impl(
-                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
-            } catch (const std::runtime_error & e) {
-                // the extra model is optional, fit the main model alone rather than giving up
-                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
-                dmds_extra = dmds_t(devs.size() + 1);
-                n_ctx_extra = cparams->n_ctx;
-                return;
-            }
-
-            dmds_extra = dmds_t(devs.size() + 1);
-            dmds_extra.back().mb = measured.back().mb;
-            for (size_t je = 0; je < devs_extra.size(); je++) {
-                for (size_t id = 0; id < devs.size(); id++) {
-                    if (devs_extra[je] == devs[id]) {
-                        dmds_extra[id].mb.model   += measured[je].mb.model;
-                        dmds_extra[id].mb.context += measured[je].mb.context;
-                        dmds_extra[id].mb.compute += measured[je].mb.compute;
-                        break;
-                    }
-                }
-            }
-            if (extra->shares_model) {
-                for (llama_device_memory_data & dmd : dmds_extra) {
-                    dmd.mb.model = 0;
-                }
-            }
-
-            n_ctx_extra = cparams->n_ctx;
-        }
-
-        for (size_t id = 0; id < dmds.size(); id++) {
-            dmds[id].mb.model   += dmds_extra[id].mb.model;
-            dmds[id].mb.context += dmds_extra[id].mb.context;
-            dmds[id].mb.compute += dmds_extra[id].mb.compute;
-        }
-    };
-
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-
-    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
-    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
-    const uint32_t n_ctx_min_total = (uint32_t) std::min<uint64_t>(uint64_t(n_ctx_min) * n_streams, UINT32_MAX);
-
-    // llama_context would use only hp_nct in total for n_ctx == 0, resolve the context before measuring anything else:
-    if (n_ctx_auto) {
-        cparams->n_ctx = n_ctx_max;
-        if (n_streams > 1) {
-            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
-                __func__, n_ctx_max, n_streams);
-            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        }
-    }
-    add_extra_memory(dmds_full);
-
+    const dmds_t dmds_full = common_get_device_memory_data_impl(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
@@ -389,8 +307,8 @@ static void common_params_fit_impl(
                     "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
                     __func__, -global_surplus/MiB);
             }
-            if (n_ctx_auto) {
-                if (n_ctx_max > n_ctx_min_total) {
+            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];
@@ -410,9 +328,8 @@ static void common_params_fit_impl(
                     }
 
                     int64_t sum_projected_used_min_ctx = 0;
-                    cparams->n_ctx = n_ctx_min_total;
-                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-                    add_extra_memory(dmds_min_ctx);
+                    cparams->n_ctx = n_ctx_min;
+                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(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 {
@@ -422,16 +339,14 @@ static void common_params_fit_impl(
                     }
                     if (sum_used_target > sum_projected_used_min_ctx) {
                         // linear interpolation between minimum and maximum context size:
-                        cparams->n_ctx += (n_ctx_max - n_ctx_min_total) * (sum_used_target - sum_projected_used_min_ctx)
+                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
                             / (sum_projected_used - sum_projected_used_min_ctx);
-                        // round down context for CUDA backend, keep it divisible by the number of streams:
-                        const uint32_t align = 256 * n_streams;
-                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % align, n_ctx_min_total);
+                        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) / (n_ctx_max - n_ctx_min_total);
-                        const int64_t memory_reduction = (n_ctx_max - cparams->n_ctx) * bytes_per_ctx;
+                        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_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                         if (nd <= 1) {
                             LOG_TRC("%s: entire model can be fit by reducing context\n", __func__);
                             return;
@@ -440,14 +355,14 @@ static void common_params_fit_impl(
                     } else {
                         const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     if (n_ctx_min == UINT32_MAX) {
-                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, n_ctx_max);
+                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
                     } else {
                         LOG_TRC("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                            __func__, n_ctx_max, n_ctx_min_total);
+                            __func__, hp_nct, n_ctx_min);
                     }
                 }
             } else {
@@ -592,9 +507,8 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        dmds_t dmd_nl = common_get_device_memory_data_impl(
+        const dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmd_nl);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
         for (size_t id = 0; id < nd; id++) {
@@ -621,9 +535,8 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
+        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
-        add_extra_memory(dmds_cpu_moe);
 
         for (size_t id = 0; id < nd; id++) {
             global_surplus_cpu_moe += dmds_cpu_moe[id].free;
@@ -883,12 +796,11 @@ enum common_params_fit_status common_fit_params(
         llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t * margins,
         uint32_t n_ctx_min,
-        const common_fit_extra_model * extra,
         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, extra, log_level);
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
         LOG_TRC("%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());

common/fit.h

@@ -11,16 +11,6 @@ enum common_params_fit_status {
     COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
 };
 
-// a second model that shares the devices of the main model, e.g. a draft model
-//   - its context follows the context of the main model, so its memory is measured again whenever that context changes
-//   - shares_model tells the fit that the weights are already counted in the main model, as for an MTP context
-struct common_fit_extra_model {
-    const char * path_model;
-    llama_model_params * mparams;
-    llama_context_params * cparams;
-    bool shares_model;
-};
-
 // 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
@@ -34,7 +24,6 @@ common_params_fit_status common_fit_params(
    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
-      const common_fit_extra_model * extra,                  // model to fit alongside the main one, nullptr if there is none
                      ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout

common/json.cpp

@@ -78,21 +78,19 @@ common_json_value::common_json_value(const common_json & val) :
 common_json_value::common_json_value(common_json && val) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
 
-// the ctors and get<T>() below are explicit specializations, giving strong symbols
-// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
 template <typename T>
-static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
+common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a set value is usable only for the types below
-#define COMMON_JSON_SET(...) template <> common_json_value::common_json_value(const std::set<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
+#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
 
 COMMON_JSON_SET(int)
 COMMON_JSON_SET(std::string)
@@ -100,45 +98,56 @@ COMMON_JSON_SET(std::string)
 #undef COMMON_JSON_SET
 
 template <typename T>
-static std::shared_ptr<common_json> map_json(const T & vals) {
+common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
     common_json out = common_json::object();
 
     for (const auto & val : vals) {
         out.set({ val.first, val.second });
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a map value is usable only for the types below
-#define COMMON_JSON_MAP(...) template <> common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
 
 COMMON_JSON_MAP(bool)
 COMMON_JSON_MAP(std::string)
 
 #undef COMMON_JSON_MAP
 
+template <typename T>
+common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
 // an unordered map value is usable only for the types below
-#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
 
 COMMON_JSON_UMAP(size_t)
 
 #undef COMMON_JSON_UMAP
 
 template <typename T>
-static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
+common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
     common_json out = common_json::array();
 
     for (const auto & val : vals) {
         out.push_back(val);
     }
 
-    return std::make_shared<common_json>(std::move(out));
+    val_json = std::make_shared<common_json>(std::move(out));
 }
 
 // a vector value is usable only for the types below
 // note: std::vector<bool> is not here, its proxy reference does not convert
-#define COMMON_JSON_VEC(...) template <> common_json_value::common_json_value(const std::vector<__VA_ARGS__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
+#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
 
 COMMON_JSON_VEC(int)
 COMMON_JSON_VEC(unsigned char)
@@ -395,6 +404,10 @@ common_json::items_view common_json::items() const {
     return items_view(const_cast<common_json *>(this), size());
 }
 
+template <typename T> T common_json::get() const {
+    return guard([&] { return as_json(this).get<T>(); });
+}
+
 // the backing library cannot build a common_json, so this one is just a copy
 template <> common_json common_json::get<common_json>() const {
     return *this;
@@ -402,7 +415,7 @@ template <> common_json common_json::get<common_json>() const {
 
 // get<T>() is usable only for the types below
 
-#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
+#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
 
 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -422,12 +435,3 @@ COMMON_JSON_GET(std::vector<size_t>)
 COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
 
 #undef COMMON_JSON_GET
-
-// must stay below the get<std::string> specialization
-common_json::operator std::string() const {
-    return get<std::string>();
-}
-
-std::string common_json::value(const std::string & key, const char * def) const {
-    return contains(key) ? at(key).get<std::string>() : std::string(def);
-}

common/json.h

@@ -221,14 +221,16 @@ class common_json {
     // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
     // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
     // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
-    operator std::string() const;
+    operator std::string() const { return get<std::string>(); }
 
     template <typename T>
     T value(const std::string & key, T def) const {
         return contains(key) ? at(key).get<T>() : def;
     }
 
-    std::string value(const std::string & key, const char * def) const;
+    std::string value(const std::string & key, const char * def) const {
+        return contains(key) ? at(key).get<std::string>() : std::string(def);
+    }
 
     // a JSON default needs no get<T>(), it is already the right type
     common_json value(const std::string & key, const common_json & def) const {

common/log.cpp

@@ -438,7 +438,7 @@ void common_log_flush(struct common_log * log) {
     log->resume();
 }
 
-int common_log_get_verbosity(enum ggml_log_level level) {
+static int common_get_verbosity(enum ggml_log_level level) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
         case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_TRACE;
@@ -452,7 +452,7 @@ int common_log_get_verbosity(enum ggml_log_level level) {
 }
 
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    auto verbosity = common_log_get_verbosity(level);
+    auto verbosity = common_get_verbosity(level);
     if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }

common/log.h

@@ -43,8 +43,6 @@ int  common_log_get_verbosity_thold(void);
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
-int common_log_get_verbosity(enum ggml_log_level level);
-
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * user_data);
 
 // the common_log uses an internal worker thread to print/write log messages

common/speculative.cpp

@@ -14,7 +14,6 @@
 
 #include <algorithm>
 #include <cassert>
-#include <cmath>
 #include <cstring>
 #include <iomanip>
 #include <map>
@@ -139,7 +138,6 @@ struct common_speculative_impl {
     const common_speculative_type type;
 
     uint32_t n_seq;
-    int32_t n_max; // maximum draft length after implementation-specific limits
 
     size_t n_call_begin  = 0; // number of times this implementation was called for refresh.
     size_t n_call_draft  = 0; // number of times this implementation was called for generation.
@@ -159,7 +157,7 @@ struct common_speculative_impl {
     int64_t t_draft_us  = 0; // total time spent in generating drafts in this implementation in microseconds.
     int64_t t_accept_us = 0; // total time spent in accumulation of this implementation in microseconds.
 
-    common_speculative_impl(common_speculative_type type, uint32_t n_seq, int32_t n_max) : type(type), n_seq(n_seq), n_max(n_max) {}
+    common_speculative_impl(common_speculative_type type, uint32_t n_seq) : type(type), n_seq(n_seq) {}
 
     virtual ~common_speculative_impl() = default;
 
@@ -184,7 +182,7 @@ struct common_speculative_impl_draft_simple : public common_speculative_impl {
     std::vector<common_sampler_ptr> smpls;
 
     common_speculative_impl_draft_simple(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, n_seq)
         , params(params.draft)
     {
         auto * ctx_dft = this->params.ctx_dft;
@@ -454,7 +452,7 @@ struct common_speculative_impl_draft_eagle3 : public common_speculative_impl {
     std::vector<float> g_embd_buf;
 
     common_speculative_impl_draft_eagle3(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, n_seq)
         , params(params.draft)
     {
         SPC_TRC("%s", "adding speculative implementation 'draft-eagle3'\n");
@@ -925,25 +923,21 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
     int32_t     block_size    = 0;
     llama_token mask_token_id = 0;
 
-    bool    is_dflash2     = false;
-    bool    is_mrope       = false;
-    int32_t selector_top_k = 0;
-
     // draft-dspark: the draft carries a Markov head and uses an anchor-first block layout
     const bool is_dspark;
 
     // dspark speculators
     bool sample_from_anchor = true;
 
-    // block-internal attention
-    bool causal_attn = false;
-
     const int32_t * target_layer_ids   = nullptr; // model_dft's extract layer indices
     uint32_t        target_layer_ids_n = 0;
 
+    // scratch buffer for concatenated target features [n_tokens, n_embd_enc]
+    std::vector<float> features_buf;
+
     common_speculative_impl_draft_dflash(const common_params_speculative & params, uint32_t n_seq,
             common_speculative_type type = COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)
-        : common_speculative_impl(type, n_seq, params.draft.n_max)
+        : common_speculative_impl(type, n_seq)
         , params(params.draft)
         , is_dspark(type == COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)
     {
@@ -972,25 +966,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             if (llama_model_meta_val_str(model_dft, "dflash.sample_from_anchor", buf, sizeof(buf)) >= 0) {
                 sample_from_anchor = std::strcmp(buf, "true") == 0;
             }
-            if (llama_model_meta_val_str(model_dft, "dflash.attention.causal", buf, sizeof(buf)) >= 0) {
-                causal_attn = std::strcmp(buf, "true") == 0;
-            }
         }
-
-        selector_top_k = llama_model_dflash_selector_top_k(model_dft);
-        is_dflash2     = selector_top_k > 0;
         mask_token_id = llama_vocab_mask(llama_model_get_vocab(model_dft));
 
-        if (is_dspark && this->params.p_min > 0.0f) {
-            char buf[16] = {};
-            const bool has_conf =
-                llama_model_meta_val_str(model_dft, "dflash.has_confidence_head", buf, sizeof(buf)) < 0 ||
-                std::strcmp(buf, "true") == 0;
-            if (!has_conf) {
-                throw std::runtime_error("DSpark draft has no confidence head: please set --spec-draft-p-min 0");
-            }
-        }
-
         LOG_INF("%s: adding speculative implementation '%s'\n", __func__, common_speculative_type_to_str(type).c_str());
         LOG_INF("%s: - n_max=%d, n_min=%d, p_min=%.2f\n", __func__, this->params.n_max, this->params.n_min, this->params.p_min);
         LOG_INF("%s: - block_size=%d, mask_token_id=%d, n_extract=%u, sample_from_anchor=%s\n", __func__,
@@ -1005,17 +983,9 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             this->params.n_max = std::min(this->params.n_max, n_draft_max);
             this->params.n_min = std::min(this->params.n_min, n_draft_max);
         }
-        this->n_max = this->params.n_max;
 
         batch        = llama_batch_init(llama_n_batch(ctx_dft), 0,          n_seq);
-        batch_inject = llama_batch_init(llama_n_ubatch(ctx_dft), n_embd_enc, n_seq);
-
-        // embd batches on an M-RoPE draft need 4 position rows per token
-        is_mrope = llama_model_rope_type(model_dft) == LLAMA_ROPE_TYPE_MROPE;
-        if (is_mrope) {
-            free(batch_inject.pos);
-            batch_inject.pos = (llama_pos *) malloc(sizeof(llama_pos) * 4 * llama_n_batch(ctx_dft));
-        }
+        batch_inject = llama_batch_init(llama_n_batch(ctx_dft), n_embd_dec, n_seq);
 
         smpls.resize(n_seq);
         for (auto & s : smpls) {
@@ -1028,7 +998,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
         // offload draft sampling to the backend
         backend_chains.assign(n_seq, nullptr);
-        if (this->params.backend_sampling && !is_dflash2) {
+        if (this->params.backend_sampling) {
             for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) n_seq; ++seq_id) {
                 llama_sampler * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
                 llama_sampler_chain_add(chain, llama_sampler_init_top_k(10));
@@ -1047,9 +1017,8 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             llama_set_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k], true);
         }
 
-        // DFlash2 reads its selector lattice from h_nextn and never consumes raw logits.
-        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ !is_dflash2);
-        llama_set_causal_attn(ctx_dft, causal_attn); // DFlash needs non-causal attention unless the model says otherwise
+        llama_set_embeddings_nextn(ctx_dft, true, /*masked*/ true);
+        llama_set_causal_attn(ctx_dft, false); // DFlash needs non-causal attention
     }
 
     ~common_speculative_impl_draft_dflash() override {
@@ -1134,34 +1103,52 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             for (int32_t offset = 0; offset < n_rows; offset += n_ubatch) {
                 const int32_t n_chunk = std::min(n_ubatch, n_rows - offset);
 
-                // gather target features per extract layer; the fused decode encodes and
-                // injects them into the K/V cache at the target positions
-                batch_inject.n_tokens = n_chunk;
+                // gather this chunk's target features, interleaved by extract layer
+                features_buf.resize((size_t) n_chunk * n_embd_enc);
                 for (uint32_t k = 0; k < target_layer_ids_n; ++k) {
                     const float * layer = llama_get_embeddings_layer_inp(ctx_tgt, (uint32_t) target_layer_ids[k]);
                     if (!layer) {
                         GGML_ABORT("DFlash: target layer %d input not extracted.", target_layer_ids[k]);
                     }
                     for (int32_t i = 0; i < n_chunk; ++i) {
-                        float       * dst = batch_inject.embd + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
+                        float       * dst = features_buf.data() + (size_t) i * n_embd_enc + k * (size_t) n_embd_tgt;
                         const float * src = layer + (size_t) (i_batch_beg[seq_id] + offset + i) * n_embd_tgt;
                         std::memcpy(dst, src, (size_t) n_embd_tgt * sizeof(float));
                     }
                 }
 
+                // fuse extracted features through DFlash encoder
+                llama_batch enc_batch = {
+                    /*.n_tokens =*/ n_chunk,
+                    /*.token    =*/ nullptr,
+                    /*.embd     =*/ features_buf.data(),
+                    /*.pos      =*/ nullptr,
+                    /*.n_seq_id =*/ nullptr,
+                    /*.seq_id   =*/ nullptr,
+                    /*.logits   =*/ nullptr,
+                };
+
+                int32_t rc = llama_encode(ctx_dft, enc_batch);
+                if (rc != 0) {
+                    LOG_ERR("%s: llama_encode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
+                            __func__, rc, (int) n_chunk, (int) offset);
+                    return false;
+                }
+
+                const float * inp_g = llama_get_embeddings_nextn(ctx_dft);
+                GGML_ASSERT(inp_g && "DFlash encoder produced no output.");
+
+                // inject the DFlash decoder K/V cache at the tokens' target positions
+                batch_inject.n_tokens = n_chunk;
+                std::memcpy(batch_inject.embd, inp_g, (size_t) n_chunk * n_embd_dec * sizeof(float));
+
                 for (int32_t i = 0; i < n_chunk; ++i) {
-                    const llama_pos p = batch_in.pos[i_batch_beg[seq_id] + offset + i];
-                    batch_inject.pos[i] = p;
-                    if (is_mrope) {
-                        batch_inject.pos[1 * n_chunk + i] = p;
-                        batch_inject.pos[2 * n_chunk + i] = p;
-                        batch_inject.pos[3 * n_chunk + i] = 0;
-                    }
+                    batch_inject.pos[i]       = batch_in.pos[i_batch_beg[seq_id] + offset + i];
                     batch_inject.n_seq_id[i]  = 1;
                     batch_inject.seq_id[i][0] = seq_id;
                     batch_inject.logits[i]    = false;
                 }
-                const int32_t rc = llama_decode(ctx_dft, batch_inject);
+                rc = llama_decode(ctx_dft, batch_inject);
                 if (rc != 0) {
                     LOG_ERR("%s: llama_decode(ctx_dft) failed rc=%d (n_tokens=%d, offset=%d)\n",
                             __func__, rc, (int) n_chunk, (int) offset);
@@ -1199,7 +1186,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
             i_block_beg[seq_id] = batch.n_tokens;
             n_block    [seq_id] = n_block_tokens;
             for (int32_t i = 0; i < n_block_tokens; ++i) {
-                common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, !is_dflash2);
+                common_batch_add(batch, i == 0 ? dp.id_last : mask_token_id, n + i, { seq_id }, true);
             }
         }
 
@@ -1227,36 +1214,6 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
             auto & result = *dp.result;
 
-            if (is_dflash2) {
-                const float * lattice = llama_get_embeddings_nextn(ctx_dft);
-                GGML_ASSERT(lattice && "DFlash2 selector produced no lattice");
-
-                int32_t predecessor = 0;
-                for (int32_t i = 1; i < n_block_tokens; ++i) {
-                    const float * row = lattice + (size_t) (beg + i) * n_embd_dec;
-                    const float * scores = row + selector_top_k + (size_t) predecessor * selector_top_k;
-
-                    predecessor = (int32_t) std::distance(scores,
-                            std::max_element(scores, scores + selector_top_k));
-                    if (params.p_min > 0.0f) {
-                        // softmax(scores) at the argmax, i.e. 1 / sum(exp(s_k - s_max))
-                        float sum = 0.0f;
-                        for (int32_t k = 0; k < selector_top_k; ++k) {
-                            sum += std::exp(scores[k] - scores[predecessor]);
-                        }
-                        if (1.0f / sum < params.p_min) {
-                            break;
-                        }
-                    }
-                    result.push_back((llama_token) row[predecessor]);
-                }
-
-                if (result.size() < (size_t) params.n_min) {
-                    result.clear();
-                }
-                continue;
-            }
-
             if (is_dspark) {
                 // DSpark: read from the first draft slot, truncate below the confidence threshold
                 const float * conf = params.p_min > 0.0f ? llama_get_embeddings_nextn(ctx_dft) : nullptr;
@@ -1358,7 +1315,7 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
     std::vector<std::vector<float>> chain_h;
 
     common_speculative_impl_draft_mtp(const common_params_speculative & params, uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq, params.draft.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, n_seq)
         , params(params.draft)
     {
         auto * ctx_tgt = this->params.ctx_tgt;
@@ -1425,7 +1382,6 @@ struct common_speculative_impl_draft_mtp : public common_speculative_impl {
                 c.reserve((size_t) (this->params.n_max + 1) * n_embd);
             }
         }
-        this->n_max = this->params.n_max;
 
         pending_h.assign(n_seq, std::vector<float>(n_embd, 0.0f));
 
@@ -1770,7 +1726,7 @@ struct common_speculative_impl_ngram_simple : public common_speculative_impl {
     common_speculative_impl_ngram_simple(
             const common_params_speculative & params, uint32_t n_seq,
             common_ngram_simple_config config)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq, params.ngram_simple.size_m)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, n_seq)
         , params(params.ngram_simple)
         , config(config)
     {
@@ -1814,7 +1770,7 @@ struct common_speculative_impl_ngram_map_k : public common_speculative_impl {
             const common_ngram_map & config,
             uint32_t n_seq)
         : common_speculative_impl(config.key_only ? COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K
-            : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq, config.size_value)
+            : COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, n_seq)
     {
         for (uint32_t i = 0; i < n_seq; i++) {
             this->config.push_back(config);
@@ -1885,7 +1841,7 @@ struct common_speculative_impl_ngram_mod : public common_speculative_impl {
     common_speculative_impl_ngram_mod(
             const common_params_speculative & params,
             uint32_t n_seq)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq, params.ngram_mod.n_max)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, n_seq)
         , params(params.ngram_mod)
         , mod(params.ngram_mod.n_match, 4*1024*1024)
         , verbose(std::getenv("LLAMA_TRACE") != nullptr) {
@@ -2061,7 +2017,7 @@ struct common_speculative_impl_ngram_cache : public common_speculative_impl {
             const std::string & path_dynamic,
             bool save_dynamic,
             bool save_static)
-        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq, n_draft)
+        : common_speculative_impl(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, n_seq)
         , params(params.ngram_cache)
         , n_draft(n_draft)
         , save_dynamic(save_dynamic)
@@ -2182,8 +2138,6 @@ struct common_speculative {
 
     // which implementaion was used for a given seq_id
     std::vector<common_speculative_impl *> impl_last;
-
-    std::vector<double> synth_probs;
 };
 
 static common_ngram_map get_common_ngram_map(
@@ -2362,101 +2316,6 @@ int32_t common_speculative_n_max(const common_params_speculative * spec) {
     return n_max;
 }
 
-int32_t common_speculative_n_max(const common_speculative * spec) {
-    int32_t n_max = 0;
-
-    if (spec == nullptr) {
-        return n_max;
-    }
-
-    for (const auto & impl : spec->impls) {
-        n_max = std::max(n_max, std::max(0, impl->n_max));
-    }
-
-    return n_max;
-}
-
-std::vector<double> common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max) {
-    const bool has_length = spec->synth_len != -1.0;
-    const bool has_rates  = !spec->synth_rates.empty();
-
-    if (!has_length && !has_rates) {
-        return {};
-    }
-    if (has_length && has_rates) {
-        throw std::invalid_argument("synthetic acceptance length and rates are mutually exclusive");
-    }
-
-    if (n_max <= 0) {
-        throw std::invalid_argument("synthetic acceptance requires at least one speculative token");
-    }
-
-    if (has_rates) {
-        const auto & rates = spec->synth_rates;
-        if (rates.size() != (size_t) n_max) {
-            throw std::invalid_argument(string_format(
-                    "synthetic acceptance rates must contain %d values, got %zu", n_max, rates.size()));
-        }
-
-        for (size_t i = 0; i < rates.size(); ++i) {
-            if (!std::isfinite(rates[i]) || rates[i] < 0.0 || rates[i] > 1.0) {
-                throw std::invalid_argument("synthetic acceptance rates must be finite and within [0, 1]");
-            }
-            if (i > 0 && rates[i] > rates[i - 1]) {
-                throw std::invalid_argument("synthetic acceptance rates must be monotonically non-increasing");
-            }
-        }
-
-        return rates;
-    }
-
-    const double length = spec->synth_len;
-    const double length_max = (double) n_max + 1.0;
-    if (!std::isfinite(length) || length < 1.0 || length > length_max) {
-        throw std::invalid_argument(string_format(
-                "synthetic acceptance length must be finite and within [1, %.0f]", length_max));
-    }
-
-    double p = 0.0;
-    if (length == length_max) {
-        p = 1.0;
-    } else if (length > 1.0) {
-        double p_min = 0.0;
-        double p_max = 1.0;
-        for (int i = 0; i < 32; ++i) {
-            const double p_mid = 0.5 * (p_min + p_max);
-            double sum = 0.0;
-            double term = p_mid;
-            for (int32_t j = 0; j < n_max; ++j) {
-                sum += term;
-                term *= p_mid;
-            }
-
-            if (sum < length - 1.0) {
-                p_min = p_mid;
-            } else {
-                p_max = p_mid;
-            }
-        }
-        p = 0.5 * (p_min + p_max);
-    }
-
-    std::vector<double> rates;
-    rates.reserve(n_max);
-    double rate = p;
-    for (int32_t i = 0; i < n_max; ++i) {
-        rates.push_back(rate);
-        rate *= p;
-    }
-
-    return rates;
-}
-
-const std::vector<double> & common_speculative_get_synth_probs(const common_speculative * spec) {
-    GGML_ASSERT(spec);
-    return spec->synth_probs;
-}
-
 common_params common_base_params_to_speculative(const common_params & params) {
     const bool has_draft = params.speculative.has_dft();
 
@@ -2529,9 +2388,6 @@ common_speculative_init_result::common_speculative_init_result(
         cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
     }
 
-    // the draft context holds as many tokens per sequence as the target context
-    cparams.n_ctx = llama_n_ctx(ctx_tgt);
-
     // note: for small models maybe we can set this to the maximum possible draft from all speculative types
     //       the extra memory for small models is likely negligible?
     cparams.n_rs_seq  = 0;
@@ -2709,39 +2565,13 @@ common_speculative * common_speculative_init(common_params_speculative & params,
         return nullptr;
     }
 
-    common_speculative_ptr result(new common_speculative {
-        /* .dparams     = */ common_speculative_draft_params_vec(n_seq),
-        /* .impls       = */ std::move(impls),
-        /* .impl_last   = */ std::vector<common_speculative_impl *>(n_seq, nullptr),
-        /* .synth_probs = */ {},
-    });
+    auto * result = new common_speculative {
+        /* .dparams   = */ common_speculative_draft_params_vec(n_seq),
+        /* .impls     = */ std::move(impls),
+        /* .impl_last = */ std::vector<common_speculative_impl *>(n_seq, nullptr)
+    };
 
-    const int32_t n_max_configured = common_speculative_n_max(&params);
-    const int32_t n_max_effective  = common_speculative_n_max(result.get());
-    const auto rates = common_speculative_synth_rates_resolve(&params, n_max_effective);
-
-    std::vector<std::string> rates_str;
-    rates_str.reserve(rates.size());
-    result->synth_probs.reserve(rates.size());
-    double rate_prev = 1.0;
-    double acceptance_length = 1.0;
-    for (const double rate : rates) {
-        result->synth_probs.push_back(rate_prev > 0.0 ? rate / rate_prev : 0.0);
-        rates_str.push_back(string_format("%.6g", rate));
-        rate_prev = rate;
-        acceptance_length += rate;
-    }
-    if (!result->synth_probs.empty()) {
-        SPC_WRN("%s", "synthetic speculative acceptance is enabled for benchmarking; generated output is not valid\n");
-        if (n_max_effective != n_max_configured) {
-            SPC_WRN("synthetic acceptance draft limit was reduced from %d to %d by the initialized speculative implementations\n",
-                    n_max_configured, n_max_effective);
-        }
-        SPC_INF("synthetic acceptance: n_max = %zu, mean length = %.6f, rates = [%s]\n",
-                rates.size(), acceptance_length, string_join(rates_str, ", ").c_str());
-    }
-
-    return result.release();
+    return result;
 }
 
 void common_speculative_free(common_speculative * spec) {

common/speculative.h

@@ -26,15 +26,6 @@ std::string common_speculative_type_to_str(enum common_speculative_type type);
 // return the max number of draft tokens based on the speculative parameters
 int32_t common_speculative_n_max(const common_params_speculative * spec);
 
-// return the max number of draft tokens from the initialized implementations
-int32_t common_speculative_n_max(const common_speculative * spec);
-
-// validate and resolve the unconditional synthetic acceptance rates
-std::vector<double> common_speculative_synth_rates_resolve(const common_params_speculative * spec, int32_t n_max);
-
-// return the conditional synthetic acceptance probabilities
-const std::vector<double> & common_speculative_get_synth_probs(const common_speculative * spec);
-
 common_params common_base_params_to_speculative(const common_params & params);
 
 struct common_speculative_output_limits {

conversion/__init__.py

@@ -54,7 +54,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "DeepseekV3ForCausalLM": "deepseek",
     "DeepseekV32ForCausalLM": "deepseek",
     "DFlashDraftModel": "qwen",
-    "DFlash2DraftModel": "qwen",
     "Qwen3DSparkModel": "qwen",
     "DSparkDraftModel": "qwen",
     "DSparkSpeculator": "qwen",
@@ -236,8 +235,6 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Qwen3_5ForConditionalGeneration": "qwen",
     "Qwen3_5MoeForCausalLM": "qwen",
     "Qwen3_5MoeForConditionalGeneration": "qwen",
-    "Qwen4ExpForCausalLM": "qwen4exp",
-    "Qwen4ExpForConditionalGeneration": "qwen4exp",
     "RND1": "qwen",
     "RWForCausalLM": "falcon",
     "RWKV6Qwen2ForCausalLM": "rwkv",
@@ -335,7 +332,6 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
     "Qwen3_5ForConditionalGeneration": "qwen3vl",
     "Qwen3_5MoeForConditionalGeneration": "qwen3vl",
-    "Qwen4ExpForConditionalGeneration": "qwen4exp",
     "RADIOModel": "nemotron",
     "Sarashina2VisionForCausalLM": "sarashina2",
     "SmolVLMForConditionalGeneration": "smolvlm",

conversion/base.py

@@ -1006,16 +1006,12 @@ class ModelBase:
                     else:
                         raise ValueError(f"Unknown file type: {self.ftype.name}")
 
-                # a chunked tensor quantizes as one chunk at a time, while it is written
-                quantize = data.quantize if isinstance(data, gguf.LazyChunkedTensor) else (
-                    lambda qtype, d=data: gguf.quants.quantize(d, qtype))
-
                 try:
-                    data = quantize(data_qtype)
+                    data = gguf.quants.quantize(data, data_qtype)
                 except gguf.QuantError as e:
                     logger.warning("%s, %s", e, "falling back to F16")
                     data_qtype = gguf.GGMLQuantizationType.F16
-                    data = quantize(data_qtype)
+                    data = gguf.quants.quantize(data, data_qtype)
 
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 

conversion/glm.py

@@ -112,38 +112,12 @@ class GlmOCRModel(Glm4Model):
 @ModelBase.example("zai-org/GLM-4.5-Air")
 class Glm4MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
-    supports_mtp_export = True
-    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
-            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None):
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._n_main_layers = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
-
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        if (titem := super().filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-
-        assert cls._n_main_layers is not None
-        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
-
-        if is_mtp and cls.no_mtp:
-            return None
-        if cls.mtp_only and not is_mtp and name not in (
-            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-        ):
-            return None
-
-        return name, gen
+        # GLM4_MOE has num_hidden_layers + 1 actual layers (including NextN layer)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
     def set_vocab(self):
         return self._set_vocab_glm()
@@ -179,22 +153,10 @@ class Glm4MoeModel(TextModel):
         if (norm_topk_prob := self.hparams.get("norm_topk_prob")) is not None:
             self.gguf_writer.add_expert_weights_norm(norm_topk_prob)
 
-        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        # NextN/MTP prediction layers
+        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
     _experts: list[dict[str, Tensor]] | None = None
 
     # note: unlike GLM4V non-MoE, we don't need to permute Q/K here since GLM4V_MOE uses Neox ordering already
@@ -386,7 +348,6 @@ class GlmMoeDsaModel(DeepseekV2Model):
 @ModelBase.example("upstage/Solar-Open-100B")
 class SolarOpenModel(Glm4MoeModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
-    supports_mtp_export = False
 
     def set_vocab(self):
         from transformers import AutoTokenizer

conversion/nemotron.py

@@ -202,10 +202,6 @@ class NemotronHModel(GraniteHybridModel):
     is_moe: bool = False
     supports_mtp_export = True
 
-    _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
-    _ATTN_LAYER_TYPES = {"attention", "full_attention"}
-    _MLP_LAYER_TYPES = {"moe"}
-
     def __init__(self, *args, **kwargs):
         # We have to determine the correct model architecture (MoE vs non-MoE) before
         # calling the parent __init__. This is because the parent constructor
@@ -246,8 +242,8 @@ class NemotronHModel(GraniteHybridModel):
             self._ssm_layers = [i for i, val in enumerate(pattern) if val == "M"]
             self._mlp_layers = [i for i, val in enumerate(pattern) if val == ("E" if self.is_moe else "-")]
         else:
-            self._ssm_layers = [i for i, val in enumerate(pattern) if val in self._SSM_LAYER_TYPES]
-            self._mlp_layers = [i for i, val in enumerate(pattern) if val in self._MLP_LAYER_TYPES]
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]
 
         # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
         self._mtp_bid: int | None = None
@@ -276,7 +272,7 @@ class NemotronHModel(GraniteHybridModel):
         if isinstance(pattern, str):
             return [i for i, val in enumerate(pattern) if val == "*"]
 
-        return [i for i, val in enumerate(pattern) if val in self._ATTN_LAYER_TYPES]
+        return [i for i, val in enumerate(pattern) if val == "attention"]
 
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
@@ -302,10 +298,6 @@ class NemotronHModel(GraniteHybridModel):
             )
             if not keep:
                 return None
-        # PEFT names adapter tensors using model.layers.*, while Nemotron-H checkpoints
-        # and the GGUF tensor map use backbone.layers.*
-        if name.startswith("model.layers.") and ".mixer." in name:
-            name = name.replace("model.layers.", "backbone.layers.", 1)
         return super().filter_tensors((name, gen))
 
     def prepare_metadata(self, vocab_only: bool):

conversion/qwen.py

@@ -639,7 +639,7 @@ class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 
-@ModelBase.register("DFlashDraftModel", "DFlash2DraftModel")
+@ModelBase.register("DFlashDraftModel")
 @ModelBase.example("z-lab/Qwen3.5-9B-DFlash")
 class DFlashModel(Qwen3Model):
     model_arch = gguf.MODEL_ARCH.DFLASH
@@ -678,98 +678,34 @@ class DFlashModel(Qwen3Model):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        dflash_config = self.hparams.get("dflash_config", {})
-        block_size = dflash_config.get("block_size", self.hparams.get("block_size", 16))
+        block_size = self.hparams.get("block_size", 16)
         self.gguf_writer.add_block_size(block_size)
-
-        if "conv_kernel_size" in dflash_config:
-            self.gguf_writer.add_conv_kernel_size(int(dflash_config["conv_kernel_size"]))
-            self.gguf_writer.add_conv_group_size(int(dflash_config["conv_group_size"]))
-            self.gguf_writer.add_selector_rank(int(dflash_config["selector_rank"]))
-            self.gguf_writer.add_selector_top_k(int(dflash_config["selector_top_k"]))
-
-        output_multiplier = dflash_config.get(
-            "output_multiplier", self.hparams.get("output_multiplier")
-        )
-        if output_multiplier is not None:
-            self.gguf_writer.add_logit_scale(float(output_multiplier))
-        softcap = dflash_config.get(
-            "final_logit_softcapping", self.hparams.get("final_logit_softcapping")
-        )
-        if softcap is not None and float(softcap) > 0:
-            self.gguf_writer.add_final_logit_softcapping(float(softcap))
-        embedding_scale = dflash_config.get(
-            "input_embedding_scale", self.hparams.get("input_embedding_scale")
-        )
-        if embedding_scale is not None:
-            self.gguf_writer.add_embedding_scale(float(embedding_scale))
+        dflash_config = self.hparams.get("dflash_config", {})
 
         target_layer_ids = dflash_config.get("target_layer_ids", [])
         if target_layer_ids:
             extract_layer_ids = [i + 1 for i in target_layer_ids]
             self.gguf_writer.add_target_layers(extract_layer_ids)
 
-        use_sliding_window = self.hparams.get("use_sliding_window", False) or dflash_config.get("use_swa", False)
-        sliding_window = dflash_config.get("swa_window_size") or self.hparams.get("sliding_window")
+        use_sliding_window = self.hparams.get("use_sliding_window", False)
+        sliding_window = self.hparams.get("sliding_window")
         layer_types = self.hparams.get("layer_types")
         if use_sliding_window and sliding_window and layer_types:
             is_swa = [lt == "sliding_attention" for lt in layer_types]
             self.gguf_writer.add_sliding_window(sliding_window)
             self.gguf_writer.add_sliding_window_pattern(is_swa)
 
-        causal = self.hparams.get("is_causal")
-        if causal is None:
-            causal = dflash_config.get("causal")
-        if causal is not None:
-            self.gguf_writer.add_causal_attention(bool(causal))
-
-        # M-RoPE target: the draft ropes on the temporal dim only, so write
-        # degenerate sections [n_rot/2, 0, 0, 0]
-        if self._target_uses_mrope():
-            head_dim = self.hparams.get("head_dim") or self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
-            self.gguf_writer.add_rope_dimension_sections([head_dim // 2, 0, 0, 0])
-
-    def _target_uses_mrope(self) -> bool:
-        if self.target_model_dir is None:
-            return False
-        with open(self.target_model_dir / "config.json", "r", encoding="utf-8") as f:
-            cfg = json.load(f)
-        cfg = cfg.get("text_config", cfg)
-        rope = cfg.get("rope_parameters") or cfg.get("rope_scaling") or {}
-        return "mrope_section" in rope
-
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
         if not name.startswith("model."):
             name = "model." + name
-        if "sink" in name and not name.endswith(".weight"):
-            name += ".weight"
         return super().filter_tensors((name, gen))
 
-    _ROPE_PERMUTE_SUFFIXES = (
-        "self_attn.q_proj.weight",
-        "self_attn.k_proj.weight",
-        "self_attn.q_norm.weight",
-        "self_attn.k_norm.weight",
-    )
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name == "model.embed_tokens.weight" and not self.hparams.get("has_embed_tokens", True):
             return
 
-        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
-        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
-            head_dim = self.hparams["head_dim"]
-            shape = data_torch.shape
-            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
-
-        if name in (
-            "model.candidate_selector.predecessor_codebook",
-            "model.candidate_selector.successor_codebook",
-        ):
-            name += ".weight"
-
         yield from super().modify_tensors(data_torch, name, bid)
 
 
@@ -823,10 +759,6 @@ class DSparkModel(DFlashModel):
         super().set_gguf_parameters()
         self.gguf_writer.add_sample_from_anchor(self._sample_from_anchor)
 
-        # confidence head is optional: vanilla-markov exports ship without it
-        has_conf = any("confidence_head.proj" in name for name in self.model_tensors)
-        self.gguf_writer.add_has_confidence_head(has_conf)
-
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         if item[0] == "t2d":  # not used at runtime
@@ -845,7 +777,7 @@ class DSparkModel(DFlashModel):
             self._d2t = data_torch
             return
 
-        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith("lm_head.weight"):
+        if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
             return
 
         # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd

conversion/qwen4exp.py

@@ -1,195 +0,0 @@
-from __future__ import annotations
-
-from typing import Iterable, cast
-
-import torch
-from torch import Tensor
-
-import gguf
-import numpy as np
-
-from .base import ModelBase
-from .qwen import _LinearAttentionVReorderBase, _Qwen35MRopeMixin
-from .qwen3vl import Qwen3VLVisionModel
-
-
-@ModelBase.register("Qwen4ExpForConditionalGeneration", "Qwen4ExpForCausalLM")
-@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
-class Qwen4ExpTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
-    """Qwen3.8-Flash-Next.
-
-    Shares the Qwen3.5 gated delta net and interleaved mrope, and adds three things:
-    hyper-connections in place of every layer norm, QSA sparse attention on the full
-    attention layers, and PLE n-gram hash embeddings on a single layer.
-    """
-
-    model_arch = gguf.MODEL_ARCH.QWEN4EXP
-
-    # the MTP block is a separate draft head; vLLM drops it too
-    supports_mtp_export = False
-    no_mtp = True
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        # only the shard names, so the table itself is never held
-        self._ple_shards: dict[int, str] = {}
-        self._ple_row_dim: int | None = None
-
-    def _read_hash_constants(self, suffix: str) -> list[int]:
-        """Read an int64 PLE constant straight from the checkpoint.
-
-        prepare_tensors() casts every non-float dtype to float32 before
-        modify_tensors() sees it (base.py), which would silently round these
-        45-bit multipliers. Reading the lazy tensor here bypasses that.
-        """
-        for name, gen in self.model_tensors.items():
-            if name.endswith(suffix):
-                t = gen()
-                if t.dtype != torch.int64:
-                    t = t.to(torch.int64)
-                return [int(x) for x in t.tolist()]
-        raise ValueError(f"PLE constant {suffix!r} missing from the checkpoint")
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hp = self.hparams
-
-        self.gguf_writer.add_hyper_connection_count(hp["hc_count"])
-        self.gguf_writer.add_hyper_connection_low_rank(hp["hc_lowrank"])
-
-        n_layer = hp["num_hidden_layers"]
-        self.gguf_writer.add_indexer_head_count(hp["indexer_n_heads"])
-        self.gguf_writer.add_indexer_key_length(hp["indexer_head_dim"])
-        self.gguf_writer.add_indexer_top_k(hp["indexer_budget"])
-        ratio = hp["indexer_compress_ratio"]
-        layer_types = hp["layer_types"]
-        self.gguf_writer.add_attention_compress_ratios(
-            [ratio if layer_types[i] == "full_attention" else 0 for i in range(n_layer)]
-        )
-
-        # ple_layer_ids is 1-based in the HF config; empty means no n-gram table,
-        # so emit no PLE keys rather than optional ones
-        ple_layers = [i - 1 for i in hp["ple_layer_ids"]]
-        if not ple_layers:
-            return
-        self.gguf_writer.add_ple_layers(ple_layers)
-        self.gguf_writer.add_ple_ngram_size(hp["ngram_size"])
-        self.gguf_writer.add_ple_heads_per_ngram(hp["heads_per_ngram"])
-        self.gguf_writer.add_ple_conv_kernel(hp["ple_conv_kernel_size"])
-        self.gguf_writer.add_ple_eos_token_id(self._eos_token_id())
-        # an image is decoded as an embeddings-only batch, so the graph has no placeholder
-        # ids to hash; carry the id and let it stand in for those positions
-        _img = self._image_token_id()
-        if _img is not None:
-            self.gguf_writer.add_ple_image_token_id(int(_img))
-        if self._ple_row_dim is not None:
-            self.gguf_writer.add_embedding_length_per_layer_input(self._ple_row_dim)
-
-        self.gguf_writer.add_ple_layer_multipliers(
-            self._read_hash_constants("ple_embedding.layer_multipliers"))
-        self.gguf_writer.add_ple_head_offsets(
-            self._read_hash_constants("ple_embedding.ngram_heads_offsets"))
-        self.gguf_writer.add_ple_head_vocab_sizes(
-            self._read_hash_constants("ple_embedding.ngram_heads_vocab_sizes"))
-
-    def _image_token_id(self) -> int | None:
-        img = self.hparams.get("image_token_id")
-        return None if img is None else int(img)
-
-    def _eos_token_id(self) -> int:
-        eos = self.hparams.get("eos_token_id")
-        if isinstance(eos, list):
-            # the PLE hash resets n-grams on the primary EOS
-            return int(eos[-1])
-        if eos is None:
-            raise ValueError("eos_token_id is required: the PLE hash resets its n-grams on it")
-        return int(eos)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # int64 hash constants must stay exact; 1-D tensors force F32, so use KV
-        if name.endswith("ple_embedding.layer_multipliers"):
-            self._ple_multipliers = [int(x) for x in data_torch.tolist()]
-            return []
-        if name.endswith("ple_embedding.ngram_heads_offsets"):
-            self._ple_head_offsets = [int(x) for x in data_torch.tolist()]
-            return []
-        if name.endswith("ple_embedding.ngram_heads_vocab_sizes"):
-            self._ple_head_vocab_sizes = [int(x) for x in data_torch.tolist()]
-            return []
-
-        if ".ngram_embedding.shard_" in name:
-            return self._place_ple_shard(data_torch, name)
-
-        # one projection feeds indexer q and k; split it, as minimax-m3 does
-        if ".indexer.index_qk_proj.weight" in name:
-            n_q = self.hparams["indexer_n_heads"] * self.hparams["indexer_head_dim"]
-            q = data_torch[:n_q]
-            k = data_torch[n_q:]
-            return [
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_PROJ, bid, ".weight"), q),
-                (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_K_PROJ, bid, ".weight"), k),
-            ]
-
-        # Gemma zero-centred gammas the inherited norm.weight rule misses
-        if name.endswith((".ple.norm_key.weight", ".ple.norm_query.weight", ".ple.norm_conv.weight",
-                          ".indexer.q_layernorm.weight", ".indexer.k_layernorm.weight")):
-            return [(self.map_tensor_name(name), data_torch + 1)]
-
-        if name.endswith(".ple.conv1d.weight"):
-            return [(self.map_tensor_name(name), data_torch.squeeze())]
-
-        return super().modify_tensors(data_torch, name, bid)
-
-    # the shards concatenate into a tensor of well over 100 GB
-    # use LazyChunkedTensor here, a single shard resident at a time
-    def _place_ple_shard(self, data_torch: Tensor, name: str) -> Iterable[tuple[str, Tensor]]:
-
-        idx = int(name.rpartition(".shard_")[2].partition(".")[0])
-        n_parts = self.hparams["split_ngram_parts"]
-
-        self._ple_shards[idx] = name
-        self._ple_row_dim = int(data_torch.shape[-1])
-
-        if len(self._ple_shards) < n_parts:
-            return []
-
-        # the checkpoint may yield the shards in any order, the row order is by index
-        shards = [self._ple_shards[i] for i in sorted(self._ple_shards)]
-        rows = 0
-        for shard in shards:
-            shape = self.model_tensors[shard]().shape
-            if int(shape[-1]) != self._ple_row_dim:
-                raise ValueError(
-                    f"PLE shard {shard} has row dim {int(shape[-1])}, expected {self._ple_row_dim}")
-            rows += int(shape[0])
-
-        table = gguf.LazyChunkedTensor(
-            [self._load_ple_shard(shard) for shard in shards],
-            shape=(rows, self._ple_row_dim),
-            dtype=np.float32,
-        )
-        gguf_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD]
-        return [(gguf_name + ".weight", cast(Tensor, table))]
-
-    def _load_ple_shard(self, name: str):
-        def load() -> np.ndarray:
-            from .base import LazyTorchTensor
-
-            # a fresh lazy tensor every call, or to_eager() memoizes every shard
-            eager = LazyTorchTensor.to_eager(self.model_tensors[name]())
-            return eager.to(torch.float32).contiguous().numpy()
-        return load
-
-    def prepare_tensors(self):
-        super().prepare_tensors()
-        n_parts = self.hparams.get("split_ngram_parts", 0)
-        if self._ple_shards and len(self._ple_shards) != n_parts:
-            raise ValueError(
-                f"got {len(self._ple_shards)} PLE embedding shards, expected {n_parts}"
-            )
-
-
-@ModelBase.register("Qwen4ExpForConditionalGeneration")
-@ModelBase.example("Qwen/Qwen3.8-Flash-Next")
-class Qwen4ExpVisionModel(Qwen3VLVisionModel):
-    """The vision tower is an unmodified Qwen3-VL ViT."""

docs/android.md

@@ -53,7 +53,7 @@ To see what it might look like visually, here's an old demo of an interactive se
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
 ## Cross-compile CLI using Android NDK
-It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK/NDK and set `ANDROID_NDK` to the NDK root). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
+It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
 
 Once you're ready and have cloned `llama.cpp`, invoke the following in the project directory:
 
@@ -62,22 +62,18 @@ $ cmake \
   -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
   -DANDROID_ABI=arm64-v8a \
   -DANDROID_PLATFORM=android-28 \
-  -DGGML_NATIVE=OFF \
+  -DCMAKE_C_FLAGS="-march=armv8.7a" \
+  -DCMAKE_CXX_FLAGS="-march=armv8.7a" \
   -DGGML_OPENMP=OFF \
   -DGGML_LLAMAFILE=OFF \
-  -DLLAMA_OPENSSL=OFF \
   -B build-android
 ```
 
 Notes:
-  - `GGML_NATIVE=OFF` is required for cross-compilation because the host CPU is not the Android target CPU
   - While later versions of Android NDK ship with OpenMP, it must still be installed by CMake as a dependency, which is not supported at this time
   - `llamafile` does not appear to support Android devices (see: https://github.com/Mozilla-Ocho/llamafile/issues/325)
-  - `LLAMA_OPENSSL=OFF` avoids depending on OpenSSL, which is not part of the Android NDK stable native API set
 
-The above command configures a portable Android `arm64-v8a` build. Do not add a global `-march` flag unless you intentionally want to raise the baseline instruction set for every compiled source.
-
-For optional KleidiAI acceleration on Android `arm64-v8a`, see the [Arm KleidiAI section in build.md](./build.md#arm-kleidiai).
+The above command should configure `llama.cpp` with the most performant options for modern devices. Even if your device is not running `armv8.7a`, `llama.cpp` includes runtime checks for available CPU features it can use.
 
 Feel free to adjust the Android ABI for your target. Once the project is configured:
 

docs/autoparser.md

@@ -443,21 +443,21 @@ Each returned parser is wrapped by `wrap_for_generation_prompt()`, which prepend
 |                                           | `wrap_for_generation_prompt()`, string helpers                                  |
 | `common/chat-peg-parser.h/cpp`            | `common_chat_peg_builder`, `common_chat_peg_mapper`, and helpers                |
 | `common/chat.cpp`                         | Entry point: `common_chat_templates_apply_jinja()`                              |
-| `tests/test-chat-auto-parser.cpp`         | Auto-parser unit tests; also a debug tool when given a template path            |
-| `tests/test-chat-analysis.cpp`            | Template differential analysis debug tool                                       |
+| `tools/parser/debug-template-parser.cpp`  | Debug tool for template analysis                                                |
+| `tools/parser/template-analysis.cpp`      | Template analysis tool                                                          |
 
 ## Testing & Debugging
 
 ### Debug Tools
 
-**Template Debugger**: `tests/test-chat-auto-parser.cpp`
+**Template Debugger**: `tools/parser/debug-template-parser.cpp`
 
-- Usage: `./bin/test-chat-auto-parser path/to/template.jinja` (without a path, it runs the automated tests)
+- Usage: `./bin/llama-debug-template-parser path/to/template.jinja`
 - Shows detected format, markers, generated parser, and GBNF grammar
 
-**Template Analysis**: `tests/test-chat-analysis.cpp`
+**Template Analysis**: `tools/parser/template-analysis.cpp`
 
-- Usage: `./bin/test-chat-analysis --template-file path/to/template.jinja` (without arguments, it runs on all templates from the test suite)
+- Usage: `./bin/llama-template-analysis path/to/template.jinja`
 
 **Debug Logging**: Enable with `LLAMA_ARG_LOG_VERBOSITY=2`
 
@@ -519,7 +519,7 @@ The following templates have active tests in `tests/test-chat.cpp`:
 
 To support a new template format:
 
-1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `test-chat-auto-parser <template_path>` to verify markers are correctly extracted.
+1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `llama-debug-template-parser` to verify markers are correctly extracted.
 2. **If differential analysis extracts incorrect markers** — Add a workaround lambda to the `workarounds` vector in `common/chat-diff-analyzer.cpp`. Inspect the template source for a unique identifying substring.
 3. **If it needs fundamentally different handling** — Add a dedicated handler function in `chat.cpp` before the auto-parser block (as done for GPT-OSS, Functionary v3.2, and Ministral).
 

docs/backend/OPENVINO.md

@@ -22,8 +22,8 @@ The OpenVINO backend is implemented in `ggml/src/ggml-openvino` and provides a t
   - [0. Prerequisites](#0-prerequisites)
   - [1. Install OpenVINO Runtime](#1-install-openvino-runtime)
   - [2. Build llama.cpp with OpenVINO Backend](#2-build-llamacpp-with-openvino-backend)
-    - [Ubuntu Build Script](#ubuntu-build-script)
-    - [Windows Build Script](#windows-build-script)
+    - [Automated Ubuntu Build Script](#automated-ubuntu-build-script)
+    - [Automated Windows Build Script](#automated-windows-build-script)
   - [3. Download Sample Model](#3-download-sample-model)
   - [4. Run Inference with OpenVINO Backend](#4-run-inference-with-openvino-backend)
   - [5. Docker Build](#5-docker-build)
@@ -96,7 +96,7 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
   - **SL** = Stateless (`GGML_OPENVINO_STATEFUL_EXECUTION=0`)
   - **SF** = Stateful (`GGML_OPENVINO_STATEFUL_EXECUTION=1`)
   - Note: The NPU operates in stateless mode only.
-- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.31.39395.13-0 | Intel NPU Driver 1.35.0.
+- **Validation system:** Intel® Core™ Ultra 5 238V (Lunar Lake) | 32 GB RAM | Ubuntu 24.04 | Intel OpenCL GPU Driver 26.18.38308.1 | Intel NPU Driver 1.33.0.
 - See [Known Limitations](#known-limitations) for context on observed failures.
 
 | Model | CPU (SL / SF) | GPU (SL / SF) | NPU (SL) |
@@ -105,32 +105,27 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [bartowski/Llama-3.2-3B-Instruct-Q4_K_M](https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Meta-Llama-3.1-8B-Instruct-Q4_K_M](https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
-| [Qwen/qwen2.5-1.5b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [Qwen/qwen2.5-coder-7b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3-0.6B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3-1.7B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-1.7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [Qwen/Qwen3-4B-Q4_K_M](https://huggingface.co/Qwen/Qwen3-4B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [lm-kit/Qwen3-8B-Q4_K_M](https://huggingface.co/lm-kit/qwen-3-8b-instruct-gguf) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Qwen_Qwen3.5-0.8B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-0.8B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [bartowski/Qwen_Qwen3.5-2B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-2B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [bartowski/Qwen_Qwen3.5-4B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3.5-4B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
-| [lmstudio-community/Qwen3.5-9B-Q4_K_M](https://huggingface.co/lmstudio-community/Qwen3.5-9B-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
+| [Qwen/qwen2.5-1.5b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [Qwen/qwen2.5-coder-7b-instruct-q4_k_m](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Qwen_Qwen3-0.6B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-0.6B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Qwen_Qwen3-1.7B-Q4_K_M](https://huggingface.co/bartowski/Qwen_Qwen3-1.7B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [Qwen/Qwen3-4B-Q4_K_M](https://huggingface.co/Qwen/Qwen3-4B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [lm-kit/Qwen3-8B-Q4_K_M](https://huggingface.co/lm-kit/qwen-3-8b-instruct-gguf) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
-| [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
+| [unsloth/gemma-3-4b-it-Q4_K_M](https://huggingface.co/unsloth/gemma-3-4b-it-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/google_gemma-4-E2B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E2B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
 | [bartowski/google_gemma-4-E4B-it-Q4_K_M](https://huggingface.co/bartowski/google_gemma-4-E4B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
-| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✓ |
+| [bartowski/gemma-4-12B-it-Q4_K_M](https://huggingface.co/bartowski/gemma-4-12B-it-GGUF) | ✓ / ✗ | ✓ / ✗ | ✗ |
 |  |  |  |  |
-| [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/microsoft_Phi-4-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/microsoft_Phi-4-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [bartowski/Phi-3-mini-4k-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3-mini-4k-instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/Phi-3.5-mini-instruct-Q4_K_M](https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [bartowski/Mistral-7B-Instruct-v0.3-Q4_K_M](https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [QuantFactory/Ministral-3b-instruct.Q4_K_M](https://huggingface.co/QuantFactory/Ministral-3b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 | [bartowski/Ministral-8B-Instruct-2410-Q4_K_M](https://huggingface.co/bartowski/Ministral-8B-Instruct-2410-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [bartowski/DeepSeek-R1-Distill-Llama-8B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Llama-8B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [bartowski/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M](https://huggingface.co/bartowski/DeepSeek-R1-Distill-Qwen-7B-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [ibm-granite/granite-4.0-350m-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-350m-GGUF) | ✓ / ✓ | ✗ / ✗ | ✓ |
 | [ibm-granite/granite-4.0-micro-Q4_K_M](https://huggingface.co/ibm-granite/granite-4.0-micro-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
@@ -138,10 +133,10 @@ Although, the validated models below were tested with `llama-cli` using the `Q4_
 | [ibm-research/granite-3.2-8b-instruct-Q4_K_M](https://huggingface.co/ibm-research/granite-3.2-8b-instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
 |  |  |  |  |
 | [HuggingFaceTB/smollm2-1.7b-instruct-q4_k_m](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [openbmb/MiniCPM-V-2_6-Q4_K_M](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/tencent_Hunyuan-7B-Instruct-Q4_K_M](https://huggingface.co/bartowski/tencent_Hunyuan-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-Q4_K_M](https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
-| [bartowski/prism-ml_Bonsai-8B-unpacked-Q4_K_M](https://huggingface.co/bartowski/prism-ml_Bonsai-8B-unpacked-GGUF) | ✓ / ✓ | ✓ / ✓ | ✓ |
+| [openbmb/MiniCPM-V-2_6-Q4_K_M](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/tencent_Hunyuan-7B-Instruct-Q4_K_M](https://huggingface.co/bartowski/tencent_Hunyuan-7B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-Q4_K_M](https://huggingface.co/LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
+| [bartowski/prism-ml_Bonsai-8B-unpacked-Q4_K_M](https://huggingface.co/bartowski/prism-ml_Bonsai-8B-unpacked-GGUF) | ✓ / ✓ | ✓ / ✗ | ✓ |
 |  |  |  |  |
 | [gpustack/bge-m3-Q4_K_M.gguf](https://huggingface.co/gpustack/bge-m3-GGUF) | ✓ | ✗ | ✗ |
 
@@ -222,18 +217,18 @@ cmake --build build\ReleaseOV --parallel
 > [!NOTE]
 > The Windows install path is `C:\Intel\openvino` (no spaces) to avoid quoting problems some CMake/Ninja toolchains have with `C:\Program Files (x86)\...`. Adjust to wherever you installed OpenVINO Runtime. From `cmd`, run `C:\Intel\openvino\setupvars.bat`; from PowerShell, run `& "C:\Intel\openvino\setupvars.ps1"` instead. Once the build is finished you can launch the binaries from any `cmd` or `PowerShell` window after sourcing the matching `setupvars` script for that shell.
 
-#### Ubuntu Build Script
+#### Automated Ubuntu Build Script
 
 For Ubuntu24 users, the following shell script automates the prerequisite installs (build tools, OpenCL ICD), the OpenVINO Runtime download/extract/setup, and the Ninja-based llama.cpp build.
-Save the following as `build-llamacpp-ov.sh` next to where you want the `llama.cpp` folder to land, then run it:
+Save the following as `ubuntu-llamacpp-ov-install.sh` next to where you want the `llama.cpp` folder to land, then run it:
 
 ```bash
-chmod +x build-llamacpp-ov.sh
-./build-llamacpp-ov.sh
+chmod +x ubuntu-llamacpp-ov-install.sh
+./ubuntu-llamacpp-ov-install.sh
 ```
 
 <details>
-<summary>Click to expand <code>build-llamacpp-ov.sh</code></summary>
+<summary>Click to expand <code>ubuntu-llamacpp-ov-install.sh</code></summary>
 
 ```bash
 #!/usr/bin/env bash
@@ -242,8 +237,8 @@ chmod +x build-llamacpp-ov.sh
 # ============================================
 set -euo pipefail
 
-OPENVINO_VERSION_MAJOR="2026.3.1"
-OPENVINO_VERSION_FULL="2026.3.1.22476.56d9685302d"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -318,9 +313,8 @@ fi
 echo "============================================"
 echo "Configuring with CMake..."
 echo "============================================"
-set +u
+# shellcheck disable=SC1091
 source "${OPENVINO_ROOT}/setupvars.sh"
-set -u
 
 cmake -B build/ReleaseOV -G Ninja \
     -DCMAKE_BUILD_TYPE=Release \
@@ -340,27 +334,27 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
 
 </details>
 
-#### Windows Build Script
+#### Automated Windows Build Script
 
 For Windows users, the following `.bat` script automates the prerequisite installs (Git, Ninja, CMake, Visual Studio 2022 Build Tools, vcpkg + OpenCL), the OpenVINO Runtime download/extract, and the Ninja-based llama.cpp build.
-Save the following as `build-llamacpp-ov.bat` next to where you want the `llama.cpp` to land, then run it from either **Command Prompt** or **PowerShell**:
+Save the following as `windows-llamacpp-ov-install.bat` next to where you want the `llama.cpp` to land, then run it from either **Command Prompt** or **PowerShell**:
 
 ```cmd
 :: Command Prompt
-build-llamacpp-ov.bat
+windows-llamacpp-ov-install.bat
 ```
 
 ```powershell
 # PowerShell
-.\build-llamacpp-ov.bat
+.\windows-llamacpp-ov-install.bat
 ```
 
 <details>
-<summary>Click to expand <code>build-llamacpp-ov.bat</code></summary>
+<summary>Click to expand <code>windows-llamacpp-ov-install.bat</code></summary>
 
 ```bat
 @echo off
@@ -370,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================
 
-set "OPENVINO_VERSION_MAJOR=2026.3.1"
-set "OPENVINO_VERSION_FULL=2026.3.1.22476.56d9685302d"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"
 
 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -459,6 +453,9 @@ if exist "%OPENVINO_INSTALL_DIR%\setupvars.bat" (
     )
 
     REM Move the single top-level folder contents into the versioned install dir.
+    REM NOTE: delayed expansion (!VAR!) is required because the surrounding else( ... )
+    REM block is parsed once up-front, so %OPENVINO_EXTRACTED% would expand to "" here
+    REM and xcopy would then treat "\*" as C:\* and fail with "Cannot perform a cyclic copy".
     set "OPENVINO_EXTRACTED="
     for /d %%i in ("%OPENVINO_EXTRACT_TMP%\*") do set "OPENVINO_EXTRACTED=%%i"
     if not defined OPENVINO_EXTRACTED (
@@ -550,7 +547,7 @@ endlocal
 ```
 
 > [!NOTE]
-> The script pins OpenVINO `2026.3.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
 
 </details>
 
@@ -715,7 +712,6 @@ Boolean flags follow a uniform convention: set to a **positive integer** (e.g. `
 | `GGML_OPENVINO_CACHE_DIR`         | String    | `not set`  | Directory for OpenVINO model caching (recommended: `/tmp/ov_cache`). Enables model caching when set. **Not supported on NPU devices.** |
 | `GGML_OPENVINO_COMPILED_MODEL_CACHE_DIR` | String | `not set` | Directory for the frontend compiled-model cache. When set, OpenVINO compiled models are exported as blobs and imported on later runs to skip weight requantization, graph conversion, and compilation for matching single-graph models. |
 | `GGML_OPENVINO_PREFILL_CHUNK_SIZE`| Integer   | `256`      | Token chunk size for **NPU** prefill (NPU-only; ignored on CPU/GPU). Must be a positive integer; otherwise the default is used. |
-| `GGML_OPENVINO_NPU_COMPILE_CONFIG` | String | `not set` | NPU-only compiler mode parameters forwarded to OpenVINO as `NPU_COMPILATION_MODE_PARAMS`, for example `optimization-level=3`. |
 | `GGML_OPENVINO_STATEFUL_EXECUTION`| Boolean   | `0`        | Enable stateful KV cache for better performance. Recommended on CPU, GPU.                                   |
 | `GGML_OPENVINO_DISABLE_CACHE`     | Boolean   | `0`        | Disable the in-process compiled-model / decoder cache (cache is on by default). Set to `1` to disable.      |
 | `GGML_OPENVINO_DISABLE_KV_SLICE`  | Boolean   | `0`        | Disable the KV-cache input-tensor slicing optimization (slicing is on by default on CPU/GPU). Set to `1` to disable. |
@@ -729,11 +725,9 @@ Boolean flags follow a uniform convention: set to a **positive integer** (e.g. `
 | `GGML_OPENVINO_DEBUG_INPUT`       | Boolean   | `0`        | Enable input debugging and print input tensor info.                                                         |
 | `GGML_OPENVINO_DEBUG_OUTPUT`      | Boolean   | `0`        | Enable output debugging and print output tensor info.                                                       |
 | `GGML_OPENVINO_PRINT_CGRAPH_TENSOR_ADDRESS` | Boolean | `0` | Print tensor address map once.                                                                           |
-| `GGML_OPENVINO_LOG_UNSUPPORTED_OPS`| Boolean   | `0`        | Log warning messages with tensor details and rejection reasons for any ops not supported by the OpenVINO backend. Emits at `WARN` level (requires `--log-verbosity >= 2`, enabled by default). |
 
 > [!NOTE]
-> - `GGML_OPENVINO_STATEFUL_EXECUTION` is an **Experimental** feature to allow stateful execution for managing the KV cache internally inside the OpenVINO model, improving performance on CPUs and GPUs. Stateful execution is not effective on NPUs, and not all models currently support this feature. This feature is experimental and has been validated only with the llama-simple, llama-cli, llama-bench, and llama-run applications and is recommended to enable for the best performance. Other applications, such as llama-server and llama-perplexity, are not yet supported.
-> - `GGML_OPENVINO_LOG_UNSUPPORTED_OPS` emits logs at `WARN` level (`GGML_LOG_WARN`), which requires application log verbosity `--log-verbosity >= 2` (or `-lv 2`).
+>`GGML_OPENVINO_STATEFUL_EXECUTION` is an **Experimental** feature to allow stateful execution for managing the KV cache internally inside the OpenVINO model, improving performance on CPUs and GPUs. Stateful execution is not effective on NPUs, and not all models currently support this feature. This feature is experimental and has been validated only with the llama-simple, llama-cli, llama-bench, and llama-run applications and is recommended to enable for the best performance. Other applications, such as llama-server and llama-perplexity, are not yet supported.
 
 ### Example Usage
 

docs/backend/SYCL.md

@@ -795,9 +795,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_FLASH_ATTN | 1 (default) or 0| Enable Flash-Attention. It can reduce memory usage. The performance impact depends on the LLM.|
 | GGML_SYCL_ENABLE_OPT | 0 or 1 (default)| Enable optimize features for Intel GPUs. (Recommended to 0 for Intel devices older than Gen 10) |
 | GGML_SYCL_ENABLE_GRAPH | 0 (default) or 1 | Enable running computations through SYCL Graphs feature. Disabled by default because SYCL Graph is still on development, no better performance. |
-| GGML_SYCL_ENABLE_HOST_PINNED_MEM | 0 or 1 (default) | Enable host pinned memory to speed up copy data from host to device. When disable it, host memory will common malloc() on CPU. Disable it when use `--load-model mlock`.|
-| GGML_SYCL_HOST_PINNED_MEM_2G | 0 (default) or 1 | Limit the max memory allocation to be no more than 2GB when enable host pinned memory. USM allocations above 2 GiB take the relaxed/large-allocation path, which serializes H2D copies with compute and prevents copy/compute overlap. It will impact the startup time. Need more test. Depend on `GGML_SYCL_ENABLE_HOST_PINNED_MEM=1`.|
-| GGML_SYCL_GET_MEM_API | 0 (default) or 1  | Set to get memory info (free, total) by Level Zero or SYCL API:<br>0 - Level Zero API: support more GPUs, only run on Level Zero running time. When there is an error, fallback to call SYCL API. Depend on GGML_SYCL_SUPPORT_LEVEL_ZERO_API.<br>1 - SYCL API: legacy, support more running time, it can't get the free size of some GPUs (like Arc770). In such case, return total size for free size.|
+| GGML_SYCL_ENABLE_HOST_PINNED_MEM | 0 or 1 (default) | Enable host pinned memory to speed up copy data from host to device. When disable it, host memory will common malloc() on CPU.|
 | GGML_SYCL_USE_LEVEL_ZERO_API | 1 (default) or 0 | Use Level Zero API for device memory allocation instead of SYCL. Reduces system RAM usage on Intel dGPUs by avoiding DMA-buf/TTM host memory staging. Requires GGML_SYCL_SUPPORT_LEVEL_ZERO_API=ON at build time. SYCL backend always runs on Level Zero running time even if it's set as OFF (The SYCL api will be usage for memory allocation).|
 | GGML_SYCL_ENABLE_DNN | 0 or 1 (default)| Enable running computations through oneDNN and always use oneMKL. |
 | GGML_SYCL_FA_ONEDNN | 1 (default) or 0 | Enable the oneDNN fused SDPA (flash-attention) path on supported GPUs. Set to 0 to always use the native SYCL flash-attention kernel. |

docs/backend/snapdragon/CMakeUserPresets.json

@@ -8,7 +8,7 @@
         "toolset":      { "value": "host=x86_64", "strategy": "external" },
         "cacheVariables": {
             "ANDROID_ABI":      "arm64-v8a",
-            "ANDROID_PLATFORM": "android-34",
+            "ANDROID_PLATFORM": "android-31",
             "CMAKE_TOOLCHAIN_FILE": "$env{ANDROID_NDK_ROOT}/build/cmake/android.toolchain.cmake",
             "CMAKE_C_FLAGS":   "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",
             "CMAKE_CXX_FLAGS": "-march=armv8.7a+fp16+dotprod+i8mm -fvectorize -ffp-model=fast -fno-finite-math-only -flto -D_GNU_SOURCE",

docs/backend/snapdragon/README.md

@@ -2,47 +2,39 @@
 
 ## Setup
 
-The cross-compilation toolchain images are provided by the
-[Qualcomm Snapdragon Toolchain registry](https://github.com/snapdragon-toolchain).
-These Docker images include the Android NDK, OpenCL SDK, Hexagon SDK, CMake, and the necessary cross-compilers:
+### Android
 
-* **Android toolchain**: `ghcr.io/snapdragon-toolchain/arm64-android:v0.7`
-* **Linux toolchain**: `ghcr.io/snapdragon-toolchain/arm64-linux:v0.7`
+The easiest way to build llama.cpp for a Snapdragon-based Android device is using the toolchain Docker image (see github.com/snapdragon-toolchain).
+This image includes Android NDK, OpenCL SDK, Hexagon SDK, CMake, etc.
 
-The unified build utility (`scripts/snapdragon/build.py`) automatically pulls
-and orchestrates these containers to perform target compilation.
-You only need to ensure that Docker (or Docker Desktop on macOS/Windows) is running on your host machine.
-Specific setup, build, and installation details for Linux and Windows on Snapdragon platforms are documented in:
-* [Linux on Snapdragon guide](linux.md)
-* [Windows on Snapdragon guide](windows.md)
+This method works on Linux, macOS, and Windows. macOS and Windows users should install Docker Desktop.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.7
+[d]/> cd /workspace
+```
+
+Note: The rest of the **Android** build process assumes that you're running inside the toolchain container.
+
+### Windows On Snapdragon
+
+Native Windows 11 arm64 builds has the following tools dependencies:
+- MS Visual Studio 2026 (Community Edition or Pro)
+  - MSVC arm64 standard and runtime libraries
+  - UCRT and Driver Kit
+- LLVM core libraries and Clang compiler (winget)
+- CMake, Git, Python (winget)
+- Hexagon SDK Community Edition 6.6 or later (see windows.md)
+- OpenCL SDK 2.3 or later (see windows.md)
+
+Note: The rest of the **Windows** build process assumes that you're running natively in Powershell.
+Adapt below build commands accordingly.
 
 ## How to Build
 
-### Using build.py script (Recommended)
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
-The easiest way to build llama.cpp is by using the `scripts/snapdragon/build.py` script. It automatically copies the CMake presets,
-launches the correct compilation Docker container, builds the libraries and tools,
-installs them, and optionally pushes them to your ADB device.
-
-Build and deploy for Android target (accepts `android` or `adb` alias):
 ```
-$ ./scripts/snapdragon/build.py --target adb --push
-```
-
-Build and deploy for Linux target (accepts `linux` or `lnx` alias):
-```
-$ ./scripts/snapdragon/build.py --target linux:user@host --push
-```
-
-### Manual CMake Build
-
-Alternatively, you can build llama.cpp manually by entering the cross-compilation Docker container and running the CMake commands:
-
-```bash
-# Start the cross-compilation container manually:
-~/src/llama.cpp$ docker run -it --rm -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-android:v0.7
-
-# Inside the container, build the project using presets:
 [d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-android-snapdragon-release -B build-snapdragon
@@ -76,19 +68,19 @@ Preset CMake variables:
 To generate an installable "package" simply use cmake --install:
 
 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-android/llama.cpp
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon/llama.cpp
 -- Install configuration: "Release"
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-cpu.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-opencl.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-hexagon.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v73.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v75.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v79.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml-htp-v81.so
--- Installing: /workspace/pkg-android/llama.cpp/lib/libggml.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-cpu.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-opencl.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-hexagon.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v73.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v75.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v79.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml-htp-v81.so
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/lib/libggml.so
 ...
--- Installing: /workspace/pkg-android/llama.cpp/bin/llama-bench
--- Installing: /workspace/pkg-android/llama.cpp/bin/llama-cli
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-cli
 ...
 ```
 
@@ -99,14 +91,14 @@ To generate an installable "package" simply use cmake --install:
 For this step, your device needs to be configured for on-device development.
 Please see https://developer.android.com/studio/debug/dev-options for details.
 
-Once ADB is enabled, use `adb push` to install `pkg-android` on the device.
+Once ADB is enabled, use `adb push` to install `pkg-snapdragon` on the device.
 **Note that the toolchain Docker image doesn't have ADB and doesn't set up the ADB bridge. Please use native ADB on the host.**
 
 ```
-~/src/llama.cpp$ adb push pkg-android/llama.cpp /data/local/tmp/
-pkg-android/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
-pkg-android/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
-pkg-android/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
+~/src/llama.cpp$ adb push pkg-snapdragon/llama.cpp /data/local/tmp/
+pkg-snapdragon/llama.cpp/bin/: 67 files pushed, 0 skipped. 190.2 MB/s (919095042 bytes in 4.607s)
+pkg-snapdragon/llama.cpp/include/: 19 files pushed, 0 skipped. 20.5 MB/s (255173 bytes in 0.012s)
+pkg-snapdragon/llama.cpp/lib/: 16 files pushed, 0 skipped. 144.4 MB/s (43801382 bytes in 0.289s)
 102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
 ```
 
@@ -123,44 +115,24 @@ Llama-3.2-1B-Instruct-Q4_0.gguf: 1 file pushed, 0 skipped. 38.3 MB/s (773025920
 
 ### Windows
 
-All artifacts are already installed in the `pkg-wos` folder.
-To run, you can use the `scripts/snapdragon/run.py` runner script (see details below).
+All artifacts are already installed in the `pkg-snapdragon` folder.
+To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/windows`.
 
 ## How to Run
 
-The easiest way to run llama.cpp cli tools is using the provided `scripts/snapdragon/run.py` wrapper script. This script automatically
-maps CLI options to environment variables, resolves executable paths, and runs the command locally, via ADB, or remotely via SSH on the
-target device.
+The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
 
-llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU.
-You can select which backend(s) to run the model on using the `--device` option of the tool (or `--devices` option in `run.py`).
+llama.cpp supports three backends on Snapdragon-based devices: CPU, Adreno GPU (GPUOpenCL), and Hexagon NPU (HTP0-4).
+You can select which backend to run the model on using the `D=` variable, which maps to the `--device` option.
 
 Hexagon NPU behaves as a "GPU" device when it comes to `-ngl` and other offload-related options.
 
-Here are some examples of running various llama.cpp tools.
+Here are some examples of running various llama.cpp tools via ADB.
 
-Generating a completion with Gemma on Android (relying on default `HTP0:0` device and default thread count `-t 6`):
+Simple question for Llama-3.2-1B
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb -- llama-completion -m models/gemma-2-2b-it-Q4_0.gguf -f prompts/sample_prompt_1024.txt --jinja -st
-...
-ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
-ggml-hex: Hexagon Arch version v79
-ggml-hex: allocating new session: HTP0:0
-...
-load_tensors: offloading output layer to GPU
-load_tensors: offloaded 27/27 layers to GPU
-load_tensors:          CPU model buffer size =   300.00 MiB
-load_tensors:      HTP0:0 model buffer size  =  1400.26 MiB
-...
-llama_perf_context_print: prompt eval time =     320.00 ms /  1024 tokens (    0.31 ms per token,  3200.00 tokens per second)
-llama_perf_context_print:        eval time =     2100.00 ms /   100 runs   (   21.00 ms per token,    47.62 tokens per second)
-```
-
-Simple question for Llama-3.2-1B:
-
-```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target android --devices HTP0 -- llama-cli -m Llama-3.2-1B-Instruct-Q4_0.gguf -p "what is the most popular cookie in the world?"
+~/src/llama.cpp$ M=Llama-3.2-1B-Instruct-Q4_0.gguf D=HTP0 ./scripts/snapdragon/adb/run-completion.sh -p "what is the most popular cookie in the world?"
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -170,7 +142,8 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 load_tensors: offloading output layer to GPU
 load_tensors: offloaded 17/17 layers to GPU
 load_tensors:          CPU model buffer size =   225.49 MiB
-load_tensors:         HTP0 model buffer size =   504.26 MiB
+load_tensors:         HTP0 model buffer size =     0.26 MiB
+load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
 ...
 I hope this helps you understand the world's most popular cookies! [end of text]
 ...
@@ -183,25 +156,60 @@ llama_perf_context_print:    graphs reused =        473
 llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
 llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
 llama_memory_breakdown_print: |   - Host               |                  439 =   225 +     136 +      77                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                  504 =   504 +       0 +       0                |
 ```
 
-Op test for MUL_MAT:
+Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --hex-hostbuf 0 --devices HTP0:0 -- test-backend-ops -b HTP0:0 -o MUL_MAT
+~/src/llama.cpp$ M=OLMoE-1B-7B-0125-Instruct-Q4_0.gguf NDEV=2 D=HTP0,HTP1 ./scripts/snapdragon/adb/run-completion.sh -f surfing.txt
 ...
-Backend 2/3: HTP0:0
+ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
+ggml-hex: Hexagon Arch version v81
+ggml-hex: allocating new session: HTP0
+ggml-hex: allocating new session: HTP1
+...
+load_tensors: offloading output layer to GPU
+load_tensors: offloaded 17/17 layers to GPU
+load_tensors:          CPU model buffer size =   143.86 MiB
+load_tensors:         HTP1 model buffer size =     0.23 MiB
+load_tensors:  HTP1-REPACK model buffer size =  1575.00 MiB
+load_tensors:         HTP0 model buffer size =     0.28 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2025.00 MiB
+...
+llama_context:        CPU  output buffer size =     0.19 MiB
+llama_kv_cache:       HTP1 KV buffer size =   238.00 MiB
+llama_kv_cache:       HTP0 KV buffer size =   306.00 MiB
+llama_kv_cache: size =  544.00 MiB (  8192 cells,  16 layers,  1/1 seqs), K (q8_0):  272.00 MiB, V (q8_0):  272.00 MiB
+llama_context:       HTP0 compute buffer size =    15.00 MiB
+llama_context:       HTP1 compute buffer size =    15.00 MiB
+llama_context:        CPU compute buffer size =    24.56 MiB
+...
+llama_perf_context_print: prompt eval time =    1730.57 ms /   212 tokens (    8.16 ms per token,   122.50 tokens per second)
+llama_perf_context_print:        eval time =    5624.75 ms /   257 runs   (   21.89 ms per token,    45.69 tokens per second)
+llama_perf_context_print:       total time =    7377.33 ms /   469 tokens
+llama_perf_context_print:    graphs reused =        255
+llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - Host               |                  742 =   144 +     544 +      54                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 1575 =  1575 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2025 =  2025 +       0 +       0                |
+```
+
+Op test for MUL_MAT
+
+```
+~/src/llama.cpp$ HB=0 ./scripts/snapdragon/adb/run-tool.sh test-backend-ops -b HTP0 -o MUL_MAT
+...
+Backend 2/3: HTP0
 Device description: Hexagon
 Device memory: 2048 MB (2048 MB free)
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
 MUL_MAT(type_a=q4_0,type_b=f32,m=16,n=3,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],v=0,o=1): OK
-```
 
-Llama benchmark:
-
-```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0 -- llama-bench -p 128 -n 64 -m Llama-3.2-1B-Instruct-Q4_0.gguf
+~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -211,20 +219,15 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
 | ---------------| ---------: | -----: | ---------- | --: | ------: | ------: | ---: | ----: | ------------: |
 | llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 | pp128 | 169.42 ± 1.75 |
 | llama 1B Q4_0  | 729.75 MiB | 1.24 B | HTP        |  99 |       4 |     128 |    0 |  tg64 |  51.54 ± 1.13 |
+
+build: 6a8cf8914 (6733)
 ```
 
 ## Environment variables
 
-- `GGML_HEXAGON_DEVICES` (default: not set, defaults to HTP0 session)
-  Controls which NPU devices and sessions to allocate. Can be configured as:
-  - A single integer `N`: Allocates `N` sessions named `HTP0`, `HTP1`, ..., `HTP<N-1>` (behaves identically to `GGML_HEXAGON_NDEV=N`).
-  - A comma-separated list of device names in `HTP<physical_idx>:<virtual_idx>` format (or legacy `HTP<idx>` format). For example, `HTP0:0,HTP0:1` creates two virtual
-    sessions on the first physical NPU (useful for memory limits). `HTP0:0,HTP1:0` allocates one session on each of the two physical NPUs
-    on a dual-NPU device.
-
-- `GGML_HEXAGON_NDEV` (deprecated)
-  Replaced by `GGML_HEXAGON_DEVICES`. Controls the number of virtual sessions to allocate on physical NPU `0`.
-  Allocates sessions named `HTP0`, `HTP1`, etc.
+- `GGML_HEXAGON_NDEV=1`
+  Controls the number of devices/sessions to allocate. The default is 1.
+  Most quantized models under 4B fit into a single session; an 8B model needs two, and a 20B model needs four.
 
 - `GGML_HEXAGON_NHVX=0`
   Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
@@ -252,17 +255,26 @@ ggml-hex: new session: HTP0 : session-id 0 domain-id 3 uri file:///libggml-htp-v
   - `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.
+  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 ./scripts/snapdragon/run.py --target adb -- llama-cli ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`
+      `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_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:
 
   Examples:
 
-      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
-      `GGML_HEXAGON_OPFILTER="ADD\|SUB" ./scripts/snapdragon/run.py --target adb -- llama-cli ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)
-
+      `GGML_HEXAGON_OPFILTER="FLASH_ATTN_EXT" llama-completion ...` - Disable Flash Attention on Hexagon (falls back to CPU or GPU)
+      `GGML_HEXAGON_OPFILTER="ADD\|SUB" llama-completion ...` - Disable ADD and SUB on Hexagon (fall back to CPU or GPU)

docs/backend/snapdragon/developer.md

@@ -39,21 +39,22 @@ the repacking.
 
 ## Large model handling
 
-Hexagon NPU sessions (aka Process Domains (PD) in the Hexagon SDK) are limited to a maximum memory mapping window of around 3.5GB.
-In llama.cpp/GGML, each Hexagon session is mapped to a single GGML backend device (e.g., `HTP0:0`, `HTP0:1`, etc. when using
-`GGML_HEXAGON_DEVICES`, or `HTP0`, `HTP1` in legacy mode).
+Hexagon NPU session (aka Process Domain (PD) in the Hexagon docs) is limited to a memory mapping of around 3.5GB.
+In llama.cpp/GGML the Hexagon session is mapped to a single GGML backend device (HTP0, HTP1, etc).
 
-To support running models larger than 3.5GB on a single device, the Hexagon backend dynamically maps and unmaps execution buffers
-during the graph execution cycle to stay within the Process Domain window. This enables large models to run successfully on a single
-NPU device.
+In order to map models larger than 3.5GB we need to allocate multiple devices and split the model.
+For this we're taking advantage of the llama.cpp/GGML multi-GPU layer-splitting support.
+Each Hexagon device behaves like a GPU from the offload and model splitting perspective.
 
-Alternatively, users can choose to use standard llama.cpp/GGML layer-splitting mode to partition and split the model across
-multiple Hexagon devices or virtual sessions (which behave like multiple GPUs from the offload and splitting perspective).
-
-Here is an example of running GPT-OSS-20B model on a Snapdragon device using 4 virtual sessions on a single NPU (physical index 0).
+Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
 
 ```
-~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --devices HTP0:0,HTP0:1,HTP0:2,HTP0:3 -- llama-cli --load-mode none -m /data/local/tmp/gguf/gpt-oss-20b-Q4_0.gguf -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 -no-cnv -f surfing.txt
+M=gpt-oss-20b-Q4_0.gguf NDEV=4 D=HTP0,HTP1,HTP2,HTP3 P=surfing.txt scripts/snapdragon/adb/run-completion.sh -f surfing.txt -n 32
+...
+LD_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+ADSP_LIBRARY_PATH=/data/local/tmp/llama.cpp/lib
+GGML_HEXAGON_NDEV=4 ./bin/llama-cli --load-mode none -m /data/local/tmp/llama.cpp/../gguf/gpt-oss-20b-Q4_0.gguf
+      -t 4 --ctx-size 8192 --batch-size 128 -ctk q8_0 -ctv q8_0 -fa on -ngl 99 --device HTP0,HTP1,HTP2,HTP3 -no-cnv -f surfing.txt
 ...
 llama_model_loader: - type  f32:  289 tensors
 llama_model_loader: - type q4_0:   96 tensors
@@ -62,29 +63,33 @@ llama_model_loader: - type mxfp4:  72 tensors
 ...
 load_tensors: offloaded 25/25 layers to GPU
 load_tensors:          CPU model buffer size =  1182.09 MiB
-load_tensors:       HTP0:1 model buffer size =  2512.58 MiB
-load_tensors:       HTP0:3 model buffer size =  2093.83 MiB
-load_tensors:       HTP0:0 model buffer size =  2931.34 MiB
-load_tensors:       HTP0:2 model buffer size =  2512.58 MiB
+load_tensors:         HTP1 model buffer size =     6.64 MiB
+load_tensors:  HTP1-REPACK model buffer size =  2505.94 MiB
+load_tensors:         HTP3 model buffer size =     5.55 MiB
+load_tensors:  HTP3-REPACK model buffer size =  2088.28 MiB
+load_tensors:         HTP0 model buffer size =     7.75 MiB
+load_tensors:  HTP0-REPACK model buffer size =  2923.59 MiB
+load_tensors:         HTP2 model buffer size =     6.64 MiB
+load_tensors:  HTP2-REPACK model buffer size =  2505.94 MiB
 ...
 llama_context: n_ctx_per_seq (8192) < n_ctx_train (131072) -- the full capacity of the model will not be utilized
 llama_context:        CPU  output buffer size =     0.77 MiB
 llama_kv_cache_iswa: creating non-SWA KV cache, size = 8192 cells
-llama_kv_cache:     HTP0:1 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =    25.50 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP1 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP3 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP0 KV buffer size =    25.50 MiB
+llama_kv_cache:       HTP2 KV buffer size =    25.50 MiB
 llama_kv_cache: size =  102.00 MiB (  8192 cells,  12 layers,  1/1 seqs), K (q8_0):   51.00 MiB, V (q8_0):   51.00 MiB
 llama_kv_cache_iswa: creating     SWA KV cache, size = 256 cells
-llama_kv_cache:     HTP0:1 KV buffer size =     0.80 MiB
-llama_kv_cache:     HTP0:3 KV buffer size =     0.53 MiB
-llama_kv_cache:     HTP0:0 KV buffer size =     1.06 MiB
-llama_kv_cache:     HTP0:2 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP1 KV buffer size =     0.80 MiB
+llama_kv_cache:       HTP3 KV buffer size =     0.53 MiB
+llama_kv_cache:       HTP0 KV buffer size =     1.06 MiB
+llama_kv_cache:       HTP2 KV buffer size =     0.80 MiB
 llama_kv_cache: size =    3.19 MiB (   256 cells,  12 layers,  1/1 seqs), K (q8_0):    1.59 MiB, V (q8_0):    1.59 MiB
-llama_context:     HTP0:0 compute buffer size =    16.06 MiB
-llama_context:     HTP0:1 compute buffer size =    16.06 MiB
-llama_context:     HTP0:2 compute buffer size =    16.06 MiB
-llama_context:     HTP0:3 compute buffer size =    16.06 MiB
+llama_context:       HTP0 compute buffer size =    16.06 MiB
+llama_context:       HTP1 compute buffer size =    16.06 MiB
+llama_context:       HTP2 compute buffer size =    16.06 MiB
+llama_context:       HTP3 compute buffer size =    16.06 MiB
 llama_context:        CPU compute buffer size =    98.19 MiB
 ...
 llama_perf_context_print: prompt eval time =    3843.67 ms /   197 tokens ( 19.51 ms per token, 51.25 tokens per second)
@@ -92,9 +97,13 @@ llama_perf_context_print:        eval time =    1686.13 ms /    31 runs   ( 54.3
 llama_perf_context_print:       total time =    6266.30 ms /   228 tokens
 llama_perf_context_print:    graphs reused =         30
 llama_memory_breakdown_print: | memory breakdown [MiB] | total   free    self   model   context   compute    unaccounted |
-llama_memory_breakdown_print: |   - HTP0:0 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:1 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:2 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
-llama_memory_breakdown_print: |   - HTP0:3 (Hexagon)   |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP0 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP1 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP2 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
+llama_memory_breakdown_print: |   - HTP3 (Hexagon)     |  2048 = 2048 + (   0 =     0 +       0 +       0) +           0 |
 llama_memory_breakdown_print: |   - Host               |                 1476 =  1208 +     105 +     162                |
+llama_memory_breakdown_print: |   - HTP1-REPACK        |                 2505 =  2505 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP3-REPACK        |                 2088 =  2088 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP0-REPACK        |                 2923 =  2923 +       0 +       0                |
+llama_memory_breakdown_print: |   - HTP2-REPACK        |                 2505 =  2505 +       0 +       0                |
 ```

docs/backend/snapdragon/linux.md

@@ -1,37 +1,25 @@
 # Snapdragon-based Linux devices
 
-The cross-compilation is performed using the Snapdragon Linux Docker toolchain image (see
-[github.com/snapdragon-toolchain](https://github.com/snapdragon-toolchain)):
+## Docker Setup
 
-* **Linux toolchain**: `ghcr.io/snapdragon-toolchain/arm64-linux:v0.7`
+The easiest way to build llama.cpp for a Snapdragon-based Linux device is using the toolchain Docker image (see [github.com/snapdragon-toolchain](https://github.com/snapdragon-toolchain)).
+This image includes OpenCL SDK, Hexagon SDK, CMake, and the ARM64 Linux cross-compilation toolchain.
 
-The unified build utility (`scripts/snapdragon/build.py`) automatically pulls
-and orchestrates this container to perform target compilation. You only need to
-ensure that Docker is running on your host machine.
+Cross-compilation is supported on **Linux X86** hosts. The resulting binaries are deployed to and run on the target **Qualcomm Snapdragon ARM64 Linux** device.
+
+```
+~/src/llama.cpp$ docker run -it -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-linux:v0.1
+[d]/> cd /workspace
+```
+
+Note: The rest of the **Linux** build process assumes that you're running inside the toolchain container.
 
 
 ## How to Build
 
-### Using build.py script (Recommended)
+Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 
-The easiest way to build llama.cpp is by using the `scripts/snapdragon/build.py` script. It automatically copies the CMake presets,
-launches the correct compilation Docker container, builds the libraries and tools,
-installs them, and optionally pushes them to your target device.
-
-Build and deploy for a Linux target (using SSH deployment alias `lnx` or `linux`):
 ```
-$ ./scripts/snapdragon/build.py --target lnx:user@host --push
-```
-
-### Manual CMake Build
-
-Alternatively, you can build llama.cpp manually by entering the cross-compilation Docker container and running the CMake commands:
-
-```bash
-# Start the cross-compilation container manually:
-~/src/llama.cpp$ docker run -it --rm -u $(id -u):$(id -g) --volume $(pwd):/workspace --platform linux/amd64 ghcr.io/snapdragon-toolchain/arm64-linux:v0.7
-
-# Inside the container, build the project using presets:
 [d]/workspace> cp docs/backend/snapdragon/CMakeUserPresets.json .
 
 [d]/workspace> cmake --preset arm64-linux-snapdragon-release -B build-snapdragon
@@ -42,19 +30,17 @@ Alternatively, you can build llama.cpp manually by entering the cross-compilatio
 To generate an installable "package" simply use cmake --install, then zip it:
 
 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-linux
-[d]/workspace> zip -r pkg-linux.zip pkg-linux
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon
+[d]/workspace> zip -r pkg-snapdragon.zip pkg-snapdragon
 ```
 
 ## How to Install
 
-For this step, you will deploy the built binaries and libraries to the target
-Linux device. Transfer `pkg-linux.zip` to the target device, then unzip it
-and set up the environment variables:
+For this step, you will deploy the built binaries and libraries to the target Linux device. Transfer `pkg-snapdragon.zip` to the target device, then unzip it and set up the environment variables:
 
 ```
-$ unzip pkg-linux.zip
-$ cd pkg-linux
+$ unzip pkg-snapdragon.zip
+$ cd pkg-snapdragon
 $ export LD_LIBRARY_PATH=./lib
 $ export ADSP_LIBRARY_PATH=./lib
 ```
@@ -66,28 +52,7 @@ $ wget https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/
 ```
 
 ## How to Run
-You can run locally on the Snapdragon Linux device:
-```
-$ ./scripts/snapdragon/run.py --devices HTP0 -- llama-cli -m Llama-3.2-3B-Instruct-Q4_0.gguf -ngl 99 -p "what is the most popular cookie in the world?"
-```
-
-Or run remotely from your host development machine using the SSH target option:
-```
-$ ./scripts/snapdragon/run.py --target lnx:user@host --devices HTP0 -- llama-cli -m Llama-3.2-3B-Instruct-Q4_0.gguf -ngl 99 -p "what is the most popular cookie in the world?"
-```
-
-For multi-NPU systems, you can run a tensor split completion command targeting a remote Linux system:
-```
-$ ./scripts/snapdragon/run.py --target ubuntu:maxk@192.168.1.87 --device HTP0:0,HTP1:0 -- llama-completion -m models/gemma-2b-it-Q4_0.gguf -f prompts/sample_prompt_1024.txt --jinja -st --split-mode tensor --ctx-size 8192
-```
-
-This translates to the following command being executed remotely via SSH:
-```
-+ ssh maxk@192.168.1.87 "cd ~/llama.cpp && ulimit -c unlimited && LD_LIBRARY_PATH=./lib ADSP_LIBRARY_PATH=./lib GGML_HEXAGON_DEVICES=HTP0:0,HTP1:0 GGML_HEXAGON_OPPOLL=1 ./bin/llama-completion -m models/gemma-2b-it-Q4_0.gguf -f prompts/sample_prompt_1024.txt --jinja -st --split-mode tensor --ctx-size 8192 -v -n 16 --device HTP0:0,HTP1:0 -ngl 99 --ubatch-size 1024 -fa on -t 6"
-```
-
-Alternatively, you can run the binary directly on the device:
+Next, since we have setup the environment variables, we can run the llama-cli with the Hexagon backends:
 ```
 $ ./bin/llama-cli -m Llama-3.2-3B-Instruct-Q4_0.gguf --device HTP0 -ngl 99 -p "what is the most popular cookie in the world?"
 ```
-

docs/backend/snapdragon/windows.md

@@ -1,18 +1,3 @@
-# Snapdragon-based Windows devices
-
-## Tool Dependencies
-
-Native Windows 11 arm64 builds have the following tool dependencies:
-- MS Visual Studio 2026 (Community Edition or Pro)
-  - MSVC arm64 standard and runtime libraries
-  - UCRT and Driver Kit
-- LLVM core libraries and Clang compiler (winget)
-- CMake, Git, Python (winget)
-- Hexagon SDK Community Edition 6.6 or later (see below)
-- OpenCL SDK 2.3 or later (see below)
-
-Note: The rest of the **Windows** build process assumes that you're running natively in Powershell.
-
 ## Overview
 
 The document covers procedures for installing the latest GPU and NPU drivers, and OpenCL and Hexagon SDKs.
@@ -24,18 +9,7 @@ must be included in the .cat file digitally signed with a trusted certificate.
 This document covers details on how to generate personal certificate files (.pfx) and how to configure the system
 to allow for test signatures (aka test-signing).
 
-## Install Windows SDKs
-
-The recommended method is `setup-sdk.py`:
-
-```
-> python scripts\snapdragon\setup-sdk.py --list-sdk-releases
-> python scripts\snapdragon\setup-sdk.py --hexagon --opencl
-```
-
-It installs the selected SDKs under `C:\Qualcomm` and sets their corresponding environment variables for the current user. Start a new terminal after it completes; native Windows builds check all SDK paths before CMake runs.
-
-Select the SDKs to install with `--hexagon` and `--opencl`; use both to prepare a dual-backend build. To select a different available version, pass it to the SDK option, for example `--hexagon 6.4.0.2`. SDK versions install side by side, so you can switch versions without deleting an existing installation. Use `--force` to reinstall the selected SDKs. Use a new CMake build directory after each switch because CMake caches the SDK paths.
+## Install the latest Adreno OpenCL SDK
 
 Either use the trimmed down version (optimized for CI) from
 
@@ -79,8 +53,7 @@ Download the driver from
 
     https://softwarecenter.qualcomm.com/catalog/item/Qualcomm_HND
 
-After the automated installation and reboot please make sure that the Hexagon NPU device shows up in the `Device Manager`
-(under `Neural Processors`).
+After the automated installation and reboot please make sure that the Hexagon NPU device shows up in the `Device Manager` (under `Neural Processors`).
 
 If the device is not available you can try installing all components (`qcnspmcdm8380`, `qcnspmcdm8380_ext`) manually.
 The components are extracted into
@@ -157,12 +130,12 @@ However, additional settings are required for generating and signing HTP Ops lib
 
 > cmake --preset arm64-windows-snapdragon-release -B build-wos
 ...
-> cmake --install build-wos --prefix pkg-wos
+> cmake --install build-wos --prefix pkg-snapdragon
 ```
 
 Once the build is complete HTP ops libraries will be installed like this
 ```
-> dir pkg-wos/lib
+> dir pkg-snapdragon/lib
 ...
 -a----         1/22/2026   6:01 PM         187656 libggml-htp-v73.so
 -a----         1/22/2026   6:01 PM         191752 libggml-htp-v75.so
@@ -174,8 +147,8 @@ Once the build is complete HTP ops libraries will be installed like this
 The .cat file, the signature and proper certificate installation can be verified with
 
 ```
-> signtool.exe verify /v /pa .\pkg-wos\lib\libggml-htp.cat
-Verifying: .\pkg-wos\lib\libggml-htp.cat
+> signtool.exe verify /v /pa .\pkg-snapdragon\lib\libggml-htp.cat
+Verifying: .\pkg-snapdragon\lib\libggml-htp.cat
 
 Signature Index: 0 (Primary Signature)
 Hash of file (sha256): 9820C664DA59D5EAE31DBB664127FCDAEF59CDC31502496BC567544EC2F401CF
@@ -183,6 +156,6 @@ Hash of file (sha256): 9820C664DA59D5EAE31DBB664127FCDAEF59CDC31502496BC567544EC
 Signing Certificate Chain:
         Issued to: GGML.HTP.v1
 ...
-Successfully verified: .\pkg-wos\lib\libggml-htp.cat
+Successfully verified: .\pkg-snapdragon\lib\libggml-htp.cat
 ...
 ```

docs/build.md

@@ -614,100 +614,30 @@ You can test with:
 For detailed information about hardware support, setup instructions, and performance optimization, refer to [llama.cpp for ZenDNN](./backend/ZenDNN.md).
 
 ## Arm® KleidiAI™
-KleidiAI provides optimized Arm CPU microkernels used by the ggml CPU backend. Enabling it at build time makes those kernels available; it does not force every operation to use KleidiAI. At runtime, llama.cpp selects the best compatible CPU kernel from the detected CPU features, tensor type, operation shape, and active backend priority.
-
-Supported targets:
-
-| Platform | Supported ABI / architecture | Notes |
-| --- | --- | --- |
-| Linux | AArch64 / arm64 | Runtime CPU feature detection is automatic. |
-| Android | `arm64-v8a` | Use the Android NDK command below for a portable build. |
-| Apple | arm64 | Runtime CPU feature detection is automatic. Non-streaming SVE vector length is treated as unavailable. |
-| Windows | arm64 | Runtime CPU feature detection is automatic. SMCU count is treated as unknown until a detection path is verified. |
-
-`GGML_CPU_KLEIDIAI=ON` is valid only for AArch64/arm64 builds. Do not enable it for x86, 32-bit Arm, or Android ABIs other than `arm64-v8a`.
-
-### Native AArch64/arm64 build
-
-From the llama.cpp source directory:
+KleidiAI is a library of optimized microkernels for AI workloads, specifically designed for Arm CPUs. These microkernels enhance performance and can be enabled for use by the CPU backend.
 
+To enable KleidiAI, go to the llama.cpp directory and build using CMake
 ```bash
-cmake -S . -B build -DGGML_CPU_KLEIDIAI=ON
+cmake -B build -DGGML_CPU_KLEIDIAI=ON
 cmake --build build --config Release
 ```
-
-### Android arm64-v8a NDK build
-
-Set `ANDROID_NDK` to the Android NDK root, then run the following from the llama.cpp source directory. This command configures a portable Android `arm64-v8a` build with KleidiAI enabled and avoids Android dependencies that are not part of the NDK stable native API set.
-
-```bash
-cmake -S . -B build-android \
-  -DCMAKE_BUILD_TYPE=Release \
-  -DCMAKE_TOOLCHAIN_FILE="$ANDROID_NDK/build/cmake/android.toolchain.cmake" \
-  -DANDROID_ABI=arm64-v8a \
-  -DANDROID_PLATFORM=android-28 \
-  -DGGML_CPU_KLEIDIAI=ON \
-  -DGGML_NATIVE=OFF \
-  -DGGML_OPENMP=OFF \
-  -DGGML_LLAMAFILE=OFF \
-  -DLLAMA_OPENSSL=OFF
-cmake --build build-android --config Release --parallel
-cmake --install build-android --prefix {install-dir} --config Release
-```
-
-Important Android options:
-
-- `GGML_CPU_KLEIDIAI=ON` enables KleidiAI for Android `arm64-v8a`.
-- `GGML_NATIVE=OFF` is required for cross-compilation because the build host CPU is not the Android target CPU.
-- `GGML_OPENMP=OFF` avoids adding an OpenMP runtime dependency to this NDK command-line build.
-- `GGML_LLAMAFILE=OFF` avoids the llamafile backend, which is not supported on Android.
-- `LLAMA_OPENSSL=OFF` avoids depending on OpenSSL, which is not part of the Android NDK stable native API set.
-
-The Android Studio project under `examples/llama.android` enables KleidiAI automatically for `arm64-v8a`. For Android command-line CMake builds on `arm64-v8a`, pass `-DGGML_CPU_KLEIDIAI=ON` explicitly.
-
-Global -march flags such as `-march=armv8.7a` flag are not required for a portable Android `arm64-v8a` build. Global `-march` flags raise the baseline instruction set for generic code. No manual architecture-specific source selection is required; llama.cpp selects compatible KleidiAI kernels at runtime. The KleidiAI libraries internal CMake handles the -march flags for each particular kernel.
-
-### Verifying the build
-
-Run an installed or in-tree binary:
-
+You can verify that KleidiAI is being used by running
 ```bash
 ./build/bin/llama-cli -m PATH_TO_MODEL -p "What is a car?"
 ```
-
-If KleidiAI is enabled, the output contains a line similar to:
-
+If KleidiAI is enabled, the output will contain a line similar to:
 ```
 load_tensors: CPU_KLEIDIAI model buffer size =  3474.00 MiB
 ```
+KleidiAI’s microkernels implement optimized tensor operations using Arm CPU features such as dotprod, int8mm, SVE, and SME. Llama.cpp selects the most efficient kernels at runtime based on detected CPU capabilities.
+On CPUs that support SME, SME microkernels are enabled automatically using runtime detection.
+The environment variable GGML_KLEIDIAI_SME can be used to control SME behavior:
+- Not set: enable SME automatically if supported and detected.
+- 0: disable SME.
+- <n> > 0: enable SME and assume <n> available SME units (override auto detection).
+If SME is not supported by the CPU, SME microkernels are always disabled.
 
-This confirms that the model has tensors allocated through the KleidiAI CPU buffer. It does not prove that every operation, or any specific SME-family operation, used a KleidiAI microkernel. Runtime CPU features, tensor type, operation shape, and backend priority still control dispatch.
-
-Depending on the build target, another backend may have higher priority than the CPU backend. To force CPU execution for a run, disable higher priority backends at build time, for example `-DGGML_METAL=OFF`, or use a runtime device option such as `--device none` where supported.
-
-### Runtime dispatch
-
-KleidiAI microkernels use Arm CPU features such as dotprod, i8mm, SVE, and SME/SME2. Build-time configuration makes the kernels available. Runtime dispatch selects a compatible kernel for the detected CPU and operation. Older or lower-feature CPUs fall back automatically to compatible kernels.
-
-KleidiAI accelerates selected `GGML_OP_MUL_MAT` paths for F32 and common quantized formats. Exact coverage depends on the bundled KleidiAI version and the llama.cpp runtime selector, so unsupported tensor types, unsupported operation shapes, or higher priority backends may bypass KleidiAI even when the CPU supports the required Arm feature. This is also why a model may not use SME-family kernels on SME-capable hardware.
-
-The current llama.cpp KleidiAI SVE selector only enables SVE kernels when the runtime SVE vector length is known to be QK8_0 bytes, currently 32 bytes. Linux and Android query this at runtime. Apple reports SVE capability separately from userspace non-streaming SVE availability, so llama.cpp treats the SVE vector length as unknown there. Windows exposes SVE feature presence but not the runtime SVE vector length used by this selector, so that value is also treated as unknown. Windows arm64 also treats SMCU count as unknown until a detection mechanism is verified.
-
-The set of available SME-family kernels depends on the bundled KleidiAI version and the detected CPU capabilities. Production configuration does not require any KleidiAI runtime environment variables.
-
-### Diagnostics and debug overrides
-
-KleidiAI runtime environment variables are diagnostics/debug overrides, not production configuration. Leave them unset for normal use.
-
-`GGML_KLEIDIAI_SME` controls SME-family kernel selection and overrides the maximum number of threads assigned to selected quantized SME-family kernels:
-
-- Not set: use automatic runtime detection.
-- `0`: disable SME-family kernels.
-- `<n> > 0`: enable compatible SME-family kernels and allow up to `<n>` threads for quantized SME-family kernels.
-
-On Windows arm64, use `GGML_KLEIDIAI_SME=<n>` as the temporary diagnostics/debug override for SME thread-cap calibration until automatic SMCU count detection is verified.
-
-If the CPU does not support the required SME-family capability for a bundled kernel, that kernel is disabled regardless of the environment variable.
+Depending on your build target, other higher priority backends may be enabled by default. To ensure the CPU backend is used, you must disable the higher priority backends either at compile time, e.g. -DGGML_METAL=OFF, or during run-time using the command line option `--device none`.
 
 ## OpenCL
 

docs/ops.md

@@ -35,8 +35,8 @@ Legend:
 |                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |

docs/ops/Vulkan.csv

@@ -19292,10 +19292,10 @@
 "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=4096,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
 "Vulkan0","FLASH_ATTN_EXT","hsk=256,hsv=256,nh=4,nr23=[6,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
 "Vulkan0","FLASH_ATTN_EXT","hsk=128,hsv=128,nh=8,nr23=[4,1],kv=16384,nb=512,mask=1,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_K=f16,type_V=f16,permute=[0,1,2,3]","support","1","yes","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
-"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","1","yes","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[10,5,4,3]","support","0","no","Vulkan"
+"Vulkan0","CROSS_ENTROPY_LOSS_BACK","type=f32,ne=[30000,1,1,1]","support","0","no","Vulkan"
 "Vulkan0","OPT_STEP_ADAMW","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
 "Vulkan0","OPT_STEP_SGD","type=f32,ne=[10,5,4,3]","support","1","yes","Vulkan"
 "Vulkan0","GATED_DELTA_NET","type=f32,head_count=32,head_size=128,n_seq_tokens=1,n_seqs=1,v_repeat=1,permuted=0,kda=0,K=1","support","1","yes","Vulkan"

docs/speculative.md

@@ -212,15 +212,6 @@ Use `--backend-sampling` to run supported target-model samplers on the model bac
 
 Unsupported samplers and device layouts fall back to CPU sampling. Tensor split mode does not support backend sampling. A fixed seed produces repeatable random draws, but stochastic CPU and backend sampling can still select different tokens because floating-point operations can differ between implementations and devices. Use greedy sampling when exact output matching is required.
 
-### Synthetic Acceptance
-
-`llama-server` and `llama-cli` can replace normal speculative verification with synthetic decisions for benchmarking. The generated output is not valid model output because accepted draft tokens do not have to match the target model.
-
-Use exactly one of these options:
-
-- `--spec-synth-rates P0,P1,...` sets unconditional per-position acceptance probabilities. Entry `i` is the probability that the first `i+1` draft tokens are all accepted. The number of entries must match the effective maximum draft length. Values must be finite, within `[0, 1]`, and monotonically non-increasing.
-- `--spec-synth-len L` sets the target mean acceptance length, including the target token. For `K` maximum draft tokens, `L` must be within `[1, K+1]`. The server finds a constant conditional probability `p` such that `p + p^2 + ... + p^K = L - 1`, then uses unconditional rates `[p, p^2, ..., p^K]`.
-
 ### General Speculative Parameters
 
 ```

ggml/CMakeLists.txt

@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 22)
+set(GGML_VERSION_MINOR 21)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
@@ -242,8 +242,6 @@ option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"
 set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                             "ggml: metal minimum macOS version")
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
-set   (GGML_METAL_TARGET_OS "macos" CACHE STRING
-                                            "ggml: metal -mtargetos OS name (macos, ios, xros, tvos)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_OPENMP_FETCH                    "ggml: fetch LLVM OpenMP"                         OFF)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
@@ -344,6 +342,9 @@ set(GGML_PUBLIC_HEADERS
     include/gguf.h)
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+#if (GGML_METAL)
+#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
+#endif()
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 

ggml/cmake/ggml-config.cmake.in

@@ -110,16 +110,6 @@ set_and_check(GGML_INCLUDE_DIR "@PACKAGE_GGML_INCLUDE_INSTALL_DIR@")
 set_and_check(GGML_LIB_DIR "@PACKAGE_GGML_LIB_INSTALL_DIR@")
 #set_and_check(GGML_BIN_DIR "@PACKAGE_GGML_BIN_INSTALL_DIR@")
 
-if (NOT GGML_SHARED_LIB AND GGML_CPU_KLEIDIAI)
-    unset(KLEIDIAI_LIBRARY CACHE)
-    unset(KLEIDIAI_LIBRARY)
-    find_library(KLEIDIAI_LIBRARY kleidiai
-        REQUIRED
-        HINTS ${GGML_LIB_DIR}
-        NO_CMAKE_FIND_ROOT_PATH)
-    list(APPEND GGML_CPU_INTERFACE_LINK_LIBRARIES ${KLEIDIAI_LIBRARY})
-endif()
-
 if(NOT TARGET ggml::ggml)
     find_package(Threads REQUIRED)
 

ggml/include/ggml-backend.h

@@ -424,10 +424,6 @@ extern "C" {
     // Compare the output of two backends
     GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor const * const * test_nodes, size_t num_test_nodes);
 
-    // returns true for ops that may require additional memory for fleeting data on some backends,
-    // i.e. the backend's get_alloc_size may return more than ggml_nbytes for the output tensor
-    GGML_API bool ggml_backend_op_alloc_size_may_expand(enum ggml_op op);
-
     // Tensor initialization
     GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor);

ggml/include/ggml-rpc.h

@@ -6,8 +6,8 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    6
-#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_MAJOR_VERSION    5
+#define RPC_PROTO_MINOR_VERSION    1
 #define RPC_PROTO_PATCH_VERSION    0
 
 #ifdef  __cplusplus

ggml/include/ggml.h

@@ -627,7 +627,6 @@ extern "C" {
         GGML_GLU_OP_SWIGLU_OAI,
         GGML_GLU_OP_GEGLU_ERF,
         GGML_GLU_OP_GEGLU_QUICK,
-        GGML_GLU_OP_SWIGLU_CLAMP,
 
         GGML_GLU_OP_COUNT,
     };
@@ -1368,12 +1367,6 @@ extern "C" {
             float                 alpha,
             float                 limit);
 
-    GGML_API struct ggml_tensor * ggml_swiglu_clamp(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            float                 limit);
-
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1731,19 +1724,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past);
 
-    GGML_API struct ggml_tensor * ggml_clamp(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            float                 min,
-            float                 max);
-
-    // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_clamp_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            float                 min,
-            float                 max);
-
     GGML_API struct ggml_tensor * ggml_soft_max(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -2010,6 +1990,14 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_offs);
 
+    // clamp
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
+
     // im2col
     // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     GGML_API struct ggml_tensor * ggml_im2col(

ggml/src/ggml-alloc.c

@@ -40,7 +40,6 @@ bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_SILU_BACK:
         case GGML_OP_RMS_NORM:
         case GGML_OP_RMS_NORM_BACK:
-        case GGML_OP_CLAMP:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_SOFT_MAX_BACK:
             return true;

ggml/src/ggml-backend-impl.h

@@ -83,7 +83,6 @@ extern "C" {
     GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
     GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
     GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API void                  ggml_backend_meta_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend (meta)
@@ -103,16 +102,6 @@ extern "C" {
     // Backend (stream)
     //
 
-    // passed to graph_optimize so the backend can add allocation dependencies:
-    // if the backend executes parts of the graph out of order (e.g. on concurrent streams),
-    // it must keep the affected tensors allocated until a node where execution is known to have joined
-    struct ggml_backend_graph_optimize_params {
-        // keep `tensor` allocated at least until `until` (a node of the same graph) has been computed
-        // can be called multiple times for the same tensor: the longest lifetime applies
-        void (*add_alloc_dep)(void * user_data, struct ggml_tensor * tensor, struct ggml_tensor * until);
-        void * user_data;
-    };
-
     struct ggml_backend_i {
         const char * (*get_name)(ggml_backend_t backend);
 
@@ -147,7 +136,7 @@ extern "C" {
         void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
 
         // (optional) sort/optimize the nodes in the graph
-        void                      (*graph_optimize)    (ggml_backend_t backend, struct ggml_cgraph * cgraph, struct ggml_backend_graph_optimize_params * params);
+        void                      (*graph_optimize)    (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     };
 
     struct ggml_backend {

ggml/src/ggml-backend-meta.cpp

@@ -592,18 +592,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
             return {assume_sync ? GGML_BACKEND_SPLIT_AXIS_MIRRORED : GGML_BACKEND_SPLIT_AXIS_PARTIAL, {0}, {1}, 1};
         }
-        if (src_ss[0].axis == src_ss[1].axis && src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 &&
-                src_ss[0].axis < GGML_MAX_DIMS) {
-            GGML_ASSERT(split_states_equal(src_ss[0], src_ss[1]));
-            return src_ss[0];
-        }
-        // batched matmul with the batches split across devices and a replicated activation
-        if (src_ss[0].axis >= GGML_BACKEND_SPLIT_AXIS_2 && src_ss[0].axis < GGML_MAX_DIMS &&
-                src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-            return src_ss[0];
-        }
-        GGML_ABORT("unsupported mul_mat split states: node=%s src0=%s axis=%d src1=%s axis=%d",
-            tensor->name, tensor->src[0]->name, (int) src_ss[0].axis, tensor->src[1]->name, (int) src_ss[1].axis);
+        GGML_ABORT("fatal error");
         //return {GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, {1}, 1};
     };
 
@@ -613,40 +602,27 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_BACKEND_SPLIT_AXIS_1:
             case GGML_BACKEND_SPLIT_AXIS_2:
             case GGML_BACKEND_SPLIT_AXIS_3: {
-                int64_t base_ne_in = 1;
-                for (int dim = 0; dim <= src_ss[0].axis; dim++) {
+                GGML_ASSERT(src_ss[0].n_segments == 1);
+                if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
+                    return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
+                }
+                int64_t base_ne_in = tensor->src[0]->ne[0];
+                for (int dim = 1; dim <= src_ss[0].axis; dim++) {
                     base_ne_in *= tensor->src[0]->ne[dim];
                 }
-                if (src_ss[0].n_segments == 1) {
-                    base_ne_in /= src_ss[0].nr[0];
-                    if (src_ss[0].axis == ggml_n_dims(tensor->src[0]) - 1 && src_ss[0].nr[0] == 1) {
-                        return {ggml_backend_meta_split_axis(ggml_n_dims(tensor) - 1), {0}, {1}, 1};
-                    }
-                    if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0 && tensor->ne[0] == tensor->src[0]->ne[0] &&
-                            tensor->ne[1] == 1 && src_ss[0].nr[0] == 1) {
-                        bool complete_rows = true;
-                        for (size_t j = 0; j < n_bufs; j++) {
-                            const int64_t ne = src_ss[0].ne[j];
-                            complete_rows = complete_rows && (ne == 0 || ne == tensor->src[0]->ne[0]);
-                        }
-                        if (complete_rows) {
-                            // Move a complete dim-0 split to the following singleton dimension.
-                            return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
-                        }
-                    }
-                }
-                // Reshape outputs use one segment; split-state propagation merges source segments.
+                base_ne_in /= src_ss[0].nr[0];
                 int64_t base_ne_out = 1;
                 for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-                    base_ne_out *= tensor->ne[dim];
-                    if (base_ne_out % base_ne_in == 0) {
-                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out/base_ne_in)}, 1};
+                    const int64_t base_ne_out_next = base_ne_out *= tensor->ne[dim];
+                    if (base_ne_out_next % base_ne_in == 0) {
+                        return {ggml_backend_meta_split_axis(dim), {0}, {uint32_t(base_ne_out_next/base_ne_in)}, 1};
                     }
-                    if (base_ne_out > base_ne_in) {
+                    if (base_ne_out_next > base_ne_in) {
                         GGML_ASSERT(src_ss[0].n_segments == 1);
                         GGML_ASSERT(src_ss[0].nr[0]      == 1);
                         return {ggml_backend_meta_split_axis(dim), {0}, {1}, 1};
                     }
+                    base_ne_out = base_ne_out_next;
                 }
                 GGML_ABORT("shape mismatch for %s", ggml_op_name(tensor->op));
             }
@@ -771,33 +747,14 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
     };
 
     auto handle_flash_attn_ext = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
-        GGML_ASSERT(tensor->src[3] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-
-        if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-            GGML_ASSERT(src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-            GGML_ASSERT(src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-            GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-            return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
-        }
-
-        GGML_ASSERT(src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
-        const bool kv_split = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2 &&
-                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2;
-        const bool kv_mirrored = src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED &&
-                src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED;
-        GGML_ASSERT(kv_split || kv_mirrored);
+        GGML_ASSERT(                             src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_2);
+        GGML_ASSERT(                             src_ss[1].axis == GGML_BACKEND_SPLIT_AXIS_2);
+        GGML_ASSERT(                             src_ss[2].axis == GGML_BACKEND_SPLIT_AXIS_2);
+        GGML_ASSERT(tensor->src[4] == nullptr || src_ss[3].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
         GGML_ASSERT(tensor->src[4] == nullptr || src_ss[4].axis == GGML_BACKEND_SPLIT_AXIS_0);
         return {GGML_BACKEND_SPLIT_AXIS_1, {0}, {1}, 1};
     };
 
-    auto handle_lightning_indexer = [&](
-            const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
-        for (size_t i = 0; i < 4; i++) {
-            GGML_ASSERT(src_ss[i].axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-        }
-        return {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
-    };
-
     auto handle_ssm_conv = [&](const std::vector<ggml_backend_meta_split_state> & src_ss) -> ggml_backend_meta_split_state {
         if (src_ss[0].axis == src_ss[1].axis) {
             if (src_ss[0].axis == GGML_BACKEND_SPLIT_AXIS_0) {
@@ -835,7 +792,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             ggml_backend_dev_t dev = ggml_backend_buft_get_device(ggml_backend_buffer_get_type(tensor->buffer));
             const ggml_backend_meta_device_context * dev_ctx = (const ggml_backend_meta_device_context *) dev->context;
             ggml_backend_meta_split_state ret = dev_ctx->get_split_state(tensor, dev_ctx->get_split_state_ud);
-            if (ret.axis >= 0 && ret.axis < GGML_MAX_DIMS) {
+            if (ret.axis >= 0 && ret.axis <= GGML_MAX_DIMS) {
                 const int64_t granularity = ret.axis == GGML_BACKEND_SPLIT_AXIS_0 ? ggml_blck_size(tensor->type) : 1;
                 int64_t ne_sum = 0;
                 for (size_t s = 0; s < ret.n_segments; s++) {
@@ -845,9 +802,6 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                     }
                 }
                 GGML_ASSERT(ne_sum == tensor->ne[ret.axis]);
-            } else if (ret.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                GGML_ASSERT(ret.n_segments == 1);
-                GGML_ASSERT(ret.nr[0] == 1);
             }
             return ret;
         }
@@ -968,7 +922,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
                 split_state = handle_rope(src_ss);
             } break;
             case GGML_OP_ROPE_BACK: {
-                split_state = handle_rope(src_ss);
+                split_state = handle_generic(src_ss, /*scalar_only =*/ true);
             } break;
             case GGML_OP_CLAMP: {
                 split_state = handle_generic(src_ss, /*scalar_only =*/ false);
@@ -1032,9 +986,6 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
             case GGML_OP_GATED_DELTA_NET: {
                 split_state = handle_gated_delta_net(src_ss);
             } break;
-            case GGML_OP_LIGHTNING_INDEXER: {
-                split_state = handle_lightning_indexer(src_ss);
-            } break;
             case GGML_OP_DSV4_HC_COMB:
             case GGML_OP_DSV4_HC_PRE:
             case GGML_OP_DSV4_HC_POST: {
@@ -1119,14 +1070,13 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
         if (buf_ctx->debug > 0) {
             std::string srcs_info;
             for (size_t i = 0; i < GGML_MAX_SRC; i++) {
-                if (tensor->src[i] == nullptr || tensor->src[i] == tensor) {
+                if (tensor->src[i] == nullptr) {
                     continue;
                 }
                 if (!srcs_info.empty()) {
                     srcs_info += ", ";
                 }
-                const ggml_backend_meta_split_state split_state =
-                        ggml_backend_meta_get_split_state(tensor->src[i], true);
+                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor->src[0], true);
                 GGML_ASSERT(split_state.n_segments == 1);
                 const char * axis_name = ggml_backend_meta_split_axis_name(split_state.axis);
                 std::string ne_info;
@@ -1168,6 +1118,7 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
 }
 
 static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(const struct ggml_tensor * tensor, bool assume_sync) {
+    GGML_ASSERT(ggml_backend_buffer_is_meta(tensor->buffer));
     ggml_backend_meta_buffer_context * buf_ctx = (ggml_backend_meta_buffer_context *) tensor->buffer->context;
     return ggml_backend_meta_get_split_state(buf_ctx->get_simple_tensor_container(tensor), tensor, assume_sync);
 }
@@ -1258,14 +1209,7 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor_impl(ggml_backend_m
             t_ij->data = (char *) ggml_backend_buffer_get_base(simple_buf)
                 + size_t(tensor->data) - size_t(ggml_backend_buffer_get_base(tensor->buffer));
         }
-
-        if (simple_buf) {
-            // the backend that owns the buffer will set .extra
-            ggml_backend_buffer_init_tensor(simple_buf, t_ij);
-        } else {
-            t_ij->extra = tensor->extra;
-        }
-
+        t_ij->extra = tensor->extra;
         for (int i = 0; i < GGML_MAX_SRC; i++) {
             t_ij->src[i] = tensor->src[i];
             if (tensor->src[i] == tensor) {
@@ -1311,108 +1255,6 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
     return ggml_backend_meta_buffer_init_tensor_impl(buf_ctx->get_simple_tensor_container(tensor), tensor);
 }
 
-static void ggml_backend_meta_buffer_memset_tensor(
-        ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
-    const ggml_backend_meta_split_state split_state =
-            ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
-    GGML_ASSERT(ggml_is_contiguous(tensor) || split_state.axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED);
-
-    if (split_state.n_segments != 1 || split_state.nr[0] != 1) {
-        GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
-        GGML_ASSERT(split_state.nr[0] != 0);
-        GGML_ASSERT(tensor->ne[3] == 1);
-
-        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 row_start = offset / row_stride;
-            const int64_t row_count = size   / row_stride;
-            GGML_ASSERT(row_start + row_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 r = 0; r < split_state.nr[s]; r++) {
-                    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];
-                        for (int64_t row = 0; row < row_count; row++) {
-                            ggml_backend_tensor_memset(simple_tensor, value,
-                                    simple_offsets[j] + (row_start + row)*simple_tensor->nb[1], nbytes);
-                        }
-                        simple_offsets[j] += nbytes;
-                    }
-                }
-            }
-            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 row_start = offset / row_stride;
-        const int64_t row_count = size   / row_stride;
-        GGML_ASSERT(row_start + row_count <= tensor->ne[2]);
-
-        for (size_t s = 0; s < split_state.n_segments; s++) {
-            for (size_t r = 0; r < split_state.nr[s]; r++) {
-                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];
-                    for (int64_t row = 0; row < row_count; row++) {
-                        ggml_backend_tensor_memset(simple_tensor, value,
-                                simple_offsets[j] + (row_start + row)*simple_tensor->nb[2], nbytes);
-                    }
-                    simple_offsets[j] += nbytes;
-                }
-            }
-        }
-        return;
-    }
-
-    switch (split_state.axis) {
-        case GGML_BACKEND_SPLIT_AXIS_0:
-        case GGML_BACKEND_SPLIT_AXIS_1:
-        case GGML_BACKEND_SPLIT_AXIS_2: {
-            const size_t chunk_size_full = tensor->nb[split_state.axis + 1];
-            GGML_ASSERT(offset % chunk_size_full == 0);
-            GGML_ASSERT(size   % chunk_size_full == 0);
-            const int64_t i_start =  offset        / chunk_size_full;
-            const int64_t i_stop  = (offset + size) / chunk_size_full;
-            for (size_t j = 0; j < n_bufs; j++) {
-                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                const size_t chunk_size = simple_tensor->nb[split_state.axis + 1];
-                if (chunk_size == 0) {
-                    continue;
-                }
-                for (int64_t i = i_start; i < i_stop; i++) {
-                    ggml_backend_tensor_memset(simple_tensor, value, i*chunk_size, chunk_size);
-                }
-            }
-        } break;
-        case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
-            GGML_ASSERT(value == 0);
-            [[fallthrough]];
-        }
-        case GGML_BACKEND_SPLIT_AXIS_MIRRORED: {
-            for (size_t j = 0; j < n_bufs; j++) {
-                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                ggml_backend_tensor_memset(simple_tensor, value, offset, size);
-            }
-        } break;
-        default: {
-            GGML_ABORT("fatal error");
-        }
-    }
-}
-
 static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     const size_t n_bufs = ggml_backend_meta_buffer_n_bufs(buffer);
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
@@ -1510,29 +1352,15 @@ static void ggml_backend_meta_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
         } break;
         case GGML_BACKEND_SPLIT_AXIS_PARTIAL: {
             GGML_ASSERT(tensor->type == GGML_TYPE_F32);
-            GGML_ASSERT(offset % sizeof(float) == 0);
-            GGML_ASSERT(size   % sizeof(float) == 0);
-            const size_t n_values = size / sizeof(float);
-            size_t n_contributors = 0;
-            for (size_t j = 0; j < n_bufs; j++) {
-                n_contributors += split_state.ne[j] != 0;
-            }
-            const bool has_contributor_mask = n_contributors != 0;
-            if (!has_contributor_mask) {
-                n_contributors = n_bufs;
-            }
-            std::vector<float> tmp(n_values);
-            for (size_t i = 0; i < n_values; i++) {
-                tmp[i] = ((const float *) data)[i] / n_contributors;
-            }
-            std::vector<float> zero;
-            if (has_contributor_mask) {
-                zero.resize(n_values, 0.0f);
+            const int64_t ne = ggml_nelements(tensor);
+            std::vector<float> tmp;
+            tmp.reserve(ne);
+            for (int64_t i = 0; i < ne; i++) {
+                tmp.push_back(((const float *) data)[i] / n_bufs);
             }
             for (size_t j = 0; j < n_bufs; j++) {
                 ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                const float * partial = has_contributor_mask && split_state.ne[j] == 0 ? zero.data() : tmp.data();
-                ggml_backend_tensor_set(simple_tensor, partial, offset, size);
+                ggml_backend_tensor_set(simple_tensor, tmp.data(), offset, size);
             }
         } break;
         default: {
@@ -1660,7 +1488,7 @@ static const ggml_backend_buffer_i ggml_backend_meta_buffer_iface = {
     /* .free_buffer     = */ ggml_backend_meta_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_meta_buffer_get_base,
     /* .init_tensor     = */ ggml_backend_meta_buffer_init_tensor,
-    /* .memset_tensor   = */ ggml_backend_meta_buffer_memset_tensor,
+    /* .memset_tensor   = */ nullptr, // TODO implement
     /* .set_tensor      = */ ggml_backend_meta_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_meta_buffer_get_tensor,
     /* .set_tensor_2d   = */ nullptr,
@@ -1674,16 +1502,6 @@ bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf) {
     return buf != nullptr && buf->iface.free_buffer == ggml_backend_meta_buffer_iface.free_buffer;
 }
 
-void ggml_backend_meta_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
-    GGML_ASSERT(ggml_backend_buffer_is_meta(buffer));
-    ggml_backend_meta_buffer_context * buf_ctx = (ggml_backend_meta_buffer_context *) buffer->context;
-    for (size_t i = 0; i < buf_ctx->bufs.size(); i++) {
-        if (buf_ctx->bufs[i]) {
-            ggml_backend_buffer_set_usage(buf_ctx->bufs[i].get(), usage);
-        }
-    }
-}
-
 static ggml_backend_buffer_t ggml_backend_meta_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
     const size_t n_simple_bufts = ggml_backend_meta_buft_n_bufts(buft);
 
@@ -2023,7 +1841,7 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
 
         {
             // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-            auto get_i_delayed_branch = [&](const int i) -> int {
+            auto get_i_delayed = [&](const int i) -> int {
                 int id = i; // i_delayed
                 int idr = i; // i_delayed return, last safe return value
 
@@ -2123,62 +1941,6 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                 return idr;
             };
 
-            // AllReduce(a) + AllReduce(b) == AllReduce(a + b) for independent partial branches.
-            auto get_i_delayed = [&](const int i) -> int {
-                const int i_delayed = get_i_delayed_branch(i);
-                ggml_tensor * node = cgraph->nodes[i_delayed];
-
-                if (ggml_node_get_use_count(cgraph, i_delayed) != 1) {
-                    return i_delayed;
-                }
-
-                for (int id = i_delayed + 1; id < cgraph->n_nodes; id++) {
-                    ggml_tensor * next = cgraph->nodes[id];
-                    if (next->view_src == node) {
-                        return i_delayed;
-                    }
-                    for (int s = 0; s < GGML_MAX_SRC; s++) {
-                        if (next->src[s] == node) {
-                            return i_delayed;
-                        }
-                    }
-
-                    if (next->view_src != nullptr && next->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(next->view_src->buffer)) {
-                        continue;
-                    }
-                    if (ggml_backend_meta_get_split_state(next, false).axis != GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                        continue;
-                    }
-
-                    const int i_other = id;
-                    const int i_other_delayed = get_i_delayed_branch(i_other);
-                    ggml_tensor * other = cgraph->nodes[i_other_delayed];
-                    if (ggml_node_get_use_count(cgraph, i_other_delayed) != 1 || i_other_delayed + 1 >= cgraph->n_nodes) {
-                        return i_delayed;
-                    }
-
-                    ggml_tensor * sum = cgraph->nodes[i_other_delayed + 1];
-                    if (sum->op != GGML_OP_ADD ||
-                            !ggml_are_same_shape(node, other) || node->type != other->type || sum->type != node->type ||
-                            !((sum->src[0] == node && sum->src[1] == other) ||
-                              (sum->src[0] == other && sum->src[1] == node)) ||
-                            ggml_backend_meta_get_split_state(sum, false).axis != GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                        return i_delayed;
-                    }
-
-                    for (size_t j = 0; j < n_backends; j++) {
-                        auto & bcj = backend_ctx->backend_configs[j];
-                        const bool compute       = bcj.nodes[i]->flags       & GGML_TENSOR_FLAG_COMPUTE;
-                        const bool compute_other = bcj.nodes[i_other]->flags & GGML_TENSOR_FLAG_COMPUTE;
-                        if (compute != compute_other) {
-                            return i_delayed;
-                        }
-                    }
-                    return i_other_delayed + 1;
-                }
-                return i_delayed;
-            };
-
             int i_start = 0;
             for (int i = 0; i < cgraph->n_nodes; i++) {
                 ggml_tensor * node = cgraph->nodes[i];

ggml/src/ggml-backend.cpp

@@ -20,7 +20,6 @@
 #include <stdlib.h>
 #include <string.h>
 #include <algorithm>
-#include <unordered_map>
 #include <vector>
 
 #ifdef __APPLE__
@@ -65,14 +64,6 @@ size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const s
     if (buft->iface.get_alloc_size) {
         size_t size = buft->iface.get_alloc_size(buft, tensor);
         assert(size >= ggml_nbytes(tensor));
-
-        // [TAG_ALLOC_SIZE_EXPAND]
-        // if you hit this assert, update ggml_backend_op_alloc_size_may_expand() accordingly
-        GGML_ASSERT(size <= ggml_nbytes(tensor) ||
-                    ggml_op_is_empty(tensor->op) ||
-                    ggml_is_quantized(tensor->type) || // [TAG_ALLOC_SIZE_EXPAND]
-                    ggml_backend_op_alloc_size_may_expand(tensor->op));
-
         return size;
     }
     return ggml_nbytes(tensor);
@@ -191,8 +182,6 @@ void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backe
     // FIXME: add a generic callback to the buffer interface
     if (ggml_backend_buffer_is_multi_buffer(buffer)) {
         ggml_backend_multi_buffer_set_usage(buffer, usage);
-    } else if (ggml_backend_buffer_is_meta(buffer)) {
-        ggml_backend_meta_buffer_set_usage(buffer, usage);
     }
 }
 
@@ -567,10 +556,10 @@ void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event)
     backend->iface.event_wait(backend, event);
 }
 
-static void ggml_backend_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * cgraph, struct ggml_backend_graph_optimize_params * params) {
+static void ggml_backend_graph_optimize(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     GGML_ASSERT(backend);
     if (backend->iface.graph_optimize != NULL) {
-        backend->iface.graph_optimize(backend, cgraph, params);
+        backend->iface.graph_optimize(backend, cgraph);
     }
 }
 
@@ -1450,40 +1439,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    // optimize the split graphs and collect the allocation dependencies added by the backends
-    // this needs to happen before we make graph_copy, so they are in sync
-    // TODO: this may create many small allocations in the scheduler, restructure to use a flat array
-    std::unordered_map<ggml_tensor *, std::vector<ggml_tensor *>> alloc_deps;
-
-    struct ggml_backend_graph_optimize_params opt_params = {
-        /* .add_alloc_dep = */ [](void * user_data, ggml_tensor * tensor, ggml_tensor * until) {
-            auto & deps = *(std::unordered_map<ggml_tensor *, std::vector<ggml_tensor *>> *) user_data;
-            std::vector<ggml_tensor *> & keep = deps[until];
-            if (std::find(keep.begin(), keep.end(), tensor) == keep.end()) {
-                keep.push_back(tensor);
-            }
-        },
-        /* .user_data     = */ &alloc_deps,
-    };
-
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &sched->splits[i];
-        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
-
-        ggml_backend_graph_optimize(sched->backends[split->backend_id], &split->graph, &opt_params);
-    }
-
-    // each dep is added to graph_copy as a GGML_OP_NONE node with the kept tensors as srcs
-    int n_dep_nodes = 0;
-    for (const auto & it : alloc_deps) {
-        n_dep_nodes += (it.second.size() + GGML_MAX_SRC - 1) / GGML_MAX_SRC;
-    }
-
     int total_inputs = sched->n_graph_inputs;
     for (int i = 0; i < sched->n_splits; i++) {
         total_inputs += sched->splits[i].n_inputs;
     }
-    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies + n_dep_nodes;
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies;
 
     // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
     sched->debug_prev_graph_size = sched->debug_graph_size;
@@ -1501,10 +1461,13 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
 
     struct ggml_cgraph * graph_copy = &sched->graph;
 
-    int n_dep_nodes_added = 0;
-
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
+
+        // Optimize this split of the graph. This needs to happen before we make graph_copy,
+        // so they are in sync.
+        ggml_backend_graph_optimize(sched->backends[split->backend_id], &split->graph);
 
         // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
@@ -1529,32 +1492,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
             assert(graph_copy->size > graph_copy->n_nodes);
             sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
             graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
-
-            if (alloc_deps.empty()) {
-                continue;
-            }
-
-            // add a dependency node so that the kept tensors are not freed before this node is computed
-            auto it = alloc_deps.find(graph->nodes[j]);
-            if (it != alloc_deps.end()) {
-                const std::vector<ggml_tensor *> & keep = it->second;
-                for (size_t k = 0; k < keep.size(); k += GGML_MAX_SRC) {
-                    struct ggml_tensor * dep = ggml_view_tensor(sched->ctx, keep[k]);
-                    for (size_t s = 0; s < GGML_MAX_SRC && k + s < keep.size(); s++) {
-                        dep->src[s] = keep[k + s];
-                    }
-                    assert(graph_copy->size > graph_copy->n_nodes);
-                    sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id;
-                    graph_copy->nodes[graph_copy->n_nodes++] = dep;
-                    n_dep_nodes_added++;
-                }
-            }
         }
     }
 
-    // a mismatch means a backend added a dep with an `until` tensor that is not a node of the optimized graph
-    GGML_ASSERT(n_dep_nodes_added == n_dep_nodes);
-
     if (sched->n_copies > 1) {
         // add input copies as leafs so that they are allocated first
         for (int i = 0; i < sched->n_graph_inputs; i++) {
@@ -2109,23 +2049,6 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
 
 // utils
 
-// [TAG_ALLOC_SIZE_EXPAND]
-// returns true for ops that may require additional memory for fleeting data on some backends,
-// i.e. the backend's get_alloc_size may return more than ggml_nbytes for the output tensor
-bool ggml_backend_op_alloc_size_may_expand(enum ggml_op op) {
-    switch (op) {
-        case GGML_OP_FLASH_ATTN_EXT:
-        case GGML_OP_MUL_MAT:
-        case GGML_OP_MUL_MAT_ID:
-        case GGML_OP_CUMSUM:
-        case GGML_OP_ARGSORT:
-        case GGML_OP_TOP_K:
-            return true;
-        default:
-            return false;
-    }
-}
-
 enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor);
     GGML_ASSERT(tensor->buffer == NULL);

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -211,50 +211,6 @@ void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
     GGML_CANN_CALL_ACLNN_OP(ctx, SwiGlu, acl_src.get(), (int64_t)2, acl_dst.get());
 }
 
-void ggml_cann_swiglu_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * src0 = dst->src[0];
-    ggml_tensor * src1 = dst->src[1];
-
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-
-    const int32_t  swapped = ggml_get_op_params_i32(dst, 1);
-    acl_tensor_ptr acl_gate;
-    acl_tensor_ptr acl_up;
-    if (src1) {
-        GGML_ASSERT(ggml_is_contiguous_1(src1));
-        GGML_ASSERT(src0->type == src1->type);
-        acl_gate = ggml_cann_create_tensor(src0);
-        acl_up   = ggml_cann_create_tensor(src1);
-    } else {
-        int64_t ne[] = { src0->ne[0] / 2, src0->ne[1], src0->ne[2], src0->ne[3] };
-        size_t  nb[] = { src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3] };
-        acl_gate     = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, 0);
-        acl_up = ggml_cann_create_tensor(src0, ne, nb, GGML_MAX_DIMS, ACL_FORMAT_ND, ne[0] * ggml_element_size(src0));
-        if (swapped) {
-            std::swap(acl_gate, acl_up);
-        }
-    }
-
-    ggml_cann_pool_alloc temp_alloc(ctx.pool(), ggml_nbytes(dst));
-    acl_tensor_ptr       acl_temp = ggml_cann_create_tensor(temp_alloc.get(), ggml_cann_type_mapping(dst->type),
-                                                            ggml_element_size(dst), dst->ne, dst->nb, GGML_MAX_DIMS);
-    acl_tensor_ptr       acl_dst  = ggml_cann_create_tensor(dst);
-
-    const float    limit        = ggml_get_op_params_f32(dst, 3);
-    float          min_gate     = -INFINITY;
-    float          min_up       = -limit;
-    float          max_value    = limit;
-    acl_scalar_ptr acl_min_gate = ggml_cann_create_scalar(&min_gate, ACL_FLOAT);
-    acl_scalar_ptr acl_min_up   = ggml_cann_create_scalar(&min_up, ACL_FLOAT);
-    acl_scalar_ptr acl_limit    = ggml_cann_create_scalar(&max_value, ACL_FLOAT);
-
-    GGML_CANN_CALL_ACLNN_OP(ctx, Clamp, acl_gate.get(), acl_min_gate.get(), acl_limit.get(), acl_temp.get());
-    GGML_CANN_CALL_ACLNN_OP(ctx, Silu, acl_temp.get(), acl_dst.get());
-    GGML_CANN_CALL_ACLNN_OP(ctx, Clamp, acl_up.get(), acl_min_up.get(), acl_limit.get(), acl_temp.get());
-    GGML_CANN_CALL_ACLNN_OP(ctx, InplaceMul, acl_dst.get(), acl_temp.get());
-}
-
 // Fused GeGLU using aclnnGeGluV3: splits input along ne[0] (CANN last dim),
 // activates the LEFT half with GELU, multiplies by right half.
 // approximate: 0=tanh, 1=none(erf). activateLeft=true matches GGML convention.
@@ -4477,3 +4433,4 @@ void ggml_cann_gated_linear_attn(ggml_backend_cann_context & ctx, ggml_tensor *
         }
     }
 }
+

ggml/src/ggml-cann/aclnn_ops.h

@@ -76,7 +76,6 @@
 void ggml_cann_repeat(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 
 void ggml_cann_swiglu(ggml_backend_cann_context & ctx, ggml_tensor * dst);
-void ggml_cann_swiglu_clamp(ggml_backend_cann_context & ctx, ggml_tensor * dst);
 void ggml_cann_geglu(ggml_backend_cann_context & ctx, ggml_tensor * dst, int64_t approximate);
 
 /**

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

@@ -1872,9 +1872,6 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context & ctx, struct gg
                 case GGML_GLU_OP_SWIGLU:
                     ggml_cann_swiglu(ctx, dst);
                     break;
-                case GGML_GLU_OP_SWIGLU_CLAMP:
-                    ggml_cann_swiglu_clamp(ctx, dst);
-                    break;
                 case GGML_GLU_OP_GEGLU_QUICK:
                     ggml_cann_geglu_quick(ctx, dst);
                     break;
@@ -2431,7 +2428,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
                 case GGML_GLU_OP_SWIGLU:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
-                case GGML_GLU_OP_SWIGLU_CLAMP:
                     return true;
                 default:
                     return false;

ggml/src/ggml-common.h

@@ -1131,7 +1131,7 @@ GGML_TABLE_END()
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f
-#if defined(GGML_COMMON_IMPL_C) || defined(GGML_COMMON_IMPL_CPP)
+#if defined(GGML_COMMON_IMPL_C)
 GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
     0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
     0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,

ggml/src/ggml-cpu/CMakeLists.txt

@@ -31,8 +31,6 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         ggml-cpu/ggml-cpu.cpp
         ggml-cpu/repack.cpp
         ggml-cpu/repack.h
-        ggml-cpu/iqp.cpp
-        ggml-cpu/iqp.h
         ggml-cpu/hbm.cpp
         ggml-cpu/hbm.h
         ggml-cpu/quants.c
@@ -455,16 +453,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ggml-cpu/spacemit/repack.h
                 ggml-cpu/spacemit/ime_env.cpp
                 ggml-cpu/spacemit/ime_env.h
+                ggml-cpu/spacemit/ime1_kernels.cpp
+                ggml-cpu/spacemit/ime2_kernels.cpp
                 ggml-cpu/spacemit/ime_kernels.h
                 ggml-cpu/spacemit/rvv_kernels.cpp
                 ggml-cpu/spacemit/rvv_kernels.h
             )
-            if ("RISCV64_SPACEMIT_IME1" IN_LIST RISCV64_SPACEMIT_IME_SPEC)
-                list(APPEND GGML_CPU_SOURCES ggml-cpu/spacemit/ime1_kernels.cpp)
-            endif()
-            if ("RISCV64_SPACEMIT_IME2" IN_LIST RISCV64_SPACEMIT_IME_SPEC)
-                list(APPEND GGML_CPU_SOURCES ggml-cpu/spacemit/ime2_kernels.cpp)
-            endif()
         endif()
         if(NOT GGML_CPU_ALL_VARIANTS)
             set(MARCH_STR "rv64gc")
@@ -582,26 +576,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     endif()
 
     if (GGML_CPU_KLEIDIAI)
-        # upstream repo requires at least cmake 3.16
-        if (CMAKE_VERSION VERSION_LESS 3.16)
-            message(FATAL_ERROR "GGML_CPU_KLEIDIAI requires CMake >= 3.16")
-        endif()
-
-        set(GGML_CPU_KLEIDIAI_AARCH64 OFF)
-        if (GGML_SYSTEM_ARCH STREQUAL "ARM" AND
-            (APPLE OR WIN32 OR CMAKE_SYSTEM_NAME MATCHES "^(Linux|Android)$") AND
-            (CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64|ARM64|arm64-v8a)$" OR
-             CMAKE_OSX_ARCHITECTURES MATCHES "arm64" OR
-             CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-             CMAKE_ANDROID_ARCH_ABI STREQUAL "arm64-v8a"))
-            set(GGML_CPU_KLEIDIAI_AARCH64 ON)
-        endif()
-        if (NOT GGML_CPU_KLEIDIAI_AARCH64)
-            message(FATAL_ERROR "GGML_CPU_KLEIDIAI requires a Linux, Android, Apple, or Windows AArch64/arm64 target")
-        endif()
-
         message(STATUS "Using KleidiAI optimized kernels if applicable")
 
+        # Disable the KleidiAI tests
+        set(KLEIDIAI_BUILD_TESTS  OFF)
+
         # Fetch KleidiAI sources:
         include(FetchContent)
         set(KLEIDIAI_COMMIT_TAG "v1.24.0")
@@ -616,49 +595,31 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             list(APPEND KLEIDIAI_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP NEW)
         endif()
 
-        FetchContent_Declare(kleidiai
-            ${KLEIDIAI_FETCH_ARGS}
-        )
-
-        # Disable tests and benchmark building
-        set(KLEIDIAI_BUILD_TESTS OFF CACHE BOOL "" FORCE)
-        set(KLEIDIAI_BUILD_BENCHMARK OFF CACHE BOOL "" FORCE)
-
-        # Use the Populate/add_subdirectory flow for compatibility with CMake 3.16.
-        FetchContent_GetProperties(kleidiai
-            SOURCE_DIR KLEIDIAI_SRC
-            BINARY_DIR KLEIDIAI_BIN
-            POPULATED KLEIDIAI_POPULATED
-        )
-        if (NOT KLEIDIAI_POPULATED)
-            FetchContent_Populate(kleidiai)
-            FetchContent_GetProperties(kleidiai
-                SOURCE_DIR KLEIDIAI_SRC
-                BINARY_DIR KLEIDIAI_BIN
-            )
-        endif()
-
-        if (NOT TARGET kleidiai)
-            add_subdirectory(
-                "${CMAKE_CURRENT_SOURCE_DIR}/ggml-cpu/kleidiai"
-                "${CMAKE_CURRENT_BINARY_DIR}/kleidiai-wrapper"
+        if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.28")
+            FetchContent_Declare(KleidiAI_Download
+                ${KLEIDIAI_FETCH_ARGS}
                 EXCLUDE_FROM_ALL
             )
-            if (NOT CMAKE_SKIP_INSTALL_RULES AND
-                (NOT DEFINED BUILD_SHARED_LIBS OR NOT BUILD_SHARED_LIBS))
-                install(TARGETS kleidiai ARCHIVE)
+
+            FetchContent_MakeAvailable(KleidiAI_Download)
+            FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
+        else()
+            FetchContent_Declare(KleidiAI_Download
+                ${KLEIDIAI_FETCH_ARGS}
+            )
+
+            FetchContent_GetProperties(KleidiAI_Download
+                SOURCE_DIR KLEIDIAI_SRC
+                POPULATED KLEIDIAI_POPULATED
+            )
+
+            if (NOT KLEIDIAI_POPULATED)
+                FetchContent_Populate(KleidiAI_Download)
+                FetchContent_GetProperties(KleidiAI_Download SOURCE_DIR KLEIDIAI_SRC)
             endif()
         endif()
 
-        if (NOT TARGET kleidiai)
-            message(FATAL_ERROR "KleidiAI target was not created")
-        endif()
-
-        set_target_properties(kleidiai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-
-        target_link_libraries(${GGML_CPU_NAME} PRIVATE kleidiai)
-
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_KLEIDIAI)
+        add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
 
         list(APPEND GGML_CPU_SOURCES
             ggml-cpu/kleidiai/kleidiai.cpp
@@ -666,6 +627,108 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ggml-cpu/kleidiai/kleidiai.h
             ggml-cpu/kleidiai/kernels.h
             )
+
+        # KleidiAI
+        include_directories(
+            ${KLEIDIAI_SRC}/
+            ${KLEIDIAI_SRC}/kai/
+            ${KLEIDIAI_SRC}/kai/ukernels/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/)
+
+        set(ARCH_FLAGS_TEMP "${ARCH_FLAGS}")
+        if (NOT ARCH_FLAGS_TEMP)
+            string(REGEX MATCH "-march=[^ ]+" ARCH_FLAGS_TEMP "${CMAKE_C_FLAGS}")
+        endif()
+        string(FIND "${ARCH_FLAGS_TEMP}" "+dotprod" DOTPROD_ENABLED)
+        string(FIND "${ARCH_FLAGS_TEMP}" "+i8mm" I8MM_ENABLED)
+        string(FIND "${ARCH_FLAGS_TEMP}" "+sme" SME_ENABLED)
+        string(FIND "${ARCH_FLAGS_TEMP}" "+sve" SVE_ENABLED)
+
+        set(PRIVATE_ARCH_FLAGS ${ARCH_FLAGS_TEMP})
+
+        list(APPEND GGML_KLEIDIAI_SOURCES
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
+            ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.c)
+
+        if (NOT DOTPROD_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.c)
+        endif()
+
+        if (NOT I8MM_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.c)
+        endif()
+
+        if (NOT SME_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SME_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa_asm.S)
+            set_source_files_properties(${GGML_KLEIDIAI_SME_SOURCES}
+                PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;${ARCH_FLAGS_TEMP}+sve+sve2+sme")
+            list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SME_SOURCES})
+
+            list(APPEND GGML_KLEIDIAI_SME2_SOURCES
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme_asm.S
+                ${KLEIDIAI_SRC}/kai/kai_common_sme_asm.S)
+            set_source_files_properties(${GGML_KLEIDIAI_SME2_SOURCES}
+                PROPERTIES COMPILE_OPTIONS "-fno-tree-vectorize;${ARCH_FLAGS_TEMP}+sve+sve2+sme2+fp16")
+            list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SME2_SOURCES})
+            set(PRIVATE_ARCH_FLAGS "-fno-tree-vectorize;${PRIVATE_ARCH_FLAGS}")
+        endif()
+
+        if (NOT SVE_ENABLED MATCHES -1)
+            list(APPEND GGML_KLEIDIAI_SOURCES
+                ${KLEIDIAI_SRC}/kai/kai_common_sve_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.c
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm_asm.S
+                ${KLEIDIAI_SRC}/kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.c)
+        endif()
+
+        set_source_files_properties(${GGML_KLEIDIAI_SOURCES} PROPERTIES COMPILE_OPTIONS "${PRIVATE_ARCH_FLAGS}")
+        list(APPEND GGML_CPU_SOURCES ${GGML_KLEIDIAI_SOURCES})
     endif()
 
     message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")

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

@@ -4,7 +4,6 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
 #include "traits.h"
-#include "iqp.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
 #include "quants.h"
@@ -1364,13 +1363,6 @@ UseGgmlGemm1:;
 
     ggml_barrier(params->threadpool);
 
-    // IQ panel gemm (see iqp.h) - must come after the barrier above, it consumes the q8_K rows
-    // of src1 from the work buffer
-    if (ggml_cpu_iqp_supports_mul_mat(dst) && !params->use_ref) {
-        ggml_compute_forward_mul_mat_iqp(params, dst);
-        return;
-    }
-
 #if GGML_USE_LLAMAFILE
     if (src1->type != vec_dot_type) {
         const void* wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
@@ -1588,16 +1580,6 @@ static void ggml_compute_forward_mul_mat_id(
     char (*atomic_current_chunk)[CACHE_LINE_SIZE] = // [n_as]
         incr_ptr_aligned(&wdata_cur, CACHE_LINE_SIZE * n_as, CACHE_LINE_SIZE);
 
-    // IQ panel gemm (see iqp.h); per expert eligibility is decided below, but the work buffer is
-    // reserved for the whole node (ggml_graph_plan sizes it without params, use_ref only skips the dispatch)
-    const bool iqp = ggml_cpu_iqp_supports_mul_mat_id(dst) && !params->use_ref;
-
-    char * iqp_panels = NULL;
-
-    if (iqp) {
-        iqp_panels = incr_ptr_aligned(&wdata_cur, nth * ggml_cpu_iqp_scratch_size(dst), 64);
-    }
-
     GGML_ASSERT(params->wsize >= (size_t)((char *) wdata_cur - (char *) params->wdata));
 
     if (src1->type != vec_dot_type) {
@@ -1669,13 +1651,6 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        if (iqp && ggml_cpu_iqp_mul_mat_id_min_batch(cne1)) {
-            ggml_compute_forward_mul_mat_id_iqp(params, dst, cur_a, cne1, (const int32_t *) &MMID_MATRIX_ROW(cur_a, 0),
-                                                iqp_panels);
-
-            continue;
-        }
-
         const char * src0_cur = (const char *) src0->data + cur_a * nb02;
         const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
@@ -2336,7 +2311,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
-                case GGML_GLU_OP_SWIGLU_CLAMP:
                     {
                         n_tasks = n_threads;
                     } break;
@@ -2883,11 +2857,6 @@ struct ggml_cplan ggml_graph_plan(
                         if (node->src[1]->type != vec_dot_type) {
                             cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
                         }
-
-                        // the IQ panel path needs one scratch panel per thread past the q8_K rows
-                        if (ggml_cpu_iqp_supports_mul_mat(node)) {
-                            cur = GGML_PAD(cur, 64) + n_tasks * ggml_cpu_iqp_scratch_size(node);
-                        }
                     } break;
                 case GGML_OP_MUL_MAT_ID:
                     {
@@ -2907,10 +2876,6 @@ struct ggml_cplan ggml_graph_plan(
                         cur += n_as*ids->ne[0]*ids->ne[1]*sizeof(struct mmid_row_mapping) + sizeof(int64_t);
                         // atomic_current_chunk
                         cur += CACHE_LINE_SIZE*n_as + CACHE_LINE_SIZE;
-                        // the IQ panel path needs one scratch panel per thread on top of that
-                        if (ggml_cpu_iqp_supports_mul_mat_id(node)) {
-                            cur += n_tasks * ggml_cpu_iqp_scratch_size(node) + 64;
-                        }
                     } break;
                 case GGML_OP_OUT_PROD:
                     {
@@ -2971,13 +2936,12 @@ struct ggml_cplan ggml_graph_plan(
                         const int64_t ne10 = node->src[1]->ne[0]; // W
                         const int64_t ne11 = node->src[1]->ne[1]; // H
                         const int64_t ne12 = node->src[1]->ne[2]; // Channels In
-                        const int64_t ne13 = node->src[1]->ne[3]; // Batch
 
                         GGML_ASSERT(node->src[0]->type == GGML_TYPE_F16 || node->src[0]->type == GGML_TYPE_F32);
                         GGML_ASSERT(node->src[1]->type == GGML_TYPE_F32);
 
                         cur += ggml_type_size(node->src[0]->type) * ne00 * ne01 * ne02 * ne03;
-                        cur += ggml_type_size(node->src[0]->type) * ne10 * ne11 * ne12 * ne13;
+                        cur += ggml_type_size(node->src[0]->type) * ne10 * ne11 * ne12;
 
                     } break;
                 case GGML_OP_TOP_K:

ggml/src/ggml-cpu/iqp.h

@@ -1,39 +0,0 @@
-#pragma once
-
-#include "ggml-cpu-impl.h"
-#include "ggml.h"
-
-// GGML internal header
-
-// batched mul_mat path for the grid based IQ types: decode 8 src0 rows at a time into per thread scratch
-// (block_iqp_x8, see iqp.cpp) and run an integer gemm over them against all src1 columns
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// whether cne1 rows of src1 are enough for the decode to pay for itself, per expert, for MUL_MAT_ID
-bool ggml_cpu_iqp_mul_mat_id_min_batch(int64_t cne1);
-
-bool ggml_cpu_iqp_supports_mul_mat(const struct ggml_tensor * dst);
-
-// node level test only - per expert eligibility is decided with ggml_cpu_iqp_mul_mat_id_min_batch
-bool ggml_cpu_iqp_supports_mul_mat_id(const struct ggml_tensor * dst);
-
-// per thread panel scratch bytes, padded
-size_t ggml_cpu_iqp_scratch_size(const struct ggml_tensor * dst);
-
-// must be called after src1 has been converted to q8_K into params->wdata and the threads have synchronized on it
-void ggml_compute_forward_mul_mat_iqp(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-
-// one expert: expert_rows points at its row of the matrix_rows table of (i1, i2) int32 pairs, panels at the base of the per thread panel scratches
-void ggml_compute_forward_mul_mat_id_iqp(const struct ggml_compute_params * params,
-                                         struct ggml_tensor *               dst,
-                                         int64_t                            cur_a,
-                                         int64_t                            cne1,
-                                         const int32_t *                    expert_rows,
-                                         void *                             panels);
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-cpu/kleidiai/CMakeLists.txt

@@ -1,14 +0,0 @@
-set(BUILD_SHARED_LIBS OFF)
-set(CMAKE_SKIP_INSTALL_RULES TRUE)
-
-add_subdirectory("${KLEIDIAI_SRC}" "${KLEIDIAI_BIN}" EXCLUDE_FROM_ALL)
-
-if (NOT TARGET kleidiai)
-    message(FATAL_ERROR "KleidiAI target was not created")
-endif()
-
-if (MSVC)
-    target_compile_options(kleidiai PRIVATE $<$<COMPILE_LANGUAGE:C,CXX>:/WX->)
-else()
-    target_compile_options(kleidiai PRIVATE $<$<COMPILE_LANGUAGE:C,CXX>:-Wno-error>)
-endif()

ggml/src/ggml-cpu/kleidiai/kernels.cpp

@@ -3,44 +3,44 @@
 //
 
 // KleidiAI micro-kernels
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp_qsi8cxp_interface.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
-#include "kai/ukernels/matmul/matmul_clamp_fp32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi8cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_qsi8d32p_qsi4c32p/kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_f16p_qsi4c32p/kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.h"
-#include "kai/ukernels/matmul/matmul_clamp_f32_f32p_f32p/kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.h"
+#include "kai_matmul_clamp_f32_qsi8d32p_qsi4c32p_interface.h"
+#include "kai_matmul_clamp_f32_qai8dxp_qsi8cxp_interface.h"
+#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h"
+#include "kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot.h"
+#include "kai_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot.h"
+#include "kai_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme_mopa.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme_dot.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x8_qsi8cxp4x8_1x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4x4_1x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp4x4_qsi8cxp4x4_16x4_neon_dotprod.h"
+#include "kai_matmul_clamp_f32_qai8dxp4x8_qsi8cxp4x8_16x4_neon_i8mm.h"
+#include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm.h"
+#include "kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod.h"
+#include "kai_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa.h"
+#include "kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.h"
+#include "kai_matmul_clamp_f32_f32p2vlx1_f32p2vlx1b_2vlx2vl_sme_mopa.h"
 
-#include "kai/ukernels/matmul/pack/kai_lhs_pack_bf16p2vlx2_f32_sme.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_pack_f32p2vlx1_f32_sme.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.h"
+#include "kai_lhs_pack_bf16p2vlx2_f32_sme.h"
+#include "kai_lhs_pack_f32p2vlx1_f32_sme.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32.h"
+#include "kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon.h"
+#include "kai_lhs_quant_pack_qsi8d32p_f32_neon.h"
+#include "kai_lhs_quant_pack_qai8dxp_f32.h"
 
-#include "kai/ukernels/matmul/pack/kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
-#include "kai/ukernels/matmul/pack/kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h"
-#include "kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
-#include "kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
-#include "kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h"
-#include "kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.h"
+#include "kai_rhs_pack_kxn_bf16p2vlx2b_f32_x32_sme.h"
+#include "kai_rhs_pack_nxk_f32p2vlx1biasf32_f32_f32_sme.h"
+#include "kai_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0.h"
+#include "kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon.h"
+#include "kai_rhs_pack_nxk_qsi8cxp_qsi8cx_neon.h"
+#include "kai_lhs_pack_f16pmrx2_f32_neon.h"
 
-#include "kai/kai_common.h"
+#include "kai_common.h"
 
 #include "simd-mappings.h"
 
@@ -328,8 +328,9 @@ static void dequantize_row_qsi8cxp(
 }
 
 static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
+#if defined(__ARM_FEATURE_SME)
     {
-        /* SME2 GEMM */
+        /* SME GEMM */
         /* .kern_info = */ {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f16p1vlx2_qsi4c32p4vlx2_1vlx4vl_sme2_mopa,
@@ -350,7 +351,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_pack_f16pmrx2_f32_neon>,
             /* .pack_func_ex          = */ &lhs_pack_void_fn10<kai_run_lhs_pack_f16pmrx2_f32_neon>,
         },
-        /* SME2 GEMV */
+        /* SME GEMV */
         /* .kern_info = */ {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme2_sdot,
@@ -377,13 +378,13 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
             /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0_neon>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_SME2 | CPU_FEATURE_FP16,
+        /* .required_cpu       = */ CPU_FEATURE_SME2,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
     {
-        /* SME2 GEMM */
+        /* SME GEMM */
         /* .kern_info = */ {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
@@ -403,7 +404,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
             /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_bf16p2vlx2_f32_sme>,
             /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_bf16p2vlx2_f32_sme>,
         },
-        /* SME2 GEMV */
+        /* SME GEMV */
         /* .kern_info = */ {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_bf16p2vlx2_bf16p2vlx2_2vlx2vl_sme2_mopa,
@@ -435,220 +436,9 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
         /* .rhs_type           = */ GGML_TYPE_F16,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
 #if defined(__APPLE__)
-    {
-        /* DOTPROD GEMM */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
-        },
-        /* .gemm_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
-        },
-        /* DOTPROD GEMV */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
-        },
-        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
-        },
-        /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-        },
-        /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
-        /* .lhs_type           = */ GGML_TYPE_F32,
-        /* .rhs_type           = */ GGML_TYPE_Q4_0,
-        /* .op_type            = */ GGML_TYPE_F32,
-    },
-    {
-        /* i8mm GEMM */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-        },
-        /* .gemm_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-        },
-        /* DOTPROD GEMV */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-        },
-        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
-        },
-        /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-        },
-        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
-        /* .lhs_type           = */ GGML_TYPE_F32,
-        /* .rhs_type           = */ GGML_TYPE_Q4_0,
-        /* .op_type            = */ GGML_TYPE_F32,
-    },
-#else
-    {
-        /* SVE i8mm GEMM */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
-        },
-        /* .gemm_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-        },
-        /* SVE dotprod GEMV */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
-        },
-        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
-        },
-        /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-        },
-        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
-        /* .lhs_type           = */ GGML_TYPE_F32,
-        /* .rhs_type           = */ GGML_TYPE_Q4_0,
-        /* .op_type            = */ GGML_TYPE_F32,
-    },
-    {
-        /* i8mm GEMM */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
-        },
-        /* .gemm_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
-        },
-        /* DOTPROD GEMV */
-        /* .kern_info = */ {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
-            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
-        },
-        /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
-            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
-            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
-        },
-        /* .rhs_info = */ {
-            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
-            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
-            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
-        },
-        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
-        /* .lhs_type           = */ GGML_TYPE_F32,
-        /* .rhs_type           = */ GGML_TYPE_Q4_0,
-        /* .op_type            = */ GGML_TYPE_F32,
-    },
+#if defined(__ARM_FEATURE_DOTPROD)
     {
         /* DOTPROD GEMM */
         /* .kern_info = */ {
@@ -702,13 +492,236 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
         /* .rhs_type           = */ GGML_TYPE_Q4_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    {
+        /* i8mm GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+        },
+        /* i8mm GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
+#else
+#if defined(__ARM_FEATURE_SVE)
+    {
+        /* SVE i8mm GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p8x8_16x8_sve_i8mm>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+        },
+        /* SVE dotprod GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p8x8_1x8_sve_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_SVE | CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    {
+        /* i8mm GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p4x8sb_f32_neon>,
+        },
+        /* i8mm GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif // __ARM_FEATURE_MATMUL_INT8
+#if defined(__ARM_FEATURE_DOTPROD)
+    {
+        /* DOTPROD GEMM */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod>,
+        },
+        /* .gemm_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* DOTPROD GEMV */
+        /* .kern_info = */ {
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn3<kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn3<kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+            /* .run_kernel_ex         = */ &kernel_run_fn11<kai_run_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod>,
+        },
+        /* .gemv_lhs_info = */ {
+            /* .get_offset            = */ kai_get_lhs_offset_lhs_quant_pack_qsi8d32p_f32,
+            /* .get_packed_offset_ex  = */ &lhs_offs_fn6<kai_get_lhs_packed_offset_lhs_quant_pack_qsi8d32p_f32>,
+            /* .packed_size_ex        = */ &lhs_ps_fn6<kai_get_lhs_packed_size_lhs_quant_pack_qsi8d32p_f32>,
+            /* .pack_func_ex          = */ &lhs_pack_float_fn10<kai_run_lhs_quant_pack_qsi8d32p_f32>,
+        },
+        /* .rhs_info = */ {
+            /* .packed_stride         = */ kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0,
+            /* .to_float              = */ dequantize_row_qsi4c32pscalef16,
+            /* .packed_size_ex        = */ &rhs_ps_fn5<kai_get_rhs_packed_size_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+            /* .pack_func_ex          = */ &rhs_pack_fn12<kai_run_rhs_pack_nxk_qsi4c32pscalef16_qsu4c32s16s0>,
+        },
+        /* .required_cpu       = */ CPU_FEATURE_DOTPROD,
+        /* .lhs_type           = */ GGML_TYPE_F32,
+        /* .rhs_type           = */ GGML_TYPE_Q4_0,
+        /* .op_type            = */ GGML_TYPE_F32,
+    },
+#endif
 #endif
     { /* Sentinel */ }
 };
 
 static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
+#if defined(__ARM_FEATURE_SME)
     {
-        /* SME2 GEMM */
+        /* SME GEMM */
         {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1vlx4_qsi8cxp4vlx4_1vlx4vl_sme2_mopa,
@@ -728,7 +741,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
             /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_quant_pack_qai8dxp_f32>,
             /* .pack_func_ex          = */ &lhs_pack_float_fn9_no_bl<kai_run_lhs_quant_pack_qai8dxp_f32>,
         },
-        /* SME2 GEMV */
+        /* SME GEMV */
         {
             /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
             /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_qai8dxp1x4_qsi8cxp4vlx4_1x4vl_sme2_dot,
@@ -813,6 +826,8 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
         /* .rhs_type           = */ GGML_TYPE_Q8_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
+#if defined(__ARM_FEATURE_MATMUL_INT8)
     {
         /* I8MM GEMM */
         {
@@ -861,11 +876,13 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
             /* .packed_stride_ex      = */ &rhs_stride_fn4<kai_get_rhs_packed_stride_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
             /* .pack_func_ex          = */ &rhs_pack_scale_fn12<kai_run_rhs_pack_nxk_qsi8cxp_qsi8cx_neon>,
         },
-        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM,
         /* .lhs_type           = */ GGML_TYPE_F32,
         /* .rhs_type           = */ GGML_TYPE_Q8_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
+#if defined(__ARM_FEATURE_DOTPROD)
     {
         /* DOTPROD GEMM */
         {
@@ -919,10 +936,12 @@ static ggml_kleidiai_kernels gemm_gemv_kernels_q8[] = {
         /* .rhs_type           = */ GGML_TYPE_Q8_0,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
     { /* Sentinel */ }
 };
 
 static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
+#if defined(__ARM_FEATURE_SME)
     {
         /* SME2 GEMM */
         {
@@ -1029,6 +1048,7 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
         /* .rhs_type           = */ GGML_TYPE_F32,
         /* .op_type            = */ GGML_TYPE_F32,
     },
+#endif
     { /* Sentinel */ }
 };
 
@@ -1036,6 +1056,10 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
     ggml_kleidiai_kernels * kernel = nullptr;
 
     if (tensor->op == GGML_OP_MUL_MAT && tensor->src[0] != nullptr && tensor->src[1] != nullptr) {
+#if defined(__ARM_FEATURE_SME)          ||  \
+    defined(__ARM_FEATURE_DOTPROD)      ||  \
+    defined(__ARM_FEATURE_MATMUL_INT8)  ||  \
+    defined(__ARM_FEATURE_SVE)
         auto try_table = [&](auto & table) {
             for (size_t i = 0; i < NELEMS(table) - 1; ++i) {
                 if ((cpu_features & table[i].required_cpu) == table[i].required_cpu &&
@@ -1056,6 +1080,12 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
         } else {
             try_table(gemm_gemv_kernels);
         }
+#else
+    GGML_UNUSED(gemm_gemv_kernels);
+    GGML_UNUSED(gemm_gemv_kernels_q8);
+    GGML_UNUSED(ggml_kleidiai_kernels_f32);
+    GGML_UNUSED(cpu_features);
+#endif
     }
 
     return kernel;
@@ -1064,13 +1094,19 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels(cpu_feature cpu_features, c
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
+#if defined(__ARM_FEATURE_SME)          ||  \
+    defined(__ARM_FEATURE_DOTPROD)      ||  \
+    defined(__ARM_FEATURE_MATMUL_INT8)  ||  \
+    defined(__ARM_FEATURE_SVE)
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels) - 1; ++i) {
-        if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
-            gemm_gemv_kernels[i].rhs_type == GGML_TYPE_Q4_0) {
+        if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu) {
             kernels = &gemm_gemv_kernels[i];
             break;
         }
     }
+#else
+    GGML_UNUSED(features);
+#endif
 
     return kernels;
 }
@@ -1078,12 +1114,16 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q4_0(cpu_feature features)
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
+#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_DOTPROD) || defined(__ARM_FEATURE_MATMUL_INT8)
     for (size_t i = 0; i < NELEMS(gemm_gemv_kernels_q8) - 1; ++i) {
         if ((features & gemm_gemv_kernels_q8[i].required_cpu) == gemm_gemv_kernels_q8[i].required_cpu) {
             kernels = &gemm_gemv_kernels_q8[i];
             break;
         }
     }
+#else
+    GGML_UNUSED(features);
+#endif
 
     return kernels;
 }
@@ -1091,11 +1131,16 @@ ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_q8_0(cpu_feature features)
 ggml_kleidiai_kernels * ggml_kleidiai_select_kernels_f32(cpu_feature features) {
     ggml_kleidiai_kernels * kernels = nullptr;
 
+#if defined(__ARM_FEATURE_SME)
     for (size_t i = 0; i < NELEMS(ggml_kleidiai_kernels_f32) - 1; ++i) {
         if ((features & ggml_kleidiai_kernels_f32[i].required_cpu) == ggml_kleidiai_kernels_f32[i].required_cpu) {
             kernels = &ggml_kleidiai_kernels_f32[i];
             break;
         }
     }
+#else
+    GGML_UNUSED(features);
+#endif
+
     return kernels;
 }

ggml/src/ggml-cpu/kleidiai/kernels.h

@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
 // SPDX-License-Identifier: MIT
 //
 
@@ -12,8 +12,7 @@ enum cpu_feature {
     CPU_FEATURE_I8MM    = 2,
     CPU_FEATURE_SVE     = 4,
     CPU_FEATURE_SME     = 8,
-    CPU_FEATURE_SME2    = 16,
-    CPU_FEATURE_FP16    = 32
+    CPU_FEATURE_SME2    = 16
 };
 
 inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {

ggml/src/ggml-cpu/kleidiai/kleidiai.cpp

@@ -48,7 +48,7 @@
 
 #include "kernels.h"
 
-#include "kai/kai_common.h"
+#include "kai_common.h"
 
 #define GGML_COMMON_DECL_CPP
 #include "ggml-common.h"
@@ -316,7 +316,6 @@ static void init_kleidiai_context(void) {
 
         ctx.features  = (runtime_feat.has_dotprod  ? CPU_FEATURE_DOTPROD : CPU_FEATURE_NONE) |
                         (runtime_feat.has_i8mm     ? CPU_FEATURE_I8MM    : CPU_FEATURE_NONE) |
-                        (runtime_feat.has_fp16     ? CPU_FEATURE_FP16    : CPU_FEATURE_NONE) |
                         (runtime_feat.sve_cnt == QK8_0 ? CPU_FEATURE_SVE : CPU_FEATURE_NONE);
 
         if (env_threads) {

ggml/src/ggml-cpu/ops.cpp

@@ -3403,139 +3403,6 @@ static void ggml_compute_forward_swiglu_oai(
     }
 }
 
-// ggml_compute_forward_swiglu_clamp
-
-static void ggml_compute_forward_swiglu_clamp_f32(const ggml_compute_params * params, ggml_tensor * dst) {
-    const ggml_tensor * src0   = dst->src[0];
-    const ggml_tensor * src1   = dst->src[1];
-    char *              src0_d = (char *) src0->data;
-    char *              src1_d = (char *) (src1 ? src1->data : src0->data);
-    const size_t        src0_o = src0->nb[1];
-    const size_t        src1_o = src1 ? src1->nb[1] : src0->nb[1];
-
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-
-    if (src1) {
-        GGML_ASSERT(ggml_is_contiguous_1(src1));
-        GGML_ASSERT(src0->type == src1->type);
-    }
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
-    const int nr = ggml_nrows(src0);
-
-    GGML_ASSERT(dst->ne[0] == nc);
-    GGML_ASSERT(ggml_nrows(dst) == nr);
-
-    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
-    const float   limit   = ggml_get_op_params_f32(dst, 3);
-
-    const int dr  = (nr + nth - 1) / nth;
-    const int ir0 = dr * ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * src0_p = (float *) (src0_d + i1 * src0_o);
-        float * src1_p = (float *) (src1_d + i1 * src1_o);
-        float * dst_p  = (float *) ((char *) dst->data + i1 * (dst->nb[1]));
-
-        if (!src1) {
-            src0_p += swapped ? nc : 0;
-            src1_p += swapped ? 0 : nc;
-        }
-
-        for (int k = 0; k < nc; k++) {
-            const float gate = std::min(src0_p[k], limit);
-            const float up   = std::clamp(src1_p[k], -limit, limit);
-            dst_p[k]         = gate / (1.f + expf(-gate)) * up;
-        }
-
-#ifndef NDEBUG
-        for (int k = 0; k < nc; k++) {
-            const float x = dst_p[k];
-            GGML_UNUSED(x);
-            assert(!isnan(x));
-            assert(!isinf(x));
-        }
-#endif  // NDEBUG
-    }
-}
-
-static void ggml_compute_forward_swiglu_clamp_f16(const ggml_compute_params * params, ggml_tensor * dst) {
-    const ggml_tensor * src0   = dst->src[0];
-    const ggml_tensor * src1   = dst->src[1];
-    char *              src0_d = (char *) src0->data;
-    char *              src1_d = (char *) (src1 ? src1->data : src0->data);
-    const size_t        src0_o = src0->nb[1];
-    const size_t        src1_o = src1 ? src1->nb[1] : src0->nb[1];
-
-    GGML_ASSERT(ggml_is_contiguous_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_1(dst));
-
-    if (src1) {
-        GGML_ASSERT(ggml_is_contiguous_1(src1));
-        GGML_ASSERT(src0->type == src1->type);
-    }
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int nc = src1 ? src0->ne[0] : src0->ne[0] / 2;
-    const int nr = ggml_nrows(src0);
-
-    GGML_ASSERT(dst->ne[0] == nc);
-    GGML_ASSERT(ggml_nrows(dst) == nr);
-
-    const int32_t swapped = ggml_get_op_params_i32(dst, 1);
-    const float   limit   = ggml_get_op_params_f32(dst, 3);
-
-    const int dr  = (nr + nth - 1) / nth;
-    const int ir0 = dr * ith;
-    const int ir1 = MIN(ir0 + dr, nr);
-
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        ggml_fp16_t * src0_p = (ggml_fp16_t *) (src0_d + i1 * src0_o);
-        ggml_fp16_t * src1_p = (ggml_fp16_t *) (src1_d + i1 * src1_o);
-        ggml_fp16_t * dst_p  = (ggml_fp16_t *) ((char *) dst->data + i1 * (dst->nb[1]));
-
-        if (!src1) {
-            src0_p += swapped ? nc : 0;
-            src1_p += swapped ? 0 : nc;
-        }
-
-        for (int k = 0; k < nc; k++) {
-            const float gate = std::min(GGML_FP16_TO_FP32(src0_p[k]), limit);
-            const float up   = std::clamp(GGML_FP16_TO_FP32(src1_p[k]), -limit, limit);
-            dst_p[k]         = GGML_FP32_TO_FP16(gate / (1.f + expf(-gate)) * up);
-        }
-
-#ifndef NDEBUG
-        for (int k = 0; k < nc; k++) {
-            const float x = GGML_FP16_TO_FP32(dst_p[k]);
-            GGML_UNUSED(x);
-            assert(!isnan(x));
-            assert(!isinf(x));
-        }
-#endif  // NDEBUG
-    }
-}
-
-static void ggml_compute_forward_swiglu_clamp(const ggml_compute_params * params, ggml_tensor * dst) {
-    switch (dst->src[0]->type) {
-        case GGML_TYPE_F32:
-            ggml_compute_forward_swiglu_clamp_f32(params, dst);
-            break;
-        case GGML_TYPE_F16:
-            ggml_compute_forward_swiglu_clamp_f16(params, dst);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-    }
-}
-
 // ggml_compute_forward_geglu_erf
 
 static void ggml_compute_forward_geglu_erf_f32(
@@ -7400,21 +7267,18 @@ static void ggml_compute_forward_conv_transpose_2d_impl(
             }
         }
 
-        // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh), for all batches
+        // permute source data (src1) from (Sw x Sh x Cin) to (Cin x Sw x Sh)
         {
             kernel_t * const wdata = (kernel_t *) params->wdata + nk;
-            for (int i13 = 0; i13 < ne13; i13++) {
-                kernel_t * const wdata_b = wdata + i13*ne10*ne11*ne12;
-                for (int i12 = 0; i12 < ne12; i12++) {
-                    for (int i11 = 0; i11 < ne11; i11++) {
-                        const float * const src = (float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
-                        kernel_t * dst_data = wdata_b + i11*ne10*ne12;
-                        for (int i10 = 0; i10 < ne10; i10++) {
-                            if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
-                                dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
-                            } else {
-                                dst_data[i10*ne12 + i12] = src[i10];
-                            }
+            for (int i12 = 0; i12 < ne12; i12++) {
+                for (int i11 = 0; i11 < ne11; i11++) {
+                    const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
+                    kernel_t * dst_data = wdata + i11*ne10*ne12;
+                    for (int i10 = 0; i10 < ne10; i10++) {
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
+                        } else {
+                            dst_data[i10*ne12 + i12] = src[i10];
                         }
                     }
                 }
@@ -7441,27 +7305,24 @@ static void ggml_compute_forward_conv_transpose_2d_impl(
     kernel_t * const wdata_src = wdata + nk;
 
     for (int i2 = ip0; i2 < ip1; i2++) { // Cout
+        float * dst_data = (float *)((char *) dst->data + i2*nb2);
         kernel_t * wdata_kernel = wdata + i2*ne01*ne00*ne03;
-        for (int i3 = 0; i3 < ne3; i3++) { // batch
-            float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2);
-            kernel_t * wdata_src_b = wdata_src + i3*ne10*ne11*ne12;
-            for (int i11 = 0; i11 < ne11; i11++) {
-                for (int i10 = 0; i10 < ne10; i10++) {
-                    const int i1n = i11*ne10*ne12 + i10*ne12;
-                    for (int i01 = 0; i01 < ne01; i01++) {
-                        for (int i00 = 0; i00 < ne00; i00++) {
-                            float v = 0;
-                            if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
-                                ggml_vec_dot_f16(ne03, &v, 0,
-                                        wdata_src_b + i1n, 0,
-                                        wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
-                            } else {
-                                ggml_vec_dot_f32(ne03, &v, 0,
-                                        wdata_src_b + i1n, 0,
-                                        wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
-                            }
-                            dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
+        for (int i11 = 0; i11 < ne11; i11++) {
+            for (int i10 = 0; i10 < ne10; i10++) {
+                const int i1n = i11*ne10*ne12 + i10*ne12;
+                for (int i01 = 0; i01 < ne01; i01++) {
+                    for (int i00 = 0; i00 < ne00; i00++) {
+                        float v = 0;
+                        if constexpr (std::is_same_v<kernel_t, ggml_fp16_t>) {
+                            ggml_vec_dot_f16(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
+                        } else {
+                            ggml_vec_dot_f32(ne03, &v, 0,
+                                    wdata_src + i1n, 0,
+                                    wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
                         }
+                        dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
             }
@@ -10269,10 +10130,6 @@ void ggml_compute_forward_glu(
             {
                 ggml_compute_forward_geglu_quick(params, dst);
             } break;
-        case GGML_GLU_OP_SWIGLU_CLAMP:
-            {
-                ggml_compute_forward_swiglu_clamp(params, dst);
-            } break;
         default:
             {
                 GGML_ABORT("fatal error");

ggml/src/ggml-cuda/common.cuh

@@ -1539,7 +1539,6 @@ struct ggml_cuda_mm_fusion_args_host {
     const ggml_tensor * x_scale = nullptr;
     const ggml_tensor * gate_scale = nullptr;
     ggml_glu_op glu_op;
-    float glu_limit = 0.0f;
 };
 struct ggml_cuda_mm_fusion_args_device {
     const void * x_bias = nullptr;
@@ -1548,7 +1547,6 @@ struct ggml_cuda_mm_fusion_args_device {
     const void * x_scale = nullptr;
     const void * gate_scale = nullptr;
     ggml_glu_op glu_op;
-    float glu_limit = 0.0f;
 };
 
 struct ggml_cuda_kernel_launch_params {
@@ -1675,3 +1673,4 @@ static __inline__ void ggml_cuda_kernel_launch(Kernel kernel, const ggml_cuda_ke
     kernel<<<launch_params.block_nums, launch_params.block_dims, launch_params.shmem, launch_params.stream>>>(std::forward<Args>(args)... );
     CUDA_CHECK(cudaGetLastError());
 }
+

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

@@ -2,7 +2,6 @@
 #include "cp-async.cuh"
 #include "mma.cuh"
 #include "fattn-common.cuh"
-#include "fattn-swizzle.cuh"
 
 using namespace ggml_cuda_mma;
 
@@ -67,7 +66,7 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(192, 128, 32, 128, 2,  32,  96,  64,  64, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(192, 128, 64, 128, 2,  32,  96,  64,  64, 2, true);
 
-    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256,  8, 128, 2,  64, 128, 128, 128, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256,  8,  64, 4,  64, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 16,  64, 4,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  32, 128, 128, 128, 2, true);
     GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  32, 128, 128, 128, 2, true);
@@ -361,7 +360,7 @@ static constexpr __device__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, c
 
 // ------------------------------------------------------------------------------------------------------------------
 
-template<int stride_tile, bool swz, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check>
+template<int stride_tile, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check>
 static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
@@ -398,12 +397,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    if constexpr (swz) {
-                        const int smem_offs_b = ggml_cuda_fattn_smem_swizzle::bytes_rc<stride_tile>(i, k*h2_per_chunk);
-                        cp_async_cg_16<preload>(tile_KV_32 + smem_offs_b, KV + i*stride_KV + k*h2_per_chunk);
-                    } else {
-                        cp_async_cg_16<preload>(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i*stride_KV + k*h2_per_chunk);
-                    }
+                    cp_async_cg_16<preload>(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i*stride_KV + k*h2_per_chunk);
                 }
             }
         };
@@ -438,13 +432,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    if constexpr (swz) {
-                        ggml_cuda_memcpy_1<16>((char *) tile_KV + ggml_cuda_fattn_smem_swizzle::bytes_rc<stride_tile>(i, k*h2_per_chunk),
-                            !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
-                    } else {
-                        ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
-                            !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
-                    }
+                    ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
+                        !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
                 }
             }
         };
@@ -579,11 +568,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     constexpr bool Q_in_reg        = ggml_cuda_fattn_mma_get_Q_in_reg (DKQ, DV, ncols);
     constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages  (DKQ, DV, ncols1, ncols2);
 
-    // swizzle the tile stride for K and V based on the batch size.
-    constexpr int stride_tile_K = ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_K2);
-    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_V2);
-    constexpr bool swz_K = ggml_cuda_fattn_smem_swizzle::enabled(nbatch_K2);
-    constexpr bool swz_V = V_is_K_view ? swz_K : ggml_cuda_fattn_smem_swizzle::enabled(nbatch_V2);
+    constexpr int stride_tile_K = nbatch_K2 + 4;
+
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
 
     const int k_VKQ_0 = kb0 * nbatch_fa;
 #if defined(TURING_MMA_AVAILABLE)
@@ -601,7 +588,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         constexpr bool use_cp_async = true;
         cp_async_wait_all();
         __syncthreads();
-        flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check>
+        flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check>
             (V_h2 + int64_t(k_VKQ_0)*stride_V, tile_V, nbatch_V2, stride_V, k_VKQ_sup);
     } else {
         constexpr bool use_cp_async = nstages == 1;
@@ -620,7 +607,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         if constexpr (nstages <= 1) {
             const int k0_diff = k0_stop - k0_start;
             constexpr bool use_cp_async = nstages == 1;
-            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
+            flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check>
                 (K_h2 + int64_t(k_VKQ_0)*stride_K + k0_start, tile_K, k0_diff, stride_K, k_VKQ_sup);
             if (use_cp_async) {
                 cp_async_wait_all();
@@ -636,7 +623,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
                 for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
                     T_A_KQ K_A;
-                    ggml_cuda_fattn_smem_swizzle::load_ldmatrix<stride_tile_K, swz_K>(K_A, tile_K, i_KQ_0, k_KQ_0 - k0_start);
+                    load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
                     if constexpr (cols_per_warp == 8) {
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
                     } else {
@@ -662,7 +649,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     const int i_KQ_0 = i_KQ_00 + (threadIdx.y % np)*T_A_KQ::I;
 
                     T_A_KQ K_A;
-                    ggml_cuda_fattn_smem_swizzle::load_ldmatrix<stride_tile_K, swz_K>(K_A, tile_K, i_KQ_0, k_KQ_0 - k0_start);
+                    load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
                     if constexpr (cols_per_warp == 8) {
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
@@ -956,7 +943,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check>
                     (mask_h + k_VKQ_0 + nbatch_fa, tile_mask, stride_mask, k_VKQ_sup, jt*ncols1, ne01);
             }
-            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
+            flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check>
                 (K_h2 + int64_t(k_VKQ_0 + nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K, k_VKQ_sup);
         }
     }
@@ -972,7 +959,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
             const int i0_diff = i0_stop - i0_start;
             if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
                 constexpr bool use_cp_async = nstages == 1;
-                flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check>
+                flash_attn_ext_f16_load_tile<stride_tile_V, nwarps, nbatch_fa, use_cp_async, oob_check>
                     (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V, k_VKQ_sup);
                 if (use_cp_async) {
                     cp_async_wait_all();
@@ -991,7 +978,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
-                ggml_cuda_fattn_smem_swizzle::load_ldmatrix_trans<stride_tile_V, swz_V>(A, tile_V, (int)(tile_V_i - tile_V) + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2);
+                load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/T_A_VKQ::I], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {
@@ -1017,7 +1004,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::I;
 
                 T_A_VKQ A; // Transposed in both SRAM and registers, load normally.
-                ggml_cuda_fattn_smem_swizzle::load_ldmatrix<stride_tile_V, swz_V>(A, tile_V, (int)(tile_V_i - tile_V) + k0*stride_tile_V + (i_VKQ_0 - i0_start)/2);
+                load_ldmatrix(A, tile_V_i + k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
                 mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::I)], A);
             }
         }
@@ -1181,12 +1168,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     static_assert(nwarps * (cols_per_warp/ncols2) % ncols1 == 0, "bad nwarps");
 
     constexpr int stride_tile_Q = DKQ/2     + 4;
-    // swizzle the tile stride for K and V based on the batch size.
-    constexpr int stride_tile_K = ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_K2);
-    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_V2);
+    constexpr int stride_tile_K = nbatch_K2 + 4;
+
+    constexpr int stride_tile_V = V_is_K_view ? stride_tile_K : nbatch_V2 + 4;
     constexpr int stride_tile_KV_max = stride_tile_K > stride_tile_V ? stride_tile_K : stride_tile_V;
-    constexpr bool swz_K = ggml_cuda_fattn_smem_swizzle::enabled(nbatch_K2);
-    constexpr bool swz_V = V_is_K_view ? swz_K : ggml_cuda_fattn_smem_swizzle::enabled(nbatch_V2);
 
     extern __shared__ half2 tile_Q[];
     half2 * tile_K    = Q_in_reg              ? tile_Q                             : tile_Q + ncols     * stride_tile_Q;
@@ -1280,7 +1265,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
             flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check>
                 (mask_h + kb0*nbatch_fa, tile_mask, stride_mask, k_VKQ_sup, jt*ncols1, ne01);
         }
-        flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
+        flash_attn_ext_f16_load_tile<stride_tile_K, nwarps, nbatch_fa, use_cp_async, oob_check>
             (K_h2 + int64_t(kb0)*nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K, k_VKQ_sup);
     }
 
@@ -1445,17 +1430,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     constexpr int tile_stride = nbatch_combine + 4;
     static_assert((DV/2) % nbatch_combine == 0, "bad nbatch_combine");
 
-    constexpr bool combine_needs_sync = swz_K || swz_V;
-
     if constexpr (cols_per_warp == 8) {
         const int jc_cwmo = (threadIdx.x % (2*T_C_VKQ::J)) / T_C_VKQ::J; // jc combine write meta offset
         const int jc_cwm = threadIdx.y*(2*T_C_VKQ::J) + 2*T_C_VKQ::get_j(-1) + jc_cwmo; // jc combine write meta
         const float2 KQ_cmr = make_float2(KQ_max[jc_cwmo], KQ_rowsum[jc_cwmo]); // KQ combine max rowsum
 
-        if constexpr (combine_needs_sync) {
-            __syncthreads();
-        }
-
         if (((!needs_fixup && !is_fixup) || np > 1) && threadIdx.x < 2*T_C_VKQ::J) {
             // Use the 16 bytes of padding in each row to store the meta data: KQ max, KQ rowsum, KQ max scale.
             ((float2 *) tile_Q)[jc_cwm*(tile_stride/2) + nbatch_combine/2] = KQ_cmr;
@@ -1492,10 +1471,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const bool thread_should_write = T_C_KQ::J == 8 || T_C_KQ::get_j(threadIdx.x & 2) < 8;
 #endif // defined(TURING_MMA_AVAILABLE)
 
-        if constexpr (combine_needs_sync) {
-            __syncthreads();
-        }
-
         if (((!needs_fixup && !is_fixup) || np > 1) && thread_should_write) {
             ((float2 *) tile_Q)[jc_cwm*(tile_stride/2) + nbatch_combine/2] = KQ_cmr;
         }
@@ -1939,11 +1914,8 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
 
     constexpr bool V_is_K_view = DKQ == 576; // Guaranteed by the kernel selection logic in fattn.cu
 
-    // KV tile strides must match flash_attn_ext_f16_iter / _process_tile.
-    const int stride_tile_K = ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_K2, cc);
-    const int stride_tile_V = V_is_K_view ? stride_tile_K : ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_V2, cc);
-    const size_t nbytes_shared_KV_1stage = nbatch_fa            * std::max(stride_tile_K,  stride_tile_V) * sizeof(half2);
-    const size_t nbytes_shared_KV_2stage = nbatch_fa            *         (stride_tile_K + stride_tile_V) * sizeof(half2);
+    const size_t nbytes_shared_KV_1stage = nbatch_fa            * std::max(nbatch_K2 + 4,  nbatch_V2 + 4) * sizeof(half2);
+    const size_t nbytes_shared_KV_2stage = nbatch_fa            *         (nbatch_K2 + 4 + nbatch_V2 + 4) * sizeof(half2);
     const size_t nbytes_shared_Q         = ncols                * (DKQ/2 + 4)                             * sizeof(half2);
     const size_t nbytes_shared_mask      = ncols1               * (nbatch_fa/2 + 4)                       * sizeof(half2);
     const size_t nbytes_shared_combine   = nwarps*cols_per_warp * (nbatch_combine + 4)                    * sizeof(half2);

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

@@ -1,126 +0,0 @@
-#pragma once
-
-#include "common.cuh"
-#include "mma.cuh"
-
-// XOR swizzle for K/V SMEM tiles to avoid bank conflicts without row padding (Turing+ only).
-// Stride must be a multiple of 32 half2 columns, otherwise we keep +4 row padding.
-
-namespace ggml_cuda_fattn_smem_swizzle {
-
-static __host__ __device__ constexpr bool bank_aligned(const int nbatch_2) {
-    return nbatch_2 >= 32 && nbatch_2 % 32 == 0;
-}
-
-static __device__ constexpr bool enabled(const int nbatch_2) {
-#if defined(TURING_MMA_AVAILABLE)
-    return bank_aligned(nbatch_2);
-#else
-    GGML_UNUSED(nbatch_2);
-    return false;
-#endif // defined(TURING_MMA_AVAILABLE)
-}
-
-static __host__ bool enabled(const int nbatch_2, const int cc) {
-#ifdef GGML_USE_HIP
-    GGML_UNUSED(nbatch_2);
-    GGML_UNUSED(cc);
-    return false;
-#else
-    return turing_mma_available(cc) && bank_aligned(nbatch_2);
-#endif // GGML_USE_HIP
-}
-
-static __device__ constexpr int tile_stride(const int nbatch_2) {
-    return enabled(nbatch_2) ? nbatch_2 : nbatch_2 + 4;
-}
-
-static __host__ int tile_stride(const int nbatch_2, const int cc) {
-    return enabled(nbatch_2, cc) ? nbatch_2 : nbatch_2 + 4;
-}
-
-// Swizzled byte offset for tile element (row, col_h2), same map used for writes and reads.
-template<int stride_h2>
-static __device__ __forceinline__ int bytes_rc(const int row, const int col_h2) {
-    static_assert(bank_aligned(stride_h2), "swizzled tile needs a stride that is a multiple of 32");
-    return ((row * stride_h2 + col_h2) * (int) sizeof(half2)) ^ ((row & 7) << 4);
-}
-
-// ldmatrix.x4 via 64-bit generic pointer.
-static __device__ __forceinline__ void ldmatrix_x4(int * xi, const half2 * addr) {
-#if defined(TURING_MMA_AVAILABLE)
-    asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
-        : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
-        : "l"(addr));
-#else
-    GGML_UNUSED_VARS(xi, addr);
-    NO_DEVICE_CODE;
-#endif // defined(TURING_MMA_AVAILABLE)
-}
-
-static __device__ __forceinline__ void ldmatrix_x4_trans(int * xi, const half2 * addr) {
-#if defined(TURING_MMA_AVAILABLE)
-    asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
-        : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
-        : "l"(addr));
-#else
-    GGML_UNUSED_VARS(xi, addr);
-    NO_DEVICE_CODE;
-#endif // defined(TURING_MMA_AVAILABLE)
-}
-
-// Per-lane swizzled address for one tile<16, 8, half2> ldmatrix: 16 rows, 4 half2 columns per lane.
-template<int stride_h2>
-static __device__ __forceinline__ const half2 * lane_addr(
-        const half2 * tile_base, const int base_row, const int base_col_h2, const int I, const int J) {
-    static_assert(bank_aligned(stride_h2), "swizzled tile needs a stride that is a multiple of 32");
-    const int lane_row = threadIdx.x % I;
-    const int lane_col = (threadIdx.x / I) * (J / 2);
-    uint32_t byte_off = (uint32_t) ((base_row + lane_row)*stride_h2 + base_col_h2 + lane_col) * (uint32_t) sizeof(half2);
-    byte_off ^= (uint32_t) (((base_row + lane_row) & 7) << 4);
-    return (const half2 *) ((const char *) tile_base + byte_off);
-}
-
-template<int stride_h2, bool swz, typename TileT>
-static __device__ __forceinline__ void load_ldmatrix(
-        TileT & t, const half2 * tile_base, const int base_row, const int base_col_h2) {
-    if constexpr (swz) {
-        static_assert(std::is_same_v<TileT, ggml_cuda_mma::tile<16, 8, half2>>,
-            "the swizzled layout is only supported for tile<16, 8, half2>");
-        ldmatrix_x4((int *) t.x, lane_addr<stride_h2>(tile_base, base_row, base_col_h2, TileT::I, TileT::J));
-    } else {
-        ggml_cuda_mma::load_ldmatrix(t, tile_base + base_row*stride_h2 + base_col_h2, stride_h2);
-    }
-}
-
-template<int stride_h2, bool swz, typename TileT>
-static __device__ __forceinline__ void load_ldmatrix(TileT & t, const half2 * tile_base, const int off_h2) {
-    if constexpr (swz) {
-        load_ldmatrix<stride_h2, swz>(t, tile_base, off_h2 / stride_h2, off_h2 % stride_h2);
-    } else {
-        ggml_cuda_mma::load_ldmatrix(t, tile_base + off_h2, stride_h2);
-    }
-}
-
-template<int stride_h2, bool swz, typename TileT>
-static __device__ __forceinline__ void load_ldmatrix_trans(
-        TileT & t, const half2 * tile_base, const int base_row, const int base_col_h2) {
-    if constexpr (swz) {
-        static_assert(std::is_same_v<TileT, ggml_cuda_mma::tile<16, 8, half2>>,
-            "the swizzled layout is only supported for tile<16, 8, half2>");
-        ldmatrix_x4_trans((int *) t.x, lane_addr<stride_h2>(tile_base, base_row, base_col_h2, TileT::I, TileT::J));
-    } else {
-        ggml_cuda_mma::load_ldmatrix_trans(t, tile_base + base_row*stride_h2 + base_col_h2, stride_h2);
-    }
-}
-
-template<int stride_h2, bool swz, typename TileT>
-static __device__ __forceinline__ void load_ldmatrix_trans(TileT & t, const half2 * tile_base, const int off_h2) {
-    if constexpr (swz) {
-        load_ldmatrix_trans<stride_h2, swz>(t, tile_base, off_h2 / stride_h2, off_h2 % stride_h2);
-    } else {
-        ggml_cuda_mma::load_ldmatrix_trans(t, tile_base + off_h2, stride_h2);
-    }
-}
-
-} // namespace ggml_cuda_fattn_smem_swizzle

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

@@ -32,14 +32,12 @@
 #include "ggml-cuda/mmq.cuh"
 #include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
-#include "ggml-cuda/moe-weighted-reduction.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
 #include "ggml-cuda/opt-step-sgd.cuh"
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
-#include "ggml-cuda/pool1d.cuh"
 #include "ggml-cuda/quantize.cuh"
 #include "ggml-cuda/rope.cuh"
 #include "ggml-cuda/roll.cuh"
@@ -916,7 +914,6 @@ static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_t
         : ggml_nbytes(tensor);
     int64_t ne0 = tensor->ne[0];
 
-    // [TAG_ALLOC_SIZE_EXPAND]
     if (ggml_is_quantized(tensor->type)) {
         if (ne0 % MATRIX_ROW_PADDING != 0) {
             GGML_ASSERT(tensor->nb[0] == ggml_element_size(tensor));
@@ -1746,7 +1743,7 @@ static bool ggml_cuda_should_fuse_mul_mat(const ggml_tensor * ffn_up,
         return false;
     }
 
-    static constexpr std::array<ggml_glu_op, 4> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI, GGML_GLU_OP_SWIGLU_CLAMP };
+    static constexpr std::array<ggml_glu_op, 3> valid_glu_ops = { GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_OAI };
 
     if (std::find(valid_glu_ops.begin(), valid_glu_ops.end(), ggml_get_glu_op(glu)) == valid_glu_ops.end()) {
         return false;
@@ -1808,7 +1805,7 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
         return false;
     }
 
-    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] > get_mmvq_mmid_max_batch(src0->type, cc)) {
+    if (tensor->op == GGML_OP_MUL_MAT_ID && dst->ne[2] != 1) {
         return false;
     }
 
@@ -2205,9 +2202,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                 case GGML_GLU_OP_GEGLU_QUICK:
                     ggml_cuda_op_geglu_quick(ctx, dst);
                     break;
-                case GGML_GLU_OP_SWIGLU_CLAMP:
-                    ggml_cuda_op_swiglu_clamp(ctx, dst);
-                    break;
                 default:
                     return false;
             }
@@ -2332,9 +2326,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_POOL_2D:
             ggml_cuda_op_pool2d(ctx, dst);
             break;
-        case GGML_OP_POOL_1D:
-            ggml_cuda_op_pool1d(ctx, dst);
-            break;
         case GGML_OP_SUM:
             ggml_cuda_op_sum(ctx, dst);
             break;
@@ -2984,10 +2975,9 @@ static bool ggml_cuda_check_fusion_memory_ranges(const ggml_cgraph * cgraph,
     };
 
     bool is_ok = true;
-    // one block reads all logits before it writes, so logits may alias the out nodes
-    const ggml_tensor * logits_may_alias = nullptr;
-    if (is_topk_moe && ggml_nrows(cgraph->nodes[node_idx]) <= TOPK_MOE_ROWS_PER_BLOCK) {
-        logits_may_alias = cgraph->nodes[node_idx]->src[0];
+    // exception for topk-moe, as each row is read entirely before writing
+    if (ggml_nrows(cgraph->nodes[node_idx]) == 1 && is_topk_moe) {
+        return true;
     }
 
     for (int i = 0; i < out_count; ++i) {
@@ -3001,7 +2991,7 @@ static bool ggml_cuda_check_fusion_memory_ranges(const ggml_cgraph * cgraph,
             for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
                 const ggml_tensor * src = cgraph->nodes[j]->src[src_idx];
 
-                if (!src || src->op == GGML_OP_NONE || src == logits_may_alias) {
+                if (!src || src->op == GGML_OP_NONE) {
                     continue;
                 }
 
@@ -3027,150 +3017,6 @@ static bool ggml_cuda_check_fusion_memory_ranges(const ggml_cgraph * cgraph,
     return is_ok;
 }
 
-// The long form spans 2*k + 1 nodes. ggml_can_fuse_subgraph() accepts at most
-// 31 nodes, so k <= 15; larger values use the per-operation path.
-static constexpr int MOE_WEIGHTED_REDUCTION_MAX_EXPERTS = 15;
-
-struct ggml_cuda_moe_weighted_reduction_match {
-    const ggml_tensor * experts      = nullptr;
-    const ggml_tensor * expert_scale = nullptr;
-    const ggml_tensor * weights      = nullptr;
-    ggml_tensor *       dst          = nullptr;
-    int                 node_count   = 0;
-};
-
-static bool ggml_cuda_match_moe_weighted_reduction(
-        const ggml_cgraph * cgraph,
-        int node_idx,
-        ggml_cuda_moe_weighted_reduction_match & match) {
-    const ggml_tensor * first = cgraph->nodes[node_idx];
-    if (first->op != GGML_OP_MUL || first->type != GGML_TYPE_F32 || !ggml_is_contiguous(first)) {
-        return false;
-    }
-
-    auto split_mul = [](const ggml_tensor * mul, const ggml_tensor *& full, const ggml_tensor *& broadcast) {
-        auto is_weights = [mul](const ggml_tensor * tensor) {
-            return tensor && tensor->type == GGML_TYPE_F32 && ggml_is_contiguous(tensor) && tensor->ne[0] == 1 &&
-                tensor->ne[1] == mul->ne[1] && tensor->ne[2] == mul->ne[2] && tensor->ne[3] == mul->ne[3];
-        };
-        auto is_experts = [mul](const ggml_tensor * tensor) {
-            return tensor && tensor->type == GGML_TYPE_F32 && ggml_is_contiguous(tensor) &&
-                ggml_are_same_shape(tensor, mul);
-        };
-
-        if (is_experts(mul->src[0]) && is_weights(mul->src[1])) {
-            full      = mul->src[0];
-            broadcast = mul->src[1];
-            return true;
-        }
-        if (is_experts(mul->src[1]) && is_weights(mul->src[0])) {
-            full      = mul->src[1];
-            broadcast = mul->src[0];
-            return true;
-        }
-        return false;
-    };
-
-    const ggml_tensor * weighted     = first;
-    const ggml_tensor * experts      = nullptr;
-    const ggml_tensor * expert_scale = nullptr;
-    const ggml_tensor * weights      = nullptr;
-    int                 mul_count    = 1;
-
-    // Match both structural forms:
-    //   (experts * expert_scale) * router_weight
-    //   experts * router_weight
-    // The matcher does not depend on the model or quantization type.
-    if (node_idx + 1 < cgraph->n_nodes) {
-        const ggml_tensor * second = cgraph->nodes[node_idx + 1];
-        const ggml_tensor * scaled = nullptr;
-        const ggml_tensor * route  = nullptr;
-        const ggml_tensor * raw    = nullptr;
-        const ggml_tensor * scale  = nullptr;
-        if (second->op == GGML_OP_MUL && second->type == GGML_TYPE_F32 && ggml_is_contiguous(second) &&
-                split_mul(second, scaled, route) && scaled == first && split_mul(first, raw, scale)) {
-            weighted     = second;
-            experts      = raw;
-            expert_scale = scale;
-            weights      = route;
-            mul_count    = 2;
-        }
-    }
-
-    if (experts == nullptr && !split_mul(first, experts, weights)) {
-        return false;
-    }
-
-    const int     n_expert_used = (int) weighted->ne[1];
-    const int64_t n_tokens      = weighted->ne[2] * weighted->ne[3];
-    if (n_expert_used < 2 || n_expert_used > MOE_WEIGHTED_REDUCTION_MAX_EXPERTS || n_tokens <= 0) {
-        return false;
-    }
-
-    const int node_count = 2 * n_expert_used + mul_count - 1;
-    if (node_idx + node_count > cgraph->n_nodes) {
-        return false;
-    }
-
-    std::vector<ggml_op> ops(node_count, GGML_OP_VIEW);
-    ops[0] = GGML_OP_MUL;
-    if (mul_count == 2) {
-        ops[1] = GGML_OP_MUL;
-    }
-    std::vector<const ggml_tensor *> views;
-    views.reserve(n_expert_used);
-    const ggml_tensor * previous = nullptr;
-    int n_adds = 0;
-    for (int offset = mul_count; offset < node_count; ++offset) {
-        const ggml_tensor * candidate = cgraph->nodes[node_idx + offset];
-        ops[offset] = candidate->op;
-
-        if (candidate->op == GGML_OP_VIEW) {
-            const int expert = (int) views.size();
-            if (expert >= n_expert_used || candidate->src[0] != weighted || candidate->view_src != weighted ||
-                    candidate->type != GGML_TYPE_F32 || candidate->ne[0] != weighted->ne[0] ||
-                    candidate->ne[1] != n_tokens || candidate->ne[2] != 1 || candidate->ne[3] != 1 ||
-                    candidate->nb[0] != weighted->nb[0] || candidate->nb[1] != weighted->nb[2] ||
-                    candidate->view_offs != (size_t) expert * weighted->nb[1]) {
-                return false;
-            }
-            views.push_back(candidate);
-            continue;
-        }
-
-        if (candidate->op != GGML_OP_ADD || views.size() < 2 || n_adds + 1 >= (int) views.size()) {
-            return false;
-        }
-        const ggml_tensor * lhs = n_adds == 0 ? views[0] : previous;
-        const ggml_tensor * rhs = views[n_adds + 1];
-        if (candidate->src[0] != lhs || candidate->src[1] != rhs || candidate->type != GGML_TYPE_F32) {
-            return false;
-        }
-        previous = candidate;
-        ++n_adds;
-    }
-
-    if ((int) views.size() != n_expert_used || n_adds != n_expert_used - 1 || previous == nullptr) {
-        return false;
-    }
-    if (!ggml_is_contiguous(previous) || previous->ne[0] != weighted->ne[0] ||
-            previous->ne[1] != n_tokens || previous->ne[2] != 1 || previous->ne[3] != 1) {
-        return false;
-    }
-
-    const int output_idx = node_idx + node_count - 1;
-    if (!ggml_can_fuse_subgraph(cgraph, node_idx, node_count, ops.data(), &output_idx, 1)) {
-        return false;
-    }
-
-    match.experts      = experts;
-    match.expert_scale = expert_scale;
-    match.weights      = weights;
-    match.dst          = cgraph->nodes[output_idx];
-    match.node_count   = node_count;
-    return true;
-}
-
 
 static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
                                int                                       node_idx,
@@ -3433,18 +3279,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
 
     ggml_tensor * node = cgraph->nodes[i];
 
-    if (node->op == GGML_OP_MUL) {
-        ggml_cuda_moe_weighted_reduction_match match;
-        if (ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
-            const int output_idx = i + match.node_count - 1;
-            if (ggml_cuda_check_fusion_memory_ranges(cgraph, i, match.node_count, &output_idx, 1)) {
-                ggml_cuda_op_moe_weighted_reduction(
-                    *cuda_ctx, match.experts, match.expert_scale, match.weights, match.dst);
-                return match.node_count - 1;
-            }
-        }
-    }
-
     // gated_delta_net -> cpy: scatter recurrent-state snapshots into the cache
     if (node->op == GGML_OP_GATED_DELTA_NET) {
         ggml_cuda_gated_delta_net_fused_cache fused_state_cpy;
@@ -3757,7 +3591,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                 fusion_data.x_scale    = up_scale;
                 fusion_data.gate_scale = gate_scale;
                 fusion_data.glu_op     = ggml_get_glu_op(glu);
-                fusion_data.glu_limit  = ggml_get_op_params_f32(glu, 3);
 
                 if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
                     ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
@@ -3851,7 +3684,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                 fusion_data.x_scale    = up_scale;
                 fusion_data.gate_scale = gate_scale;
                 fusion_data.glu_op     = ggml_get_glu_op(glu);
-                fusion_data.glu_limit  = ggml_get_op_params_f32(glu, 3);
 
                 if (ggml_cuda_should_fuse_mul_mat_vec_q(up_n)) {
                     ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, cgraph->nodes[glu_idx], &fusion_data);
@@ -3908,7 +3740,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                 fusion_data.x_bias    = up_bias_tensor;
                 fusion_data.gate_bias = gate_bias_tensor;
                 fusion_data.glu_op    = ggml_get_glu_op(glu);
-                fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
 
                 ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
                 fused_mul_mat_vec = true;
@@ -3922,7 +3753,6 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
                 fusion_data.x_bias    = up_bias_tensor;
                 fusion_data.gate_bias = gate_bias_tensor;
                 fusion_data.glu_op    = ggml_get_glu_op(glu);
-                fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
 
                 ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
                 fused_mul_mat_vec = true;
@@ -3947,9 +3777,8 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
 
             if (ggml_cuda_should_fuse_mul_mat_vec_f(up)) {
                 ggml_cuda_mm_fusion_args_host fusion_data{};
-                fusion_data.gate      = gate->src[0];
-                fusion_data.glu_op    = ggml_get_glu_op(glu);
-                fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
+                fusion_data.gate   = gate->src[0];
+                fusion_data.glu_op = ggml_get_glu_op(glu);
 
                 ggml_cuda_mul_mat_vec_f(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
                 fused_mul_mat_vec = true;
@@ -3959,9 +3788,8 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph
 
             if (ggml_cuda_should_fuse_mul_mat_vec_q(up)) {
                 ggml_cuda_mm_fusion_args_host fusion_data{};
-                fusion_data.gate      = gate->src[0];
-                fusion_data.glu_op    = ggml_get_glu_op(glu);
-                fusion_data.glu_limit = ggml_get_op_params_f32(glu, 3);
+                fusion_data.gate   = gate->src[0];
+                fusion_data.glu_op = ggml_get_glu_op(glu);
 
                 ggml_cuda_mul_mat_vec_q(*cuda_ctx, src0, src1, ids, glu, &fusion_data);
                 fused_mul_mat_vec = true;
@@ -4496,31 +4324,9 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
     }
 }
 
-static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph, ggml_backend_graph_optimize_params * params) {
+static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
     ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
 
-    static const bool disable_fusion = getenv("GGML_CUDA_DISABLE_FUSION") != nullptr && std::atoi(getenv("GGML_CUDA_DISABLE_FUSION"));
-    if (!disable_fusion) {
-        for (int i = 0; i < cgraph->n_nodes; ++i) {
-            if (cgraph->nodes[i]->op != GGML_OP_MUL) {
-                continue;
-            }
-
-            ggml_cuda_moe_weighted_reduction_match match;
-            if (!ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
-                continue;
-            }
-
-            params->add_alloc_dep(params->user_data, const_cast<ggml_tensor *>(match.experts), match.dst);
-            params->add_alloc_dep(params->user_data, const_cast<ggml_tensor *>(match.weights), match.dst);
-            if (match.expert_scale != nullptr) {
-                params->add_alloc_dep(
-                    params->user_data, const_cast<ggml_tensor *>(match.expert_scale), match.dst);
-            }
-            i += match.node_count - 1;
-        }
-    }
-
 #ifdef USE_CUDA_GRAPH
     const void * graph_key = ggml_cuda_graph_get_key(cgraph);
     const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
@@ -4801,8 +4607,8 @@ static std::string ggml_cuda_device_description(int device) {
     const ggml_cuda_device_info & info = ggml_cuda_info();
     std::string description = prop.name;
     if (info.device_count > info.physical_device_count) {
-        description += " (dev p" + std::to_string(info.devices[device].physical_device) +
-                       "/v" + std::to_string(info.devices[device].virtual_index) + ")";
+        description += " (physical device " + std::to_string(info.devices[device].physical_device) +
+                       ", virtual device " + std::to_string(info.devices[device].virtual_index) + ")";
     }
     return description;
 }
@@ -5107,7 +4913,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 case GGML_GLU_OP_SWIGLU_OAI:
                 case GGML_GLU_OP_GEGLU_ERF:
                 case GGML_GLU_OP_GEGLU_QUICK:
-                case GGML_GLU_OP_SWIGLU_CLAMP:
                     return ggml_is_contiguous_1(op->src[0]);
                 default:
                     return false;
@@ -5440,7 +5245,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CONV_2D_DW:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_CONV_TRANSPOSE_2D:
-        case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
             return true;
         case GGML_OP_ACC:
@@ -5450,11 +5254,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SUM:
             return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_TOP_K:
-#if defined(GGML_USE_HIP) || defined(GGML_CUDA_USE_CUB)
-            return true;
-#else
-            return op->src[0]->ne[0] <= 1024;
-#endif // defined(GGML_USE_HIP) || defined(GGML_CUDA_USE_CUB)
         case GGML_OP_ARGSORT:
 #ifndef GGML_CUDA_USE_CUB
             return op->src[0]->ne[0] <= 1024;

ggml/src/ggml-cuda/mmid.cu

@@ -19,11 +19,6 @@ struct mm_ids_helper_store {
 };
 static_assert(sizeof(mm_ids_helper_store) == 4, "unexpected size for mm_ids_helper_store");
 
-// the generic path passes 0, which needs no padding since it never groups lanes by token
-template <int n> struct mm_ids_pow2 { static constexpr int value = 2*mm_ids_pow2<(n + 1)/2>::value; };
-template <>      struct mm_ids_pow2<1> { static constexpr int value = 1; };
-template <>      struct mm_ids_pow2<0> { static constexpr int value = 1; };
-
 // Helper function for mul_mat_id, converts ids to a more convenient format.
 // ids_src1 describes how to permute the flattened column indices of src1 in order to get a compact src1 tensor sorted by expert.
 // ids_dst describes the same mapping but for the dst tensor.
@@ -37,9 +32,6 @@ static __global__ void mm_ids_helper(
     const int n_expert_used = n_expert_used_template == 0 ? n_expert_used_var : n_expert_used_template;
     const int expert = blockIdx.x;
 
-    // token slots per warp lane group, padded to a power of 2 so a warp divides evenly
-    constexpr int neu_padded = mm_ids_pow2<n_expert_used_template>::value;
-
     extern __shared__ char data_mm_ids_helper[];
     mm_ids_helper_store * store = (mm_ids_helper_store *) data_mm_ids_helper;
 
@@ -68,8 +60,8 @@ static __global__ void mm_ids_helper(
         }
     } else {
         // Implementation optimized for specific numbers of experts used:
-        // a warp holds a whole number of token slots, so the slot count is padded to a power of 2
-        static_assert(neu_padded <= warp_size && warp_size % neu_padded == 0, "bad n_expert_used");
+        static_assert(n_expert_used == 6 || warp_size % n_expert_used == 0, "bad n_expert_used");
+        const int neu_padded = n_expert_used == 6 ? 8 : n_expert_used; // Padded to next higher power of 2.
         for (int it0 = 0; it0 < n_tokens; it0 += warp_size/neu_padded) {
             const int it = it0 + threadIdx.x / neu_padded;
 
@@ -164,9 +156,6 @@ void ggml_cuda_launch_mm_ids_helper(
         case  8:
             launch_mm_ids_helper< 8>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
             break;
-        case 10:
-            launch_mm_ids_helper<10>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
-            break;
         case 16:
             launch_mm_ids_helper<16>(ids, ids_src1, ids_dst, expert_bounds, n_experts, n_tokens, n_expert_used, nchannels_y, si1, sis1, write_inverse, stream);
             break;

ggml/src/ggml-cuda/mmq-config-pascal-older.cuh

@@ -1,273 +0,0 @@
-static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal_older(ggml_type type, int J, bool fallback) {
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-// ---------------------------------------------------------------------------------------------
-
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 256, 1, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-
-// ---------------------------------------------------------------------------------------------
-
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
-// ---------------------------------------------------------------------------------------------
-
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  24, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  40, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-
-    return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);
-}

ggml/src/ggml-cuda/mmq-config-pascal.cuh

@@ -1,4 +1,4 @@
-static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal_dp4a(ggml_type type, int J, bool fallback) {
+static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal(ggml_type type, int J, bool fallback) {
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,   8, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);

ggml/src/ggml-cuda/mmq-config-rdna3.cuh

@@ -1,273 +1,289 @@
 static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_rdna3(ggml_type type, int J, bool fallback) {
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q1_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q1_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q1_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-
     CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q2_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q2_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_0, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q2_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q2_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_0, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q4_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 128, 4,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_0, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_1, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_0, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 128, 4,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_0, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_1, 128, 4,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_1, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_1, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q8_0, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q8_0, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
 // ---------------------------------------------------------------------------------------------
 
     CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q2_K, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 1, 128,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q2_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q2_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q2_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q2_K, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q3_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q3_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q3_K, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q3_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q4_K, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q4_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q5_K, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 128, 4,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 128, 4,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q5_K, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q5_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_Q6_K, 128, 4,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_Q6_K, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_Q6_K, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_Q6_K, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q6_K, MMQ_ITER_K, false, false);
 
 // ---------------------------------------------------------------------------------------------
 
     CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 128, 4,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 128, 4,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 128, 1,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ1_S, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ1_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 4,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 4,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XXS, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 128, 4,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 128, 4,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_XS, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 256, 1, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 128, 4,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ2_S, 128, 4,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ2_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q3_K, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 128, 4,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 128, 4,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_XXS, 128, 1,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_XXS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 128, 4,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 128, 4,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ3_S, 128, 2,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ3_S, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_XS, 128, 4,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_XS, 128, 2,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_XS, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 128, 1,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_IQ4_NL, 128, 2,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_IQ4_NL, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, MMQ_ITER_K, false, false);
 
 // ---------------------------------------------------------------------------------------------
 
     CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_MXFP4, 128, 1,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 128, 1,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 128, 1,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_MXFP4, 128, 2,  64, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_MXFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_1, MMQ_ITER_K, false, false);
 
     CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
-    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, true);
     CASE(GGML_TYPE_NVFP4, 128, 2,  64,  16, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 128, 2,  64,  32, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
-    CASE(GGML_TYPE_NVFP4, 128, 2,  64,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  48, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  64, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128,  80, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128,  96, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
+    CASE(GGML_TYPE_NVFP4, 256, 2, 128, 112, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
     CASE(GGML_TYPE_NVFP4, 256, 2, 128, 128, GGML_CUDA_MMQ_SRAM_LAYOUT_NVFP4, MMQ_ITER_K, false, false);
 
     return ggml_cuda_mmq_config(GGML_TYPE_COUNT, 256, 2, 128, 64, GGML_CUDA_MMQ_SRAM_LAYOUT_Q8_0, 256, false, true);

ggml/src/ggml-cuda/mmq-load-tiles.cuh

@@ -138,20 +138,12 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
         for (int j = 0; j < 4; ++j) {
             const int q  = qxi[j];
 
-#if defined(GGML_USE_HIP)
-            const uint32_t qx_indices = (q & 0x03) | ((q & 0x0C) << 6) | ((q & 0x30) << 12) | ((q & 0xC0) << 18);
-            const uint32_t qy_bits    = q >> 8;
-            const uint32_t qy_indices = (qy_bits & 0x03) | ((qy_bits & 0x0C) << 6) | ((qy_bits & 0x30) << 12) | ((qy_bits & 0xC0) << 18);
-            const int qx = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qx_indices);
-            const int qy = __builtin_amdgcn_perm(0x020100FF, 0x020100FF, qy_indices);
-#else
             // unpack even and odd crumbs into byte values
             const int qe = __byte_perm(0x020100FF, 0x020100FF, q >> 0);
             const int qo = __byte_perm(0x020100FF, 0x020100FF, q >> 2);
             // unshuffle values
             const int qx = __byte_perm(qe, qo, 0x5140);
             const int qy = __byte_perm(qe, qo, 0x7362);
-#endif // defined(GGML_USE_HIP)
 
 #if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
             x_qs[i*sram_stride           + dst_offset + j*2+0] = qx;

ggml/src/ggml-cuda/mmq.cu

@@ -314,9 +314,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     }
 
     if (ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_DP4A) {
-        // for MoE, mmq is faster even without native dp4a
-        // TODO: check if cards older than pascal might benefit from this as well
-        return cc >= GGML_CUDA_CC_PASCAL && n_experts > 0;
+        return false;
     }
 
 #ifdef GGML_CUDA_FORCE_MMQ