test: move tools/parser to tests

.devops/openvino.Dockerfile

@@ -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-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:
@@ -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
 
@@ -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-cpu.yml

@@ -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' }}
 
@@ -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@v1.2.21
         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@v1.2.21
         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@v1.2.21
         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-opencl.yml

@@ -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

@@ -167,13 +167,3 @@ jobs:
 
           cd build
           ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -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' }}

.github/workflows/build-sycl.yml

@@ -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
 
@@ -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

@@ -55,7 +55,7 @@ jobs:
       - name: ccache
         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
 
@@ -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
 
@@ -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

@@ -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

@@ -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
 
@@ -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/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

@@ -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:
@@ -96,11 +119,12 @@ 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: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -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
@@ -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
@@ -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
@@ -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
@@ -429,11 +460,12 @@ 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: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -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
@@ -544,11 +577,12 @@ 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: ccache
         uses: ggml-org/ccache-action@v1.2.21
@@ -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
@@ -668,11 +702,12 @@ 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: Install Ninja
         run: |
@@ -714,10 +749,8 @@ jobs:
         with:
           key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
 
-  # TODO: build only the ggml-hip backend like the other windows backend jobs
-  #       (windows-cuda, windows-sycl), then drop the ui-build dependency
   windows-rocm:
-    needs: [check-release, ui-build]
+    needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
     runs-on: windows-2022
@@ -736,12 +769,6 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Download UI build
-        uses: actions/download-artifact@v7
-        with:
-          name: llama-ui.zip
-          path: tools/ui/dist
-
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
@@ -852,8 +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' }}
@@ -884,6 +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: ${{ matrix.backend == 'vulkan' }}
@@ -946,8 +978,6 @@ jobs:
           path: llama-bin-win-${{ 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]
@@ -976,6 +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:
@@ -1047,8 +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' }}
@@ -1083,6 +1118,13 @@ 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@v1.2.21
         with:
@@ -1153,7 +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:
@@ -1195,11 +1237,12 @@ 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@v1.2.21
@@ -1244,11 +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
@@ -1267,11 +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
@@ -1281,7 +1325,7 @@ jobs:
       - name: ccache
         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"
 
@@ -1344,6 +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)
 
@@ -1369,10 +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
 
@@ -1400,7 +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
@@ -1523,9 +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' }}
@@ -1540,13 +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
@@ -1579,27 +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
@@ -1623,7 +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

@@ -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
 
@@ -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)

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>
 

ci/run.sh

@@ -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 {
@@ -785,8 +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

@@ -2644,27 +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",
@@ -2720,19 +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(
-        {"--tensor-read-lazy"}, "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.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_ON;   }
-            else if (value == "auto") { params.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_AUTO; }
-            else if (value == "off")  { params.tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_OFF;  }
-            else { throw std::invalid_argument("invalid value"); }
-        }
-    ).set_env("LLAMA_ARG_TENSOR_READ_LAZY"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"
@@ -2784,20 +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",
@@ -4124,7 +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"));
 
@@ -4145,38 +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);
@@ -1688,7 +1687,6 @@ 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.tensor_read_lazy = params.tensor_read_lazy;
     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_tensor_read_lazy tensor_read_lazy = LLAMA_TENSOR_READ_LAZY_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;
@@ -1123,30 +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/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/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");
@@ -939,7 +937,7 @@ struct common_speculative_impl_draft_dflash : public common_speculative_impl {
 
     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)
     {
@@ -985,7 +983,6 @@ 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_batch(ctx_dft), n_embd_dec, n_seq);
@@ -1318,7 +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;
@@ -1385,7 +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));
 
@@ -1730,7 +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)
     {
@@ -1774,7 +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);
@@ -1845,7 +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) {
@@ -2021,7 +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)
@@ -2142,8 +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(
@@ -2322,101 +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();
 
@@ -2669,39 +2568,13 @@ 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/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):

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.
@@ -68,8 +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
@@ -146,12 +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
@@ -163,8 +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
@@ -172,6 +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/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}")
 
@@ -342,6 +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-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

@@ -1724,19 +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);
@@ -2003,6 +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)

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

@@ -182,8 +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);
     }
 }
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -576,26 +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")
@@ -610,49 +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
@@ -660,6 +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/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-cuda/ggml-cuda.cu

@@ -38,7 +38,6 @@
 #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"
@@ -2327,9 +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;
@@ -4611,8 +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;
 }
@@ -5249,7 +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:

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.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

ggml/src/ggml-cuda/mmq.cuh

@@ -213,8 +213,7 @@ struct ggml_cuda_mmq_config {
         return ggml_cuda_mmq_config((type_), (nthreads_), (occupancy_), (I_), (J_), (sram_layout_), (K_vram_), (stream_k_), (fallback_)); \
     }                                                                                                                                     \
 
-#include "mmq-config-pascal-older.cuh"
-#include "mmq-config-pascal-dp4a.cuh"
+#include "mmq-config-pascal.cuh"
 #include "mmq-config-ampere.cuh"
 #include "mmq-config-blackwell.cuh"
 
@@ -248,10 +247,7 @@ static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type ty
     if (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) {
         return ggml_cuda_mmq_get_config_ampere(type, J, fallback);
     }
-    if (ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_DP4A) {
-        return ggml_cuda_mmq_get_config_pascal_dp4a(type, J, fallback);
-    }
-    return ggml_cuda_mmq_get_config_pascal_older(type, J, fallback);
+    return ggml_cuda_mmq_get_config_pascal(type, J, fallback);
 }
 
 static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_type type, int J, bool fallback) {
@@ -272,10 +268,8 @@ static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_t
     return ggml_cuda_mmq_get_config_blackwell(type, J, fallback);
 #elif __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     return ggml_cuda_mmq_get_config_ampere(type, J, fallback);
-#elif __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
-    return ggml_cuda_mmq_get_config_pascal_dp4a(type, J, fallback);
 #else
-    return ggml_cuda_mmq_get_config_pascal_older(type, J, fallback);
+    return ggml_cuda_mmq_get_config_pascal(type, J, fallback);
 #endif // BLACKWELL_MMA_AVAILABLE
 #endif // GGML_USE_HIP
     GGML_UNUSED_VARS(type, J, fallback);

ggml/src/ggml-cuda/pool1d.cu

@@ -1,85 +0,0 @@
-#include "pool1d.cuh"
-
-static __global__ void pool1d_nchw_kernel(
-        const int iw, const int ow,
-        const int kw, const int sw, const int pw,
-        const int parallel_elements,
-        const float * src, float * dst, const enum ggml_op_pool op) {
-    const int idx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (idx >= parallel_elements) {
-        return;
-    }
-
-    const int nc = idx / ow;
-    const int cur_ow = idx % ow;
-
-    const float * i_ptr = src + nc * iw;
-    float * o_ptr = dst + nc * ow;
-
-    const int start = cur_ow * sw - pw;
-    const int b = max(0, start);
-    const int e = min(iw, start + kw);
-
-    float res;
-    switch (op) {
-        case GGML_OP_POOL_AVG: res = 0.0f;     break;
-        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
-        default: return;
-    }
-
-    int count = 0;
-    for (int i = b; i < e; i++) {
-#if __CUDA_ARCH__ >= 350
-        float cur = __ldg(i_ptr + i);
-#else
-        float cur = i_ptr[i];
-#endif
-        switch (op) {
-            case GGML_OP_POOL_AVG: res += cur;                break;
-            case GGML_OP_POOL_MAX: res = max(res, cur);       break;
-            default: break;
-        }
-        count++;
-    }
-
-    if (op == GGML_OP_POOL_AVG) {
-        res = (count > 0) ? (res / count) : 0.0f;
-    }
-
-    o_ptr[cur_ow] = res;
-}
-
-static void pool1d_nchw_kernel_f32_f32_cuda(
-        const int iw, const int ow,
-        const int kw, const int sw, const int pw,
-        const int parallel_elements,
-        const float * src, float * dst, const enum ggml_op_pool op,
-        cudaStream_t stream) {
-    const int num_blocks = (parallel_elements + CUDA_POOL1D_BLOCK_SIZE - 1) / CUDA_POOL1D_BLOCK_SIZE;
-    dim3 block_nums(num_blocks);
-    pool1d_nchw_kernel<<<block_nums, CUDA_POOL1D_BLOCK_SIZE, 0, stream>>>(iw, ow, kw, sw, pw, parallel_elements, src, dst, op);
-}
-
-void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int32_t * opts = (const int32_t *)dst->op_params;
-    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
-    const int k0 = opts[1];
-    const int s0 = opts[2];
-    const int p0 = opts[3];
-
-    const int64_t IW = src0->ne[0];
-    const int64_t OW = dst->ne[0];
-    const int64_t nr = ggml_nrows(src0);
-
-    const int parallel_elements = (int)(nr * OW);
-
-    pool1d_nchw_kernel_f32_f32_cuda(IW, OW, k0, s0, p0, parallel_elements, src0_d, dst_d, op, stream);
-}

ggml/src/ggml-cuda/pool1d.cuh

@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_POOL1D_BLOCK_SIZE 256
-
-void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -8,107 +8,60 @@
 #include <algorithm>
 #include <string>
 #include <vector>
-#include <memory>
 #include <stdio.h>
 #include "htp-ops.h"
 #include "htp/matmul-ops.h"
 #include "htp/flash-attn-ops.h"
 #include "htp/unary-ops.h"
-#include "htp/allreduce-ops.h"
 
 struct htp_opnode {
-    ggml_tensor * node   { nullptr };
-    htp_op_code   opcode { HTP_OP_INVALID };
-    int32_t       kernel_params[HTP_OP_MAX_KERN_PARAMS] {0};
+    ggml_tensor * node = nullptr;
 
-    std::vector<ggml_tensor *>                fused;
-    std::vector<std::shared_ptr<ggml_tensor>> dummy;
+    std::vector<ggml_tensor *> fused;
 
-    std::vector<const ggml_tensor *> inputs;
-    std::vector<const ggml_tensor *> outputs;
-    std::string                      name;
+    htp_op_code opcode = HTP_OP_INVALID;
 
-    int n_active_src(const ggml_tensor * t) const {
-        if (!t) return 0;
-        for (int i = GGML_MAX_SRC - 1; i >= 0; i--) {
-            if (t->src[i]) {
-                return i + 1;
-            }
-        }
-        return 0;
+    std::vector<ggml_tensor *> extra_dsts;
+
+    int32_t kernel_params[HTP_OP_MAX_KERN_PARAMS] = {0};
+
+    htp_opnode(ggml_tensor * node = nullptr, std::vector<ggml_tensor *> fused = {}, htp_op_code opcode = HTP_OP_INVALID, std::vector<ggml_tensor *> extra_dsts = {})
+        : node(node), fused(std::move(fused)), opcode(opcode), extra_dsts(std::move(extra_dsts)) {}
+
+    ggml_op op() const {
+        return node->op;
     }
 
-    void init(ggml_tensor * node) {
-        this->node = node;
-        if (this->node) {
-            this->name = ggml_op_desc(this->node);
-
-            // Build inputs (preserving optional nullptrs)
-            int n_inputs = n_active_src(this->node);
-            this->inputs.resize(n_inputs, nullptr);
-            for (int i = 0; i < n_inputs; i++) {
-                this->inputs[i] = this->node->src[i];
-            }
-
-            // Build outputs
-            this->outputs.push_back(this->dst());
-        }
-    }
-
-    htp_opnode(htp_op_code opcode = HTP_OP_INVALID, ggml_tensor * node = nullptr) : opcode(opcode) {
-        init(node);
-    }
-
-    ggml_op             op()   const { return node->op; }
-    const ggml_tensor * src0() const { return node->src[0]; }
-    const ggml_tensor * src1() const { return node->src[1]; }
-    const ggml_tensor * dst()  const { return outputs.empty() ? node : outputs.back(); }
-
-    ggml_tensor * add_dummy(const ggml_tensor & t) {
-        dummy.push_back(std::make_shared<ggml_tensor>(t));
-        return dummy.back().get();
+    const ggml_tensor * dst() const {
+        return fused.empty() ? node : fused.back();
     }
 
     void add_fused(ggml_tensor * t, bool extra_dst = false) {
         fused.push_back(t);
-
-        name += "+";
-        name += ggml_op_desc(t);
-
         if (extra_dst) {
-            outputs.push_back(t);
-        } else {
-            outputs.clear();
-            outputs.push_back(t);
+            extra_dsts.push_back(t);
         }
+    }
 
-        // Remove the newly fused intermediate output tensor t from inputs (if it was there)
-        inputs.erase(std::remove(inputs.begin(), inputs.end(), t), inputs.end());
-
-        // Append new inputs from t, preserving middle nullptrs
-        int n_inputs = n_active_src(t);
-        for (int i = 0; i < n_inputs; i++) {
-            const auto * src = t->src[i];
-            if (!src) {
-                inputs.push_back(nullptr);
-            } else if (src != node &&
-                       std::find(fused.begin(), fused.end(), src) == fused.end() &&
-                       std::find(inputs.begin(), inputs.end(), src) == inputs.end()) {
-                inputs.push_back(src);
+    std::vector<const ggml_tensor *> get_outputs() const {
+        std::vector<const ggml_tensor *> res;
+        if (extra_dsts.empty()) {
+            res.push_back(dst());
+        } else {
+            res.push_back(node);
+            for (const auto * x : extra_dsts) {
+                res.push_back(x);
             }
         }
+        return res;
     }
 
-    const std::vector<const ggml_tensor *> & get_inputs() const {
-        return inputs;
+    const ggml_tensor * src0() const {
+        return node->src[0];
     }
 
-    const std::vector<const ggml_tensor *> & get_outputs() const {
-        return outputs;
-    }
-
-    std::string op_name() const {
-        return name;
+    const ggml_tensor * src1() const {
+        return node->src[1];
     }
 
     bool is_empty() const {
@@ -128,6 +81,75 @@ struct htp_opnode {
     bool same_input(const htp_opnode& n) const {
         return n.src1() == this->src1();
     }
+
+    std::vector<const ggml_tensor *> get_inputs() const {
+        if (fused.empty()) {
+            int last_non_null = -1;
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (node->src[i]) {
+                    last_non_null = i;
+                }
+            }
+            std::vector<const ggml_tensor *> inputs(last_non_null + 1, nullptr);
+            for (int i = 0; i <= last_non_null; i++) {
+                inputs[i] = node->src[i];
+            }
+            return inputs;
+        }
+
+        std::vector<const ggml_tensor *> inputs(GGML_MAX_SRC, nullptr);
+        std::vector<const ggml_tensor *> outputs;
+        outputs.push_back(node);
+        for (const auto * f : fused) {
+            outputs.push_back(f);
+        }
+
+        auto contains = [&](const std::vector<const ggml_tensor *> & vec, const ggml_tensor * t) {
+            for (const auto * x : vec) {
+                if (x == t) return true;
+            }
+            return false;
+        };
+
+        int count = 0;
+        auto add_input = [&](const ggml_tensor * t) {
+            if (t && !contains(outputs, t) && !contains(inputs, t)) {
+                if (count < (int)inputs.size()) {
+                    inputs[count++] = t;
+                } else {
+                    inputs.push_back(t);
+                }
+            }
+        };
+
+        for (int i = 0; i < GGML_MAX_SRC; i++) {
+            if (node->src[i]) {
+                add_input(node->src[i]);
+            }
+        }
+        for (const auto * f : fused) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                if (f->src[i]) {
+                    add_input(f->src[i]);
+                }
+            }
+        }
+
+        inputs.resize(count);
+        return inputs;
+    }
+
+    std::string op_name() const {
+        if (fused.empty()) {
+            return ggml_op_desc(node);
+        }
+        std::string name = ggml_op_desc(node);
+        for (const auto * f : fused) {
+            name += "+";
+            name += ggml_op_desc(f);
+        }
+        return name;
+    }
 };
 
 struct htp_opformat {
@@ -315,7 +337,8 @@ struct htp_opformat {
     }
     void format_kernel_params(char * str, size_t max_size, const htp_opnode & node) {
         if (node.opcode == HTP_OP_MUL_MAT || node.opcode == HTP_OP_MUL_MAT_ID ||
-            node.opcode == HTP_OP_MUL_MAT_NX || node.opcode == HTP_OP_MUL_MAT_ADD) {
+            node.opcode == HTP_OP_MUL_MAT_QKV || node.opcode == HTP_OP_MUL_MAT_FFN ||
+            node.opcode == HTP_OP_MUL_MAT_ADD) {
             const auto * kparams = (const struct htp_mm_kernel_params *) node.kernel_params;
             const char * path = "unknown";
             int32_t type = kparams->kernel_type;

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

@@ -43,7 +43,6 @@ add_library(${HTP_LIB} SHARED
     pad-ops.c
     argsort-ops.c
     im2col-ops.c
-    allreduce-ops.c
 )
 
 target_compile_definitions(${HTP_LIB} PRIVATE

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

@@ -183,53 +183,6 @@ static void swiglu_oai_f32(const float * restrict src0,
 static const float GELU_COEF_A     = 0.044715f;
 static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
 
-static inline HVX_Vector hvx_vec_fast_sigmoid_f32_2it(HVX_Vector v) {
-    v = Q6_Vqf32_vmpy_VsfVsf(v, Q6_V_vsplat_R(FAST_SIGMOID_LOG2F));
-    v = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v), Q6_V_vsplat_R(FAST_SIGMOID_C3));
-
-    HVX_Vector in_int = hvx_vec_truncate_f32(Q6_Vsf_equals_Vqf32(v));
-    HVX_Vector x      = Q6_Vqf32_vsub_Vqf32Vsf(v, Q6_Vsf_equals_Vw(in_int));
-    HVX_Vector xx     = Q6_Vqf32_vmpy_Vqf32Vqf32(x, x);
-
-    HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(xx), Q6_V_vsplat_R(FAST_SIGMOID_C2));
-    v1            = Q6_Vqf32_vadd_Vqf32Vsf(v1, Q6_V_vsplat_R(FAST_SIGMOID_LOG2F));
-
-    HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(x), Q6_V_vsplat_R(FAST_SIGMOID_C1));
-    v2            = Q6_Vqf32_vmpy_Vqf32Vqf32(v2, xx);
-    v2            = Q6_Vqf32_vadd_Vqf32Vqf32(v2, x);
-
-    HVX_Vector v3 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vqf32(v2, v1));
-    v3            = Q6_Vw_vaslacc_VwVwR(v3, in_int, 24);
-
-    HVX_Vector v4 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_Vqf32Vqf32(v2, v1));
-    HVX_Vector v5 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v3, v4));
-
-    // Newton-Raphson with 2 iterations
-    HVX_Vector two_sf = hvx_vec_splat_f32(2.0f);
-    HVX_Vector i_sf   = Q6_Vw_vsub_VwVw(Q6_V_vsplat_R(0x7EEEEBB3), v5);
-    HVX_Vector r_qf   = Q6_Vqf32_vmpy_VsfVsf(
-        i_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(two_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(i_sf, v5)))));
-    r_qf = Q6_Vqf32_vmpy_Vqf32Vqf32(
-        r_qf, Q6_Vqf32_vsub_VsfVsf(two_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(r_qf), v5))));
-    HVX_Vector res = Q6_Vsf_equals_Vqf32(r_qf);
-
-    res = Q6_Vqf32_vmpy_VsfVsf(v3, res);
-
-    return Q6_Vsf_equals_Vqf32(res);
-}
-
-static inline HVX_Vector hvx_vec_fast_sigmoid_f32_guard_2it(HVX_Vector v,
-                                                            HVX_Vector one,
-                                                            HVX_Vector max_exp,
-                                                            HVX_Vector min_exp) {
-    const HVX_VectorPred pred_max = Q6_Q_vcmp_gt_VsfVsf(max_exp, v);
-    const HVX_VectorPred pred_min = Q6_Q_vcmp_gt_VsfVsf(v, min_exp);
-
-    HVX_Vector out = hvx_vec_fast_sigmoid_f32_2it(v);
-    out            = Q6_V_vmux_QVV(pred_max, out, one);
-    return Q6_V_vmux_QVV(pred_min, out, Q6_V_vzero());
-}
-
 static inline void hvx_geglu_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) {
     assert((unsigned long) dst  % 128 == 0);
     assert((unsigned long) src0 % 128 == 0);
@@ -247,13 +200,20 @@ static inline void hvx_geglu_f32_aa(uint8_t * restrict dst, const uint8_t * rest
 
     const HVX_Vector v_coef_a_times_sqrt = hvx_vec_splat_f32(GELU_COEF_A_TIMES_SQRT);
     const HVX_Vector v_sqrt_2_pi         = hvx_vec_splat_f32(SQRT_2_OVER_PI);
+    const HVX_Vector v_half              = hvx_vec_splat_f32(0.5f);
     const HVX_Vector v_one               = hvx_vec_splat_f32(1.0f);
+    const HVX_Vector v_two               = hvx_vec_splat_f32(2.0f);
+
+    // Hoisted fast sigmoid / inverse constants to avoid loop-internal overhead
+    const HVX_Vector v_log2f             = Q6_V_vsplat_R(FAST_SIGMOID_LOG2F);
+    const HVX_Vector v_c1                = Q6_V_vsplat_R(FAST_SIGMOID_C1);
+    const HVX_Vector v_c2                = Q6_V_vsplat_R(FAST_SIGMOID_C2);
+    const HVX_Vector v_inv_aprox         = Q6_V_vsplat_R(0x7EEEEBB3);
     const HVX_Vector v_max_exp           = hvx_vec_splat_f32(87.0f);
     const HVX_Vector v_min_exp           = hvx_vec_splat_f32(-87.0f);
 
     uint32_t i = 0;
 
-    _Pragma("unroll(4)")
     for (; i < nvec; i++) {
         HVX_Vector x = vsrc0[i];
         HVX_Vector g = vsrc1[i];
@@ -263,13 +223,56 @@ static inline void hvx_geglu_f32_aa(uint8_t * restrict dst, const uint8_t * rest
         coef = hvx_vec_add_f32_f32(coef, v_sqrt_2_pi);
         HVX_Vector inner = hvx_vec_mul_f32_f32(x, coef);
 
-        // y2 = 2 * inner = inner + inner
-        HVX_Vector y2 = hvx_vec_add_f32_f32(inner, inner);
+        // y2 = 2 * inner
+        HVX_Vector y2 = hvx_vec_mul_f32_f32(inner, v_two);
 
-        // Fast sigmoid approximation (2 iterations)
-        HVX_Vector sig2y = hvx_vec_fast_sigmoid_f32_guard_2it(y2, v_one, v_max_exp, v_min_exp);
+        // Sigmoid guard check predicates
+        HVX_VectorPred pred_max = Q6_Q_vcmp_gt_VsfVsf(v_max_exp, y2);
+        HVX_VectorPred pred_min = Q6_Q_vcmp_gt_VsfVsf(y2, v_min_exp);
+
+        // Fast sigmoid approximation
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(y2, v_log2f);
+        v = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v), v_half);
+
+        HVX_Vector in_int = hvx_vec_truncate_f32(Q6_Vsf_equals_Vqf32(v));
+        HVX_Vector x_sig  = Q6_Vqf32_vsub_Vqf32Vsf(v, Q6_Vsf_equals_Vw(in_int));
+        HVX_Vector xx_sig = Q6_Vqf32_vmpy_Vqf32Vqf32(x_sig, x_sig);
+
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(xx_sig), v_c2);
+        v1 = Q6_Vqf32_vadd_Vqf32Vsf(v1, v_log2f);
+
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(x_sig), v_c1);
+        v2 = Q6_Vqf32_vmpy_Vqf32Vqf32(v2, xx_sig);
+        v2 = Q6_Vqf32_vadd_Vqf32Vqf32(v2, x_sig);
+
+        HVX_Vector v3 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vqf32(v2, v1));
+        v3 = Q6_Vw_vaslacc_VwVwR(v3, in_int, 24);
+
+        HVX_Vector v4 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_Vqf32Vqf32(v2, v1));
+        HVX_Vector v5 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v3, v4));
+
+        // Fast division (Newton-Raphson with 2 iterations)
+        HVX_Vector i_sf = Q6_Vw_vsub_VwVw(v_inv_aprox, v5);
+        HVX_Vector r_qf = Q6_Vqf32_vmpy_VsfVsf(
+            i_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v_two, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(i_sf, v5)))));
+        r_qf = Q6_Vqf32_vmpy_Vqf32Vqf32(
+            r_qf, Q6_Vqf32_vsub_VsfVsf(v_two, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(r_qf), v5))));
+        HVX_Vector res_inv = Q6_Vsf_equals_Vqf32(r_qf);
+
+        HVX_Vector sig2y = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(v3, res_inv));
+
+        // Sigmoid guards
+        sig2y = Q6_V_vmux_QVV(pred_max, sig2y, v_one);
+        sig2y = Q6_V_vmux_QVV(pred_min, sig2y, Q6_V_vzero());
+
+        // tanh(inner) = 2 * sigmoid(2 * inner) - 1
+        HVX_Vector tanh_val = hvx_vec_mul_f32_f32(sig2y, v_two);
+        tanh_val = hvx_vec_sub_f32_f32(tanh_val, v_one);
+
+        HVX_Vector tanh_plus_one = hvx_vec_add_f32_f32(tanh_val, v_one);
+        HVX_Vector half_x = hvx_vec_mul_f32_f32(x, v_half);
+        HVX_Vector gelu_x = hvx_vec_mul_f32_f32(half_x, tanh_plus_one);
 
-        HVX_Vector gelu_x = hvx_vec_mul_f32_f32(x, sig2y);
         vdst[i] = hvx_vec_mul_f32_f32(gelu_x, g);
     }
 
@@ -282,11 +285,50 @@ static inline void hvx_geglu_f32_aa(uint8_t * restrict dst, const uint8_t * rest
         coef = hvx_vec_add_f32_f32(coef, v_sqrt_2_pi);
         HVX_Vector inner = hvx_vec_mul_f32_f32(x, coef);
 
-        HVX_Vector y2 = hvx_vec_add_f32_f32(inner, inner);
+        HVX_Vector y2 = hvx_vec_mul_f32_f32(inner, v_two);
 
-        HVX_Vector sig2y = hvx_vec_fast_sigmoid_f32_guard_2it(y2, v_one, v_max_exp, v_min_exp);
+        HVX_VectorPred pred_max = Q6_Q_vcmp_gt_VsfVsf(v_max_exp, y2);
+        HVX_VectorPred pred_min = Q6_Q_vcmp_gt_VsfVsf(y2, v_min_exp);
+
+        HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(y2, v_log2f);
+        v = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(v), v_half);
+
+        HVX_Vector in_int = hvx_vec_truncate_f32(Q6_Vsf_equals_Vqf32(v));
+        HVX_Vector x_sig  = Q6_Vqf32_vsub_Vqf32Vsf(v, Q6_Vsf_equals_Vw(in_int));
+        HVX_Vector xx_sig = Q6_Vqf32_vmpy_Vqf32Vqf32(x_sig, x_sig);
+
+        HVX_Vector v1 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(xx_sig), v_c2);
+        v1 = Q6_Vqf32_vadd_Vqf32Vsf(v1, v_log2f);
+
+        HVX_Vector v2 = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(x_sig), v_c1);
+        v2 = Q6_Vqf32_vmpy_Vqf32Vqf32(v2, xx_sig);
+        v2 = Q6_Vqf32_vadd_Vqf32Vqf32(v2, x_sig);
+
+        HVX_Vector v3 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vqf32(v2, v1));
+        v3 = Q6_Vw_vaslacc_VwVwR(v3, in_int, 24);
+
+        HVX_Vector v4 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_Vqf32Vqf32(v2, v1));
+        HVX_Vector v5 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v3, v4));
+
+        HVX_Vector i_sf = Q6_Vw_vsub_VwVw(v_inv_aprox, v5);
+        HVX_Vector r_qf = Q6_Vqf32_vmpy_VsfVsf(
+            i_sf, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vsub_VsfVsf(v_two, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(i_sf, v5)))));
+        r_qf = Q6_Vqf32_vmpy_Vqf32Vqf32(
+            r_qf, Q6_Vqf32_vsub_VsfVsf(v_two, Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(r_qf), v5))));
+        HVX_Vector res_inv = Q6_Vsf_equals_Vqf32(r_qf);
+
+        HVX_Vector sig2y = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(v3, res_inv));
+
+        sig2y = Q6_V_vmux_QVV(pred_max, sig2y, v_one);
+        sig2y = Q6_V_vmux_QVV(pred_min, sig2y, Q6_V_vzero());
+
+        HVX_Vector tanh_val = hvx_vec_mul_f32_f32(sig2y, v_two);
+        tanh_val = hvx_vec_sub_f32_f32(tanh_val, v_one);
+
+        HVX_Vector tanh_plus_one = hvx_vec_add_f32_f32(tanh_val, v_one);
+        HVX_Vector half_x = hvx_vec_mul_f32_f32(x, v_half);
+        HVX_Vector gelu_x = hvx_vec_mul_f32_f32(half_x, tanh_plus_one);
 
-        HVX_Vector gelu_x = hvx_vec_mul_f32_f32(x, sig2y);
         HVX_Vector res = hvx_vec_mul_f32_f32(gelu_x, g);
         hvx_vec_store_a((void *) &vdst[i], nloe * sizeof(float), res);
     }

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

@@ -1,398 +0,0 @@
-#pragma clang diagnostic ignored "-Wunused-variable"
-#pragma clang diagnostic ignored "-Wunused-function"
-#pragma clang diagnostic ignored "-Wunused-but-set-variable"
-
-#include <HAP_farf.h>
-#include <HAP_perf.h>
-#include <stdatomic.h>
-#include <math.h>
-#include <string.h>
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-#include "htp-ctx.h"
-#include "htp-ops.h"
-#include "hvx-utils.h"
-#include "htp-tensor.h"
-#include "hex-dma.h"
-#include "hex-profile.h"
-#include "allreduce-ops.h"
-
-struct htp_allreduce_context {
-    struct htp_ops_context * octx;
-    uint32_t n_ranks;
-    uint32_t n_dsts;
-    uint32_t nelem;
-    uint32_t ne0;
-    uint32_t ne1;
-    uint32_t row_size_aligned;
-    uint32_t rank_elem_start;
-    uint32_t rank_nelem;
-    uint32_t elems_per_thread;
-    uint32_t block_elems;
-    uint32_t vtcm_size_per_thread;
-    bool     is_row_bcast;
-    uint8_t * src_spad_base[HTP_ALLREDUCE_MAX_RANKS];
-    uint8_t * dst_spad_base;
-    uint8_t * res_spad_base;
-};
-
-#define DEFINE_ALLREDUCE_THREAD_DMA_1D(SUFFIX, TYPE, HVX_ADD_FN, HAS_ADD)                                             \
-static void allreduce_thread_dma_1d_##SUFFIX(unsigned int nth, unsigned int ith, void * data) {                       \
-    struct htp_allreduce_context * actx = (struct htp_allreduce_context *) data;                                      \
-    struct htp_ops_context * octx = actx->octx;                                                                       \
-                                                                                                                      \
-    const uint32_t n_ranks     = actx->n_ranks;                                                                       \
-    const uint32_t n_dsts      = actx->n_dsts;                                                                        \
-    const uint32_t block_elems = actx->block_elems;                                                                   \
-                                                                                                                      \
-    const uint32_t dr  = actx->elems_per_thread;                                                                      \
-    const uint32_t ir0 = actx->rank_elem_start + dr * ith;                                                            \
-    const uint32_t ir1 = MIN(ir0 + dr, actx->rank_elem_start + actx->rank_nelem);                                     \
-    if (ir0 >= ir1) return;                                                                                           \
-                                                                                                                      \
-    struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                            \
-    dma_queue * q = octx->ctx->dma[ith];                                                                              \
-                                                                                                                      \
-    uint8_t * src_spad_base[HTP_ALLREDUCE_MAX_RANKS];                                                                 \
-    for (uint32_t s = 0; s < n_ranks; s++) {                                                                          \
-        src_spad_base[s] = actx->src_spad_base[s] + (ith * actx->vtcm_size_per_thread);                               \
-    }                                                                                                                 \
-    uint8_t * dst_spad_base = actx->dst_spad_base + (ith * actx->vtcm_size_per_thread);                               \
-    uint8_t * res_spad_base = HAS_ADD ? (actx->res_spad_base + (ith * actx->vtcm_size_per_thread)) : NULL;            \
-                                                                                                                      \
-    const size_t spad_half = actx->vtcm_size_per_thread / 2;                                                          \
-    uint32_t ir_prefetch = ir0;                                                                                       \
-    int spad_idx = 0;                                                                                                 \
-                                                                                                                      \
-    for (int k = 0; k < 2 && ir_prefetch < ir1; k++) {                                                                \
-        uint32_t cur_elems = MIN(block_elems, ir1 - ir_prefetch);                                                     \
-        size_t   cur_bytes = cur_elems * sizeof(TYPE);                                                                \
-        uint8_t * d_spad = dst_spad_base + spad_idx * spad_half;                                                      \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                       \
-            uint8_t * d_ddr = (uint8_t *) octx->dsts[d]->data + ir_prefetch * sizeof(TYPE);                           \
-            dma_queue_push(q, dma_make_ptr(d_ddr, d_spad), cur_bytes, cur_bytes, cur_bytes, 0);                       \
-        }                                                                                                             \
-        for (uint32_t s = 0; s < n_ranks; s++) {                                                                      \
-            uint8_t * s_spad = src_spad_base[s] + spad_idx * spad_half;                                               \
-            const uint8_t * s_ddr = (const uint8_t *) octx->src[s]->data + ir_prefetch * sizeof(TYPE);                \
-            dma_queue_push(q, dma_make_ptr(s_spad, s_ddr), cur_bytes, cur_bytes, cur_bytes, 1);                       \
-        }                                                                                                             \
-        if (HAS_ADD) {                                                                                                \
-            uint8_t * r_spad = res_spad_base + spad_idx * spad_half;                                                  \
-            const uint8_t * r_ddr = (const uint8_t *) octx->src[2 * n_ranks]->data + ir_prefetch * sizeof(TYPE);      \
-            dma_queue_push(q, dma_make_ptr(r_spad, r_ddr), cur_bytes, cur_bytes, cur_bytes, 1);                       \
-        }                                                                                                             \
-        ir_prefetch += cur_elems;                                                                                     \
-        spad_idx ^= 1;                                                                                                \
-    }                                                                                                                 \
-                                                                                                                      \
-    for (uint32_t ir = ir0; ir < ir1; ) {                                                                             \
-        uint32_t cur_elems = MIN(block_elems, ir1 - ir);                                                              \
-        size_t   cur_bytes = cur_elems * sizeof(TYPE);                                                                \
-        uint8_t * d_spad = NULL;                                                                                      \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                       \
-            d_spad = (uint8_t *) dma_queue_pop(q).src;                                                                \
-        }                                                                                                             \
-        uint8_t * s_spad[HTP_ALLREDUCE_MAX_RANKS];                                                                    \
-        for (uint32_t s = 0; s < n_ranks; s++) {                                                                      \
-            s_spad[s] = (uint8_t *) dma_queue_pop(q).dst;                                                             \
-        }                                                                                                             \
-        uint8_t * r_spad = HAS_ADD ? (uint8_t *) dma_queue_pop(q).dst : NULL;                                         \
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);                                             \
-        HVX_ADD_FN(d_spad, s_spad[0], s_spad[1], cur_elems);                                                          \
-        for (uint32_t s = 2; s < n_ranks; s++) {                                                                      \
-            HVX_ADD_FN(d_spad, d_spad, s_spad[s], cur_elems);                                                         \
-        }                                                                                                             \
-        if (HAS_ADD) {                                                                                                \
-            HVX_ADD_FN(d_spad, d_spad, r_spad, cur_elems);                                                            \
-        }                                                                                                             \
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) ir);                                              \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                       \
-            uint8_t * d_ddr = (uint8_t *) octx->dsts[d]->data + ir * sizeof(TYPE);                                    \
-            dma_queue_push(q, dma_make_ptr(d_ddr, d_spad), cur_bytes, cur_bytes, cur_bytes, 1);                       \
-        }                                                                                                             \
-        if (ir_prefetch < ir1) {                                                                                      \
-            uint32_t next_elems = MIN(block_elems, ir1 - ir_prefetch);                                                \
-            size_t   next_bytes = next_elems * sizeof(TYPE);                                                          \
-            for (uint32_t s = 0; s < n_ranks; s++) {                                                                  \
-                const uint8_t * s_next = (const uint8_t *) octx->src[s]->data + ir_prefetch * sizeof(TYPE);           \
-                dma_queue_push(q, dma_make_ptr(s_spad[s], s_next), next_bytes, next_bytes, next_bytes, 1);            \
-            }                                                                                                         \
-            if (HAS_ADD) {                                                                                            \
-                const uint8_t * r_next = (const uint8_t *) octx->src[2 * n_ranks]->data + ir_prefetch * sizeof(TYPE); \
-                dma_queue_push(q, dma_make_ptr(r_spad, r_next), next_bytes, next_bytes, next_bytes, 1);               \
-            }                                                                                                         \
-            ir_prefetch += next_elems;                                                                                \
-        }                                                                                                             \
-        ir += cur_elems;                                                                                              \
-    }                                                                                                                 \
-    dma_queue_flush(q);                                                                                               \
-}
-
-DEFINE_ALLREDUCE_THREAD_DMA_1D(f16,     __fp16, hvx_add_f16_aaa, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_1D(f32,     float,  hvx_add_f32_aaa, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_1D(add_f16, __fp16, hvx_add_f16_aaa, 1)
-DEFINE_ALLREDUCE_THREAD_DMA_1D(add_f32, float,  hvx_add_f32_aaa, 1)
-
-#define DEFINE_ALLREDUCE_THREAD_DMA_2D(SUFFIX, TYPE, HVX_ADD_FN, HAS_ADD, IS_ROW_BCAST)                                                           \
-static void allreduce_thread_dma_2d_##SUFFIX(unsigned int nth, unsigned int ith, void * data) {                                                   \
-    struct htp_allreduce_context * actx = (struct htp_allreduce_context *) data;                                                                  \
-    struct htp_ops_context * octx = actx->octx;                                                                                                   \
-                                                                                                                                                  \
-    const uint32_t n_ranks          = actx->n_ranks;                                                                                              \
-    const uint32_t n_dsts           = actx->n_dsts;                                                                                               \
-    const uint32_t ne0              = actx->ne0;                                                                                                  \
-    const uint32_t block_rows       = actx->block_elems;                                                                                          \
-    const uint32_t row_size_aligned = actx->row_size_aligned;                                                                                     \
-    const uint32_t row_bytes        = ne0 * sizeof(TYPE);                                                                                         \
-                                                                                                                                                  \
-    const uint32_t dr  = actx->elems_per_thread;                                                                                                  \
-    const uint32_t r0  = actx->rank_elem_start + dr * ith;                                                                                        \
-    const uint32_t r1  = MIN(r0 + dr, actx->rank_elem_start + actx->rank_nelem);                                                                  \
-    if (r0 >= r1) return;                                                                                                                         \
-                                                                                                                                                  \
-    struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                                                        \
-    dma_queue * q = octx->ctx->dma[ith];                                                                                                          \
-                                                                                                                                                  \
-    uint8_t * src_spad_base[HTP_ALLREDUCE_MAX_RANKS];                                                                                             \
-    for (uint32_t s = 0; s < n_ranks; s++) {                                                                                                      \
-        src_spad_base[s] = actx->src_spad_base[s] + (ith * actx->vtcm_size_per_thread);                                                           \
-    }                                                                                                                                             \
-    uint8_t * dst_spad_base = actx->dst_spad_base + (ith * actx->vtcm_size_per_thread);                                                           \
-    uint8_t * res_spad_base = HAS_ADD ? (IS_ROW_BCAST ? actx->res_spad_base : (actx->res_spad_base + (ith * actx->vtcm_size_per_thread))) : NULL; \
-                                                                                                                                                  \
-    const size_t spad_half = actx->vtcm_size_per_thread / 2;                                                                                      \
-    uint32_t r_prefetch = r0;                                                                                                                     \
-    int spad_idx = 0;                                                                                                                             \
-                                                                                                                                                  \
-    for (int k = 0; k < 2 && r_prefetch < r1; k++) {                                                                                              \
-        uint32_t cur_rows = MIN(block_rows, r1 - r_prefetch);                                                                                     \
-        uint8_t * d_spad = dst_spad_base + spad_idx * spad_half;                                                                                  \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                                                   \
-            uint8_t * d_ddr  = (uint8_t *) octx->dsts[d]->data + r_prefetch * octx->dsts[d]->nb[1];                                               \
-            dma_queue_push(q, dma_make_ptr(d_ddr, d_spad), octx->dsts[d]->nb[1], row_size_aligned, row_bytes, 0);                                 \
-        }                                                                                                                                         \
-        for (uint32_t s = 0; s < n_ranks; s++) {                                                                                                  \
-            uint8_t * s_spad = src_spad_base[s] + spad_idx * spad_half;                                                                           \
-            const uint8_t * s_ddr = (const uint8_t *) octx->src[s]->data + r_prefetch * octx->src[s]->nb[1];                                      \
-            dma_queue_push(q, dma_make_ptr(s_spad, s_ddr), row_size_aligned, octx->src[s]->nb[1], row_bytes, cur_rows);                           \
-        }                                                                                                                                         \
-        if (HAS_ADD && !IS_ROW_BCAST) {                                                                                                           \
-            uint8_t * r_spad = res_spad_base + spad_idx * spad_half;                                                                              \
-            const uint8_t * r_ddr = (const uint8_t *) octx->src[2 * n_ranks]->data + r_prefetch * octx->src[2 * n_ranks]->nb[1];                  \
-            dma_queue_push(q, dma_make_ptr(r_spad, r_ddr), row_size_aligned, octx->src[2 * n_ranks]->nb[1], row_bytes, cur_rows);                 \
-        }                                                                                                                                         \
-        r_prefetch += cur_rows;                                                                                                                   \
-        spad_idx ^= 1;                                                                                                                            \
-    }                                                                                                                                             \
-                                                                                                                                                  \
-    for (uint32_t r = r0; r < r1; ) {                                                                                                             \
-        uint32_t cur_rows = MIN(block_rows, r1 - r);                                                                                              \
-        uint8_t * d_spad = NULL;                                                                                                                  \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                                                   \
-            d_spad = (uint8_t *) dma_queue_pop(q).src;                                                                                            \
-        }                                                                                                                                         \
-        uint8_t * s_spad[HTP_ALLREDUCE_MAX_RANKS];                                                                                                \
-        for (uint32_t s = 0; s < n_ranks; s++) {                                                                                                  \
-            s_spad[s] = (uint8_t *) dma_queue_pop(q).dst;                                                                                         \
-        }                                                                                                                                         \
-        uint8_t * r_spad = (HAS_ADD && !IS_ROW_BCAST) ? (uint8_t *) dma_queue_pop(q).dst : NULL;                                                  \
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) r);                                                                          \
-        for (uint32_t row = 0; row < cur_rows; row++) {                                                                                           \
-            uint8_t * d_row = d_spad + row * row_size_aligned;                                                                                    \
-            const uint8_t * s0_row = s_spad[0] + row * row_size_aligned;                                                                          \
-            const uint8_t * s1_row = s_spad[1] + row * row_size_aligned;                                                                          \
-            HVX_ADD_FN(d_row, s0_row, s1_row, ne0);                                                                                               \
-            for (uint32_t s = 2; s < n_ranks; s++) {                                                                                              \
-                const uint8_t * ss_row = s_spad[s] + row * row_size_aligned;                                                                      \
-                HVX_ADD_FN(d_row, d_row, ss_row, ne0);                                                                                            \
-            }                                                                                                                                     \
-            if (HAS_ADD) {                                                                                                                        \
-                const uint8_t * res_row = IS_ROW_BCAST ? res_spad_base : (r_spad + row * row_size_aligned);                                       \
-                HVX_ADD_FN(d_row, d_row, res_row, ne0);                                                                                           \
-            }                                                                                                                                     \
-        }                                                                                                                                         \
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) r);                                                                           \
-        for (uint32_t d = 0; d < n_dsts; d++) {                                                                                                   \
-            uint8_t * d_ddr = (uint8_t *) octx->dsts[d]->data + r * octx->dsts[d]->nb[1];                                                         \
-            dma_queue_push(q, dma_make_ptr(d_ddr, d_spad), octx->dsts[d]->nb[1], row_size_aligned, row_bytes, cur_rows);                          \
-        }                                                                                                                                         \
-        if (r_prefetch < r1) {                                                                                                                    \
-            uint32_t next_rows = MIN(block_rows, r1 - r_prefetch);                                                                                \
-            for (uint32_t s = 0; s < n_ranks; s++) {                                                                                              \
-                const uint8_t * s_next = (const uint8_t *) octx->src[s]->data + r_prefetch * octx->src[s]->nb[1];                                 \
-                dma_queue_push(q, dma_make_ptr(s_spad[s], s_next), row_size_aligned, octx->src[s]->nb[1], row_bytes, next_rows);                  \
-            }                                                                                                                                     \
-            if (HAS_ADD && !IS_ROW_BCAST) {                                                                                                       \
-                const uint8_t * r_next = (const uint8_t *) octx->src[2 * n_ranks]->data + r_prefetch * octx->src[2 * n_ranks]->nb[1];             \
-                dma_queue_push(q, dma_make_ptr(r_spad, r_next), row_size_aligned, octx->src[2 * n_ranks]->nb[1], row_bytes, next_rows);           \
-            }                                                                                                                                     \
-            r_prefetch += next_rows;                                                                                                              \
-        }                                                                                                                                         \
-        r += cur_rows;                                                                                                                            \
-    }                                                                                                                                             \
-    dma_queue_flush(q);                                                                                                                           \
-}
-
-DEFINE_ALLREDUCE_THREAD_DMA_2D(f16,           __fp16, hvx_add_f16_aaa, 0, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_2D(f32,           float,  hvx_add_f32_aaa, 0, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_2D(add_f16,       __fp16, hvx_add_f16_aaa, 1, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_2D(add_f32,       float,  hvx_add_f32_aaa, 1, 0)
-DEFINE_ALLREDUCE_THREAD_DMA_2D(add_bcast_f16, __fp16, hvx_add_f16_aaa, 1, 1)
-DEFINE_ALLREDUCE_THREAD_DMA_2D(add_bcast_f32, float,  hvx_add_f32_aaa, 1, 1)
-
-int op_allreduce(struct htp_ops_context * octx) {
-    const struct htp_allreduce_kernel_params * kparams = (const struct htp_allreduce_kernel_params *) octx->kernel_params;
-    const struct htp_tensor * dst = octx->dst;
-
-    const uint32_t rank    = (uint32_t) kparams->rank;
-    const uint32_t n_ranks = (uint32_t) kparams->n_ranks;
-
-    if (n_ranks < 2 || n_ranks > HTP_ALLREDUCE_MAX_RANKS || rank >= n_ranks) {
-        return HTP_STATUS_INVAL_PARAMS;
-    }
-
-    if (dst->type != HTP_TYPE_F16 && dst->type != HTP_TYPE_F32) {
-        return HTP_STATUS_NO_SUPPORT;
-    }
-
-    const uint32_t nelem = dst->ne[0] * dst->ne[1] * dst->ne[2] * dst->ne[3];
-    const uint32_t fence_seq_entry = (uint32_t) octx->op_params[0];
-    const uint32_t fence_seq_exit  = (uint32_t) octx->op_params[1];
-
-    // 1. Entry Barrier: Synchronize all ranks before reading
-    struct htp_thread_trace * tr0 = &octx->ctx->trace[0];
-    htp_trace_event_start(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_entry);
-
-    const struct htp_tensor * my_sync = octx->src[n_ranks + rank];
-    atomic_uint * my_fence = (atomic_uint *) my_sync->data;
-
-    atomic_store(&my_fence[0], fence_seq_entry);
-    asm volatile ("syncht" : : : "memory");
-    Q6_dccleaninva_A((void *) my_fence);
-
-    for (uint32_t j = 0; j < n_ranks; j++) {
-        if (j == rank) continue;
-        const struct htp_tensor * peer_sync = octx->src[n_ranks + j];
-        atomic_uint * peer_fence = (atomic_uint *) peer_sync->data;
-        uint64_t spins = 0;
-        while (1) {
-            Q6_dccleaninva_A((void *) peer_fence);
-            uint32_t val = atomic_load(&peer_fence[0]);
-            if (val == fence_seq_entry || val == fence_seq_exit) {
-                break;
-            }
-            if (++spins > HTP_FENCE_TIMEOUT) {
-                FARF(ERROR, "ggml-hex: allreduce entry fence-wait TIMEOUT: rank %u waiting on %u (fence %p seq %u)\n", rank, j, peer_fence, fence_seq_entry);
-                return HTP_STATUS_INTERNAL_ERR;
-            }
-            hex_pause();
-        }
-    }
-    asm volatile ("syncht" : : : "memory");
-
-    htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_entry);
-
-    // 2. Multi-threaded Reduction across assigned rank chunk
-    if (nelem > 0) {
-        const uint32_t n_threads            = (uint32_t) kparams->n_threads;
-        const uint32_t block_elems          = (uint32_t) kparams->block_elems;
-        const uint32_t elems_per_thread     = (uint32_t) kparams->elems_per_thread;
-        const uint32_t vtcm_size_per_thread = (uint32_t) kparams->vtcm_size_per_thread;
-
-        const bool has_add = (octx->op == HTP_OP_ALLREDUCE_ADD);
-
-        struct htp_allreduce_context actx;
-        actx.octx                 = octx;
-        actx.n_ranks              = n_ranks;
-        actx.n_dsts               = (uint32_t) kparams->n_dsts ? (uint32_t) kparams->n_dsts : n_ranks;
-        actx.nelem                = nelem;
-        actx.ne0                  = (uint32_t) kparams->ne0;
-        actx.ne1                  = (uint32_t) kparams->ne1;
-        actx.row_size_aligned     = (uint32_t) kparams->row_size_aligned;
-        actx.rank_elem_start      = (uint32_t) kparams->rank_elem_start;
-        actx.rank_nelem           = (uint32_t) kparams->rank_nelem;
-        actx.elems_per_thread     = elems_per_thread;
-        actx.block_elems          = block_elems;
-        actx.vtcm_size_per_thread = vtcm_size_per_thread;
-        actx.is_row_bcast         = (kparams->is_row_bcast != 0);
-
-        work_queue_func_t reduce_fun = NULL;
-        switch (kparams->kernel_type) {
-            case HTP_ALLREDUCE_KERNEL_DMA_1D:
-                if (has_add) {
-                    reduce_fun = (dst->type == HTP_TYPE_F16) ? allreduce_thread_dma_1d_add_f16 : allreduce_thread_dma_1d_add_f32;
-                } else {
-                    reduce_fun = (dst->type == HTP_TYPE_F16) ? allreduce_thread_dma_1d_f16 : allreduce_thread_dma_1d_f32;
-                }
-                break;
-            case HTP_ALLREDUCE_KERNEL_DMA_2D:
-                if (has_add) {
-                    if (kparams->is_row_bcast) {
-                        reduce_fun = (dst->type == HTP_TYPE_F16) ? allreduce_thread_dma_2d_add_bcast_f16 : allreduce_thread_dma_2d_add_bcast_f32;
-                    } else {
-                        reduce_fun = (dst->type == HTP_TYPE_F16) ? allreduce_thread_dma_2d_add_f16 : allreduce_thread_dma_2d_add_f32;
-                    }
-                } else {
-                    reduce_fun = (dst->type == HTP_TYPE_F16) ? allreduce_thread_dma_2d_f16 : allreduce_thread_dma_2d_f32;
-                }
-                break;
-            default:
-                return HTP_STATUS_NO_SUPPORT;
-        }
-
-        uint8_t * vtcm_ptr = (uint8_t *) octx->ctx->vtcm_base;
-        for (uint32_t s = 0; s < n_ranks; s++) {
-            actx.src_spad_base[s] = vtcm_ptr;
-            vtcm_ptr += n_threads * vtcm_size_per_thread;
-        }
-        actx.dst_spad_base = vtcm_ptr;
-        vtcm_ptr += n_threads * vtcm_size_per_thread;
-        if (has_add) {
-            actx.res_spad_base = vtcm_ptr;
-            vtcm_ptr += (actx.is_row_bcast ? 1 : n_threads) * vtcm_size_per_thread;
-        }
-
-        if (has_add && actx.is_row_bcast) {
-            const uint8_t * r_ddr = (const uint8_t *) octx->src[2 * n_ranks]->data;
-            const uint32_t row_bytes = actx.ne0 * (dst->type == HTP_TYPE_F16 ? sizeof(__fp16) : sizeof(float));
-            dma_queue * q = octx->ctx->dma[0];
-            dma_queue_push(q, dma_make_ptr(actx.res_spad_base, r_ddr), actx.row_size_aligned, 0, row_bytes, 1);
-            dma_queue_pop(q);
-        }
-
-        work_queue_run(octx->ctx->work_queue, reduce_fun, &actx, n_threads);
-    }
-
-    // 4. Exit Barrier: Synchronize all ranks after writing
-    htp_trace_event_start(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
-
-    atomic_store(&my_fence[0], fence_seq_exit);
-    asm volatile ("syncht" : : : "memory");
-    Q6_dccleaninva_A((void *) my_fence);
-
-    for (uint32_t j = 0; j < n_ranks; j++) {
-        if (j == rank) continue;
-        const struct htp_tensor * peer_sync = octx->src[n_ranks + j];
-        atomic_uint * peer_fence = (atomic_uint *) peer_sync->data;
-        uint64_t spins = 0;
-        while (1) {
-            Q6_dccleaninva_A((void *) peer_fence);
-            uint32_t val = atomic_load(&peer_fence[0]);
-            if (val == fence_seq_exit) {
-                break;
-            }
-            if (++spins > HTP_FENCE_TIMEOUT) {
-                FARF(ERROR, "ggml-hex: allreduce exit fence-wait TIMEOUT: rank %u waiting on %u (fence %p seq %u)\n", rank, j, peer_fence, fence_seq_exit);
-                return HTP_STATUS_INTERNAL_ERR;
-            }
-            hex_pause();
-        }
-    }
-    asm volatile ("syncht" : : : "memory");
-
-    htp_trace_event_stop(tr0, HTP_TRACE_EVT_FENCE, (uint16_t) fence_seq_exit);
-
-    return HTP_STATUS_OK;
-}

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

@@ -1,40 +0,0 @@
-#ifndef ALLREDUCE_OPS_H
-#define ALLREDUCE_OPS_H
-
-#include <stdint.h>
-
-#define HTP_ALLREDUCE_MAX_RANKS 4
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-enum htp_allreduce_kernel_type {
-    HTP_ALLREDUCE_KERNEL_UNSUPPORTED = 0,
-    HTP_ALLREDUCE_KERNEL_DMA_1D,
-    HTP_ALLREDUCE_KERNEL_DMA_2D,
-};
-
-struct htp_allreduce_kernel_params {
-    int32_t rank;
-    int32_t n_ranks;
-    int32_t n_threads;
-    int32_t block_elems;          // 1D: block_elems, 2D: block_rows
-    int32_t elems_per_thread;     // 1D: nelem_per_thread, 2D: nrows_per_thread
-    int32_t vtcm_size_per_thread;
-    int32_t vtcm_size;
-    int32_t kernel_type;
-    int32_t ne0;
-    int32_t ne1;
-    int32_t row_size_aligned;
-    int32_t rank_elem_start;
-    int32_t rank_nelem;
-    int32_t n_dsts;
-    int32_t is_row_bcast;
-};
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /* ALLREDUCE_OPS_H */

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

@@ -4,7 +4,6 @@
 
 #include <HAP_farf.h>
 #include <HAP_perf.h>
-#include <qurt_memory.h>
 
 #include <math.h>
 #include <string.h>
@@ -15,7 +14,6 @@
 #include "htp-ops.h"
 #include "htp-ops.h"
 #include "hvx-utils.h"
-#include "htp-tensor.h"
 
 struct htp_copy_context {
     struct htp_ops_context * octx;
@@ -80,7 +78,7 @@ static void cpy_thread_##NAME##_sameshape(unsigned int nth, unsigned int ith, vo
     }                                                                                                          \
 }
 
-DEFINE_CPY_SAMESHAPE(f32,  float, 4)
+DEFINE_CPY_SAMESHAPE(f32, float, 4)
 DEFINE_CPY_SAMESHAPE(f16, __fp16, 2)
 
 #define DEFINE_CPY_RESHAPE(NAME, ELEM_TYPE, ELEM_SIZE)                                                         \
@@ -181,7 +179,7 @@ static void cpy_thread_##NAME##_reshape(unsigned int nth, unsigned int ith, void
     }                                                                                                          \
 }
 
-DEFINE_CPY_RESHAPE(f32,  float, 4)
+DEFINE_CPY_RESHAPE(f32, float, 4)
 DEFINE_CPY_RESHAPE(f16, __fp16, 2)
 
 static void cpy_thread_f16_f32_sameshape(unsigned int nth, unsigned int ith, void * data) {
@@ -234,41 +232,6 @@ static void cpy_thread_f32_f16_sameshape(unsigned int nth, unsigned int ith, voi
     }
 }
 
-static inline void cpy_dma_sametype_sameshape(
-    struct htp_ops_context * octx,
-    const struct htp_tensor * dst,
-    const struct htp_tensor * src0,
-    uint32_t elem_size,
-    uint32_t ne00, uint32_t ne01, uint32_t ne02, uint32_t ne03,
-    uint32_t nb01, uint32_t nb02, uint32_t nb03,
-    uint32_t  nb1, uint32_t  nb2, uint32_t nb3
-) {
-    const bool contiguous_outer =
-        (ne02 == 1 || (nb02 == ne01 * nb01 && nb2 == ne01 * nb1)) &&
-        (ne03 == 1 || (nb03 == ne02 * nb02 && nb3 == ne02 * nb2));
-
-    dma_queue * q = octx->ctx->dma[0];
-
-    if (contiguous_outer) {
-        dma_queue_push(q, dma_make_ptr((void *) dst->data, (const void *) src0->data), nb1, nb01, ne00 * elem_size, ne01 * ne02 * ne03);
-        dma_queue_pop(q);
-        return;
-    }
-
-    for (uint32_t i03 = 0; i03 < ne03; i03++) {
-        for (uint32_t i02 = 0; i02 < ne02; i02++) {
-            uint8_t* dst_ptr  = (uint8_t*) dst->data  + i02*nb2  + i03*nb3;
-            uint8_t* src0_ptr = (uint8_t*) src0->data + i02*nb02 + i03*nb03;
-            if (!dma_queue_push(q, dma_make_ptr(dst_ptr, src0_ptr), nb1, nb01, ne00 * elem_size, ne01)) {
-                dma_queue_flush(q);
-                dma_queue_push(q, dma_make_ptr(dst_ptr, src0_ptr), nb1, nb01, ne00 * elem_size, ne01);
-            }
-        }
-    }
-
-    dma_queue_flush(q);
-}
-
 int op_cpy(struct htp_ops_context * octx) {
     cpy_preamble;
 
@@ -301,11 +264,14 @@ int op_cpy(struct htp_ops_context * octx) {
 
     ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
 
-    worker_callback_t copy_fun = NULL;
-    bool use_dma = false;
+    worker_callback_t copy_fun;
 
     if (sametype && sameshape) {
-        use_dma = true;
+        if (src0->type == HTP_TYPE_F32) {
+            copy_fun = cpy_thread_f32_sameshape;
+        } else {
+            copy_fun = cpy_thread_f16_sameshape;
+        }
     } else if (sameshape) {
         /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
             copy_fun = cpy_thread_f16_f32_sameshape;
@@ -323,28 +289,7 @@ int op_cpy(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    if (use_dma) {
-        cpy_dma_sametype_sameshape(octx, dst, src0, ct.src0_type_size, ne00, ne01, ne02, ne03, nb01, nb02, nb03, nb1, nb2, nb3);
-    } else {
-        worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
-    }
-
-    const struct htp_tensor *sync = octx->src[1];
-    if (sync) {
-        if (!use_dma) {
-            // htp_tensor_flush_all(octx->ctx, octx->dsts, 1);
-            qurt_mem_cache_clean((qurt_addr_t) 0, 0, QURT_MEM_CACHE_FLUSH_INVALIDATE_ALL, QURT_MEM_DCACHE);
-        }
-
-        atomic_uint * sync_fence = (atomic_uint *) sync->data;
-        const uint32_t seq = (uint32_t) octx->op_params[0];
-
-        atomic_store(&sync_fence[0], seq);
-        asm volatile ("syncht" : : : "memory");
-        Q6_dccleaninva_A((void *) sync_fence);
-
-        FARF(HIGH, "ggml-hex: sync-release : fence %p seq %u\n", sync_fence, seq);
-    }
+    worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
 
     return HTP_STATUS_OK;
 }

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

@@ -244,18 +244,17 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
         return dptr;
     }
 
-    dptr = r->dptr[r->pop_idx];
-
-    volatile dma_descriptor_2d * desc = &r->desc[r->pop_idx];
+    dma_descriptor_2d * desc = &r->desc[r->pop_idx];
 
     // Wait for desc to complete
     if (!desc->done) {
-        // FARF(ALWAYS, "dma-poll: idx %u dst %p src %p", r->pop_idx, dptr.dst, dptr.src);
         while (!desc->done) {
             dmpoll();
         }
     }
 
+    dptr = r->dptr[r->pop_idx];
+
     htp_trace_event_stop(r->trace, HTP_TRACE_EVT_DMA, r->pop_idx);
 
     r->pop_idx = (r->pop_idx + 1) & r->idx_mask;

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

@@ -30,8 +30,6 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-tensor.h"
-#include "hvx-quant.h"
 
 #include "flash-attn-ops.h"
 #include "hvx-fa-kernels.h"
@@ -87,17 +85,12 @@ struct htp_fa_context {
     uint8_t * spad_m;
     uint8_t * spad_a;
 
-    const struct htp_tensor * k;
-    const struct htp_tensor * v;
-
     uint64_t t_start;
 };
 
 struct hmx_fa_context {
     const struct htp_ops_context * octx;
     const struct htp_tensor *      sinks;  // attention sinks (src[4]), NULL if absent
-    const struct htp_tensor *      k;
-    const struct htp_tensor *      v;
     bool         pipeline;  // true when n_kv_blocks >= FA_MIN_KV_BLOCKS && n_threads >= 2
     uint32_t     n_threads;
 
@@ -221,8 +214,8 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
     const uint32_t DV = nev0;
 
     const size_t size_q_row = DK * ((q->type == HTP_TYPE_F32) ? 4 : 2);
-    const size_t size_k_row = htp_tensor_get_row_size(k->type, DK);
-    const size_t size_v_row = htp_tensor_get_row_size(v->type, DV);
+    const size_t size_k_row = DK * sizeof(__fp16);
+    const size_t size_v_row = DV * sizeof(__fp16);
 
     // Scratchpad buffers for Q, K, V, Mask, and VKQ32 accumulator
     uint8_t * spad_q = factx->spad_q + factx->size_q_block * ith;
@@ -371,23 +364,6 @@ static void flash_attn_ext_f16_thread(unsigned int nth, unsigned int ith, void *
             uint8_t * v_base = dma_queue_pop(dma).dst; // V
             __fp16  * m_base = mask ? dma_queue_pop(dma).dst : NULL; // M
 
-            if (factx->k->type == HTP_TYPE_Q8_0) {
-                htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_FA_K_PREP, ir);
-                for (uint32_t r = 0; r < current_block_size; ++r) {
-                    __fp16 * row_k = (__fp16 *)(k_base + r * factx->size_k_row_padded);
-                    hvx_dequantize_row_q8_0_f16(row_k, row_k, DK);
-                }
-                htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_K_PREP, ir);
-            }
-            if (factx->v->type == HTP_TYPE_Q8_0) {
-                htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_FA_V_PREP, ir);
-                for (uint32_t r = 0; r < current_block_size; ++r) {
-                    __fp16 * row_v = (__fp16 *)(v_base + r * factx->size_v_row_padded);
-                    hvx_dequantize_row_q8_0_f16(row_v, row_v, DV);
-                }
-                htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_V_PREP, ir);
-            }
-
             htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_FA_QK, ir);
 
             // Inner loop processing the block from VTCM
@@ -649,12 +625,6 @@ static void fa_k_interleave_thread(unsigned int n, unsigned int i, void * data)
 
     struct htp_thread_trace * tr = &factx->octx->ctx->trace[i];
     htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_FA_K_PREP, (uint16_t) (args->kv_start + start));
-    if (factx->k->type == HTP_TYPE_Q8_0) {
-        for (uint32_t r = start; r < end; ++r) {
-            __fp16 * row_k = (__fp16 *)((char *)args->curr_k + r * args->src_stride * sizeof(__fp16));
-            hvx_dequantize_row_q8_0_f16(row_k, row_k, factx->DK);
-        }
-    }
     hmx_interleave_rows_to_tiles(factx->vtcm_k_tiles[args->buf_idx], (const __fp16 *) args->curr_k, total_rows, factx->DK,
                              args->src_stride, start, end);
     htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_K_PREP, (uint16_t) (args->kv_start + start));
@@ -703,12 +673,6 @@ static void fa_v_interleave_thread(unsigned int n, unsigned int i, void * data)
 
     struct htp_thread_trace * tr = &factx->octx->ctx->trace[i];
     htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_FA_V_PREP, (uint16_t) (args->kv_start + start));
-    if (factx->v->type == HTP_TYPE_Q8_0) {
-        for (uint32_t r = start; r < end; ++r) {
-            __fp16 * row_v = (__fp16 *)((char *)args->v_src + r * args->src_stride * sizeof(__fp16));
-            hvx_dequantize_row_q8_0_f16(row_v, row_v, factx->DV);
-        }
-    }
     hmx_interleave_cols_to_tiles(v_tiles_dst, (const __fp16 *) args->v_src, total_rows, factx->DV,
                              args->src_stride, (uint32_t) args->n_col_tiles, start, end);
     htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_FA_V_PREP, (uint16_t) (args->kv_start + start));
@@ -1845,8 +1809,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     memset(&factx, 0, sizeof(factx));
     factx.octx           = octx;
     factx.sinks          = octx->src[4];  // NULL if this op has no attention sinks
-    factx.k              = k;
-    factx.v              = v;
     factx.n_threads      = kparams->n_threads;
     factx.DK             = DK;
     factx.DV             = DV;
@@ -1891,10 +1853,10 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     // ======== VTCM allocation (GQA-aware) ========
     // K/V row sizes drive the DMA descriptors (not the VTCM layout) and are used
     // throughout the KV loop below.
-    const size_t size_k_row        = htp_tensor_get_row_size(k->type, DK);
-    const size_t size_v_row        = htp_tensor_get_row_size(v->type, DV);
-    const size_t size_k_row_padded = hex_round_up(DK * sizeof(__fp16), 128);
-    const size_t size_v_row_padded = hex_round_up(DV * sizeof(__fp16), 128);
+    const size_t size_k_row        = DK * sizeof(__fp16);
+    const size_t size_v_row        = DV * sizeof(__fp16);
+    const size_t size_k_row_padded = hex_round_up(size_k_row, 128);
+    const size_t size_v_row_padded = hex_round_up(size_v_row, 128);
 
     // Build the VTCM layout once (shared with the host estimator) and place every
     // scratch buffer at its computed offset.
@@ -2386,9 +2348,7 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
     const struct htp_tensor * dst  = octx->dst;
 
     // Check support
-    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) ||
-        (k->type != HTP_TYPE_F16 && k->type != HTP_TYPE_Q8_0) ||
-        (v->type != HTP_TYPE_F16 && v->type != HTP_TYPE_Q8_0)) {
+    if ((q->type != HTP_TYPE_F16 && q->type != HTP_TYPE_F32) || k->type != HTP_TYPE_F16 || v->type != HTP_TYPE_F16) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
@@ -2404,8 +2364,6 @@ int op_flash_attn_ext(struct htp_ops_context * octx) {
 
     struct htp_fa_context factx;
     factx.octx = octx;
-    factx.k = k;
-    factx.v = v;
 
     factx.t_start = HAP_perf_get_qtimer_count();
 

ggml/src/ggml-hexagon/htp/get-rows-ops.c

@@ -12,17 +12,18 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-tensor.h"
+#include "htp-ops.h"
 #include "hvx-utils.h"
-#include "hvx-quant.h"
-#include "get-rows-ops.h"
-#include "work-queue.h"
 
 struct get_rows_context {
     struct htp_ops_context * octx;
-    const struct htp_get_rows_kernel_params * kparams;
-    struct htp_get_rows_vtcm_layout vtcm_layout;
-    uint8_t * vtcm_base;
+    uint32_t tasks_per_thread;
+    uint32_t total_tasks;
+    uint32_t chunks_per_row;
+    uint32_t chunk_size;
+    struct fastdiv_values get_rows_div_ne10;
+    struct fastdiv_values get_rows_div_ne10_ne11;
+    struct fastdiv_values get_rows_div_chunks_per_row;
 };
 
 #define get_rows_preamble \
@@ -55,161 +56,102 @@ struct get_rows_context {
                                                \
     const uint32_t nr = ne10 * ne11 * ne12;
 
-#define GET_ROWS_THREAD_ST_FN(IDX_TYPE)                                                                                \
-static void get_rows_thread_st_##IDX_TYPE(unsigned int nth, unsigned int ith, void *data) {                            \
-    struct get_rows_context * grctx = (struct get_rows_context *)data;                                                 \
-    struct htp_ops_context * octx = grctx->octx;                                                                       \
-    const struct htp_get_rows_kernel_params * kparams = grctx->kparams;                                                \
-    get_rows_preamble;                                                                                                 \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                                    \
-    const uint32_t ir0 = dr * ith;                                                                                     \
-    if (ir0 >= kparams->total_tasks) {                                                                                 \
-        return;                                                                                                        \
-    }                                                                                                                  \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                          \
-    const uint32_t row_size_bytes = htp_tensor_get_row_size(octx->src[0]->type, ne00);                                 \
-    dma_queue * dma_queue = octx->ctx->dma[ith];                                                                       \
-    for (uint32_t i = ir0; i < ir1; ++i) {                                                                             \
-        const uint32_t i12 = fastdiv(i, &kparams->div_ne10_ne11);                                                      \
-        const uint32_t rem = i - i12 * ne11 * ne10;                                                                    \
-        const uint32_t i11 = fastdiv(rem, &kparams->div_ne10);                                                         \
-        const uint32_t i10 = rem - i11 * ne10;                                                                         \
-        const IDX_TYPE * src1_ptr = (const IDX_TYPE *)(octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12);           \
-        const uint32_t i01 = (uint32_t)*src1_ptr;                                                                      \
-        assert(i01 < ne01);                                                                                            \
-        const uint32_t q02 = fastdiv(i11, &kparams->div_ne02);                                                         \
-        const uint32_t i02 = i11 - q02 * ne02;                                                                         \
-        const uint32_t q03 = fastdiv(i12, &kparams->div_ne03);                                                         \
-        const uint32_t i03 = i12 - q03 * ne03;                                                                         \
-        const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i02*nb02 + i03*nb03;                                \
-        const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3;                                 \
-        while (!dma_queue_push(dma_queue, dma_make_ptr((void *)dst_ptr, (const void *)src0_ptr), nb1, nb01,            \
-                               row_size_bytes, 1)) {                                                                   \
-            dma_queue_pop(dma_queue);                                                                                  \
-        }                                                                                                              \
-    }                                                                                                                  \
-    dma_queue_flush(dma_queue);                                                                                        \
+static void get_rows_thread_f32_f32_dma(unsigned int nth, unsigned int ith, void *data) {
+    struct get_rows_context * grctx = (struct get_rows_context *)data;
+    struct htp_ops_context * octx = grctx->octx;
+    get_rows_preamble;
+
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
+    const uint32_t dr  = grctx->tasks_per_thread;
+    const uint32_t ir0 = dr * ith;
+    if (ir0 >= grctx->total_tasks) {
+        return;
+    }
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->total_tasks);
+
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+
+    dma_queue * dma_queue = octx->ctx->dma[ith];
+    for (uint32_t i = ir0; i < ir1; ++i) {
+        const uint32_t i12 = fastdiv(i, &grctx->get_rows_div_ne10_ne11);
+        const uint32_t rem = i - i12 * ne11 * ne10;
+        const uint32_t i11 = fastdiv(rem, &grctx->get_rows_div_ne10);
+        const uint32_t i10 = rem - i11 * ne10;
+
+        const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
+        uint32_t i01 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
+
+        if (i01 >= ne01) {
+            continue;
+        }
+
+        const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i11*nb02 + i12*nb03;
+        const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3;
+
+        while (!dma_queue_push(dma_queue, dma_make_ptr((void *)dst_ptr, (const void *)src0_ptr), nb1, nb01, ne00 * sizeof(float), 1)) {
+            dma_queue_pop(dma_queue);
+        }
+    }
+    dma_queue_flush(dma_queue);
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "get-rows-f32-f32-dma %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
 }
 
-GET_ROWS_THREAD_ST_FN(int32_t)
-GET_ROWS_THREAD_ST_FN(int64_t)
+static void get_rows_thread_f32_f32_hvx(unsigned int nth, unsigned int ith, void *data) {
+    struct get_rows_context * grctx = (struct get_rows_context *)data;
+    struct htp_ops_context * octx = grctx->octx;
+    get_rows_preamble;
 
-#define GET_ROWS_THREAD_DT_FN(TYPE_NAME, SRC0_SIZE_EXPR, IDX_TYPE, COMPUTE_EXPR)                                       \
-static void get_rows_thread_##TYPE_NAME##_##IDX_TYPE(unsigned int nth, unsigned int ith, void *data) {                 \
-    struct get_rows_context * grctx = (struct get_rows_context *)data;                                                 \
-    struct htp_ops_context * octx = grctx->octx;                                                                       \
-    const struct htp_get_rows_kernel_params * kparams = grctx->kparams;                                                \
-    get_rows_preamble;                                                                                                 \
-    struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                             \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                                    \
-    const uint32_t ir0 = dr * ith;                                                                                     \
-    if (ir0 >= kparams->total_tasks) {                                                                                 \
-        return;                                                                                                        \
-    }                                                                                                                  \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                          \
-    const uint32_t chunks_per_row = kparams->chunks_per_row;                                                           \
-    const uint32_t chunk_size     = kparams->chunk_size;                                                               \
-    dma_queue * dma_queue = octx->ctx->dma[ith];                                                                       \
-    const struct htp_get_rows_vtcm_layout * vtcm_layout = &grctx->vtcm_layout;                                         \
-    uint8_t * vtcm_src0 = grctx->vtcm_base + vtcm_layout->off_src0 + ith * vtcm_layout->src0_bytes_per_thread;         \
-    uint8_t * vtcm_dst  = grctx->vtcm_base + vtcm_layout->off_dst  + ith * vtcm_layout->dst_bytes_per_thread;          \
-    for (uint32_t step = 0, spad_idx = 0; step < ir1 - ir0 && spad_idx < 2; ++step, spad_idx++) {                      \
-        const uint32_t i = ir0 + step;                                                                                 \
-        const uint32_t row_idx   = fastdiv(i, &kparams->div_chunks_per_row);                                           \
-        const uint32_t chunk_idx = i - row_idx * chunks_per_row;                                                       \
-        const uint32_t i12 = fastdiv(row_idx, &kparams->div_ne10_ne11);                                                \
-        const uint32_t rem = row_idx - i12 * ne11 * ne10;                                                              \
-        const uint32_t i11 = fastdiv(rem, &kparams->div_ne10);                                                         \
-        const uint32_t i10 = rem - i11 * ne10;                                                                         \
-        const IDX_TYPE * src1_ptr = (const IDX_TYPE *)(octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12);           \
-        const uint32_t i01 = (uint32_t)*src1_ptr;                                                                      \
-        assert(i01 < ne01);                                                                                            \
-        const uint32_t q02 = fastdiv(i11, &kparams->div_ne02);                                                         \
-        const uint32_t i02 = i11 - q02 * ne02;                                                                         \
-        const uint32_t q03 = fastdiv(i12, &kparams->div_ne03);                                                         \
-        const uint32_t i03 = i12 - q03 * ne03;                                                                         \
-        const uint32_t offset = chunk_idx * chunk_size;                                                                \
-        const uint32_t cur_elems = (offset < ne00) ? MIN(chunk_size, ne00 - offset) : 0;                               \
-        const uint32_t cur_src0_bytes = SRC0_SIZE_EXPR(cur_elems);                                                     \
-        const uint32_t cur_dst_bytes  = cur_elems * sizeof(float);                                                     \
-        const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i02*nb02 + i03*nb03 + SRC0_SIZE_EXPR(offset);       \
-        dma_queue_push(dma_queue,                                                                                      \
-                       dma_make_ptr((void *)(uintptr_t)octx->dst->data,                                                \
-                                    vtcm_dst + spad_idx * vtcm_layout->dst_spad_half_size),                            \
-                       cur_dst_bytes, vtcm_layout->dst_spad_half_size, cur_dst_bytes, 0);                              \
-        dma_queue_push(dma_queue,                                                                                      \
-                       dma_make_ptr((void *)(vtcm_src0 + spad_idx * vtcm_layout->src0_spad_half_size),                 \
-                                    (const void *)src0_ptr),                                                           \
-                       vtcm_layout->src0_spad_half_size, cur_src0_bytes, cur_src0_bytes, 1);                           \
-    }                                                                                                                  \
-    for (uint32_t step = 0; step < ir1 - ir0; ++step) {                                                                \
-        const uint32_t i = ir0 + step;                                                                                 \
-        void * dst_spad = (void *) dma_queue_pop(dma_queue).src;                                                       \
-        void * src_spad = (void *) dma_queue_pop(dma_queue).dst;                                                       \
-        const uint32_t row_idx   = fastdiv(i, &kparams->div_chunks_per_row);                                           \
-        const uint32_t chunk_idx = i - row_idx * chunks_per_row;                                                       \
-        const uint32_t i12 = fastdiv(row_idx, &kparams->div_ne10_ne11);                                                \
-        const uint32_t rem = row_idx - i12 * ne11 * ne10;                                                              \
-        const uint32_t i11 = fastdiv(rem, &kparams->div_ne10);                                                         \
-        const uint32_t i10 = rem - i11 * ne10;                                                                         \
-        const uint32_t offset = chunk_idx * chunk_size;                                                                \
-        const uint32_t cur_elems = (offset < ne00) ? MIN(chunk_size, ne00 - offset) : 0;                               \
-        const uint32_t cur_dst_bytes  = cur_elems * sizeof(float);                                                     \
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, i);                                                          \
-        COMPUTE_EXPR;                                                                                                  \
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, i);                                                           \
-        const uintptr_t dst_ptr  = octx->dst->data + i10*nb1 + i11*nb2 + i12*nb3 + offset * sizeof(float);             \
-        dma_queue_push(dma_queue,                                                                                      \
-                       dma_make_ptr((void *)dst_ptr, (const void *)dst_spad),                                          \
-                       cur_dst_bytes, vtcm_layout->dst_spad_half_size, cur_dst_bytes, 1);                              \
-        const uint32_t next_step = step + 2;                                                                           \
-        if (next_step < ir1 - ir0) {                                                                                   \
-            const uint32_t pi = ir0 + next_step;                                                                       \
-            const uint32_t prow_idx   = fastdiv(pi, &kparams->div_chunks_per_row);                                     \
-            const uint32_t pchunk_idx = pi - prow_idx * chunks_per_row;                                                \
-            const uint32_t pi12 = fastdiv(prow_idx, &kparams->div_ne10_ne11);                                          \
-            const uint32_t prem = prow_idx - pi12 * ne11 * ne10;                                                       \
-            const uint32_t pi11 = fastdiv(prem, &kparams->div_ne10);                                                   \
-            const uint32_t pi10 = prem - pi11 * ne10;                                                                  \
-            const IDX_TYPE * psrc1_ptr = (const IDX_TYPE *)(octx->src[1]->data + pi10*nb10 + pi11*nb11 + pi12*nb12);   \
-            const uint32_t pi01 = (uint32_t)*psrc1_ptr;                                                                \
-            assert(pi01 < ne01);                                                                                       \
-            const uint32_t pq02 = fastdiv(pi11, &kparams->div_ne02);                                                   \
-            const uint32_t pi02 = pi11 - pq02 * ne02;                                                                  \
-            const uint32_t pq03 = fastdiv(pi12, &kparams->div_ne03);                                                   \
-            const uint32_t pi03 = pi12 - pq03 * ne03;                                                                  \
-            const uint32_t poffset = pchunk_idx * chunk_size;                                                          \
-            const uint32_t pcur_elems = (poffset < ne00) ? MIN(chunk_size, ne00 - poffset) : 0;                        \
-            const uint32_t pcur_src0_bytes = SRC0_SIZE_EXPR(pcur_elems);                                               \
-            const uintptr_t psrc0_ptr =                                                                                \
-                octx->src[0]->data + pi01*nb01 + pi02*nb02 + pi03*nb03 + SRC0_SIZE_EXPR(poffset);                      \
-            dma_queue_push(dma_queue,                                                                                  \
-                           dma_make_ptr((void *)src_spad, (const void *)psrc0_ptr),                                    \
-                           vtcm_layout->src0_spad_half_size, pcur_src0_bytes, pcur_src0_bytes, 1);                     \
-        }                                                                                                              \
-    }                                                                                                                  \
-    dma_queue_flush(dma_queue);                                                                                        \
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
+    const uint32_t dr  = grctx->tasks_per_thread;
+    const uint32_t ir0 = dr * ith;
+    if (ir0 >= grctx->total_tasks) {
+        return;
+    }
+    const uint32_t ir1 = MIN(ir0 + dr, grctx->total_tasks);
+
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+
+    const uint32_t chunks_per_row = grctx->chunks_per_row;
+    const uint32_t chunk_size     = grctx->chunk_size;
+    for (uint32_t i = ir0; i < ir1; ++i) {
+        const uint32_t row_idx   = fastdiv(i, &grctx->get_rows_div_chunks_per_row);
+        const uint32_t chunk_idx = i - row_idx * chunks_per_row;
+
+        const uint32_t i12 = fastdiv(row_idx, &grctx->get_rows_div_ne10_ne11);
+        const uint32_t rem = row_idx - i12 * ne11 * ne10;
+        const uint32_t i11 = fastdiv(rem, &grctx->get_rows_div_ne10);
+        const uint32_t i10 = rem - i11 * ne10;
+
+        const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
+        uint32_t i01 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
+
+        if (i01 >= ne01) {
+            continue;
+        }
+
+        const uint32_t offset = chunk_idx * chunk_size;
+        if (offset < ne00) {
+            const uint32_t copy_size = MIN(chunk_size, ne00 - offset);
+            const uintptr_t src0_ptr = octx->src[0]->data + i01*nb01 + i11*nb02 + i12*nb03 + offset * sizeof(float);
+            const uintptr_t dst_ptr  = octx->dst->data    + i10*nb1  + i11*nb2  + i12*nb3  + offset * sizeof(float);
+            hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, copy_size);
+        }
+    }
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "get-rows-f32-f32-hvx %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
 }
 
-#define F32_BYTES(n)  ((n) * sizeof(float))
-#define F16_BYTES(n)  ((n) * sizeof(__fp16))
-#define Q8_0_BYTES(n) (((n) / 32) * sizeof(block_q8_0))
-
-GET_ROWS_THREAD_DT_FN(f32,  F32_BYTES,  int32_t, { if (cur_elems > 0) hvx_copy_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, cur_elems); })
-GET_ROWS_THREAD_DT_FN(f32,  F32_BYTES,  int64_t, { if (cur_elems > 0) hvx_copy_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, cur_elems); })
-
-GET_ROWS_THREAD_DT_FN(f16,  F16_BYTES,  int32_t, { hvx_dequantize_row_f16_f32((float *)dst_spad, src_spad, ne00); })
-GET_ROWS_THREAD_DT_FN(f16,  F16_BYTES,  int64_t, { hvx_dequantize_row_f16_f32((float *)dst_spad, src_spad, ne00); })
-
-GET_ROWS_THREAD_DT_FN(q8_0, Q8_0_BYTES, int32_t, { hvx_dequantize_row_q8_0_f32((float *)dst_spad, src_spad, ne00); })
-GET_ROWS_THREAD_DT_FN(q8_0, Q8_0_BYTES, int64_t, { hvx_dequantize_row_q8_0_f32((float *)dst_spad, src_spad, ne00); })
-
 int op_get_rows(struct htp_ops_context * octx) {
-    const struct htp_get_rows_kernel_params * kparams = (const struct htp_get_rows_kernel_params *) octx->kernel_params;
+    get_rows_preamble;
 
-    if (octx->src[0]->type != HTP_TYPE_F32 &&
-        octx->src[0]->type != HTP_TYPE_F16 &&
-        octx->src[0]->type != HTP_TYPE_Q8_0) {
+    if (octx->src[0]->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
@@ -225,28 +167,52 @@ int op_get_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_OK;
     }
 
+    const uint32_t nb00 = octx->src[0]->nb[0];
+    const uint32_t nb0  = octx->dst->nb[0];
+
+    const bool can_use_dma = (nb00 == sizeof(float)) && (nb0 == sizeof(float));
+    const bool use_dma = can_use_dma && (ne00 >= 2048);
+
     struct get_rows_context grctx;
     grctx.octx = octx;
-    grctx.kparams = kparams;
-    grctx.vtcm_base = (uint8_t *)octx->ctx->vtcm_base;
+    grctx.get_rows_div_ne10      = init_fastdiv_values(octx->src[1]->ne[0]);
+    grctx.get_rows_div_ne10_ne11 = init_fastdiv_values(octx->src[1]->ne[0] * octx->src[1]->ne[1]);
 
-    const uint32_t ne00 = octx->src[0]->ne[0];
-    htp_get_rows_vtcm_layout_build(&grctx.vtcm_layout, octx->src[0]->type, ne00, kparams->n_threads);
+    if (use_dma) {
+        grctx.chunks_per_row = 1;
+        grctx.chunk_size = ne00;
+        grctx.total_tasks = nr;
+        grctx.get_rows_div_chunks_per_row = init_fastdiv_values(1);
 
-    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+        const uint32_t n_threads = MIN(nr, octx->n_threads);
+        grctx.tasks_per_thread = (nr + n_threads - 1) / n_threads;
 
-    work_queue_func_t q_func = NULL;
-    if (kparams->use_dma) {
-        q_func = (work_queue_func_t)(is_i32 ? get_rows_thread_st_int32_t : get_rows_thread_st_int64_t);
+        worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32_dma, &grctx, n_threads);
     } else {
-        switch (octx->src[0]->type) {
-            case HTP_TYPE_F32:  q_func = (work_queue_func_t)(is_i32 ? get_rows_thread_f32_int32_t  : get_rows_thread_f32_int64_t);  break;
-            case HTP_TYPE_F16:  q_func = (work_queue_func_t)(is_i32 ? get_rows_thread_f16_int32_t  : get_rows_thread_f16_int64_t);  break;
-            case HTP_TYPE_Q8_0: q_func = (work_queue_func_t)(is_i32 ? get_rows_thread_q8_0_int32_t : get_rows_thread_q8_0_int64_t); break;
-            default:            return HTP_STATUS_NO_SUPPORT;
-        }
-    }
+        uint32_t chunks_per_row = 1;
+        uint32_t chunk_size = ne00;
+        uint32_t total_tasks = nr;
 
-    work_queue_run(octx->ctx->work_queue, q_func, &grctx, kparams->n_threads);
+        if (nr < octx->n_threads) {
+            const uint32_t min_chunk_size = 1024;
+            uint32_t max_chunks = ne00 / min_chunk_size;
+            if (max_chunks == 0) {
+                max_chunks = 1;
+            }
+            chunks_per_row = MIN((octx->n_threads + nr - 1) / nr, max_chunks);
+            chunk_size = (ne00 + chunks_per_row - 1) / chunks_per_row;
+            total_tasks = nr * chunks_per_row;
+        }
+
+        grctx.chunks_per_row = chunks_per_row;
+        grctx.chunk_size = chunk_size;
+        grctx.total_tasks = total_tasks;
+        grctx.get_rows_div_chunks_per_row = init_fastdiv_values(chunks_per_row);
+
+        const uint32_t n_threads = MIN(total_tasks, octx->n_threads);
+        grctx.tasks_per_thread = (total_tasks + n_threads - 1) / n_threads;
+
+        worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32_hvx, &grctx, n_threads);
+    }
     return HTP_STATUS_OK;
 }

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

@@ -1,77 +0,0 @@
-#ifndef HTP_GET_ROWS_OPS_H
-#define HTP_GET_ROWS_OPS_H
-
-#include "hex-fastdiv.h"
-
-struct htp_get_rows_kernel_params {
-    int32_t  n_threads;
-    int32_t  use_dma;
-    int32_t  chunks_per_row;
-    int32_t  chunk_size;
-    int32_t  total_tasks;
-    int32_t  tasks_per_thread;
-    int32_t  vtcm_size;
-
-    // Fastdiv helpers
-    struct fastdiv_values div_ne10;
-    struct fastdiv_values div_ne10_ne11;
-    struct fastdiv_values div_chunks_per_row;
-    struct fastdiv_values div_ne02;
-    struct fastdiv_values div_ne03;
-};
-
-struct htp_get_rows_vtcm_layout {
-    size_t total_bytes;
-    size_t off_src0;
-    size_t off_dst;
-
-    size_t src0_bytes_per_thread;
-    size_t dst_bytes_per_thread;
-
-    size_t src0_spad_half_size;
-    size_t dst_spad_half_size;
-};
-
-static inline void htp_get_rows_vtcm_layout_build(
-    struct htp_get_rows_vtcm_layout * vtcm_layout,
-    int type,
-    uint32_t ne00,
-    uint32_t n_threads) {
-
-    uint32_t src0_row_size = 0;
-    switch (type) {
-        case 0: // HTP_TYPE_F32
-            src0_row_size = ne00 * 4;
-            break;
-        case 1: // HTP_TYPE_F16
-            src0_row_size = ne00 * 2;
-            break;
-        case 8: // HTP_TYPE_Q8_0
-            src0_row_size = (ne00 / 32) * 34;
-            break;
-        default:
-            src0_row_size = 0;
-            break;
-    }
-
-    size_t src0_row_size_aligned = (src0_row_size + 255) & ~255;
-    size_t dst_row_size_aligned  = (ne00 * sizeof(float) + 255) & ~255;
-
-    vtcm_layout->src0_spad_half_size = src0_row_size_aligned;
-    vtcm_layout->dst_spad_half_size  = dst_row_size_aligned;
-
-    vtcm_layout->src0_bytes_per_thread = src0_row_size_aligned * 2;
-    vtcm_layout->dst_bytes_per_thread  = dst_row_size_aligned * 2;
-
-    vtcm_layout->off_src0 = 0;
-    vtcm_layout->off_dst  = vtcm_layout->off_src0 + vtcm_layout->src0_bytes_per_thread * n_threads;
-    vtcm_layout->total_bytes = vtcm_layout->off_dst + vtcm_layout->dst_bytes_per_thread * n_threads;
-}
-
-#if defined(__cplusplus)
-static_assert(sizeof(struct htp_get_rows_kernel_params) <= 128, "htp_get_rows_kernel_params is too large for kernel_params blob");
-#else
-_Static_assert(sizeof(struct htp_get_rows_kernel_params) <= 128, "htp_get_rows_kernel_params is too large for kernel_params blob");
-#endif
-
-#endif // HTP_GET_ROWS_OPS_H

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

@@ -39,22 +39,17 @@ static inline void hex_l2fetch_block(const void * addr, size_t size) {
 
 #define HEX_L2_LINE_SIZE           128
 #define HEX_L2_BLOCK_SIZE          (HEX_L2_LINE_SIZE * 4) // flush granularity (lines per loop iteration)
-#define HEX_L2_FLUSH_IL_THRESHOLD  1024                   // inline flush threshold
 #define HEX_L2_FLUSH_WQ_THRESHOLD  (4 * 1024)
 #define HEX_L2_FLUSH_ALL_THRESHOLD (4 * 1024 * 1024)
 
 static inline void hex_l2flush(void * addr, size_t size) {
     const uint32_t s = ((uint32_t) addr) & ~(HEX_L2_LINE_SIZE - 1);
     const uint32_t e = (((uint32_t) addr) + size + HEX_L2_LINE_SIZE - 1) & ~(HEX_L2_LINE_SIZE - 1);
-    const uint32_t eb = s + ((e - s) & ~(HEX_L2_BLOCK_SIZE - 1));
-    for (uint32_t i = s; i < eb; i += HEX_L2_BLOCK_SIZE) {
-        Q6_dccleaninva_A((void *) (i + HEX_L2_LINE_SIZE * 0));
-        Q6_dccleaninva_A((void *) (i + HEX_L2_LINE_SIZE * 1));
-        Q6_dccleaninva_A((void *) (i + HEX_L2_LINE_SIZE * 2));
-        Q6_dccleaninva_A((void *) (i + HEX_L2_LINE_SIZE * 3));
-    }
-    for (uint32_t i = eb; i < e; i += HEX_L2_LINE_SIZE) {
-        Q6_dccleaninva_A((void *) i);
+    for (uint32_t i = s; i < e; i += HEX_L2_BLOCK_SIZE) {
+        Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 0);
+        Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 1);
+        Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 2);
+        Q6_dccleaninva_A((void *) i + HEX_L2_LINE_SIZE * 3);
     }
 }
 

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

@@ -117,7 +117,8 @@ struct htp_context {
 
 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
-int op_matmul_nx(struct htp_ops_context * octx);
+int op_matmul_qkv(struct htp_ops_context * octx);
+int op_matmul_ffn(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
 int op_sum_rows(struct htp_ops_context * octx);
@@ -140,6 +141,5 @@ int op_solve_tri(struct htp_ops_context * octx);
 int op_gated_delta_net(struct htp_ops_context * octx);
 int op_pad(struct htp_ops_context * octx);
 int op_im2col(struct htp_ops_context * octx);
-int op_allreduce(struct htp_ops_context * octx);
 
 #endif /* HTP_CTX_H */

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

@@ -43,6 +43,13 @@ enum htp_data_type {
 
 
 
+// Mask to enable various stages of the Ops.
+// Used for debugging and profiling.
+enum htp_op_stage {
+    HTP_OPSTAGE_QUEUE    = (1 << 0),  // Enable Queueing (ie calls into NPU)
+    HTP_OPSTAGE_COMPUTE  = (1 << 1),  // Enable Compute
+};
+
 // Do not reorder first 4 (used as an index)
 enum htp_op_code {
     HTP_OP_MUL = 0,
@@ -51,7 +58,8 @@ enum htp_op_code {
     HTP_OP_DIV = 3,
     HTP_OP_MUL_MAT,
     HTP_OP_MUL_MAT_ID,
-    HTP_OP_MUL_MAT_NX,
+    HTP_OP_MUL_MAT_QKV,
+    HTP_OP_MUL_MAT_FFN,
     HTP_OP_MUL_MAT_ADD,
     HTP_OP_RMS_NORM,
     HTP_OP_RMS_NORM_MUL,
@@ -91,15 +99,12 @@ enum htp_op_code {
     HTP_OP_CONCAT,
     HTP_OP_CLAMP,
     HTP_OP_IM2COL,
-    HTP_OP_FENCE,
-    HTP_OP_ALLREDUCE,
-    HTP_OP_ALLREDUCE_ADD,
 
     HTP_OP_INVALID
 };
 
 #define HTP_OP_MAX_DIMS    4    // aka GGML_MAX_DIMS
-#define HTP_OP_MAX_INPUTS  10   // aka GGML_MAX_SRCS
+#define HTP_OP_MAX_INPUTS  6    // aka GGML_MAX_SRCS
 #define HTP_OP_MAX_OUTPUTS 4
 #define HTP_OP_MAX_PARAMS  16   // aka GGML_MAX_OP_PARAMS
 #define HTP_OP_MAX_KERN_PARAMS 32
@@ -107,16 +112,13 @@ enum htp_op_code {
 #define HTP_OP_MAX_BUFS    16
 #define HTP_OP_MAX_TENSORS 8192 // must stay under 64K (uint16)
 
-#define HTP_FENCE_TIMEOUT  (1000000000ULL)
-
 #define HTP_OP_MAX_VMEM_DEFAULT (3355443200u)
 
 #define HTP_MMAP_MAX_VMEM  (2147483648u)
 
 enum htp_tensor_flags {
-    HTP_TENSOR_WEIGHT  = (1U << 0), // Tensor buffer model weight data (not compute)
-    HTP_TENSOR_REPACK  = (1U << 1), // Tensor is in repacked tiled format
-    HTP_TENSOR_FENCE   = (1U << 2)  // Tensor is synchronization fence (explicitly managed)
+    HTP_TENSOR_COMPUTE = (1U << 0), // Tensor buffer temporal compute data (not weights)
+    HTP_TENSOR_DIRTY   = (1U << 1)  // Tensor buffer is dirty and needs to be flushed
 };
 
 // Tensor descriptor
@@ -173,7 +175,6 @@ enum htp_trace_event_id {
     HTP_TRACE_EVT_L2FLUSH             = 1,
     HTP_TRACE_EVT_INIT                = 2,
     HTP_TRACE_EVT_BUFF                = 3,
-    HTP_TRACE_EVT_FENCE               = 4,
 
     HTP_TRACE_EVT_HVX_COMP            = 20,
     HTP_TRACE_EVT_HVX_A_QUANT         = 21,
@@ -214,7 +215,6 @@ struct htp_opbatch_req {
     uint32_t n_ops;       // Number of ops
     uint32_t n_traces;    // Number of trace descriptors per thread
     uint32_t pad;         // unused
-    uint64_t seq;         // Sequence number
     // struct htp_buf_desc  bufs[];    -- dspqueue buf 0
     // struct htp_tensor    tensors[]; -- dspqueue buf 0
     // struct htp_op_desc   ops[];     -- dspqueue buf 0
@@ -231,7 +231,6 @@ struct htp_opbatch_rsp {
     uint32_t pad;            // align to 8 bytes
     uint64_t cycles_start;   // Start cycle counter
     uint64_t cycles_stop;    // Stop cycle counter
-    uint64_t seq;            // Sequence number
     // struct htp_prof_desc profs[];  -- dspqueue buf 0
 };
 

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

@@ -79,14 +79,7 @@ void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * co
 
     for (uint32_t i = 0; i < n; i++) {
         const struct htp_tensor * t = tensors[i];
-        if (!t || (t->flags & (HTP_TENSOR_WEIGHT | HTP_TENSOR_FENCE))) {
-            continue;
-        }
-
-        if (t->size <= HEX_L2_FLUSH_IL_THRESHOLD) {
-            hex_l2flush((void *) (uintptr_t) t->data, t->size);
-            continue;
-        }
+        if (!t) continue;
 
         uint32_t t_start = t->data;
         uint32_t t_end   = t_start + t->size;
@@ -249,7 +242,7 @@ void htp_tensor_flush_all(struct htp_context * ctx, const struct htp_tensor * co
 
     for (uint32_t i = 0; i < n; i++) {
         const struct htp_tensor * t = tensors[i];
-        if (t && !(t->flags & (HTP_TENSOR_WEIGHT | HTP_TENSOR_FENCE)) && is_tensor_dirty(ctx, t)) {
+        if (t && (t->flags & HTP_TENSOR_COMPUTE) && is_tensor_dirty(ctx, t)) {
             dirty_tensors[n_dirty++] = t;
             total_dirty += t->size;
         }

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

@@ -13,15 +13,6 @@ static inline uint32_t * htp_tensor_flags(const struct htp_tensor * t) {
     return (uint32_t *) &t->flags;
 }
 
-static inline uint32_t htp_tensor_get_row_size(int type, uint32_t ne00) {
-    switch (type) {
-        case HTP_TYPE_F32:  return ne00 * 4;
-        case HTP_TYPE_F16:  return ne00 * 2;
-        case HTP_TYPE_Q8_0: return (ne00 / 32) * 34;
-        default:            return 0;
-    }
-}
-
 struct htp_context;
 void htp_tensor_flush_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n);
 void htp_tensor_dirty_all(struct htp_context * ctx, const struct htp_tensor * const * tensors, uint32_t n);

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

@@ -17,9 +17,9 @@
 
 #define hvx_arith_loop_body(dst_type, src0_type, src1_type, elem_size, vec_store, vec_op) \
     do {                                                                       \
-        dst_type * vdst  = (dst_type *) dst;                                   \
-        src0_type * vsrc0 = (src0_type *) src0;                                \
-        src1_type * vsrc1 = (src1_type *) src1;                                \
+        dst_type * restrict vdst  = (dst_type *) dst;                          \
+        src0_type * restrict vsrc0 = (src0_type *) src0;                       \
+        src1_type * restrict vsrc1 = (src1_type *) src1;                       \
                                                                                \
         const uint32_t epv  = 128 / (elem_size);                               \
         const uint32_t nvec = n / epv;                                         \
@@ -57,40 +57,40 @@
 
 // Generic macro to define alignment permutations for an op
 #define DEFINE_HVX_BINARY_OP_VARIANTS(OP_NAME, OP_MACRO, ELEM_TYPE) \
-static inline void OP_NAME##_aaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_aaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
     hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
-static inline void OP_NAME##_aau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_aau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src0 % 128 == 0); \
     hvx_arith_loop_body(HVX_Vector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
-static inline void OP_NAME##_aua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_aua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
     hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
-static inline void OP_NAME##_auu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_auu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) dst % 128 == 0); \
     hvx_arith_loop_body(HVX_Vector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_a, OP_MACRO); \
 } \
-static inline void OP_NAME##_uaa(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_uaa(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
     assert((uintptr_t) src1 % 128 == 0); \
     hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
-static inline void OP_NAME##_uau(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_uau(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src0 % 128 == 0); \
     hvx_arith_loop_body(HVX_UVector, HVX_Vector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
-static inline void OP_NAME##_uua(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_uua(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     assert((uintptr_t) src1 % 128 == 0); \
     hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_Vector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
-static inline void OP_NAME##_uuu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
+static inline void OP_NAME##_uuu(uint8_t * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
     hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \
 

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

@@ -1,165 +0,0 @@
-#ifndef HVX_QUANT_H
-#define HVX_QUANT_H
-
-#include <math.h>
-#include <stdint.h>
-#include <string.h>
-
-#include "hvx-arith.h"
-#include "hvx-base.h"
-#include "hvx-reduce.h"
-#include "hvx-repl.h"
-#include "hvx-utils.h"
-
-#ifndef GGML_COMMON_DECL_C
-#define GGML_COMMON_DECL_C
-#endif
-#include "ggml-common.h"
-#include "ggml-impl.h"
-
-static inline void hvx_quantize_row_q8_0_f32(void * restrict dst_ptr, const float * restrict src_ptr, int n) {
-    const int nb = n / QK8_0;
-    block_q8_0 * dst = (block_q8_0 *) dst_ptr;
-    HVX_Vector zero = Q6_V_vzero();
-
-    int i = 0;
-    for (; i + 3 < nb; i += 4) {
-        HVX_Vector * vx = (HVX_Vector *) (src_ptr + i * QK8_0);
-
-        HVX_Vector vmax0_sf = hvx_vec_reduce_max_f32(hvx_vec_abs_f32(vx[0]));
-        HVX_Vector vmax1_sf = hvx_vec_reduce_max_f32(hvx_vec_abs_f32(vx[1]));
-        HVX_Vector vmax2_sf = hvx_vec_reduce_max_f32(hvx_vec_abs_f32(vx[2]));
-        HVX_Vector vmax3_sf = hvx_vec_reduce_max_f32(hvx_vec_abs_f32(vx[3]));
-
-        HVX_Vector vx0_qf = Q6_Vqf32_vsub_VsfVsf(vx[0], zero);
-        HVX_Vector vx1_qf = Q6_Vqf32_vsub_VsfVsf(vx[1], zero);
-        HVX_Vector vx2_qf = Q6_Vqf32_vsub_VsfVsf(vx[2], zero);
-        HVX_Vector vx3_qf = Q6_Vqf32_vsub_VsfVsf(vx[3], zero);
-
-        HVX_Vector vmax0_qf = Q6_Vqf32_vsub_VsfVsf(vmax0_sf, zero);
-        HVX_Vector vmax1_qf = Q6_Vqf32_vsub_VsfVsf(vmax1_sf, zero);
-        HVX_Vector vmax2_qf = Q6_Vqf32_vsub_VsfVsf(vmax2_sf, zero);
-        HVX_Vector vmax3_qf = Q6_Vqf32_vsub_VsfVsf(vmax3_sf, zero);
-
-        HVX_Vector vmax01_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vmax1_qf, vmax0_qf)));
-        HVX_Vector vmax23_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vmax3_qf, vmax2_qf)));
-
-        HVX_Vector vx01_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vx1_qf, vx0_qf)));
-        HVX_Vector vx23_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(vx3_qf, vx2_qf)));
-
-        HVX_Vector vd01_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax01_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
-        HVX_Vector vd23_qf16 = Q6_Vqf16_vmpy_VhfVhf(vmax23_hf, Q6_Vh_vsplat_R(0x2008));  // 1.0 / 127.0
-        HVX_Vector vd01_hf   = Q6_Vhf_equals_Vqf16(vd01_qf16);
-        HVX_Vector vd23_hf   = Q6_Vhf_equals_Vqf16(vd23_qf16);
-
-        HVX_Vector vd01_inv_hf = hvx_vec_inverse_f16(vd01_hf);
-        HVX_Vector vd23_inv_hf = hvx_vec_inverse_f16(vd23_hf);
-        vx01_hf              = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(vx01_hf, vd01_inv_hf));
-        vx23_hf              = Q6_Vhf_equals_Vqf16(Q6_Vqf16_vmpy_VhfVhf(vx23_hf, vd23_inv_hf));
-
-        HVX_Vector vx01_i16 = hvx_vec_i16_from_hf_rnd_sat(vx01_hf);
-        HVX_Vector vx23_i16 = hvx_vec_i16_from_hf_rnd_sat(vx23_hf);
-        HVX_Vector vx_i8    = Q6_Vb_vpack_VhVh_sat(vx23_i16, vx01_i16);
-
-        hvx_vec_store_u(&dst[i + 0].d, 2, vd01_hf);
-        hvx_vec_store_u(dst[i + 0].qs, 32, vx_i8);
-
-        hvx_vec_store_u(&dst[i + 1].d, 2, Q6_V_vror_VR(vd01_hf, 64));
-        hvx_vec_store_u(dst[i + 1].qs, 32, Q6_V_vror_VR(vx_i8, 32));
-
-        hvx_vec_store_u(&dst[i + 2].d, 2, vd23_hf);
-        hvx_vec_store_u(dst[i + 2].qs, 32, Q6_V_vror_VR(vx_i8, 64));
-
-        hvx_vec_store_u(&dst[i + 3].d, 2, Q6_V_vror_VR(vd23_hf, 64));
-        hvx_vec_store_u(dst[i + 3].qs, 32, Q6_V_vror_VR(vx_i8, 96));
-    }
-
-    for (; i < nb; i++) {
-        const float * block_src = src_ptr + i * QK8_0;
-        HVX_Vector vx = *(const HVX_UVector *) block_src;
-        HVX_Vector v_abs = hvx_vec_abs_f32(vx);
-        HVX_Vector v_max = hvx_vec_reduce_max_f32(v_abs);
-        float amax = hvx_vec_get_f32(v_max);
-
-        const float d = amax / 127.0f;
-        const float id = d ? (1.0f / d) : 0.0f;
-        dst[i].d = GGML_FP32_TO_FP16(d);
-
-        HVX_Vector vid = hvx_vec_splat_f32(id);
-        HVX_Vector v_scaled = hvx_vec_mul_f32_f32(vx, vid);
-        HVX_Vector v_scaled_qf = Q6_Vqf32_vsub_VsfVsf(v_scaled, zero);
-        HVX_Vector v_scaled_hf = Q6_Vh_vdeal_Vh(Q6_Vhf_equals_Wqf32(Q6_W_vcombine_VV(zero, v_scaled_qf)));
-        HVX_Vector v_i16 = hvx_vec_i16_from_hf_rnd_sat(v_scaled_hf);
-        HVX_Vector v_i8  = Q6_Vb_vpack_VhVh_sat(zero, v_i16);
-
-        hvx_vec_store_u(dst[i].qs, 32, v_i8);
-    }
-}
-
-static inline void hvx_dequantize_row_q8_0_f32(float * restrict dst_ptr, const void * restrict src_ptr, int n) {
-    const int nb = n / QK8_0;
-    const block_q8_0 * src = (const block_q8_0 *) src_ptr;
-
-    for (int i = 0; i < nb; i++) {
-        HVX_Vector vd_f16     = Q6_Vh_vsplat_R(*(const int16_t *) &src[i].d);
-        HVX_VectorPair vp_f32 = hvx_vec_f16_to_f32(vd_f16);
-        HVX_Vector vd         = Q6_V_lo_W(vp_f32);
-
-        HVX_Vector vq_i8 = *(const HVX_UVector *) src[i].qs;
-
-        HVX_VectorPair p16   = Q6_Wh_vunpack_Vb(vq_i8);
-        HVX_Vector     v_i16 = Q6_V_lo_W(p16);
-        HVX_VectorPair p32   = Q6_Ww_vunpack_Vh(v_i16);
-        HVX_Vector     v_i32 = Q6_V_lo_W(p32);
-
-        HVX_Vector v_f32 = Q6_Vsf_equals_Vw(v_i32);
-        HVX_Vector res   = hvx_vec_mul_f32_f32(v_f32, vd);
-
-        float * block_dst = dst_ptr + i * QK8_0;
-        hvx_vmem(block_dst) = res;
-    }
-}
-
-static inline void hvx_dequantize_row_q8_0_f16(__fp16 * restrict dst_ptr, const void * restrict src_ptr, int n) {
-    const int nb = n / QK8_0;
-    const block_q8_0 * src = (const block_q8_0 *) src_ptr;
-
-    for (int i = nb - 1; i >= 0; i--) {
-        HVX_Vector vd_f16     = Q6_Vh_vsplat_R(*(const int16_t *) &src[i].d);
-        HVX_VectorPair vp_f32 = hvx_vec_f16_to_f32(vd_f16);
-        HVX_Vector vd         = Q6_V_lo_W(vp_f32);
-
-        HVX_Vector vq_i8 = *(const HVX_UVector *) src[i].qs;
-
-        HVX_VectorPair p16   = Q6_Wh_vunpack_Vb(vq_i8);
-        HVX_Vector     v_i16 = Q6_V_lo_W(p16);
-        HVX_VectorPair p32   = Q6_Ww_vunpack_Vh(v_i16);
-        HVX_Vector     v_i32 = Q6_V_lo_W(p32);
-
-        HVX_Vector v_f32 = Q6_Vsf_equals_Vw(v_i32);
-        HVX_Vector res_f32 = hvx_vec_mul_f32_f32(v_f32, vd);
-
-        HVX_Vector res_f16 = hvx_vec_f32_to_f16(res_f32, Q6_V_vzero());
-
-        __fp16 * block_dst = dst_ptr + i * QK8_0;
-        hvx_vec_store_u(block_dst, QK8_0 * sizeof(__fp16), res_f16);
-    }
-}
-
-static inline void hvx_dequantize_row_f16_f32(float * restrict dst_ptr, const void * restrict src_ptr, int n) {
-    const int nb = n / 32;
-    const _Float16 * src = (const _Float16 *) src_ptr;
-
-    for (int i = 0; i < nb; i++) {
-        HVX_Vector v_f16 = *(const HVX_UVector *) (src + i * 32);
-        HVX_VectorPair vp_f32 = hvx_vec_f16_to_f32(v_f16);
-        HVX_Vector res = Q6_V_lo_W(vp_f32);
-
-        float * block_dst = dst_ptr + i * 32;
-        hvx_vmem(block_dst) = res;
-    }
-}
-
-
-
-#endif // HVX_QUANT_H

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

@@ -18,7 +18,6 @@
 #include <qurt_memory.h>
 #include <remote.h>
 #include <string.h>
-#include <stdatomic.h>
 
 #include "hex-utils.h"
 #include "hex-dma.h"
@@ -33,7 +32,6 @@
 #include "htp_iface.h"
 #include "work-queue.h"
 #include "hex-profile.h"
-#include "allreduce-ops.h"
 
 #define HMX_QUEUE_CAPACITY     16
 #define HMX_QUEUE_STACK_SIZE   16384
@@ -48,36 +46,6 @@ struct htp_handle {
     struct htp_context * ctx;
 };
 
-static inline void * htp_mmap(uint32_t fd, uint32_t size) {
-    void * va = (void *)-1;
-    for (int retry = 0; retry < 2; retry++) {
-#if __HVX_ARCH__ > 73
-        va = HAP_mmap2(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
-#else
-        if (size > HTP_MMAP_MAX_VMEM) {
-            FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) size);
-            abort();
-        }
-        va = HAP_mmap(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
-#endif
-        if (va != (void *)-1 && va != NULL) {
-            return va;
-        }
-        if (retry == 0) {
-            FARF(HIGH, "mmap failed first try (va %p fd %u size %u), retrying...", va, fd, size);
-        }
-    }
-    return NULL;
-}
-
-static inline void htp_munmap(void * va, uint32_t size) {
-#if __HVX_ARCH__ > 73
-    HAP_munmap2(va, size);
-#else
-    HAP_munmap(va, size);
-#endif
-}
-
 AEEResult htp_iface_open(const char * uri, remote_handle64 * handle) {
     (void) uri;
     struct htp_handle * h = calloc(1, sizeof(*h));
@@ -159,7 +127,11 @@ AEEResult htp_iface_close(remote_handle64 handle) {
         // release the mmaps (if any)
         for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) {
             if (ctx->mmap[i].size) {
-                htp_munmap((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#if __HVX_ARCH__ > 73
+                HAP_munmap2((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#else
+                HAP_munmap((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#endif
                 ctx->mmap[i].size = 0;
                 ctx->mmap[i].base = NULL;
                 ctx->mmap[i].fd   = -1;
@@ -203,9 +175,18 @@ AEEResult htp_iface_mmap(remote_handle64 handle, uint32_t fd, uint32_t size) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
             FARF(HIGH, "mmap : fd %u size %u", fd, size);
-            void *va = htp_mmap(fd, size);
-            if (va == NULL) {
-                FARF(ERROR, "mmap failed : fd %u size %u", fd, (uint32_t) size);
+#if __HVX_ARCH__ > 73
+            void *va = HAP_mmap2(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#else
+            if (size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#endif
+            if (va == (void*)-1) {
+                FARF(ERROR, "mmap failed : va %p fd %u size %u", va, fd, (uint32_t) size);
                 return AEE_EFAILED;
             }
 
@@ -231,7 +212,11 @@ AEEResult htp_iface_munmap(remote_handle64 handle, uint32 fd) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (fd < 0 || m->fd == fd) {
             FARF(HIGH, "unmmap : base %p fd %u size %u", (void*) m->base, m->fd, (uint32_t) m->size);
-            htp_munmap((void *) m->base, m->size);
+#if __HVX_ARCH__ > 73
+            HAP_munmap2((void *) m->base, m->size);
+#else
+            HAP_munmap((void *) m->base, m->size);
+#endif
             m->size   = 0;
             m->base   = NULL;
             m->fd     = -1;
@@ -243,7 +228,7 @@ AEEResult htp_iface_munmap(remote_handle64 handle, uint32 fd) {
 
 static void vtcm_acquire(struct htp_context * ctx) {
     if (!ctx->vtcm_valid) {
-        int err = HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 10000000u);
+        int err = HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 1000000u);
         if (err != 0) {
             FARF(ERROR, "ggml-hex: failed to acquire VTCM: 0x%08x", (unsigned)err);
             abort();
@@ -707,45 +692,8 @@ static inline void profile_stop(uint32_t mode, struct profile_data * d) {
     }
 }
 
-static int op_fence(struct htp_ops_context * octx) {
-    struct htp_context *ctx = octx->ctx;
-    struct htp_thread_trace * tr = &ctx->trace[0];
-    const uint32_t seq = (uint32_t) octx->op_params[0];
-
-    htp_trace_event_start(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
-
-    const struct htp_tensor * sync = octx->src[0];
-    atomic_uint * sync_fence = (atomic_uint *) sync->data;
-    uint64_t spins = 0;
-    while (1) {
-        Q6_dccleaninva_A((void *) sync_fence);
-        asm volatile ("syncht" : : : "memory");
-        uint32_t val = atomic_load(&sync_fence[0]);
-        if ((int32_t)(val - seq) >= 0) {
-            break;
-        }
-        if (++spins > HTP_FENCE_TIMEOUT) {
-            FARF(ERROR, "ggml-hex: sync-wait TIMEOUT : fence %p spins %llu seq %u\n", sync_fence, spins, seq);
-            break;
-        }
-        hex_pause();
-    }
-
-    htp_trace_event_stop(tr, HTP_TRACE_EVT_FENCE, (uint16_t) seq);
-
-    FARF(HIGH, "ggml-hex: sync-done : fence %p spins %llu seq %u\n", sync_fence, spins, seq);
-    return HTP_STATUS_OK;
-}
-
 static int execute_op(struct htp_ops_context * octx) {
     switch (octx->op) {
-        case HTP_OP_FENCE:
-            return op_fence(octx);
-
-        case HTP_OP_ALLREDUCE:
-        case HTP_OP_ALLREDUCE_ADD:
-            return op_allreduce(octx);
-
         case HTP_OP_MUL_MAT:
         case HTP_OP_MUL_MAT_ADD:
             return op_matmul(octx);
@@ -753,8 +701,11 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_MUL_MAT_ID:
             return op_matmul_id(octx);
 
-        case HTP_OP_MUL_MAT_NX:
-            return op_matmul_nx(octx);
+        case HTP_OP_MUL_MAT_QKV:
+            return op_matmul_qkv(octx);
+
+        case HTP_OP_MUL_MAT_FFN:
+            return op_matmul_ffn(octx);
 
         case HTP_OP_MUL:
         case HTP_OP_ADD:
@@ -867,8 +818,12 @@ static inline bool reuse_buf(struct htp_context *ctx, uint32_t *m_reuse, struct
 
 static inline void drop_mmap(struct htp_context *ctx, struct htp_mmap *m) {
     if (m->size) {
-        FARF(ALWAYS, "unmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
-        htp_munmap((void *) m->base, m->size);
+        FARF(HIGH, "unmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
+#if __HVX_ARCH__ > 73
+        HAP_munmap2((void *) m->base, m->size);
+#else
+        HAP_munmap((void *) m->base, m->size);
+#endif
         m->size = 0;
         m->base = 0;
         m->fd   = -1;
@@ -882,9 +837,18 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
     for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
-            void *va = htp_mmap(b->fd, b->size);
-            if (va == NULL) {
-                FARF(ERROR, "mmap failed : fd %u size %u", b->fd, (uint32_t) b->size);
+#if __HVX_ARCH__ > 73
+            void *va = HAP_mmap2(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#else
+            if (b->size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) b->size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#endif
+            if (va == (void*)-1) {
+                FARF(ERROR, "mmap failed : va %p fd %u size %u", va, b->fd, (uint32_t) b->size);
                 abort(); // can't do much else at this point
             }
 
@@ -892,13 +856,10 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
             m->fd     = b->fd;
             m->size   = b->size;
 
-            FARF(ALWAYS, "mmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
+            FARF(HIGH, "mmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
             return;
         }
     }
-
-    FARF(ERROR, "mmap failed : exceeded mapping capacity limit of %u", HTP_MAX_MMAPS);
-    abort();
 }
 
 static void prep_op_bufs(struct htp_context *ctx, struct htp_buf_desc *bufs, uint32_t n_bufs) {
@@ -1120,7 +1081,6 @@ static void process_opbatch(struct htp_context * ctx, const struct htp_opbatch_r
     rsp.usecs        = batch_prof.usecs;
     rsp.cycles_start = batch_prof.cycles_start;
     rsp.cycles_stop  = batch_prof.cycles_stop;
-    rsp.seq          = req->seq;
 
     if (ctx->profiler == HTP_PROF_TRACE) {
         for (int t = 0; t <= HTP_MAX_NTHREADS; t++) {

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

@@ -88,7 +88,6 @@ struct htp_mm_kernel_params {
     int32_t  vtcm_src2_size;     // src2 scratchpad size in VTCM (fused only)
     int32_t  vtcm_src3_size;     // src3 scratchpad size in VTCM (fused only)
     int32_t  vtcm_dst_size;      // dst scratchpad size in VTCM
-    int32_t  n_weights;          // Number of weights for fused NX
 
     // Precomputed division values
     struct fastdiv_values div_ne12_ne1;
@@ -464,7 +463,8 @@ static inline void htp_mm_hvx_vtcm_layout_build(
     size_t src2_row_size,
     uint32_t n_prefetch,
     bool is_matmul_id,
-    bool is_fused_nx
+    bool is_fused_qkv,
+    bool is_fused_ffn
 ) {
     size_t src0_sz = 0;
     size_t src1_sz = 0;
@@ -476,33 +476,44 @@ static inline void htp_mm_hvx_vtcm_layout_build(
                             wtype == HTP_TYPE_Q8_0 || wtype == HTP_TYPE_IQ4_NL ||
                             wtype == HTP_TYPE_MXFP4);
 
-    if (is_fused_nx) {
+    if (is_fused_qkv || is_fused_ffn) {
         const size_t src0_row_size_padded = hex_round_up(src0_row_size, 128);
         const size_t quant_scratch_size = hex_round_up(ne10 * sizeof(float), QK_Q8_0_TILED * sizeof(float)) * n_threads;
 
-        size_t weight_sz_per_thread = 0;
+        size_t src0_sz_per_thread = 0;
+        size_t src2_sz_per_thread = 0;
+        size_t src3_sz_per_thread = 0;
 
         if (is_repack) {
             uint32_t aligned_tile_size = htp_mm_get_weight_aligned_tile_size(wtype);
             uint32_t n_k_tiles = hex_round_up(ne10, 32) / 32;
             uint32_t tile_row_size = n_k_tiles * aligned_tile_size;
 
-            weight_sz_per_thread = hex_round_up(n_prefetch * tile_row_size, 128);
+            src0_sz_per_thread = hex_round_up(n_prefetch * tile_row_size, 128);
+            src2_sz_per_thread = hex_round_up(n_prefetch * tile_row_size, 128);
+            if (is_fused_qkv) {
+                src3_sz_per_thread = hex_round_up(n_prefetch * tile_row_size, 128);
+            }
         } else {
-            weight_sz_per_thread = hex_round_up(n_prefetch * src0_row_size_padded, 128);
+            src0_sz_per_thread = hex_round_up(n_prefetch * src0_row_size_padded, 128);
+            src2_sz_per_thread = hex_round_up(n_prefetch * src0_row_size_padded, 128);
+            if (is_fused_qkv) {
+                src3_sz_per_thread = hex_round_up(n_prefetch * src0_row_size_padded, 128);
+            }
         }
 
-        size_t flat_act_row_size  = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10)  : htp_mm_q8_0_flat_row_size(ne10);
-        size_t tiled_act_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
+        size_t flat_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10) : htp_mm_q8_0_flat_row_size(ne10);
+        size_t tiled_src1_row_size = (wtype == HTP_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
 
-        size_t act_sz = (kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT)
-            ? hex_round_up(flat_act_row_size  * src1_nrows, 128)
-            : hex_round_up(tiled_act_row_size * src1_nrows, 128);
+        if (kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT) {
+            src1_sz = hex_round_up(flat_src1_row_size * src1_nrows, 128);
+        } else {
+            src1_sz = hex_round_up(tiled_src1_row_size * src1_nrows, 128);
+        }
 
-        src0_sz = weight_sz_per_thread * n_threads; // shared single-weight prefetch buffer
-        src1_sz = act_sz;                           // quantized activation buffer
-        src2_sz = 0;
-        src3_sz = 0;
+        src0_sz = src0_sz_per_thread * n_threads;
+        src2_sz = src2_sz_per_thread * n_threads;
+        src3_sz = src3_sz_per_thread * n_threads;
         dst_sz  = quant_scratch_size;
     } else if (is_matmul_id) {
         const size_t src0_row_size_padded = htp_mm_round_up(src0_row_size, 128);
@@ -605,8 +616,8 @@ static inline void htp_mm_hvx_vtcm_layout_build(
     }
 
     size_t off = 0;
-    VTCM_LAYOUT_ALLOC(off, off_src0, src0_sz);
     VTCM_LAYOUT_ALLOC(off, off_src1, src1_sz);
+    VTCM_LAYOUT_ALLOC(off, off_src0, src0_sz);
     VTCM_LAYOUT_ALLOC(off, off_src2, src2_sz);
     VTCM_LAYOUT_ALLOC(off, off_src3, src3_sz);
     VTCM_LAYOUT_ALLOC(off, off_dst,  dst_sz);

ggml/src/ggml-hexagon/htp/set-rows-ops.c

@@ -8,20 +8,14 @@
 #include <math.h>
 #include <string.h>
 
-#include "dma-queue.h"
-#include "work-queue.h"
+#include "hex-dma.h"
 #include "hvx-utils.h"
-#include "hex-utils.h"
-#include "hvx-copy.h"
-#include "hvx-quant.h"
 
 #define GGML_COMMON_DECL_C
 #include "ggml-common.h"
-
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-tensor.h"
-#include "htp/set-rows-ops.h"
+#include "htp-ops.h"
 
 #define set_rows_preamble                      \
     const uint32_t ne00 = octx->src[0]->ne[0]; \
@@ -53,142 +47,116 @@
                                                \
     const uint32_t nr  = ne01;
 
-struct set_rows_context {
+struct htp_set_rows_context {
     struct htp_ops_context * octx;
-    const struct htp_set_rows_kernel_params * kparams;
-    struct htp_set_rows_vtcm_layout vtcm_layout;
-    uint8_t * vtcm_base;
+    struct fastdiv_values div_ne12;
+    struct fastdiv_values div_ne11;
+    uint32_t src0_nrows_per_thread;
 };
 
-#define SET_ROWS_THREAD_DMA_FN(TYPE_NAME, IDX_TYPE, COMPUTE_EXPR)                                                \
-static void set_rows_thread_dma_##TYPE_NAME##_##IDX_TYPE(unsigned int nth, unsigned int ith, void *data) {       \
-    struct set_rows_context * srctx = (struct set_rows_context *)data;                                           \
-    struct htp_ops_context * octx = srctx->octx;                                                                 \
-    const struct htp_set_rows_kernel_params * kparams = srctx->kparams;                                          \
-    set_rows_preamble;                                                                                           \
-    struct htp_thread_trace * tr = &octx->ctx->trace[ith];                                                       \
-    const uint32_t dr  = kparams->tasks_per_thread;                                                              \
-    const uint32_t ir0 = dr * ith;                                                                               \
-    if (ir0 >= kparams->total_tasks) {                                                                           \
-        return;                                                                                                  \
-    }                                                                                                            \
-    const uint32_t ir1 = MIN(ir0 + dr, kparams->total_tasks);                                                    \
-    dma_queue * dma_queue = octx->ctx->dma[ith];                                                                 \
-    const struct htp_set_rows_vtcm_layout * vtcm_layout = &srctx->vtcm_layout;                                   \
-    uint8_t * vtcm_src0 = srctx->vtcm_base + vtcm_layout->off_src0 + ith * vtcm_layout->src0_bytes_per_thread;   \
-    uint8_t * vtcm_dst  = srctx->vtcm_base + vtcm_layout->off_dst  + ith * vtcm_layout->dst_bytes_per_thread;    \
-    const uint32_t src0_row_size = ne00 * sizeof(float);                                                         \
-    const uint32_t dst_row_size  = htp_tensor_get_row_size(octx->dst->type, ne00);                               \
-    const uint32_t nrows_per_thread = ir1 - ir0;                                                                 \
-    const uint32_t total_steps = ne03 * ne02 * nrows_per_thread;                                                 \
-    uint32_t pi_step = 0;                                                                                        \
-    uint32_t pi02 = 0;                                                                                           \
-    uint32_t pi03 = 0;                                                                                           \
-    for (uint32_t step = 0, spad_idx = 0; step < total_steps && spad_idx < 2; ++step, spad_idx++) {              \
-        uint32_t i = ir0 + pi_step;                                                                              \
-        const uintptr_t src0_ptr = octx->src[0]->data + i*nb01 + pi02*nb02 + pi03*nb03;                          \
-        dma_queue_push(dma_queue,                                                                                \
-                       dma_make_ptr((void *)octx->dst->data,                                                     \
-                                    vtcm_dst + spad_idx * vtcm_layout->dst_spad_half_size),                      \
-                       dst_row_size, vtcm_layout->dst_spad_half_size, dst_row_size, 0);                          \
-        dma_queue_push(dma_queue,                                                                                \
-                       dma_make_ptr((void *)(vtcm_src0 + spad_idx * vtcm_layout->src0_spad_half_size),           \
-                                    (const void *)src0_ptr),                                                     \
-                       vtcm_layout->src0_spad_half_size, src0_row_size, src0_row_size, 1);                       \
-        pi_step++;                                                                                               \
-        if (pi_step == nrows_per_thread) {                                                                       \
-            pi_step = 0;                                                                                         \
-            pi02++;                                                                                              \
-            if (pi02 == ne02) {                                                                                  \
-                pi02 = 0;                                                                                        \
-                pi03++;                                                                                          \
-            }                                                                                                    \
-        }                                                                                                        \
-    }                                                                                                            \
-    uint32_t ci_step = 0;                                                                                        \
-    uint32_t ci02 = 0;                                                                                           \
-    uint32_t ci03 = 0;                                                                                           \
-    uint32_t ci11_base = 0;                                                                                      \
-    uint32_t ci12_base = 0;                                                                                      \
-    for (uint32_t step = 0; step < total_steps; ++step) {                                                        \
-        void * dst_spad = (void *) dma_queue_pop(dma_queue).src;                                                 \
-        void * src_spad = (void *) dma_queue_pop(dma_queue).dst;                                                 \
-        uint32_t i = ir0 + ci_step;                                                                              \
-        const uintptr_t src1_addr = octx->src[1]->data + i*nb10 + ci11_base*nb11 + ci12_base*nb12;               \
-        const IDX_TYPE i1 = *(const IDX_TYPE *)src1_addr;                                                        \
-        const bool valid_i1 = ((uint64_t)i1 < (uint64_t)ne1);                                                    \
-        const uint32_t target_i1 = (uint32_t)i1;                                                                 \
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, step);                                                 \
-        if (valid_i1) {                                                                                          \
-            COMPUTE_EXPR;                                                                                        \
-        }                                                                                                        \
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, step);                                                  \
-        if (valid_i1) {                                                                                          \
-            const uintptr_t dst_ptr = octx->dst->data + target_i1*nb1 + ci02*nb2 + ci03*nb3;                     \
-            dma_queue_push(dma_queue,                                                                            \
-                           dma_make_ptr((void *)dst_ptr, (const void *)dst_spad),                                \
-                           dst_row_size, vtcm_layout->dst_spad_half_size, dst_row_size, 1);                      \
-        } else {                                                                                                 \
-            dma_queue_push(dma_queue,                                                                            \
-                           dma_make_ptr((void *)octx->dst->data, (const void *)dst_spad),                        \
-                           dst_row_size, vtcm_layout->dst_spad_half_size, dst_row_size, 0);                      \
-        }                                                                                                        \
-        const uint32_t next_step = step + 2;                                                                     \
-        if (next_step < total_steps) {                                                                           \
-            uint32_t ni = ir0 + pi_step;                                                                         \
-            const uintptr_t psrc0_ptr = octx->src[0]->data + ni*nb01 + pi02*nb02 + pi03*nb03;                    \
-            dma_queue_push(dma_queue,                                                                            \
-                           dma_make_ptr((void *)src_spad, (const void *)psrc0_ptr),                              \
-                           vtcm_layout->src0_spad_half_size, src0_row_size, src0_row_size, 1);                   \
-            pi_step++;                                                                                           \
-            if (pi_step == nrows_per_thread) {                                                                   \
-                pi_step = 0;                                                                                     \
-                pi02++;                                                                                          \
-                if (pi02 == ne02) {                                                                              \
-                    pi02 = 0;                                                                                    \
-                    pi03++;                                                                                      \
-                }                                                                                                \
-            }                                                                                                    \
-        }                                                                                                        \
-        ci_step++;                                                                                               \
-        if (ci_step == nrows_per_thread) {                                                                       \
-            ci_step = 0;                                                                                         \
-            ci02++;                                                                                              \
-            ci11_base++;                                                                                         \
-            if (ci11_base == ne11) {                                                                             \
-                ci11_base = 0;                                                                                   \
-            }                                                                                                    \
-            if (ci02 == ne02) {                                                                                  \
-                ci02 = 0;                                                                                        \
-                ci03++;                                                                                          \
-                ci12_base++;                                                                                     \
-                if (ci12_base == ne12) {                                                                         \
-                    ci12_base = 0;                                                                               \
-                }                                                                                                \
-            }                                                                                                    \
-        }                                                                                                        \
-    }                                                                                                            \
-    dma_queue_flush(dma_queue);                                                                                  \
+static void set_rows_thread_f32_f32(unsigned int nth, unsigned int ith, void *data) {
+    struct htp_set_rows_context * srctx = (struct htp_set_rows_context *)data;
+    struct htp_ops_context * octx = srctx->octx;
+
+    set_rows_preamble;
+
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
+    // parallelize by rows of src0
+    const uint32_t dr  = srctx->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    if (ir0 >= nr) {
+        return;
+    }
+    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
+
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+
+    for (uint32_t i03 = 0; i03 < ne03; ++i03) {
+        for (uint32_t i02 = 0; i02 < ne02; ++i02) {
+            for (uint32_t i = ir0; i < ir1; ++i) {
+                const uint32_t i12 = fastmodulo(i03, ne12, &srctx->div_ne12);
+                const uint32_t i11 = fastmodulo(i02, ne11, &srctx->div_ne11);
+                const uint32_t i10 = i;
+
+                const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
+
+                uint32_t i1 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
+                if (i1 >= ne1) {
+                    // ignore invalid indices
+                    continue;
+                }
+
+                const uintptr_t src0_ptr = octx->src[0]->data + i*nb01 + i02*nb02 + i03*nb03;
+                const uintptr_t dst_ptr  = octx->dst->data  + i1*nb1 + i02*nb2  + i03*nb3;
+
+                // copy row
+                hvx_copy_f32_uu((uint8_t *)dst_ptr, (const uint8_t *)src0_ptr, ne00);
+            }
+        }
+    }
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "set-rows-f32-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
 }
 
-SET_ROWS_THREAD_DMA_FN(f32,  int32_t, { hvx_copy_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, ne00); })
-SET_ROWS_THREAD_DMA_FN(f32,  int64_t, { hvx_copy_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, ne00); })
+static void set_rows_thread_f16_f32(unsigned int nth, unsigned int ith, void *data) {
+    struct htp_set_rows_context * srctx = (struct htp_set_rows_context *)data;
+    struct htp_ops_context * octx = srctx->octx;
 
-SET_ROWS_THREAD_DMA_FN(f16,  int32_t, { hvx_copy_f16_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, ne00); })
-SET_ROWS_THREAD_DMA_FN(f16,  int64_t, { hvx_copy_f16_f32_uu((uint8_t *)dst_spad, (const uint8_t *)src_spad, ne00); })
+    set_rows_preamble;
 
-SET_ROWS_THREAD_DMA_FN(q8_0, int32_t, { hvx_quantize_row_q8_0_f32(dst_spad, (const float *)src_spad, ne00); })
-SET_ROWS_THREAD_DMA_FN(q8_0, int64_t, { hvx_quantize_row_q8_0_f32(dst_spad, (const float *)src_spad, ne00); })
+    uint64_t qt = HAP_perf_get_qtimer_count();
+
+    // parallelize by rows of src0
+    const uint32_t dr  = srctx->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    if (ir0 >= nr) {
+        return;
+    }
+    const uint32_t ir1 = (ir0 + dr < nr) ? (ir0 + dr) : nr;
+
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
+
+    for (uint32_t i03 = 0; i03 < ne03; ++i03) {
+        for (uint32_t i02 = 0; i02 < ne02; ++i02) {
+            for (uint32_t i = ir0; i < ir1; ++i) {
+                const uint32_t i12 = fastmodulo(i03, ne12, &srctx->div_ne12);
+                const uint32_t i11 = fastmodulo(i02, ne11, &srctx->div_ne11);
+                const uint32_t i10 = i;
+
+                const uintptr_t src1_addr = octx->src[1]->data + i10*nb10 + i11*nb11 + i12*nb12;
+
+                uint32_t i1 = is_i32 ? *(int32_t *)src1_addr : *(int64_t *)src1_addr;
+                if (i1 >= ne1) {
+                    // ignore invalid indices
+                    continue;
+                }
+
+                const uint8_t* src0_ptr = (const uint8_t *) octx->src[0]->data + i*nb01 + i02*nb02 + i03*nb03;
+                uint8_t*       dst_ptr  = (uint8_t *)       octx->dst->data  + i1*nb1 + i02*nb2  + i03*nb3;
+
+                hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
+            }
+        }
+    }
+
+    qt = HAP_perf_qtimer_count_to_us(HAP_perf_get_qtimer_count() - qt);
+    FARF(HIGH, "set-rows-f16-f32 %d/%d: %ux%ux%ux%u (%u:%u) x %ux%ux%ux%u -> %ux%ux%ux%u usec %u\n", ith, nth,
+         ne00, ne01, ne02, ne03, ir0, ir1, ne10, ne11, ne12, ne13, ne0, ne1, ne2, ne3, (unsigned) qt);
+}
 
 int op_set_rows(struct htp_ops_context * octx) {
-    const struct htp_set_rows_kernel_params * kparams = (const struct htp_set_rows_kernel_params *)octx->kernel_params;
     set_rows_preamble;
 
+    const uint32_t n_threads = MIN(nr, octx->n_threads);
+
     if (octx->src[0]->type != HTP_TYPE_F32) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    if (octx->dst->type != HTP_TYPE_F32 && octx->dst->type != HTP_TYPE_F16 && octx->dst->type != HTP_TYPE_Q8_0) {
+    if (octx->dst->type != HTP_TYPE_F32 && octx->dst->type != HTP_TYPE_F16) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
@@ -196,27 +164,27 @@ int op_set_rows(struct htp_ops_context * octx) {
         return HTP_STATUS_NO_SUPPORT;
     }
 
-    // l2fetch the src1 (indices) tensor in the main thread
-    hex_l2fetch_block((const void *)octx->src[1]->data, octx->src[1]->ne[3] * octx->src[1]->nb[3]);
-
-    struct set_rows_context srctx;
-    srctx.octx = octx;
-    srctx.kparams = kparams;
-
-    htp_set_rows_vtcm_layout_build(&srctx.vtcm_layout, octx->dst->type, ne00, kparams->n_threads);
-    srctx.vtcm_base = (uint8_t *)octx->ctx->vtcm_base;
-
-    work_queue_func_t q_func = NULL;
-    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);
-
-    switch (octx->dst->type) {
-        case HTP_TYPE_F32:  q_func = is_i32 ? set_rows_thread_dma_f32_int32_t  : set_rows_thread_dma_f32_int64_t;  break;
-        case HTP_TYPE_F16:  q_func = is_i32 ? set_rows_thread_dma_f16_int32_t  : set_rows_thread_dma_f16_int64_t;  break;
-        case HTP_TYPE_Q8_0: q_func = is_i32 ? set_rows_thread_dma_q8_0_int32_t : set_rows_thread_dma_q8_0_int64_t; break;
-        default:            return HTP_STATUS_NO_SUPPORT;
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
     }
 
-    work_queue_run(octx->ctx->work_queue, q_func, &srctx, kparams->n_threads);
+    struct htp_set_rows_context srctx;
+    srctx.octx = octx;
+    srctx.div_ne12 = init_fastdiv_values(ne12);
+    srctx.div_ne11 = init_fastdiv_values(ne11);
+
+    srctx.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
+
+    switch(octx->dst->type) {
+    case HTP_TYPE_F32:
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f32_f32, &srctx, n_threads);
+        break;
+    case HTP_TYPE_F16:
+        worker_pool_run_func(octx->ctx->worker_pool, set_rows_thread_f16_f32, &srctx, n_threads);
+        break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
 
     return HTP_STATUS_OK;
 }

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

@@ -1,74 +0,0 @@
-#ifndef HTP_SET_ROWS_OPS_H
-#define HTP_SET_ROWS_OPS_H
-
-#include "hex-fastdiv.h"
-
-struct htp_set_rows_kernel_params {
-    int32_t  n_threads;
-    int32_t  total_tasks;
-    int32_t  tasks_per_thread;
-    int32_t  vtcm_size;
-
-    // Fastdiv helpers
-    struct fastdiv_values div_ne11;
-    struct fastdiv_values div_ne12;
-    struct fastdiv_values div_tasks_per_thread;
-    struct fastdiv_values div_ne02;
-};
-
-struct htp_set_rows_vtcm_layout {
-    size_t total_bytes;
-    size_t off_src0;
-    size_t off_dst;
-
-    size_t src0_bytes_per_thread;
-    size_t dst_bytes_per_thread;
-
-    size_t src0_spad_half_size;
-    size_t dst_spad_half_size;
-};
-
-static inline void htp_set_rows_vtcm_layout_build(
-    struct htp_set_rows_vtcm_layout * vtcm_layout,
-    int dst_type,
-    uint32_t ne00,
-    uint32_t n_threads) {
-
-    size_t src0_row_size = ne00 * 4;
-    size_t dst_row_size = 0;
-    switch (dst_type) {
-        case 0: // HTP_TYPE_F32
-            dst_row_size = ne00 * 4;
-            break;
-        case 1: // HTP_TYPE_F16
-            dst_row_size = ne00 * 2;
-            break;
-        case 8: // HTP_TYPE_Q8_0
-            dst_row_size = (ne00 / 32) * 34;
-            break;
-        default:
-            dst_row_size = 0;
-            break;
-    }
-
-    size_t src0_row_size_aligned = (src0_row_size + 255) & ~255;
-    size_t dst_row_size_aligned  = (dst_row_size + 255) & ~255;
-
-    vtcm_layout->src0_spad_half_size = src0_row_size_aligned;
-    vtcm_layout->dst_spad_half_size  = dst_row_size_aligned;
-
-    vtcm_layout->src0_bytes_per_thread = src0_row_size_aligned * 2;
-    vtcm_layout->dst_bytes_per_thread  = dst_row_size_aligned * 2;
-
-    vtcm_layout->off_src0 = 0;
-    vtcm_layout->off_dst  = vtcm_layout->off_src0 + vtcm_layout->src0_bytes_per_thread * n_threads;
-    vtcm_layout->total_bytes = vtcm_layout->off_dst + vtcm_layout->dst_bytes_per_thread * n_threads;
-}
-
-#if defined(__cplusplus)
-static_assert(sizeof(struct htp_set_rows_kernel_params) <= 128, "htp_set_rows_kernel_params is too large for kernel_params blob");
-#else
-_Static_assert(sizeof(struct htp_set_rows_kernel_params) <= 128, "htp_set_rows_kernel_params is too large for kernel_params blob");
-#endif
-
-#endif // HTP_SET_ROWS_OPS_H

ggml/src/ggml-metal/CMakeLists.txt

@@ -11,7 +11,6 @@ ggml_add_backend_library(ggml-metal
                          ggml-metal-common.cpp
                          ggml-metal-context.m
                          ggml-metal-ops.cpp
-                         ggml-metal-tuning.cpp
                         )
 
 target_link_libraries(ggml-metal PRIVATE
@@ -25,119 +24,62 @@ if (GGML_METAL_NDEBUG)
 endif()
 
 set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
-set(METALLIB_KERNELS_COMMON     "${CMAKE_CURRENT_SOURCE_DIR}/kernels/common.h")
-set(METALLIB_KERNELS_DEQUANTIZE "${CMAKE_CURRENT_SOURCE_DIR}/kernels/dequantize.h")
-set(METALLIB_KERNELS_QUANTIZE   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/quantize.h")
-
-set(METALLIB_KERNEL_SOURCES
-    kernels/fa.metal
-    kernels/mul_mv.metal
-    kernels/mul_mm.metal
-    kernels/quantize.metal
-    kernels/softmax.metal
-    kernels/norm.metal
-    kernels/unary.metal
-    kernels/binbcast.metal
-    kernels/reduce.metal
-    kernels/tri.metal
-    kernels/ssm.metal
-    kernels/wkv.metal
-    kernels/gated_delta_net.metal
-    kernels/solve_tri.metal
-    kernels/rope.metal
-    kernels/conv.metal
-    kernels/upscale.metal
-    kernels/argsort.metal
-    kernels/pool.metal
-    kernels/misc.metal
-)
-
 if (GGML_METAL_EMBED_LIBRARY)
     enable_language(ASM)
 
     add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
 
     file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
 
-    set(METALLIB_EMBED_ASM_FILES "")
-    foreach(src ${METALLIB_KERNEL_SOURCES})
-        get_filename_component(kind ${src} NAME_WE)
-        # symbol names must be valid C identifiers ('-' is not allowed)
-        string(REPLACE "-" "_" kind_sym ${kind})
+    # merge ggml-common.h and ggml-metal.metal into a single file
+    set(METALLIB_EMBED_ASM        "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+    set(METALLIB_SOURCE_EMBED     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
 
-        set(SRC   "${CMAKE_CURRENT_SOURCE_DIR}/kernels/${kind}.metal")
-        set(EMBED "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.metal")
-        set(ASM   "${CMAKE_CURRENT_BINARY_DIR}/autogenerated/ggml-metal-embed-${kind}.s")
+    add_custom_command(
+        OUTPUT "${METALLIB_EMBED_ASM}"
+        COMMAND echo "Embedding Metal library"
+        COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}"       -e "/__embed_ggml-common.h__/d"         < "${METALLIB_SOURCE}"           > "${METALLIB_SOURCE_EMBED_TMP}"
+        COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${METALLIB_SOURCE_EMBED_TMP}" > "${METALLIB_SOURCE_EMBED}"
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_start"              >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_start:"                    >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo .incbin "\"${METALLIB_SOURCE_EMBED}\""     >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo ".globl _ggml_metallib_end"                >> "${METALLIB_EMBED_ASM}"
+        COMMAND echo "_ggml_metallib_end:"                      >> "${METALLIB_EMBED_ASM}"
+        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
+        COMMENT "Generate assembly for embedded Metal library"
+        VERBATIM
+    )
 
-        # only prepend headers that this source actually includes
-        set(HEADERS_FOR_SRC ${METALLIB_KERNELS_COMMON})
-        file(STRINGS ${SRC} _has_dequantize REGEX "#include \"dequantize\\.h\"")
-        file(STRINGS ${SRC} _has_quantize   REGEX "#include \"quantize\\.h\"")
-        if(_has_dequantize)
-            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_DEQUANTIZE})
-        endif()
-        if(_has_quantize)
-            list(APPEND HEADERS_FOR_SRC ${METALLIB_KERNELS_QUANTIZE})
-        endif()
-
-        add_custom_command(
-            OUTPUT "${ASM}"
-            # Step 1: concatenate shared headers + this kernel source
-            COMMAND cat ${HEADERS_FOR_SRC} ${SRC} > "${EMBED}.tmp1"
-            # Step 2: remove internal #include and #pragma once
-            COMMAND sed -e "/\#include \"common.h\"/d" -e "/\#include \"dequantize.h\"/d" -e "/\#include \"quantize.h\"/d" -e "/\#pragma once/d" < "${EMBED}.tmp1" > "${EMBED}.tmp2"
-            # Step 3: inline ggml-common.h (replacing __embed_ggml-common.h__ sentinel)
-            COMMAND sed -e "/__embed_ggml-common.h__/r ${METALLIB_COMMON}" -e "/__embed_ggml-common.h__/d" < "${EMBED}.tmp2" > "${EMBED}.tmp3"
-            # Step 4: inline ggml-metal-impl.h
-            COMMAND sed -e "/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}" -e "/\#include \"ggml-metal-impl.h\"/d" < "${EMBED}.tmp3" > "${EMBED}"
-            # Step 5: emit an asm chunk with kind-specific start/end symbols
-            #   note: '-' is illegal in C symbols, so we use kind_sym; the macOS
-            #   section name is limited to 16 chars so we keep it shared
-            #   across kinds (__ggml_metallib) and only vary the global symbols.
-            COMMAND echo ".section __DATA,__ggml_metallib"                       >  "${ASM}"
-            COMMAND echo ".globl _ggml_metallib_${kind_sym}_start"               >> "${ASM}"
-            COMMAND echo "_ggml_metallib_${kind_sym}_start:"                     >> "${ASM}"
-            COMMAND echo .incbin "\"${EMBED}\""                                  >> "${ASM}"
-            COMMAND echo ".globl _ggml_metallib_${kind_sym}_end"                 >> "${ASM}"
-            COMMAND echo "_ggml_metallib_${kind_sym}_end:"                       >> "${ASM}"
-            DEPENDS ../ggml-common.h ggml-metal-impl.h
-                    kernels/common.h kernels/dequantize.h kernels/quantize.h
-                    kernels/${kind}.metal
-            COMMENT "Generate embedded Metal library for ${kind}"
-            VERBATIM
-        )
-
-        list(APPEND METALLIB_EMBED_ASM_FILES "${ASM}")
-    endforeach()
-
-    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM_FILES})
+    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
 else()
-    # copy header files to bin directory
+    # copy metal files to bin directory
     configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
+    configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
     configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
 
-    file(MAKE_DIRECTORY "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels")
-    configure_file(kernels/common.h     ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/common.h     COPYONLY)
-    configure_file(kernels/dequantize.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/dequantize.h COPYONLY)
-    configure_file(kernels/quantize.h   ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/quantize.h   COPYONLY)
-
-    foreach(src ${METALLIB_KERNEL_SOURCES})
-        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
-    endforeach()
-
     if (GGML_METAL_SHADER_DEBUG)
-        # note: disabling fast math is needed in order to pass tests/test-backend-ops
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
         # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
         # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
         #       ref: https://github.com/ggml-org/whisper.cpp/issues/1720
         # note: adding -g causes segmentation fault during compile
+        #set(XC_FLAGS -fno-fast-math -fno-inline -g)
         set(XC_FLAGS -fno-fast-math -fno-inline)
     else()
         set(XC_FLAGS -O3)
     endif()
 
+    # Append macOS metal versioning flags
     if (GGML_METAL_MACOSX_VERSION_MIN)
         message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
         list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
@@ -148,46 +90,35 @@ else()
         list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
     endif()
 
-    # Compile each kernel source to .air, then link into default.metallib
-    set(AIR_FILES "")
-    foreach(src ${METALLIB_KERNEL_SOURCES})
-        get_filename_component(name ${src} NAME_WE)
-        set(AIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${name}.air")
-        list(APPEND AIR_FILES ${AIR})
-        add_custom_command(
-            OUTPUT ${AIR}
-            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
-            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
-            COMMENT "Compiling ${src}"
-            VERBATIM
-        )
-    endforeach()
-
     add_custom_command(
         OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
+                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
-        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
-        DEPENDS ${AIR_FILES}
-        COMMENT "Linking Metal kernels into default.metallib"
-    )
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
+        COMMENT "Compiling Metal kernels"
+        )
 
+    # FIXME: only add to the ggml-metal target?
     add_custom_target(
         ggml-metal-lib ALL
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-    )
+        )
 endif() # GGML_METAL_EMBED_LIBRARY
 
 if (NOT GGML_METAL_EMBED_LIBRARY)
     install(
-        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
-        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
-        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
-    )
+        FILES src/ggml-metal/ggml-metal.metal
+        PERMISSIONS
+            OWNER_READ
+            OWNER_WRITE
+            GROUP_READ
+            WORLD_READ
+        DESTINATION ${CMAKE_INSTALL_BINDIR})
 
-    install(
-        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        DESTINATION ${CMAKE_INSTALL_BINDIR}
-    )
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
 endif()

ggml/src/ggml-metal/ggml-metal-context.m

@@ -84,108 +84,106 @@ struct ggml_metal {
 ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
     GGML_LOG_INFO("%s: allocating\n", __func__);
 
-    @autoreleasepool {
 #if TARGET_OS_OSX && !GGML_METAL_NDEBUG
-        // Show all the Metal device instances in the system
-        NSArray * devices = MTLCopyAllDevices();
-        for (id<MTLDevice> device in devices) {
-            GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
-        }
-        [devices release]; // since it was created by a *Copy* C method
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
+    for (id<MTLDevice> device in devices) {
+        GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
+    }
+    [devices release]; // since it was created by a *Copy* C method
 #endif
 
-        // init context
-        ggml_metal_t res = calloc(1, sizeof(struct ggml_metal));
+    // init context
+    ggml_metal_t res = calloc(1, sizeof(struct ggml_metal));
 
-        id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
 
-        GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
+
+    // TODO: would it be better to have one queue for the backend and one queue for the device?
+    //       the graph encoders and async ops would use the backend queue while the sync ops would use the device queue?
+    //res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND]
+    id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
+    if (queue == nil) {
+        GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
+        return NULL;
+    }
+
+    res->dev = dev;
+    res->lib = ggml_metal_device_get_library(dev);
+    if (res->lib == NULL) {
+        GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__);
+        GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__);
+
+        res->lib = ggml_metal_library_init(dev);
+        if (res->lib == NULL) {
+            GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__);
+
+            free(res);
 
-        // TODO: would it be better to have one queue for the backend and one queue for the device?
-        //       the graph encoders and async ops would use the backend queue while the sync ops would use the device queue?
-        //res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND]
-        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
-        if (queue == nil) {
-            GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
             return NULL;
         }
-
-        res->dev = dev;
-        res->lib = ggml_metal_device_get_library(dev);
-        if (res->lib == NULL) {
-            GGML_LOG_WARN("%s: the device does not have a precompiled Metal library - this is unexpected\n", __func__);
-            GGML_LOG_WARN("%s: will try to compile it on the fly\n", __func__);
-
-            res->lib = ggml_metal_library_init(dev);
-            if (res->lib == NULL) {
-                GGML_LOG_ERROR("%s: error: failed to initialize the Metal library\n", __func__);
-
-                free(res);
-
-                return NULL;
-            }
-        }
-
-        res->ev_cpy = ggml_metal_device_event_init(dev);
-
-        const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
-
-        snprintf(res->name, sizeof(res->name), "%s", props_dev->name);
-
-        res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
-
-        res->use_fusion      = getenv("GGML_METAL_FUSION_DISABLE") == nil;
-        res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
-
-        {
-            const char * val = getenv("GGML_METAL_GRAPH_DEBUG");
-            res->debug_graph = val ? atoi(val) : 0;
-        }
-
-        {
-            const char * val = getenv("GGML_METAL_FUSION_DEBUG");
-            res->debug_fusion = val ? atoi(val) : 0;
-        }
-
-        res->use_graph_optimize = true;
-
-        if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) {
-            res->use_graph_optimize = false;
-        }
-
-        memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
-
-        GGML_LOG_INFO("%s: use fusion         = %s\n", __func__, res->use_fusion         ? "true" : "false");
-        GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
-        GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
-
-        res->capture_compute = 0;
-        res->capture_started = false;
-        res->capture_scope = nil;
-
-        {
-            const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
-            if (val) {
-                res->capture_compute = atoi(val);
-            }
-        }
-
-        res->has_error = false;
-
-        res->gf = nil;
-        res->encode_async = nil;
-        for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
-            res->cmd_bufs[i].obj = nil;
-        }
-
-        res->cmd_bufs_ext = [[NSMutableArray alloc] init];
-
-        res->cmd_buf_last = nil;
-
-        res->pipelines_ext = ggml_metal_pipelines_init();
-
-        return res;
     }
+
+    res->ev_cpy = ggml_metal_device_event_init(dev);
+
+    const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
+
+    snprintf(res->name, sizeof(res->name), "%s", props_dev->name);
+
+    res->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
+
+    res->use_fusion      = getenv("GGML_METAL_FUSION_DISABLE") == nil;
+    res->use_concurrency = getenv("GGML_METAL_CONCURRENCY_DISABLE") == nil;
+
+    {
+        const char * val = getenv("GGML_METAL_GRAPH_DEBUG");
+        res->debug_graph = val ? atoi(val) : 0;
+    }
+
+    {
+        const char * val = getenv("GGML_METAL_FUSION_DEBUG");
+        res->debug_fusion = val ? atoi(val) : 0;
+    }
+
+    res->use_graph_optimize = true;
+
+    if (getenv("GGML_METAL_GRAPH_OPTIMIZE_DISABLE") != NULL) {
+        res->use_graph_optimize = false;
+    }
+
+    memset(res->fuse_cnt, 0, sizeof(res->fuse_cnt));
+
+    GGML_LOG_INFO("%s: use fusion         = %s\n", __func__, res->use_fusion         ? "true" : "false");
+    GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
+    GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
+
+    res->capture_compute = 0;
+    res->capture_started = false;
+    res->capture_scope = nil;
+
+    {
+        const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
+        if (val) {
+            res->capture_compute = atoi(val);
+        }
+    }
+
+    res->has_error = false;
+
+    res->gf = nil;
+    res->encode_async = nil;
+    for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
+        res->cmd_bufs[i].obj = nil;
+    }
+
+    res->cmd_bufs_ext = [[NSMutableArray alloc] init];
+
+    res->cmd_buf_last = nil;
+
+    res->pipelines_ext = ggml_metal_pipelines_init();
+
+    return res;
 }
 
 void ggml_metal_free(ggml_metal_t ctx) {

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

@@ -1,7 +1,6 @@
 #include "ggml-metal-device.h"
 
 #include "ggml-metal-impl.h"
-#include "ggml-metal-tuning.h"
 
 #include "ggml-impl.h"
 
@@ -18,10 +17,10 @@ struct ggml_metal_device_deleter {
 
 typedef std::unique_ptr<ggml_metal_device, ggml_metal_device_deleter> ggml_metal_device_ptr;
 
-ggml_metal_device_t ggml_metal_device_get(int device, int n_devices) {
+ggml_metal_device_t ggml_metal_device_get(int device) {
     static std::vector<ggml_metal_device_ptr> devs;
 
-    devs.emplace_back(ggml_metal_device_init(device, n_devices));
+    devs.emplace_back(ggml_metal_device_init(device));
 
     return devs.back().get();
 }
@@ -572,7 +571,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched
     return res;
 }
 
-ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_library_t lib, const ggml_tensor * op, bool tail)  {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_library_t lib, const ggml_tensor * op)  {
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
 
     char base[256];
@@ -580,7 +579,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_me
 
     const int nsg = (ne00 + 31)/32;
 
-    snprintf(base, 256, "kernel_ssm_scan_%s%s", ggml_type_name(op->src[0]->type), tail ? "_tail" : "");
+    snprintf(base, 256, "kernel_ssm_scan_%s", ggml_type_name(op->src[0]->type));
     snprintf(name, 256, "%s_nsg=%d", base, nsg);
 
     ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
@@ -598,27 +597,6 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_me
     return res;
 }
 
-ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan_ssd_mma(ggml_metal_library_t lib, const ggml_tensor * op)  {
-    char base[256];
-    char name[256];
-
-    snprintf(base, 256, "kernel_ssm_scan_ssd_mma_%s", ggml_type_name(op->src[0]->type));
-    snprintf(name, 256, "%s", base);
-
-    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
-    if (!res.pipeline) {
-        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
-    }
-
-    // acs/exp(acs)/state-decay vectors + dtX + SAM rows + two 8x8 tiles per simdgroup
-    res.smem = (3*OP_SSM_SCAN_SSD_CS +
-                OP_SSM_SCAN_SSD_CS*OP_SSM_SCAN_SSD_HD +
-                OP_SSM_SCAN_SSD_NSG*8*OP_SSM_SCAN_SSD_CS +
-                OP_SSM_SCAN_SSD_NSG*2*8*8)*sizeof(float);
-
-    return res;
-}
-
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_library_t lib, const ggml_tensor * op) {
     char base[256];
     char name[256];
@@ -1566,8 +1544,6 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
-        int32_t nqpsg,
-        int32_t ne,
         int32_t nsg,
         int32_t nwg,
         bool    use_kv_f16,
@@ -1583,17 +1559,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
 
     const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);
 
-    char qne_suffix[16] = {0};
-    if (!(nqpsg == 1 && ne == ggml_metal_tuning::fa_vec_baseline_ne(dk, dv))) {
-        snprintf(qne_suffix, sizeof(qne_suffix), "_q%d_ne%d", nqpsg, ne);
-    }
-
-    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d%s",
+    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
             "flash_attn_ext_vec",
             type,
             dk,
-            dv,
-            qne_suffix);
+            dv);
 
     snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
             base,

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

@@ -129,8 +129,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
-struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op, bool tail);
-struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan_ssd_mma  (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net   (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -208,8 +207,6 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
-        int32_t nqpsg,
-        int32_t ne,
         int32_t nsg,
         int32_t nwg,
         bool    use_kv_f16,
@@ -260,12 +257,8 @@ enum ggml_metal_device_id {
     GGML_METAL_DEVICE_M5_ULTRA,
 };
 
-const char * ggml_metal_device_id_token(enum ggml_metal_device_id id);
-
 struct ggml_metal_device_props {
     int device;
-    int device_phys;
-    int device_virt;
     char name[128];
     char desc[128];
 
@@ -284,7 +277,6 @@ struct ggml_metal_device_props {
     bool supports_gpu_family_apple7;
 
     enum ggml_metal_device_id device_id;
-    int gpu_family;
 
     int op_offload_min_batch_size;
 };
@@ -294,10 +286,10 @@ typedef struct ggml_metal_event * ggml_metal_event_t;
 void ggml_metal_event_encode_signal(ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
 void ggml_metal_event_encode_wait  (ggml_metal_event_t ev, ggml_metal_cmd_buf_t cmd_buf);
 
-ggml_metal_device_t ggml_metal_device_init(int device, int n_devices);
+ggml_metal_device_t ggml_metal_device_init(int device);
 void ggml_metal_device_free(ggml_metal_device_t dev);
 
-ggml_metal_device_t ggml_metal_device_get(int device, int n_devices);
+ggml_metal_device_t ggml_metal_device_get(int device);
 
 void * ggml_metal_device_get_obj  (ggml_metal_device_t dev); // id<MTLDevice>
 void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQueue>

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

@@ -158,10 +158,6 @@
 #define OP_SUM_ROWS_NUM_SUM_ROWS 10
 #define OP_SUM_ROWS_NUM_MEAN     11
 
-#define OP_SSM_SCAN_SSD_CS  64 // Metal-specific; Chunk Size; 64 is largest multiple of 8 (simdgroup tile) fitting into 32 KiB Metal threadgroup mem limit (~26.75 KiB shared mem; see smem layout comment in kernel_ssm_scan_ssd_mma_f32)
-#define OP_SSM_SCAN_SSD_HD  64 // Metal-specific; Head Dim the MMA kernel is specialized for (Mamba-2); use_mma gates on d_inner == this
-#define OP_SSM_SCAN_SSD_NSG 4  // Metal-specific; Number of SimdGroups per threadgroup; NSG*32 == threads dispatched per threadgroup
-
 // kernel argument structs
 //
 // - element counters (e.g. ne00) typically use int32_t to reduce register usage
@@ -897,8 +893,6 @@ typedef struct {
     int64_t  n_head;
     int64_t  n_group;
     int64_t  n_seq_tokens;
-    int64_t  n_seq_tokens_total;
-    int64_t  token_offset;
     int64_t  n_seqs;
     int64_t  K;
     uint64_t s_off;

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

@@ -7,7 +7,6 @@
 #include "ggml-metal-impl.h"
 #include "ggml-metal-common.h"
 #include "ggml-metal-device.h"
-#include "ggml-metal-tuning.h"
 
 #include <cassert>
 #include <algorithm>
@@ -1677,7 +1676,6 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
 
     ggml_metal_library_t lib = ctx->lib;
     ggml_metal_encoder_t enc = ctx->enc;
-    const ggml_metal_device_props * props_dev = ggml_metal_device_get_props(ctx->dev);
 
     GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
     GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
@@ -1723,8 +1721,6 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
         /*.n_head       =*/ n_head,
         /*.n_group      =*/ n_group,
         /*.n_seq_tokens =*/ n_seq_tokens,
-        /*.n_seq_tokens_total =*/ n_seq_tokens,
-        /*.token_offset =*/ 0,
         /*.n_seqs       =*/ n_seqs,
         /*.K            =*/ K,
         /*.s_off        =*/ ggml_nelements(op->src[1]) * sizeof(float),
@@ -1754,53 +1750,26 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
         /*.nb0          =*/ nb0,
     };
 
-    constexpr int64_t CHUNK = OP_SSM_SCAN_SSD_CS;
+    auto pipeline = ggml_metal_library_get_pipeline_ssm_scan(lib, op);
 
-    const int64_t snap_reserve = K > 1 ? K : 0; // tokens reserved for sequential kernel rollback snapshots
-    const int64_t mma_tokens = ((n_seq_tokens - snap_reserve) / CHUNK) * CHUNK; // largest multiple of CHUNK that leaves snap_reserve for the tail
-    const bool use_mma =
-        mma_tokens > 0 &&
-        ne30 == 1 && // checks that A tensor is set to scalar decay per head (A shape {1, n_head})
-        props_dev->has_simdgroup_mm && // hardware check for M1 or newer
-        d_state % 8 == 0 && // d_state must be multiple of 8 to align with simdgroup_float 8x8 tiles
-        d_inner == OP_SSM_SCAN_SSD_HD; // mma kernel is specialized for the Mamba-2 head dim; this checks it
+    GGML_ASSERT(d_state <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
-    const auto dispatch = [&](ggml_metal_pipeline_with_params pipeline, int64_t nth, int64_t n_tg_x) {
-        GGML_ASSERT(nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
-        GGML_ASSERT(pipeline.smem <= props_dev->max_theadgroup_memory_size);
+    const size_t smem = pipeline.smem;
 
-        ggml_metal_encoder_set_pipeline(enc, pipeline);
-        ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), 4);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), 5);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[5]), 6);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[6]), 7);
-        ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         8);
-        ggml_metal_encoder_set_threadgroup_memory_size(enc, pipeline.smem, 0);
-        ggml_metal_encoder_dispatch_threadgroups(enc, n_tg_x, n_head, n_seqs, nth, 1, 1);
-    };
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), 2);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), 3);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), 4);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), 5);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[5]), 6);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[6]), 7);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         8);
 
-    if (!use_mma) {
-        dispatch(ggml_metal_library_get_pipeline_ssm_scan(lib, op, false), d_state, d_inner);
-        return 1;
-    }
+    ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
 
-    args.n_seq_tokens = mma_tokens;
-    dispatch(
-        ggml_metal_library_get_pipeline_ssm_scan_ssd_mma(lib, op),
-        OP_SSM_SCAN_SSD_NSG*32,
-        1);
-
-    if (mma_tokens < n_seq_tokens) {
-        ggml_metal_op_concurrency_reset(ctx);
-
-        args.n_seq_tokens = n_seq_tokens - mma_tokens;
-        args.token_offset = mma_tokens;
-        dispatch(ggml_metal_library_get_pipeline_ssm_scan(lib, op, true), d_state, d_inner);
-    }
+    ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
 
     return 1;
 }
@@ -3377,18 +3346,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
     } else {
         // half4x4 kernel
-        auto cfg = ggml_metal_tuning::fa_vec_pick(
-                props_dev->device_id,
-                props_dev->gpu_family,
-                (int) op->src[1]->type,
-                (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
-                ne11, ne01);
-        int nqptg = cfg.Q;                             // queries per threadgroup
+        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
         const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
         const int nhptg = 1;                           // heads per threadgroup
 
         GGML_ASSERT(nqptg <= 32);
-        GGML_ASSERT(nqptg == 1 || nqptg == 2 || nqptg == 4);  // only instantiated Q values
+        GGML_ASSERT(nqptg  % 1  == 0);
         GGML_ASSERT(ncpsg  % 32 == 0);
 
         bool need_sync = false;
@@ -3447,7 +3410,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
         // ne20*(nsg)
         // each simdgroup has a full f32 head vector in shared mem to accumulate results
         //
-#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg)*nqptg)*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg))*(sizeof(float)/2), 16))
 
         int64_t nsg = 1;
 
@@ -3467,12 +3430,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             }
         }
 
-        // fall back to baseline (Q=1) if the tuned config exceeds threadgroup memory
-        if ((size_t) FATTN_SMEM(nsg) > props_dev->max_theadgroup_memory_size) {
-            cfg   = ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20);
-            nqptg = cfg.Q;  // = 1
-        }
-
         const int32_t ns10 = nb11_attn/nb10_attn;
         const int32_t ns20 = nb21_attn/nb20_attn;
 
@@ -3511,7 +3468,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.logit_softcap =*/ logit_softcap,
         };
 
-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nqptg, cfg.NE, nsg, nwg, use_kv_f16, ns10, ns20);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg, use_kv_f16, ns10, ns20);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 

ggml/src/ggml-metal/ggml-metal-tuning.h

@@ -1,77 +0,0 @@
-#pragma once
-
-#include "ggml-metal-device.h"  // enum ggml_metal_device_id
-#include "ggml.h"
-
-#include <cstdint>
-#include <vector>
-
-namespace ggml_metal_tuning {
-
-// FA vec selection buckets. ne01 (query rows) splits decode (==1) from batch (>=2), the
-// batch side refined into {2,3,4,5}: Q>1 reuses one K/V load across rows, so it only pays
-// off once ne01 aligns with Q. ne11 (KV length) is bucketed too, as the Q>1 crossover is
-// head-size dependent (small dk crosses late, large dk wins even at short KV).
-constexpr int FA_VEC_NE11_BUCKETS[] = { 1024, 4096, 16384 };
-constexpr int FA_VEC_NE01_BUCKETS[] = { 2, 3, 4, 5 };
-
-int fa_vec_ne11_bucket(int64_t ne11);
-int fa_vec_ne01_bucket(int64_t ne01);
-
-// NE baked into each (dk,dv) baseline instantiation in kernels/fa.metal.
-// Hand-maintained mirror; keep in sync with those instantiations.
-// The Metal test slice covers every legal config for dk=128 and dk=576.
-int fa_vec_baseline_ne(int dk, int dv);
-
-// Tuned table has two row kinds. Exact rows key a (ne11_b, ne01_b) bucket. Default rows
-// collapse ne11 over one ne01 domain: ne11_b == FA_VEC_NE11_DEFAULT and ne01_b holds the
-// domain. fa_vec_pick tries exact bucket -> domain default -> baseline; short KV
-// (ne11 < FA_VEC_NE11_BUCKETS[0]) always uses baseline.
-constexpr int8_t FA_VEC_NE11_DEFAULT  = -1;
-constexpr int8_t FA_VEC_DOMAIN_DECODE = 0;  // ne01 == 1
-constexpr int8_t FA_VEC_DOMAIN_BATCH  = 1;  // ne01 >= 2
-
-struct fa_vec_key_t {
-    int8_t  device_id;
-    int8_t  dtype;
-    int16_t dk;
-    int16_t dv;
-    int8_t  ne11_b;
-    int8_t  ne01_b;
-};
-
-static_assert(sizeof(fa_vec_key_t) == 8, "fa_vec_key_t must be tightly packed for memcmp");
-
-struct fa_vec_cfg_t {
-    int8_t Q;
-    int8_t NE;
-};
-
-struct fa_vec_entry_t {
-    fa_vec_key_t key;
-    fa_vec_cfg_t cfg;
-};
-
-// legal NE values for a (dk,dv): NL = 32/NE, require (dk/4)%NL==0 && (dv/4)%NL==0.
-// single source shared by the offline tuner and test-backend-ops.
-inline std::vector<int> fa_vec_legal_ne(int dk, int dv) {
-    std::vector<int> r;
-    for (int ne : { 1, 2, 4 }) {
-        const int nl = 32 / ne;
-        if ((dk / 4) % nl == 0 && (dv / 4) % nl == 0) {
-            r.push_back(ne);
-        }
-    }
-    return r;
-}
-
-// test/tune-only override; when set, fa_vec_pick returns it directly.
-void         fa_vec_set_override(fa_vec_cfg_t cfg);
-void         fa_vec_clear_override();
-fa_vec_cfg_t fa_vec_baseline_cfg(int dk, int dv);
-
-// device_id selects a per-SKU row; on a miss, gpu_family (0 if unknown) maps to a representative
-// SKU and the table is retried. No match -> baseline.
-fa_vec_cfg_t fa_vec_pick(enum ggml_metal_device_id device_id, int gpu_family, int dtype, int dk, int dv, int64_t ne11, int64_t ne01);
-
-}  // namespace ggml_metal_tuning

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

@@ -6,7 +6,6 @@
 #include "ggml-metal-device.h"
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
-#include "ggml-metal-tuning.h"
 
 #include <mutex>
 #include <string>
@@ -204,11 +203,6 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     ggml_metal_device_t ctx_dev = (ggml_metal_device_t)buft->device->context;
     ggml_metal_buffer_t res = ggml_metal_buffer_init(ctx_dev, size, shared);
 
-    if (res == NULL) {
-        GGML_LOG_ERROR("%s: failed to allocate Metal buffer of %zu bytes (out of memory)\n", __func__, size);
-        return NULL;
-    }
-
     ggml_backend_buffer_i buf_i = ggml_metal_buffer_is_shared(res)
         ? ggml_backend_metal_buffer_shared_i
         : ggml_backend_metal_buffer_private_i;
@@ -876,55 +870,10 @@ static ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t
     GGML_UNUSED(reg);
 }
 
-// test/tune-only override for the FA vec (Q, NE) selection, reached via proc_address.
-static void ggml_backend_metal_tuning_set_fa_vec_override(int Q, int NE) {
-    ggml_metal_tuning::fa_vec_set_override({ (int8_t) Q, (int8_t) NE });
-}
-
-static void ggml_backend_metal_tuning_clear_fa_vec_override(void) {
-    ggml_metal_tuning::fa_vec_clear_override();
-}
-
-static int ggml_backend_metal_tuning_fa_vec_ne11_bucket(int64_t ne11) {
-    return ggml_metal_tuning::fa_vec_ne11_bucket(ne11);
-}
-
-static int ggml_backend_metal_tuning_fa_vec_ne01_bucket(int64_t ne01) {
-    return ggml_metal_tuning::fa_vec_ne01_bucket(ne01);
-}
-
-static int ggml_backend_metal_tuning_fa_vec_baseline_ne(int dk, int dv) {
-    return ggml_metal_tuning::fa_vec_baseline_ne(dk, dv);
-}
-
-static const char * ggml_backend_metal_tuning_device_token(ggml_backend_dev_t dev) {
-    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
-
-    return ggml_metal_device_id_token(ggml_metal_device_get_props(ctx_dev)->device_id);
-}
-
 static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     if (strcmp(name, "ggml_backend_get_features") == 0) {
         return (void *)ggml_backend_metal_get_features;
     }
-    if (strcmp(name, "ggml_backend_metal_tuning_set_fa_vec_override") == 0) {
-        return (void *)ggml_backend_metal_tuning_set_fa_vec_override;
-    }
-    if (strcmp(name, "ggml_backend_metal_tuning_clear_fa_vec_override") == 0) {
-        return (void *)ggml_backend_metal_tuning_clear_fa_vec_override;
-    }
-    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne11_bucket") == 0) {
-        return (void *)ggml_backend_metal_tuning_fa_vec_ne11_bucket;
-    }
-    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_ne01_bucket") == 0) {
-        return (void *)ggml_backend_metal_tuning_fa_vec_ne01_bucket;
-    }
-    if (strcmp(name, "ggml_backend_metal_tuning_fa_vec_baseline_ne") == 0) {
-        return (void *)ggml_backend_metal_tuning_fa_vec_baseline_ne;
-    }
-    if (strcmp(name, "ggml_backend_metal_tuning_device_token") == 0) {
-        return (void *)ggml_backend_metal_tuning_device_token;
-    }
 
     return NULL;
 
@@ -942,7 +891,7 @@ static ggml_backend_dev_t ggml_backend_metal_device_init(ggml_backend_reg_t reg,
     return new ggml_backend_device {
         /* .iface   = */ ggml_backend_metal_device_i,
         /* .reg     = */ reg,
-        /* .context = */ ggml_metal_device_get(device, g_devices),
+        /* .context = */ ggml_metal_device_get(device),
     };
 }
 

ggml/src/ggml-metal/kernels/argsort.metal

@@ -1,232 +0,0 @@
-#include "common.h"
-
-// bitonic sort implementation following the CUDA kernels as reference
-typedef void (argsort_t)(
-        constant   ggml_metal_kargs_argsort & args,
-        device   const char * src0,
-        device      int32_t * dst,
-        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]);
-
-template<ggml_sort_order order>
-kernel void kernel_argsort_f32_i32(
-        constant   ggml_metal_kargs_argsort & args,
-        device   const char * src0,
-        device      int32_t * dst,
-        threadgroup int32_t * shmem_i32 [[threadgroup(0)]],
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-    // bitonic sort
-    const int col = tpitg[0];
-    const int ib  = tgpig[0] / args.ne01;
-
-    const int i00 = ib*ntg.x;
-    const int i01 = tgpig[0] % args.ne01;
-    const int i02 = tgpig[1];
-    const int i03 = tgpig[2];
-
-    device const float * src0_row = (device const float *) (src0 + args.nb01*i01 + args.nb02*i02 + args.nb03*i03);
-
-    // initialize indices
-    shmem_i32[col] = i00 + col;
-
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
-    for (int k = 2; k <= ntg.x; k *= 2) {
-        for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (shmem_i32[col] >= args.ne00 ||
-                       (shmem_i32[ixj] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
-                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]] :
-                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]]))
-                    ) {
-                        SWAP(shmem_i32[col], shmem_i32[ixj]);
-                    }
-                } else {
-                    if (shmem_i32[ixj] >= args.ne00 ||
-                       (shmem_i32[col] <  args.ne00 && (order == GGML_SORT_ORDER_ASC ?
-                            src0_row[shmem_i32[col]] < src0_row[shmem_i32[ixj]] :
-                            src0_row[shmem_i32[col]] > src0_row[shmem_i32[ixj]]))
-                    ) {
-                        SWAP(shmem_i32[col], shmem_i32[ixj]);
-                    }
-                }
-            }
-
-            threadgroup_barrier(mem_flags::mem_threadgroup);
-        }
-    }
-
-    const int64_t i0 = ib*args.top_k;
-
-    // copy the result to dst without the padding
-    if (i0 + col < args.ne0 && col < args.top_k) {
-        dst += i0 + args.ne0*i01 + args.ne0*args.ne1*i02 + args.ne0*args.ne1*args.ne2*i03;
-
-        dst[col] = shmem_i32[col];
-    }
-}
-
-template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
-template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
-
-typedef void (argsort_merge_t)(
-        constant   ggml_metal_kargs_argsort_merge & args,
-        device const char    * src0,
-        device const int32_t * tmp,
-        device       int32_t * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]);
-
-template<ggml_sort_order order>
-kernel void kernel_argsort_merge_f32_i32(
-        constant   ggml_metal_kargs_argsort_merge & args,
-        device const char    * src0,
-        device const int32_t * tmp,
-        device       int32_t * dst,
-        uint3   tgpig[[threadgroup_position_in_grid]],
-        ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
-
-    const int im  = tgpig[0] / args.ne01;
-    const int i01 = tgpig[0] % args.ne01;
-    const int i02 = tgpig[1];
-    const int i03 = tgpig[2];
-
-    const int start = im * (2 * args.len);
-
-    const int len0 = MIN(args.len, MAX(0, args.ne0 - (int)(start)));
-    const int len1 = MIN(args.len, MAX(0, args.ne0 - (int)(start + args.len)));
-
-    const int total = len0 + len1;
-
-    device const int32_t * tmp0 = tmp + start
-        + i01*args.ne0
-        + i02*args.ne0*args.ne01
-        + i03*args.ne0*args.ne01*args.ne02;
-
-    device const int32_t * tmp1 = tmp0 + args.len;
-
-    dst += start
-        + i01*args.top_k
-        + i02*args.top_k*args.ne01
-        + i03*args.top_k*args.ne01*args.ne02;
-
-    device const float * src0_row = (device const float *)(src0
-        + args.nb01*i01
-        + args.nb02*i02
-        + args.nb03*i03);
-
-    if (total == 0) {
-        return;
-    }
-
-    const int chunk = (total + ntg.x - 1) / ntg.x;
-
-    const int k0 = tpitg.x * chunk;
-    const int k1 = MIN(MIN(k0 + chunk, total), args.top_k);
-
-    if (k0 >= args.top_k) {
-        return;
-    }
-
-    if (k0 >= total) {
-        return;
-    }
-
-    int low  = k0 > len1 ? k0 - len1 : 0;
-    int high = MIN(k0, len0);
-
-    // binary-search partition (i, j) such that i + j = k
-    while (low < high) {
-        const int mid = (low + high) >> 1;
-
-        const int32_t idx0 = tmp0[mid];
-        const int32_t idx1 = tmp1[k0 - mid - 1];
-
-        const float val0 = src0_row[idx0];
-        const float val1 = src0_row[idx1];
-
-        bool take_left;
-        if (order == GGML_SORT_ORDER_ASC) {
-            take_left = (val0 <= val1);
-        } else {
-            take_left = (val0 >= val1);
-        }
-
-        if (take_left) {
-            low = mid + 1;
-        } else {
-            high = mid;
-        }
-    }
-
-    int i = low;
-    int j = k0 - i;
-
-    // keep the merge fronts into registers
-    int32_t idx0 = 0;
-    float   val0 = 0.0f;
-    if (i < len0) {
-        idx0 = tmp0[i];
-        val0 = src0_row[idx0];
-    }
-
-    int32_t idx1 = 0;
-    float   val1 = 0.0f;
-    if (j < len1) {
-        idx1 = tmp1[j];
-        val1 = src0_row[idx1];
-    }
-
-    for (int k = k0; k < k1; ++k) {
-        int32_t out_idx;
-
-        if (i >= len0) {
-            while (k < k1) {
-                dst[k++] = tmp1[j++];
-            }
-            break;
-        } else if (j >= len1) {
-            while (k < k1) {
-                dst[k++] = tmp0[i++];
-            }
-            break;
-        } else {
-            bool take_left;
-
-            if (order == GGML_SORT_ORDER_ASC) {
-                take_left = (val0 <= val1);
-            } else {
-                take_left = (val0 >= val1);
-            }
-
-            if (take_left) {
-                out_idx = idx0;
-                ++i;
-                if (i < len0) {
-                    idx0 = tmp0[i];
-                    val0 = src0_row[idx0];
-                }
-            } else {
-                out_idx = idx1;
-                ++j;
-                if (j < len1) {
-                    idx1 = tmp1[j];
-                    val1 = src0_row[idx1];
-                }
-            }
-        }
-
-        dst[k] = out_idx;
-    }
-}
-
-template [[host_name("kernel_argsort_merge_f32_i32_asc")]]  kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_ASC>;
-template [[host_name("kernel_argsort_merge_f32_i32_desc")]] kernel argsort_merge_t kernel_argsort_merge_f32_i32<GGML_SORT_ORDER_DESC>;