hexagon: support for multi-NPU devices (IQ9, IQ10) and fully asynchronous backend (#26501 )

* hexagon: use non-host bufs by default and make the backend fully async * hex-hb: remove optional hostbuf support and fix async copy * hex-unary: relax supported unary check * hex-bufs: use same get_alignment for host bufs * snapdragon: bump android_platform to 34 * hex-rows: super hacky get/set rows for q8_0 * hex-get-rows: fix q8_0 * hex-get-rows: supprot for f16 and cleanup for q8_0 * hex-get-rows: generic macros and specialized thread funcs * hex-get-rows: add DMA pipeline, vtcm_layout and kernel params * hex-set-rows: fix q8_0 support, add dma and tracing * hex-tests: override nmse threshold for HTP of Q8_0 quants * hex-fa: add support for Q8_0 with inplace dequantizers * hex-get-rows: simplify type dispatch * hex-rows: simplify GET/SET_ROWS DMA pipeline * hex-async: add events, set/get-tensor-async and rest of the async api support * hex-repack: use slice instead of expert in repack functions * hex-cpy: update event/async-cpy logging * hex-set-rows: optimize smaller tensors * hex-geglu: fix perf regression with larger tensors * hex-get-rows: add missing header * hex-set-rows: add missing header * hex-bufs: ressurect GGML_HEXAGON_HOSTBUF but disable it by default * hexagon: do not reject ops with non-heaxon buffers * hex-get-rows: apply >=32 restriction only for q8_0 * hex-res: bump vtcm acquire timeout to 10 seconds * hex-bufs: add support for cloning buffers between sessions to speed up tensor copies * hex-async: rework event recording and batch flushing and integrate with meta backend * hex-bufs: improved handling of repacked tensors * hex-repack: handle get_tensor_2d offsets * hex-dev: add support for devices with multiple NPUs * hex-sync: add support for sync tokens to synchronize npu devices for async splits * hex-mmap: cleanup mmap calls and add a retry for robustness * hex-sync: add failsafe if sync wait gets stuck * hex-sync: use sync_seq to check for completed events * hex-sync: rotate tokens for extra robustness * hex-devs: add supprot for legacy device names for now * hex-bufs: add support for auto-cloning buffers from diff sessions * hex-fusion: simplify and optimize htp-opnode fusion handling * hex-sync: override opnode name so that it shows up in the profiles * hex-trace: update scripts to handle multiple devices * hex-sync: bump the size of the opbatch queue and number of sync tokens * hex-cpy-sync: do not explicitly flush opbatches in cpy_tensor_async and add support for cpy-dma * hex-sync: add graph-flush threshold to avoid single op batches * hex-sync: add sync_peer so that we can flush peers we depend on during cross-device ops * hex-bufs: introduce tensor->extra and shadow_bufs for repacking * hex-l2: flush tiny tensors inline * hex-sync: use explicit l2flush for sync tokens * hex-extra: track weight flags via tensor extra * hex-fence: rename sync to fence * hex-repack: proper handling of set-tensor-2d in the shadow_buf * hex-trace: remove obsolete opstage mask that we used for profiling * hex-env: remove obsolete use_hmx variable * hexagon: new unified run.py and build.py and updated docs * snapdragon: update run script to auto-escapt test-backend-op -p argument * hex-scripts: fix trailing spaces * hex-scripts: fix flake8 warnings * snapdragon: cleanup dst lib/bin dirs before copying new build * hex-ops: add support for allreduce * hex-ar: improved allreduce with dma pipeline * hex-ar: align macros * hex-ar: consistent use of fence_seq * hex-ar: add AR_SELECT env var to select ALLREDUCE kernel or fallback * hex-ar: add proper synchronize handling for ALLREDUCE * hex-opbatch: looks like we now just rely on backend.synchronise to flush the batches, no need to flush them by threshold * hex-ar: bump block size to improve dma efficiency * hex-ar: fused ALLREDUCE+ADD * hex-ar: cleaner fence buffer management * hex-ar: futher allreduce tweaking to remove race conditions * hex-ar: add simple solver and remove non-dma kernels * hex-ar: add row-broadcast to fuse with bias ADD * hex-fence: pass seq numbers via op_params * hex-ar: allow for both entry/exit seq for completing entry wait * hex-ar: align macros * hex-ar: do not refetch broadcast row * hex-fusion: move all fusion into opbatch::add_op for consistency with ALLREDUCE and things * hex-fusion: fix incorrect MUL_MAT reordering * hex-mm: make fused 2x and 3x matmuls more generic * hex-fusion: move tensor fusion tagging to graph_compute * hexagon: make sure to copy tensor->extra by value * hex-get-rows: fix offset calc with row-chunking * hex-repack: get_tensor_2d fixes for non-zero offsets * snapdragon: make profile/trace scripts more robust and donot mix stdout/stderr by default * hex-devices: use legacy device nameing by default to ease the transition * hex-devices: hardcode CDSP domain IDs for current devices for now * hex-optrace: improve multi-NPU timestamp alignment and overall handling of cycle values * hex-optrace: more robust handling of the fence events
llama: add token ID tracking to KV cell (#27762 )
2026-08-27 10:37:33 +02:00 · 2026-08-26 18:46:50 -07:00 · 2026-08-26 23:34:28 +02:00 · 2026-08-26 21:07:24 +02:00 · 2026-08-26 20:19:19 +02:00 · 2026-08-26 19:02:06 +03:00
507 changed files with 28901 additions and 19151 deletions
@@ -90,6 +90,9 @@ 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 "

@@ -0,0 +1,95 @@
+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
@@ -21,68 +21,30 @@ 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: |
-        # 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 <<< "$CACHES"
+        bash scripts/ccache-clear.sh \
+          --key "${{ inputs.key }}" \
+          --older "${{ inputs.older }}" \
+          --min "${{ inputs.min }}" \
+          ${{ inputs.dry-run == 'true' && '--dry-run' || '' }}
@@ -22,7 +22,8 @@ on:
    types: [opened, synchronize, reopened]
    paths: [
      '.github/workflows/build-apple.yml',
-      'ggml/src/ggml-metal/**'
+      'ggml/src/ggml-metal/**',
+      'ggml/src/ggml-rpc/**'
    ]

 concurrency:
@@ -73,6 +74,16 @@ 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

@@ -109,6 +120,16 @@ 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

@@ -163,14 +184,6 @@ 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: |
@@ -196,14 +209,6 @@ 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: |
@@ -234,14 +239,6 @@ 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:
@@ -125,7 +125,7 @@ jobs:
          GH_TOKEN: ${{ github.token }}
        with:
          key: cpu-${{ matrix.os }}
-          older: 1h
+          older: 5m
          min: 1
          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

@@ -215,3 +215,13 @@ 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' }}
@@ -50,14 +50,22 @@ jobs:
          DEBIAN_FRONTEND: noninteractive
        run: |
          apt update
-          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev
+          apt install -y cmake build-essential ninja-build libgomp1 git libssl-dev jq python3 python3-venv python3-pip

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
          key: cuda-ubuntu-24.04-cuda
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          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

      - 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
@@ -72,6 +80,18 @@ 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
@@ -85,14 +105,22 @@ 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
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libssl-dev rocwmma-dev jq python3-venv

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
          key: cuda-ubuntu-22.04-hip
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          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

      - name: Build with native CMake HIP support
        id: cmake_build
@@ -103,6 +131,18 @@ 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
@@ -116,14 +156,22 @@ jobs:
        id: depends
        run: |
          apt-get update
-          apt-get install -y build-essential git cmake libssl-dev
+          apt-get install -y build-essential git cmake libssl-dev jq

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
          key: cuda-ubuntu-22.04-musa
-          evict-old-files: 1d
-          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+          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

      - name: Build with native CMake MUSA support
        id: cmake_build
@@ -131,3 +179,15 @@ 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: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: cuda-ubuntu-22.04-musa
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
@@ -80,3 +80,13 @@ jobs:
        run: |
          cmake -S . -B build -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$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' }}
@@ -167,3 +167,13 @@ 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: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
@@ -96,6 +96,16 @@ jobs:
            -DGGML_SYCL_F16=${{ 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

@@ -139,3 +149,13 @@ 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' }}
@@ -55,7 +55,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: vulkan-ubuntu-24.04-arm-new
+          key: vulkan-ubuntu-24.04-arm
          variant: ccache
          evict-old-files: 1d
          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
@@ -73,6 +73,16 @@ 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

@@ -128,6 +138,16 @@ 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

@@ -180,3 +200,13 @@ 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' }}
@@ -88,3 +88,13 @@ 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' }}
@@ -101,6 +101,16 @@ 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

@@ -153,3 +163,13 @@ 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' }}
@@ -64,7 +64,7 @@ jobs:
    needs: create_tag
    uses: ./.github/workflows/ui-build.yml
    with:
-      hf_ui_version: ${{ needs.create_tag.outputs.source_tag }}
+      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: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist

      - name: Set up QEMU
@@ -84,3 +84,13 @@ 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' }}
@@ -84,11 +84,13 @@ jobs:

            New version has been released.

+            ## Assets
+
            ${{ steps.desc.outputs.nightly }}

-            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+            ## More info

-            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+            - [Releases and versioning of `ggml-org` projects](https://github.com/ggml-org/ggml/discussions/1579)

            ## ${{ steps.desc.outputs.changelog_title }}

@@ -61,31 +61,8 @@ 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, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    strategy:
      matrix:
@@ -119,12 +96,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -141,7 +117,6 @@ 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)

@@ -167,7 +142,7 @@ jobs:
          key: release-${{ matrix.os }}-${{ matrix.arch }}

  ubuntu-cpu:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    strategy:
      matrix:
@@ -191,12 +166,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Dependencies
        id: depends
@@ -227,7 +201,6 @@ 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)

@@ -254,7 +227,7 @@ jobs:
          key: release-${{ matrix.os }}-cpu

  ubuntu-vulkan:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    strategy:
@@ -277,12 +250,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Dependencies
        id: depends
@@ -314,7 +286,6 @@ 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)

@@ -340,7 +311,7 @@ jobs:
          key: release-${{ matrix.os }}-vulkan

  android-arm64:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: ubuntu-latest
@@ -358,12 +329,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Set up JDK
        uses: actions/setup-java@v5
@@ -407,7 +377,6 @@ 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)

@@ -433,7 +402,7 @@ jobs:
          name: llama-bin-android-arm64.tar.gz

  ubuntu-24-openvino:
-    needs: [check-release, get-version]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: ubuntu-24.04
@@ -460,12 +429,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -508,7 +476,6 @@ 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

@@ -552,7 +519,7 @@ jobs:
          key: release-ubuntu-24.04-openvino-release-no-preset-v1

  windows-openvino:
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2022
@@ -577,12 +544,11 @@ jobs:
        with:
            fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -682,7 +648,7 @@ jobs:

  windows-cpu:
    name: windows-cpu / ${{ matrix.arch }}
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2025-vs2026
@@ -702,12 +668,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Install Ninja
        run: |
@@ -749,8 +714,10 @@ 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]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    runs-on: windows-2022
@@ -769,6 +736,12 @@ 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:
@@ -879,6 +852,8 @@ 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' }}
@@ -909,13 +884,6 @@ 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: ${{ matrix.backend == 'vulkan' }}
@@ -978,6 +946,8 @@ jobs:
          path: llama-bin-win-${{ 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]
@@ -1006,13 +976,6 @@ 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:
@@ -1084,6 +1047,8 @@ 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' }}
@@ -1118,13 +1083,6 @@ 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:
@@ -1195,7 +1153,7 @@ jobs:
          key: release-windows-2022-x64-sycl

  ubuntu-24-sycl:
-    needs: [check-release]
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

    strategy:
@@ -1237,12 +1195,11 @@ 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: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
@@ -1287,11 +1244,11 @@ jobs:
        with:
          key: release-ubuntu-24.04-sycl-${{ matrix.build }}

-  ubuntu-22-rocm:
-    needs: [check-release, get-version]
+  ubuntu-24-rocm:
+    needs: [check-release, ui-build]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}

-    runs-on: ubuntu-22.04
+    runs-on: ubuntu-24.04

    permissions:
      actions: write
@@ -1310,12 +1267,11 @@ jobs:
        with:
          fetch-depth: 0

-      - name: Setup Node.js
-        uses: actions/setup-node@v6
+      - name: Download UI build
+        uses: actions/download-artifact@v7
        with:
-          node-version: "24"
-          cache: "npm"
-          cache-dependency-path: "tools/ui/package-lock.json"
+          name: llama-ui.zip
+          path: tools/ui/dist

      - name: Free up disk space
        uses: ggml-org/free-disk-space@v1.3.1
@@ -1325,7 +1281,7 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
          evict-old-files: 1d
          max-size: "1G"

@@ -1388,7 +1344,6 @@ 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)

@@ -1414,10 +1369,10 @@ jobs:
      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          key: release-ubuntu-24.04-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}

  ios-xcode:
-    needs: [check-release, get-version]
+    needs: [check-release]
    if: ${{ needs.check-release.outputs.should_release == 'true' }}
    runs-on: macos-26

@@ -1445,8 +1400,7 @@ jobs:
            -DLLAMA_BUILD_SERVER=OFF \
            -DCMAKE_SYSTEM_NAME=iOS \
            -DCMAKE_OSX_DEPLOYMENT_TARGET=16.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml \
-            -DHF_UI_VERSION=${{ needs.get-version.outputs.ui_version }}
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO

      - name: xcodebuild for swift package
@@ -1569,11 +1523,9 @@ jobs:
 #          name: llama-bin-${{ matrix.chip_type }}-openEuler-${{ matrix.arch }}${{ matrix.use_acl_graph == 'on' && '-aclgraph' || '' }}.tar.gz

  ui-build:
-    needs: [check-release, get-version]
+    needs: [check-release]
    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' }}
@@ -1588,14 +1540,13 @@ jobs:
    runs-on: ubuntu-slim

    needs:
-      - get-version
      - windows
      - windows-cpu
      - windows-cuda
      - windows-sycl
      - windows-rocm
      - windows-openvino
-      - ubuntu-22-rocm
+      - ubuntu-24-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
@@ -1628,24 +1579,27 @@ jobs:
          path: ./artifact
          merge-multiple: true

-      - name: Move artifacts
+      - name: Merge artifacts
        id: move_artifacts
        run: |
          mkdir -p release

-          echo "Adding CPU backend files to existing zips..."
+          # 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..."
          for arch in x64 arm64; do
            cpu_zip="artifact/llama-bin-win-cpu-${arch}.zip"
            temp_dir=$(mktemp -d)
-            echo "Extracting CPU backend for $arch..."
+            echo "Extracting windows-cpu-${arch} package..."
            unzip "$cpu_zip" -d "$temp_dir"

-            echo "Adding CPU files to $arch zips..."
+            echo "Merging into $arch zips..."
            for target_zip in artifact/llama-bin-win-*-${arch}.zip; do
              if [[ "$target_zip" == "$cpu_zip" ]]; then
                continue
              fi
-              echo "Adding CPU backend to $(basename "$target_zip")"
+              echo "Injecting into $(basename "$target_zip")"
              realpath_target_zip=$(realpath "$target_zip")
              (cd "$temp_dir" && zip -r "$realpath_target_zip" .)
            done
@@ -1669,7 +1623,7 @@ jobs:
        id: download_ui
        uses: actions/download-artifact@v7
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: ./ui-dist

      - name: Package UI
@@ -73,13 +73,6 @@ 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: |
@@ -128,6 +128,16 @@ 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

@@ -181,3 +191,13 @@ 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' }}
@@ -31,6 +31,6 @@ jobs:
      - name: Upload built UI
        uses: actions/upload-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/
          retention-days: 1
@@ -3,8 +3,8 @@ name: UI Build
 on:
  workflow_call:
    inputs:
-      hf_ui_version:
-        description: 'Version string for version.json (e.g. 12345)'
+      ui_version:
+        description: 'Version string embedded in build.json (e.g. b1234); defaults to b<commit-count>'
        required: false
        type: string

@@ -17,6 +17,17 @@ 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
@@ -31,8 +42,7 @@ jobs:

      - name: Build application
        env:
-          HF_UI_VERSION: ${{ inputs.hf_ui_version || '' }}
-          LLAMA_BUILD_NUMBER: ${{ inputs.hf_ui_version || 'b0000' }}
+          LLAMA_BUILD_NUMBER: ${{ steps.version.outputs.ui_version }}
        run: npm run build
        working-directory: tools/ui

@@ -43,6 +53,6 @@ jobs:
      - name: Upload built UI
        uses: actions/upload-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/
          retention-days: 1
@@ -37,7 +37,7 @@ jobs:
      - name: Download UI build artifact
        uses: actions/download-artifact@v7
        with:
-          name: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Create distribution archive
@@ -64,7 +64,7 @@ jobs:
      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          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: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Build Storybook
@@ -63,7 +63,7 @@ jobs:
      - name: Download built UI artifacts
        uses: actions/download-artifact@v6
        with:
-          name: ui-build
+          name: llama-ui.zip
          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: ui-build
+          name: llama-ui.zip
          path: tools/ui/dist/

      - name: Install Playwright browsers
@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_MINOR 3)
 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"                                       ON)
-option(LLAMA_USE_PREBUILT_UI "llama: use prebuilt UI from HF Bucket when available (requires LLAMA_BUILD_UI=ON)" ON)
+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_TOOLS_INSTALL "llama: install tools" ${LLAMA_TOOLS_INSTALL_DEFAULT})
 option(LLAMA_TESTS_INSTALL "llama: install tests" ON)
@@ -74,6 +74,7 @@ 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.
@@ -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)

-[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)
+[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)

 </div>

@@ -300,6 +300,40 @@ 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 {
@@ -751,6 +785,8 @@ 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
@@ -1177,6 +1177,8 @@ 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.
@@ -1217,13 +1219,15 @@ static common_chat_params common_chat_params_init_qwen3_coder(const common_chat_

    std::vector<std::string> tool_call_starts = { "<tool_call>" };

-    // 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 + ">");
-    });
+    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 + ">");
+        });
+    }

    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
        auto generation_prompt = p.literal(GEN_PREFIX);
@@ -1288,10 +1292,13 @@ 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_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 tool_call_first = is_qwen3_coder ?
+                p.rule("tool-call-first", p.optional(p.literal("<tool_call>\n")) + tool_call_body) :
+                tool_call;

            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));
@@ -402,10 +402,11 @@ 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);

-    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, common_log_get_verbosity_thold());
+    const int verbosity = common_log_get_verbosity_thold();
+    COM_INF("%s: verbosity = %d (adjust with the `-lv N` CLI arg)\n", __func__, verbosity);

    // device enumeration creates a primary context on CUDA backends, skip it when the caller does not own any device
-    if (print_devices) {
+    if (print_devices && verbosity >= LOG_LEVEL_TRACE) {
        COM_TRC("%s", "device_info:\n");
        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
            auto * dev = ggml_backend_dev_get(i);
@@ -78,19 +78,21 @@ 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>
-common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
    common_json out = common_json::array();

    for (const auto & val : vals) {
        out.push_back(val);
    }

-    val_json = std::make_shared<common_json>(std::move(out));
+    return 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__> &);
+#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)) {}

 COMMON_JSON_SET(int)
 COMMON_JSON_SET(std::string)
@@ -98,56 +100,45 @@ COMMON_JSON_SET(std::string)
 #undef COMMON_JSON_SET

 template <typename T>
-common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> map_json(const T & vals) {
    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));
+    return 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__> &);
+#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)) {}

 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__> &);
+#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)) {}

 COMMON_JSON_UMAP(size_t)

 #undef COMMON_JSON_UMAP

 template <typename T>
-common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
+static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
    common_json out = common_json::array();

    for (const auto & val : vals) {
        out.push_back(val);
    }

-    val_json = std::make_shared<common_json>(std::move(out));
+    return 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__> &);
+#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)) {}

 COMMON_JSON_VEC(int)
 COMMON_JSON_VEC(unsigned char)
@@ -404,10 +395,6 @@ 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;
@@ -415,7 +402,7 @@ 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;
+#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }

 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -435,3 +422,12 @@ 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);
+}
@@ -221,16 +221,14 @@ 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 { return get<std::string>(); }
+    operator std::string() const;

    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 {
-        return contains(key) ? at(key).get<std::string>() : std::string(def);
-    }
+    std::string value(const std::string & key, const char * def) const;

    // 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 {
@@ -112,12 +112,38 @@ 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)
-        # 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)
+        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

    def set_vocab(self):
        return self._set_vocab_glm()
@@ -153,10 +179,22 @@ 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)

-        # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (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
@@ -348,6 +386,7 @@ 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
@@ -202,6 +202,10 @@ 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
@@ -242,8 +246,8 @@ 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 == "mamba"]
-            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]
+            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]

        # `--no-mtp` drops it entirely; `--mtp` exports only the MTP head
        self._mtp_bid: int | None = None
@@ -272,7 +276,7 @@ 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 == "attention"]
+        return [i for i, val in enumerate(pattern) if val in self._ATTN_LAYER_TYPES]

    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
@@ -443,21 +443,21 @@ Each returned parser is wrapped by `wrap_for_generation_prompt()`, which prepend
 |                                           | `wrap_for_generation_prompt()`, string helpers                                  |
 | `common/chat-peg-parser.h/cpp`            | `common_chat_peg_builder`, `common_chat_peg_mapper`, and helpers                |
 | `common/chat.cpp`                         | Entry point: `common_chat_templates_apply_jinja()`                              |
-| `tools/parser/debug-template-parser.cpp`  | Debug tool for template analysis                                                |
-| `tools/parser/template-analysis.cpp`      | Template analysis tool                                                          |
+| `tests/test-chat-auto-parser.cpp`         | Auto-parser unit tests; also a debug tool when given a template path            |
+| `tests/test-chat-analysis.cpp`            | Template differential analysis debug tool                                       |

 ## Testing & Debugging

 ### Debug Tools

-**Template Debugger**: `tools/parser/debug-template-parser.cpp`
+**Template Debugger**: `tests/test-chat-auto-parser.cpp`

- Usage: `./bin/llama-debug-template-parser path/to/template.jinja`
+- Usage: `./bin/test-chat-auto-parser path/to/template.jinja` (without a path, it runs the automated tests)
 - Shows detected format, markers, generated parser, and GBNF grammar

-**Template Analysis**: `tools/parser/template-analysis.cpp`
+**Template Analysis**: `tests/test-chat-analysis.cpp`

- Usage: `./bin/llama-template-analysis path/to/template.jinja`
+- Usage: `./bin/test-chat-analysis --template-file path/to/template.jinja` (without arguments, it runs on all templates from the test suite)

 **Debug Logging**: Enable with `LLAMA_ARG_LOG_VERBOSITY=2`

@@ -519,7 +519,7 @@ The following templates have active tests in `tests/test-chat.cpp`:

 To support a new template format:

-1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `llama-debug-template-parser` to verify markers are correctly extracted.
+1. **If it follows standard patterns** — The auto-parser should detect it automatically. Run `test-chat-auto-parser <template_path>` to verify markers are correctly extracted.
 2. **If differential analysis extracts incorrect markers** — Add a workaround lambda to the `workarounds` vector in `common/chat-diff-analyzer.cpp`. Inspect the template source for a unique identifying substring.
 3. **If it needs fundamentally different handling** — Add a dedicated handler function in `chat.cpp` before the auto-parser block (as done for GPT-OSS, Functionary v3.2, and Ministral).

@@ -8,7 +8,7 @@
        "toolset":      { "value": "host=x86_64", "strategy": "external" },
        "cacheVariables": {
            "ANDROID_ABI":      "arm64-v8a",
-            "ANDROID_PLATFORM": "android-31",
+            "ANDROID_PLATFORM": "android-34",
            "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",
@@ -2,39 +2,47 @@

 ## Setup

-### Android
+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:

-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.
+* **Android toolchain**: `ghcr.io/snapdragon-toolchain/arm64-android:v0.7`
+* **Linux toolchain**: `ghcr.io/snapdragon-toolchain/arm64-linux:v0.7`

-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.
+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)

 ## How to Build

-Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
+### Using build.py script (Recommended)

+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
@@ -68,19 +76,19 @@ Preset CMake variables:
 To generate an installable "package" simply use cmake --install:

 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon/llama.cpp
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-android/llama.cpp
 -- Install configuration: "Release"
-- 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/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/bin/llama-bench
-- Installing: /workspace/pkg-snapdragon/llama.cpp/bin/llama-cli
+-- Installing: /workspace/pkg-android/llama.cpp/bin/llama-bench
+-- Installing: /workspace/pkg-android/llama.cpp/bin/llama-cli
 ...
 ```

@@ -91,14 +99,14 @@ 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-snapdragon` on the device.
+Once ADB is enabled, use `adb push` to install `pkg-android` 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-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)
+~/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)
 102 files pushed, 0 skipped. 186.9 MB/s (963151597 bytes in 4.914s)
 ```

@@ -115,24 +123,44 @@ 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-snapdragon` folder.
-To run, adapt below instructions to use Powershell scripts in `scripts/snapdragon/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).

 ## How to Run

-The easiest way to run llama.cpp cli tools is using provided wrapper scripts that properly set up all required environment variables.
+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.

-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.
+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`).

 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 via ADB.
+Here are some examples of running various llama.cpp tools.

-Simple question for Llama-3.2-1B
+Generating a completion with Gemma on Android (relying on default `HTP0:0` device and default thread count `-t 6`):

 ```
-~/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?"
+~/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?"
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -142,8 +170,7 @@ 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 =     0.26 MiB
-load_tensors:  HTP0-REPACK model buffer size =   504.00 MiB
+load_tensors:         HTP0 model buffer size =   504.26 MiB
 ...
 I hope this helps you understand the world's most popular cookies! [end of text]
 ...
@@ -156,60 +183,25 @@ 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                |
 ```

-Summary request for OLMoE-1B-7B. This is a large model that requires two HTP sessions/devices
+Op test for 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
+~/src/llama.cpp$ ./scripts/snapdragon/run.py --target adb --hex-hostbuf 0 --devices HTP0:0 -- test-backend-ops -b HTP0:0 -o MUL_MAT
 ...
-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
+Backend 2/3: HTP0:0
 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
+```

-~/src/llama.cpp-hexagon$ M=Llama-3.2-1B-Instruct-Q4_0.gguf ./scripts/snapdragon/adb/run-bench.sh -p 128 -n 64
+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
 ...
 ggml-hex: Hexagon backend (experimental) : allocating new registry : ndev 1
 ggml-hex: Hexagon Arch version v79
@@ -219,15 +211,20 @@ 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_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_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_NHVX=0`
  Controls the number of HVX hardware threads to use. The default is all (actual number varies depending on the hardware version).
@@ -255,26 +252,17 @@ build: 6a8cf8914 (6733)
  - `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 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_PROFILE=1 ./scripts/snapdragon/run.py --target adb -- llama-cli ... |& ./scripts/snapdragon/ggml-hexagon-profile.py -`

 - `GGML_HEXAGON_OPFILTER=regex`
  Allows filtering (disabling) Ops that match the regex pattern:

  Examples:

-      `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)
+      `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)
+
@@ -39,22 +39,21 @@ the repacking.

 ## Large model handling

-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).
+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).

-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.
+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.

-Here is an example of running GPT-OSS-20B model on a newer Snapdragon device with 16GB of DDR.
+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).

 ```
-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
+~/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
 ...
 llama_model_loader: - type  f32:  289 tensors
 llama_model_loader: - type q4_0:   96 tensors
@@ -63,33 +62,29 @@ 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:         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
+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
 ...
 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:       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:     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: 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:       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:     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: 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 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:     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:        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)
@@ -97,13 +92,9 @@ 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 (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: |   - 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: |   - 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                |
 ```
@@ -1,25 +1,37 @@
 # Snapdragon-based Linux devices

-## Docker Setup
+The cross-compilation is performed using the Snapdragon Linux Docker toolchain image (see
+[github.com/snapdragon-toolchain](https://github.com/snapdragon-toolchain)):

-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.
+* **Linux toolchain**: `ghcr.io/snapdragon-toolchain/arm64-linux:v0.7`

-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.
+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.


 ## How to Build

-Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
+### Using build.py script (Recommended)

+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
@@ -30,17 +42,19 @@ Let's build llama.cpp with CPU, OpenCL, and Hexagon backends via CMake presets:
 To generate an installable "package" simply use cmake --install, then zip it:

 ```
-[d]/workspace> cmake --install build-snapdragon --prefix pkg-snapdragon
-[d]/workspace> zip -r pkg-snapdragon.zip pkg-snapdragon
+[d]/workspace> cmake --install build-snapdragon --prefix pkg-linux
+[d]/workspace> zip -r pkg-linux.zip pkg-linux
 ```

 ## How to Install

-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:
+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:

 ```
-$ unzip pkg-snapdragon.zip
-$ cd pkg-snapdragon
+$ unzip pkg-linux.zip
+$ cd pkg-linux
 $ export LD_LIBRARY_PATH=./lib
 $ export ADSP_LIBRARY_PATH=./lib
 ```
@@ -52,7 +66,28 @@ $ wget https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF/resolve/main/
 ```

 ## How to Run
-Next, since we have setup the environment variables, we can run the llama-cli with the Hexagon backends:
+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:
 ```
 $ ./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?"
 ```
+
@@ -1,3 +1,18 @@
+# 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.
@@ -53,7 +68,8 @@ 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
@@ -130,12 +146,12 @@ 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-snapdragon
+> cmake --install build-wos --prefix pkg-wos
 ```

 Once the build is complete HTP ops libraries will be installed like this
 ```
-> dir pkg-snapdragon/lib
+> dir pkg-wos/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
@@ -147,8 +163,8 @@ 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-snapdragon\lib\libggml-htp.cat
-Verifying: .\pkg-snapdragon\lib\libggml-htp.cat
+> signtool.exe verify /v /pa .\pkg-wos\lib\libggml-htp.cat
+Verifying: .\pkg-wos\lib\libggml-htp.cat

 Signature Index: 0 (Primary Signature)
 Hash of file (sha256): 9820C664DA59D5EAE31DBB664127FCDAEF59CDC31502496BC567544EC2F401CF
@@ -156,6 +172,6 @@ Hash of file (sha256): 9820C664DA59D5EAE31DBB664127FCDAEF59CDC31502496BC567544EC
 Signing Certificate Chain:
        Issued to: GGML.HTP.v1
 ...
-Successfully verified: .\pkg-snapdragon\lib\libggml-htp.cat
+Successfully verified: .\pkg-wos\lib\libggml-htp.cat
 ...
 ```
@@ -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 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -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","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","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","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"
@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_MINOR 22)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

@@ -342,9 +342,6 @@ 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)

@@ -110,6 +110,16 @@ 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)

@@ -6,8 +6,8 @@
 extern "C" {
 #endif

-#define RPC_PROTO_MAJOR_VERSION    5
-#define RPC_PROTO_MINOR_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    6
+#define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0

 #ifdef  __cplusplus
@@ -1724,6 +1724,19 @@ 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);
@@ -1990,14 +2003,6 @@ 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(
@@ -40,6 +40,7 @@ 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;
@@ -83,6 +83,7 @@ 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)
@@ -592,7 +592,18 @@ 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};
        }
-        GGML_ABORT("fatal error");
+        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);
        //return {GGML_BACKEND_SPLIT_AXIS_UNKNOWN, {0}, {1}, 1};
    };

@@ -602,27 +613,40 @@ 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: {
-                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++) {
+                int64_t base_ne_in = 1;
+                for (int dim = 0; dim <= src_ss[0].axis; dim++) {
                    base_ne_in *= tensor->src[0]->ne[dim];
                }
-                base_ne_in /= src_ss[0].nr[0];
+                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.
                int64_t base_ne_out = 1;
                for (int dim = 0; dim < GGML_MAX_DIMS; dim++) {
-                    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};
+                    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};
                    }
-                    if (base_ne_out_next > base_ne_in) {
+                    if (base_ne_out > 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));
            }
@@ -747,14 +771,33 @@ 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(                             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[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(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) {
@@ -792,7 +835,7 @@ 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++) {
@@ -802,6 +845,9 @@ 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;
        }
@@ -922,7 +968,7 @@ 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_generic(src_ss, /*scalar_only =*/ true);
+                split_state = handle_rope(src_ss);
            } break;
            case GGML_OP_CLAMP: {
                split_state = handle_generic(src_ss, /*scalar_only =*/ false);
@@ -986,6 +1032,9 @@ 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: {
@@ -1070,13 +1119,14 @@ 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) {
+                if (tensor->src[i] == nullptr || tensor->src[i] == tensor) {
                    continue;
                }
                if (!srcs_info.empty()) {
                    srcs_info += ", ";
                }
-                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor->src[0], true);
+                const ggml_backend_meta_split_state split_state =
+                        ggml_backend_meta_get_split_state(tensor->src[i], 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;
@@ -1118,7 +1168,6 @@ 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);
 }
@@ -1209,7 +1258,14 @@ 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));
        }
-        t_ij->extra = tensor->extra;
+
+        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;
+        }
+
        for (int i = 0; i < GGML_MAX_SRC; i++) {
            t_ij->src[i] = tensor->src[i];
            if (tensor->src[i] == tensor) {
@@ -1255,6 +1311,108 @@ 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);
@@ -1352,15 +1510,29 @@ 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);
-            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);
+            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);
            }
            for (size_t j = 0; j < n_bufs; j++) {
                ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
-                ggml_backend_tensor_set(simple_tensor, tmp.data(), offset, size);
+                const float * partial = has_contributor_mask && split_state.ne[j] == 0 ? zero.data() : tmp.data();
+                ggml_backend_tensor_set(simple_tensor, partial, offset, size);
            }
        } break;
        default: {
@@ -1488,7 +1660,7 @@ 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   = */ nullptr, // TODO implement
+    /* .memset_tensor   = */ ggml_backend_meta_buffer_memset_tensor,
    /* .set_tensor      = */ ggml_backend_meta_buffer_set_tensor,
    /* .get_tensor      = */ ggml_backend_meta_buffer_get_tensor,
    /* .set_tensor_2d   = */ nullptr,
@@ -1502,6 +1674,16 @@ 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);

@@ -1841,7 +2023,7 @@ 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 = [&](const int i) -> int {
+            auto get_i_delayed_branch = [&](const int i) -> int {
                int id = i; // i_delayed
                int idr = i; // i_delayed return, last safe return value

@@ -1941,6 +2123,62 @@ 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];
@@ -182,6 +182,8 @@ 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);
    }
 }

@@ -576,10 +576,25 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
    endif()

    if (GGML_CPU_KLEIDIAI)
-        message(STATUS "Using KleidiAI optimized kernels if applicable")
+        # 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()

-        # Disable the KleidiAI tests
-        set(KLEIDIAI_BUILD_TESTS  OFF)
+        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")

        # Fetch KleidiAI sources:
        include(FetchContent)
@@ -595,31 +610,49 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            list(APPEND KLEIDIAI_FETCH_ARGS DOWNLOAD_EXTRACT_TIMESTAMP NEW)
        endif()

-        if (CMAKE_VERSION VERSION_GREATER_EQUAL "3.28")
-            FetchContent_Declare(KleidiAI_Download
-                ${KLEIDIAI_FETCH_ARGS}
+        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"
                EXCLUDE_FROM_ALL
            )
-
-            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)
+            if (NOT CMAKE_SKIP_INSTALL_RULES AND
+                (NOT DEFINED BUILD_SHARED_LIBS OR NOT BUILD_SHARED_LIBS))
+                install(TARGETS kleidiai ARCHIVE)
            endif()
        endif()

-        add_compile_definitions(GGML_USE_CPU_KLEIDIAI)
+        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)

        list(APPEND GGML_CPU_SOURCES
            ggml-cpu/kleidiai/kleidiai.cpp
@@ -627,108 +660,6 @@ 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} ${ARCH_DEFINITIONS}")
@@ -0,0 +1,14 @@
+set(BUILD_SHARED_LIBS OFF)
+set(CMAKE_SKIP_INSTALL_RULES TRUE)
+
+add_subdirectory("${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()
@@ -3,44 +3,44 @@
 //

 // KleidiAI micro-kernels
-#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/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_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_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_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/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_common.h"
+#include "kai/kai_common.h"

 #include "simd-mappings.h"

@@ -328,9 +328,8 @@ static void dequantize_row_qsi8cxp(
 }

 static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
-#if defined(__ARM_FEATURE_SME)
    {
-        /* SME GEMM */
+        /* SME2 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,
@@ -351,7 +350,7 @@ 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>,
        },
-        /* SME GEMV */
+        /* SME2 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,
@@ -378,13 +377,13 @@ 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,
+        /* .required_cpu       = */ CPU_FEATURE_SME2 | CPU_FEATURE_FP16,
        /* .lhs_type           = */ GGML_TYPE_F32,
        /* .rhs_type           = */ GGML_TYPE_Q4_0,
        /* .op_type            = */ GGML_TYPE_F32,
    },
    {
-        /* SME GEMM */
+        /* SME2 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,
@@ -404,7 +403,7 @@ 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>,
        },
-        /* SME GEMV */
+        /* SME2 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,
@@ -436,9 +435,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
        /* .rhs_type           = */ GGML_TYPE_F16,
        /* .op_type            = */ GGML_TYPE_F32,
    },
-#endif
 #if defined(__APPLE__)
-#if defined(__ARM_FEATURE_DOTPROD)
    {
        /* DOTPROD GEMM */
        /* .kern_info = */ {
@@ -492,8 +489,6 @@ 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 = */ {
@@ -515,7 +510,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .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 */
+        /* 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,
@@ -542,14 +537,12 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .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,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
        /* .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 = */ {
@@ -603,8 +596,6 @@ 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 = */ {
@@ -626,7 +617,7 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .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 */
+        /* 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,
@@ -653,13 +644,11 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
            /* .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,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
        /* .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 = */ {
@@ -713,15 +702,13 @@ static ggml_kleidiai_kernels gemm_gemv_kernels[] = {
        /* .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)
    {
-        /* SME GEMM */
+        /* SME2 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,
@@ -741,7 +728,7 @@ 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>,
        },
-        /* SME GEMV */
+        /* SME2 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,
@@ -826,8 +813,6 @@ 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 */
        {
@@ -876,13 +861,11 @@ 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,
+        /* .required_cpu       = */ CPU_FEATURE_I8MM | CPU_FEATURE_DOTPROD,
        /* .lhs_type           = */ GGML_TYPE_F32,
        /* .rhs_type           = */ GGML_TYPE_Q8_0,
        /* .op_type            = */ GGML_TYPE_F32,
    },
-#endif
-#if defined(__ARM_FEATURE_DOTPROD)
    {
        /* DOTPROD GEMM */
        {
@@ -936,12 +919,10 @@ 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 */
        {
@@ -1048,7 +1029,6 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
        /* .rhs_type           = */ GGML_TYPE_F32,
        /* .op_type            = */ GGML_TYPE_F32,
    },
-#endif
    { /* Sentinel */ }
 };

@@ -1056,10 +1036,6 @@ 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 &&
@@ -1080,12 +1056,6 @@ 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;
@@ -1094,19 +1064,13 @@ 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) {
+        if ((features & gemm_gemv_kernels[i].required_cpu) == gemm_gemv_kernels[i].required_cpu &&
+            gemm_gemv_kernels[i].rhs_type == GGML_TYPE_Q4_0) {
            kernels = &gemm_gemv_kernels[i];
            break;
        }
    }
-#else
-    GGML_UNUSED(features);
-#endif

    return kernels;
 }
@@ -1114,16 +1078,12 @@ 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;
 }
@@ -1131,16 +1091,11 @@ 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;
 }
@@ -1,4 +1,4 @@
-// SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-FileCopyrightText: Copyright 2025-2026 Arm Limited and/or its affiliates <open-source-office@arm.com>
 // SPDX-License-Identifier: MIT
 //

@@ -12,7 +12,8 @@ enum cpu_feature {
    CPU_FEATURE_I8MM    = 2,
    CPU_FEATURE_SVE     = 4,
    CPU_FEATURE_SME     = 8,
-    CPU_FEATURE_SME2    = 16
+    CPU_FEATURE_SME2    = 16,
+    CPU_FEATURE_FP16    = 32
 };

 inline cpu_feature& operator|=(cpu_feature& lhs, cpu_feature rhs) {
@@ -48,7 +48,7 @@

 #include "kernels.h"

-#include "kai_common.h"
+#include "kai/kai_common.h"

 #define GGML_COMMON_DECL_CPP
 #include "ggml-common.h"
@@ -316,6 +316,7 @@ 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) {
@@ -38,6 +38,7 @@
 #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"
@@ -2326,6 +2327,9 @@ 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;
@@ -4607,8 +4611,8 @@ 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 += " (physical device " + std::to_string(info.devices[device].physical_device) +
-                       ", virtual device " + std::to_string(info.devices[device].virtual_index) + ")";
+        description += " (dev p" + std::to_string(info.devices[device].physical_device) +
+                       "/v" + std::to_string(info.devices[device].virtual_index) + ")";
    }
    return description;
 }
@@ -5245,6 +5249,7 @@ 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:
@@ -1,4 +1,4 @@
-static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal(ggml_type type, int J, bool fallback) {
+static constexpr __host__ __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config_pascal_dp4a(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);
@@ -0,0 +1,273 @@
+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);
+}
@@ -314,7 +314,9 @@ 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) {
-        return false;
+        // 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;
    }

 #ifdef GGML_CUDA_FORCE_MMQ
@@ -213,7 +213,8 @@ 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.cuh"
+#include "mmq-config-pascal-older.cuh"
+#include "mmq-config-pascal-dp4a.cuh"
 #include "mmq-config-ampere.cuh"
 #include "mmq-config-blackwell.cuh"

@@ -247,7 +248,10 @@ 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);
    }
-    return ggml_cuda_mmq_get_config_pascal(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);
 }

 static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_type type, int J, bool fallback) {
@@ -268,8 +272,10 @@ 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(type, J, fallback);
+    return ggml_cuda_mmq_get_config_pascal_older(type, J, fallback);
 #endif // BLACKWELL_MMA_AVAILABLE
 #endif // GGML_USE_HIP
    GGML_UNUSED_VARS(type, J, fallback);
@@ -0,0 +1,85 @@
+#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);
+}
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
@@ -8,60 +8,107 @@
 #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;
+    ggml_tensor * node   { nullptr };
+    htp_op_code   opcode { HTP_OP_INVALID };
+    int32_t       kernel_params[HTP_OP_MAX_KERN_PARAMS] {0};

-    std::vector<ggml_tensor *> fused;
+    std::vector<ggml_tensor *>                fused;
+    std::vector<std::shared_ptr<ggml_tensor>> dummy;

-    htp_op_code opcode = HTP_OP_INVALID;
+    std::vector<const ggml_tensor *> inputs;
+    std::vector<const ggml_tensor *> outputs;
+    std::string                      name;

-    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;
+    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;
    }

-    const ggml_tensor * dst() const {
-        return fused.empty() ? node : fused.back();
+    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();
    }

    void add_fused(ggml_tensor * t, bool extra_dst = false) {
        fused.push_back(t);
-        if (extra_dst) {
-            extra_dsts.push_back(t);
-        }
-    }

-    std::vector<const ggml_tensor *> get_outputs() const {
-        std::vector<const ggml_tensor *> res;
-        if (extra_dsts.empty()) {
-            res.push_back(dst());
+        name += "+";
+        name += ggml_op_desc(t);
+
+        if (extra_dst) {
+            outputs.push_back(t);
        } else {
-            res.push_back(node);
-            for (const auto * x : extra_dsts) {
-                res.push_back(x);
+            outputs.clear();
+            outputs.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);
            }
        }
-        return res;
    }

-    const ggml_tensor * src0() const {
-        return node->src[0];
+    const std::vector<const ggml_tensor *> & get_inputs() const {
+        return inputs;
    }

-    const ggml_tensor * src1() const {
-        return node->src[1];
+    const std::vector<const ggml_tensor *> & get_outputs() const {
+        return outputs;
+    }
+
+    std::string op_name() const {
+        return name;
    }

    bool is_empty() const {
@@ -81,75 +128,6 @@ 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 {
@@ -337,8 +315,7 @@ 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_QKV || node.opcode == HTP_OP_MUL_MAT_FFN ||
-            node.opcode == HTP_OP_MUL_MAT_ADD) {
+            node.opcode == HTP_OP_MUL_MAT_NX || 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;
@@ -43,6 +43,7 @@ add_library(${HTP_LIB} SHARED
    pad-ops.c
    argsort-ops.c
    im2col-ops.c
+    allreduce-ops.c
 )

 target_compile_definitions(${HTP_LIB} PRIVATE
@@ -183,6 +183,53 @@ 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);
@@ -200,20 +247,13 @@ 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];
@@ -223,56 +263,13 @@ 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
-        HVX_Vector y2 = hvx_vec_mul_f32_f32(inner, v_two);
+        // y2 = 2 * inner = inner + inner
+        HVX_Vector y2 = hvx_vec_add_f32_f32(inner, inner);

-        // 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);
+        // Fast sigmoid approximation (2 iterations)
+        HVX_Vector sig2y = hvx_vec_fast_sigmoid_f32_guard_2it(y2, v_one, v_max_exp, v_min_exp);

+        HVX_Vector gelu_x = hvx_vec_mul_f32_f32(x, sig2y);
        vdst[i] = hvx_vec_mul_f32_f32(gelu_x, g);
    }

@@ -285,50 +282,11 @@ 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_mul_f32_f32(inner, v_two);
+        HVX_Vector y2 = hvx_vec_add_f32_f32(inner, inner);

-        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 sig2y = hvx_vec_fast_sigmoid_f32_guard_2it(y2, v_one, v_max_exp, v_min_exp);

+        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);
    }
@@ -0,0 +1,398 @@
+#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;
+}
@@ -0,0 +1,40 @@
+#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 */
@@ -4,6 +4,7 @@

 #include <HAP_farf.h>
 #include <HAP_perf.h>
+#include <qurt_memory.h>

 #include <math.h>
 #include <string.h>
@@ -14,6 +15,7 @@
 #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;
@@ -78,7 +80,7 @@ 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)                                                         \
@@ -179,7 +181,7 @@ 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) {
@@ -232,6 +234,41 @@ 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;

@@ -264,14 +301,11 @@ int op_cpy(struct htp_ops_context * octx) {

    ct.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;

-    worker_callback_t copy_fun;
+    worker_callback_t copy_fun = NULL;
+    bool use_dma = false;

    if (sametype && sameshape) {
-        if (src0->type == HTP_TYPE_F32) {
-            copy_fun = cpy_thread_f32_sameshape;
-        } else {
-            copy_fun = cpy_thread_f16_sameshape;
-        }
+        use_dma = true;
    } else if (sameshape) {
        /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
            copy_fun = cpy_thread_f16_f32_sameshape;
@@ -289,7 +323,28 @@ int op_cpy(struct htp_ops_context * octx) {
        return HTP_STATUS_NO_SUPPORT;
    }

-    worker_pool_run_func(octx->ctx->worker_pool, copy_fun, &ct, n_threads);
+    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);
+    }

    return HTP_STATUS_OK;
 }
@@ -244,17 +244,18 @@ static inline dma_ptr dma_queue_pop(dma_queue * q) {
        return dptr;
    }

-    dma_descriptor_2d * desc = &r->desc[r->pop_idx];
+    dptr = r->dptr[r->pop_idx];
+
+    volatile 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;
@@ -30,6 +30,8 @@
 #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"
@@ -85,12 +87,17 @@ 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;

@@ -214,8 +221,8 @@ 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 = DK * sizeof(__fp16);
-    const size_t size_v_row = DV * sizeof(__fp16);
+    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);

    // Scratchpad buffers for Q, K, V, Mask, and VKQ32 accumulator
    uint8_t * spad_q = factx->spad_q + factx->size_q_block * ith;
@@ -364,6 +371,23 @@ 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
@@ -625,6 +649,12 @@ 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));
@@ -673,6 +703,12 @@ 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));
@@ -1809,6 +1845,8 @@ 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;
@@ -1853,10 +1891,10 @@ 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        = 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);
+    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);

    // Build the VTCM layout once (shared with the host estimator) and place every
    // scratch buffer at its computed offset.
@@ -2348,7 +2386,9 @@ 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 || v->type != HTP_TYPE_F16) {
+    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)) {
        return HTP_STATUS_NO_SUPPORT;
    }

@@ -2364,6 +2404,8 @@ 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();

@@ -12,18 +12,17 @@
 #include "ggml-common.h"
 #include "htp-ctx.h"
 #include "htp-ops.h"
-#include "htp-ops.h"
+#include "htp-tensor.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;
-    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;
+    const struct htp_get_rows_kernel_params * kparams;
+    struct htp_get_rows_vtcm_layout vtcm_layout;
+    uint8_t * vtcm_base;
 };

 #define get_rows_preamble \
@@ -56,102 +55,161 @@ struct get_rows_context {
                                               \
    const uint32_t nr = ne10 * ne11 * ne12;

-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);
+#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_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;
+GET_ROWS_THREAD_ST_FN(int32_t)
+GET_ROWS_THREAD_ST_FN(int64_t)

-    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 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);                                                                                        \
 }

+#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) {
-    get_rows_preamble;
+    const struct htp_get_rows_kernel_params * kparams = (const struct htp_get_rows_kernel_params *) octx->kernel_params;

-    if (octx->src[0]->type != HTP_TYPE_F32) {
+    if (octx->src[0]->type != HTP_TYPE_F32 &&
+        octx->src[0]->type != HTP_TYPE_F16 &&
+        octx->src[0]->type != HTP_TYPE_Q8_0) {
        return HTP_STATUS_NO_SUPPORT;
    }

@@ -167,52 +225,28 @@ 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.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]);
+    grctx.kparams = kparams;
+    grctx.vtcm_base = (uint8_t *)octx->ctx->vtcm_base;

-    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 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);

-        const uint32_t n_threads = MIN(nr, octx->n_threads);
-        grctx.tasks_per_thread = (nr + n_threads - 1) / n_threads;
+    const bool is_i32 = (octx->src[1]->type == HTP_TYPE_I32);

-        worker_pool_run_func(octx->ctx->worker_pool, get_rows_thread_f32_f32_dma, &grctx, 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);
    } else {
-        uint32_t chunks_per_row = 1;
-        uint32_t chunk_size = ne00;
-        uint32_t total_tasks = nr;
-
-        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;
+        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;
        }
-
-        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);
    }
+
+    work_queue_run(octx->ctx->work_queue, q_func, &grctx, kparams->n_threads);
    return HTP_STATUS_OK;
 }
@@ -0,0 +1,77 @@
+#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
@@ -39,17 +39,22 @@ 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);
-    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);
+    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);
    }
 }

@@ -117,8 +117,7 @@ struct htp_context {

 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(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_matmul_nx(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);
@@ -141,5 +140,6 @@ 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 */
@@ -43,13 +43,6 @@ 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,
@@ -58,8 +51,7 @@ enum htp_op_code {
    HTP_OP_DIV = 3,
    HTP_OP_MUL_MAT,
    HTP_OP_MUL_MAT_ID,
-    HTP_OP_MUL_MAT_QKV,
-    HTP_OP_MUL_MAT_FFN,
+    HTP_OP_MUL_MAT_NX,
    HTP_OP_MUL_MAT_ADD,
    HTP_OP_RMS_NORM,
    HTP_OP_RMS_NORM_MUL,
@@ -99,12 +91,15 @@ 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  6    // aka GGML_MAX_SRCS
+#define HTP_OP_MAX_INPUTS  10   // 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
@@ -112,13 +107,16 @@ 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_COMPUTE = (1U << 0), // Tensor buffer temporal compute data (not weights)
-    HTP_TENSOR_DIRTY   = (1U << 1)  // Tensor buffer is dirty and needs to be flushed
+    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)
 };

 // Tensor descriptor
@@ -175,6 +173,7 @@ 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,
@@ -215,6 +214,7 @@ 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
@@ -231,6 +231,7 @@ 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
 };

@@ -79,7 +79,14 @@ 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) continue;
+        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;
+        }

        uint32_t t_start = t->data;
        uint32_t t_end   = t_start + t->size;
@@ -242,7 +249,7 @@ 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_COMPUTE) && is_tensor_dirty(ctx, t)) {
+        if (t && !(t->flags & (HTP_TENSOR_WEIGHT | HTP_TENSOR_FENCE)) && is_tensor_dirty(ctx, t)) {
            dirty_tensors[n_dirty++] = t;
            total_dirty += t->size;
        }
@@ -13,6 +13,15 @@ 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);
@@ -17,9 +17,9 @@

 #define hvx_arith_loop_body(dst_type, src0_type, src1_type, elem_size, vec_store, vec_op) \
    do {                                                                       \
-        dst_type * restrict vdst  = (dst_type *) dst;                          \
-        src0_type * restrict vsrc0 = (src0_type *) src0;                       \
-        src1_type * restrict vsrc1 = (src1_type *) src1;                       \
+        dst_type * vdst  = (dst_type *) dst;                                   \
+        src0_type * vsrc0 = (src0_type *) src0;                                \
+        src1_type * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_aaa(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_aau(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_aua(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_auu(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_uaa(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_uau(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_uua(uint8_t * dst, const uint8_t * src0, const uint8_t * 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 * restrict dst, const uint8_t * restrict src0, const uint8_t * restrict src1, uint32_t n) { \
+static inline void OP_NAME##_uuu(uint8_t * dst, const uint8_t * src0, const uint8_t * src1, uint32_t n) { \
    hvx_arith_loop_body(HVX_UVector, HVX_UVector, HVX_UVector, sizeof(ELEM_TYPE), hvx_vec_store_u, OP_MACRO); \
 } \

@@ -0,0 +1,165 @@
+#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
@@ -18,6 +18,7 @@
 #include <qurt_memory.h>
 #include <remote.h>
 #include <string.h>
+#include <stdatomic.h>

 #include "hex-utils.h"
 #include "hex-dma.h"
@@ -32,6 +33,7 @@
 #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
@@ -46,6 +48,36 @@ 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));
@@ -127,11 +159,7 @@ 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) {
-#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
+                htp_munmap((void *) ctx->mmap[i].base, ctx->mmap[i].size);
                ctx->mmap[i].size = 0;
                ctx->mmap[i].base = NULL;
                ctx->mmap[i].fd   = -1;
@@ -175,18 +203,9 @@ 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);
-#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);
+            void *va = htp_mmap(fd, size);
+            if (va == NULL) {
+                FARF(ERROR, "mmap failed : fd %u size %u", fd, (uint32_t) size);
                return AEE_EFAILED;
            }

@@ -212,11 +231,7 @@ 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);
-#if __HVX_ARCH__ > 73
-            HAP_munmap2((void *) m->base, m->size);
-#else
-            HAP_munmap((void *) m->base, m->size);
-#endif
+            htp_munmap((void *) m->base, m->size);
            m->size   = 0;
            m->base   = NULL;
            m->fd     = -1;
@@ -228,7 +243,7 @@ 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, 1000000u);
+        int err = HAP_compute_res_acquire_cached(ctx->vtcm_rctx, 10000000u);
        if (err != 0) {
            FARF(ERROR, "ggml-hex: failed to acquire VTCM: 0x%08x", (unsigned)err);
            abort();
@@ -692,8 +707,45 @@ 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);
@@ -701,11 +753,8 @@ 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_QKV:
-            return op_matmul_qkv(octx);
-
-        case HTP_OP_MUL_MAT_FFN:
-            return op_matmul_ffn(octx);
+        case HTP_OP_MUL_MAT_NX:
+            return op_matmul_nx(octx);

        case HTP_OP_MUL:
        case HTP_OP_ADD:
@@ -818,12 +867,8 @@ 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(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
+        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);
        m->size = 0;
        m->base = 0;
        m->fd   = -1;
@@ -837,18 +882,9 @@ 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) {
-#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);
+            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);
                abort(); // can't do much else at this point
            }

@@ -856,10 +892,13 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
            m->fd     = b->fd;
            m->size   = b->size;

-            FARF(HIGH, "mmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
+            FARF(ALWAYS, "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) {
@@ -1081,6 +1120,7 @@ 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++) {
@@ -88,6 +88,7 @@ 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;
@@ -463,8 +464,7 @@ 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_qkv,
-    bool is_fused_ffn
+    bool is_fused_nx
 ) {
    size_t src0_sz = 0;
    size_t src1_sz = 0;
@@ -476,44 +476,33 @@ 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_qkv || is_fused_ffn) {
+    if (is_fused_nx) {
        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 src0_sz_per_thread = 0;
-        size_t src2_sz_per_thread = 0;
-        size_t src3_sz_per_thread = 0;
+        size_t weight_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;

-            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);
-            }
+            weight_sz_per_thread = hex_round_up(n_prefetch * tile_row_size, 128);
        } else {
-            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);
-            }
+            weight_sz_per_thread = hex_round_up(n_prefetch * src0_row_size_padded, 128);
        }

-        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 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);

-        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);
-        }
+        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);

-        src0_sz = src0_sz_per_thread * n_threads;
-        src2_sz = src2_sz_per_thread * n_threads;
-        src3_sz = src3_sz_per_thread * n_threads;
+        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;
        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);
@@ -616,8 +605,8 @@ static inline void htp_mm_hvx_vtcm_layout_build(
    }

    size_t off = 0;
-    VTCM_LAYOUT_ALLOC(off, off_src1, src1_sz);
    VTCM_LAYOUT_ALLOC(off, off_src0, src0_sz);
+    VTCM_LAYOUT_ALLOC(off, off_src1, src1_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);
@@ -8,14 +8,20 @@
 #include <math.h>
 #include <string.h>

-#include "hex-dma.h"
+#include "dma-queue.h"
+#include "work-queue.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-ops.h"
+#include "htp-tensor.h"
+#include "htp/set-rows-ops.h"

 #define set_rows_preamble                      \
    const uint32_t ne00 = octx->src[0]->ne[0]; \
@@ -47,116 +53,142 @@
                                               \
    const uint32_t nr  = ne01;

-struct htp_set_rows_context {
+struct set_rows_context {
    struct htp_ops_context * octx;
-    struct fastdiv_values div_ne12;
-    struct fastdiv_values div_ne11;
-    uint32_t src0_nrows_per_thread;
+    const struct htp_set_rows_kernel_params * kparams;
+    struct htp_set_rows_vtcm_layout vtcm_layout;
+    uint8_t * vtcm_base;
 };

-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);
+#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_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(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); })

-    set_rows_preamble;
+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); })

-    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);
-}
+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); })

 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) {
+    if (octx->dst->type != HTP_TYPE_F32 && octx->dst->type != HTP_TYPE_F16 && octx->dst->type != HTP_TYPE_Q8_0) {
        return HTP_STATUS_NO_SUPPORT;
    }

@@ -164,27 +196,27 @@ int op_set_rows(struct htp_ops_context * octx) {
        return HTP_STATUS_NO_SUPPORT;
    }

-    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
-        return HTP_STATUS_OK;
-    }
+    // 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 htp_set_rows_context srctx;
+    struct set_rows_context srctx;
    srctx.octx = octx;
-    srctx.div_ne12 = init_fastdiv_values(ne12);
-    srctx.div_ne11 = init_fastdiv_values(ne11);
+    srctx.kparams = kparams;

-    srctx.src0_nrows_per_thread = (nr + n_threads - 1) / n_threads;
+    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;

-    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;
+    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;
    }

+    work_queue_run(octx->ctx->work_queue, q_func, &srctx, kparams->n_threads);
+
    return HTP_STATUS_OK;
 }
@@ -0,0 +1,74 @@
+#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
@@ -11,6 +11,7 @@ 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
@@ -24,62 +25,119 @@ 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_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+    set(METALLIB_IMPL "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")

    file(MAKE_DIRECTORY "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")

-    # 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(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})

-    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
-    )
+        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")

-    target_sources(ggml-metal PRIVATE "${METALLIB_EMBED_ASM}")
+        # 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})
 else()
-    # copy metal files to bin directory
+    # copy header 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)
-        # 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: 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})
@@ -90,35 +148,46 @@ 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 metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o - |
-                xcrun -sdk macosx metallib        - -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -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.metal
-        DEPENDS ggml-metal.metal ${METALLIB_COMMON}
-        COMMENT "Compiling Metal kernels"
-        )
+        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"
+    )

-    # 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(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
+        DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/kernels/
+        DESTINATION ${CMAKE_INSTALL_BINDIR}/kernels
+        FILES_MATCHING PATTERN "*.metal" PATTERN "*.h"
+    )

-        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()
@@ -1,6 +1,7 @@
 #include "ggml-metal-device.h"

 #include "ggml-metal-impl.h"
+#include "ggml-metal-tuning.h"

 #include "ggml-impl.h"

@@ -17,10 +18,10 @@ 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) {
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices) {
    static std::vector<ggml_metal_device_ptr> devs;

-    devs.emplace_back(ggml_metal_device_init(device));
+    devs.emplace_back(ggml_metal_device_init(device, n_devices));

    return devs.back().get();
 }
@@ -571,7 +572,7 @@ 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)  {
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan(ggml_metal_library_t lib, const ggml_tensor * op, bool tail)  {
    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);

    char base[256];
@@ -579,7 +580,7 @@ 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", ggml_type_name(op->src[0]->type));
+    snprintf(base, 256, "kernel_ssm_scan_%s%s", ggml_type_name(op->src[0]->type), tail ? "_tail" : "");
    snprintf(name, 256, "%s_nsg=%d", base, nsg);

    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
@@ -597,6 +598,27 @@ 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];
@@ -1544,6 +1566,8 @@ 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,
@@ -1559,11 +1583,17 @@ 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);

-    snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
+    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",
            "flash_attn_ext_vec",
            type,
            dk,
-            dv);
+            dv,
+            qne_suffix);

    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
            base,
@@ -129,7 +129,8 @@ 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);
+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_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);
@@ -207,6 +208,8 @@ 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,
@@ -257,8 +260,12 @@ 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];

@@ -277,6 +284,7 @@ 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;
 };
@@ -286,10 +294,10 @@ 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);
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices);
 void ggml_metal_device_free(ggml_metal_device_t dev);

-ggml_metal_device_t ggml_metal_device_get(int device);
+ggml_metal_device_t ggml_metal_device_get(int device, int n_devices);

 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>
@@ -95,8 +95,63 @@ int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_wi
    return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }

+//
+// MTLLibrary collection (one library per op-source, compiled separately)
+//
+
+// Single source of truth for the per-kind metal libraries. The order here
+// defines the enum values and every per-kind table below, so adding a library
+// is a one-line change here (plus adding its source to CMakeLists.txt).
+//   X(suffix, name): name is both the kernels/<name>.metal basename and the
+//   ggml_metallib_<name>_{start,end} embed-symbol stem.
+#define GGML_METAL_LIBS \
+    X(FA,              fa)             \
+    X(MUL_MV,          mul_mv)         \
+    X(MUL_MM,          mul_mm)         \
+    X(QUANTIZE,        quantize)       \
+    X(SOFTMAX,         softmax)        \
+    X(NORM,            norm)           \
+    X(UNARY,           unary)          \
+    X(BINBCAST,        binbcast)       \
+    X(REDUCE,          reduce)         \
+    X(TRI,             tri)            \
+    X(SSM,             ssm)            \
+    X(WKV,             wkv)            \
+    X(GATED_DELTA_NET, gated_delta_net)\
+    X(SOLVE_TRI,       solve_tri)      \
+    X(ROPE,            rope)           \
+    X(CONV,            conv)           \
+    X(UPSCALE,         upscale)        \
+    X(ARGSORT,         argsort)        \
+    X(POOL,            pool)           \
+    X(MISC,            misc)
+
+enum ggml_metal_lib_kind {
+#define X(e, s) GGML_METAL_LIB_##e,
+    GGML_METAL_LIBS
+#undef X
+    GGML_METAL_LIB_COUNT,
+};
+
+static const char * const k_lib_names[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = #s,
+    GGML_METAL_LIBS
+#undef X
+};
+
 struct ggml_metal_library {
-    id<MTLLibrary> obj;
+    // Per-kind compiled libraries. When single_library is true, the whole library
+    // (e.g. a pre-compiled default.metallib or a from-source build) lives at
+    // objs[0] and the remaining slots are nil.
+    id<MTLLibrary> objs[GGML_METAL_LIB_COUNT];
+    bool single_library; // true: combined library at objs[0]; false: per-kind libs in objs[*]
+
+    // Routing table: kernel function name -> objs[] index, populated from each
+    // compiled library's -[MTLLibrary functionNames]. The actual compiled
+    // libraries are the single source of truth for which library owns a kernel,
+    // so adding kernels later requires no manual routing maintenance.
+    // nil in single_library mode (everything resolves to objs[0]).
+    NSMutableDictionary<NSString *, NSNumber *> * fn_to_lib;

    ggml_metal_device_t dev;
    ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
@@ -104,160 +159,376 @@ struct ggml_metal_library {
    NSLock * lock;
 };

-ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
-    id<MTLLibrary> library = nil;
-    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+// Build the fn_to_lib routing table by querying each compiled library's public
+// function names. Call once after all per-kind libraries have been compiled.
+static void ggml_metal_library_build_index(ggml_metal_library_t lib) {
+    @autoreleasepool {
+        NSMutableDictionary<NSString *, NSNumber *> * index = [[NSMutableDictionary alloc] init];
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            for (NSString * fname in [lib->objs[kind] functionNames]) {
+                index[fname] = @(kind);
+            }
+        }
+        lib->fn_to_lib = index;
+    }
+}

-    // load library
-    //
-    // - first check if the library is embedded
-    // - then check if the library is in the bundle
-    // - if not found, load the source and compile it
-    // - if that fails, return NULL
-    //
-    // TODO: move to a function
-    {
-        const int64_t t_start = ggml_time_us();
+// Parse a `#include "name"` line. Returns the quoted name in *include_name on
+// success. Whitespace-tolerant; ignores `#include <...>` (system headers).
+static bool ggml_metal_library_parse_quoted_include(NSString * line, NSString ** include_name) {
+    NSScanner * scanner = [NSScanner scannerWithString:line];
+    scanner.charactersToBeSkipped = [NSCharacterSet whitespaceCharacterSet];

-        NSError * error = nil;
-        NSString * src = nil;
+    if (![scanner scanString:@"#" intoString:NULL] ||
+        ![scanner scanString:@"include" intoString:NULL] ||
+        ![scanner scanString:@"\"" intoString:NULL]) {
+        return false;
+    }

-#if GGML_METAL_EMBED_LIBRARY
-        GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+    NSString * name = nil;
+    if (![scanner scanUpToString:@"\"" intoString:&name]) {
+        return false;
+    }

-        extern const char ggml_metallib_start[];
-        extern const char ggml_metallib_end[];
+    if (include_name) {
+        *include_name = name;
+    }
+    return true;
+}

-        src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
+// Recursively inline `#include "name"` directives. System includes (<...>),
+// `#if/#else/#endif`, and other preprocessor lines are passed through to the
+// Metal compiler unchanged. `#pragma once` is dropped since `seen` already
+// guards against double-inclusion.
+static bool ggml_metal_library_flatten_file(NSMutableString * dst, NSString * path,
+                                            NSArray<NSString *> * search_paths,
+                                            NSMutableSet<NSString *> * seen, NSError ** error) {
+    NSString * key = [path stringByStandardizingPath];
+    if ([seen containsObject:key]) {
+        return true;
+    }
+    [seen addObject:key];

-#ifdef SWIFT_PACKAGE
-        NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
+    NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:error];
+    if (!src) {
+        return false;
+    }

-        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (path_lib == nil) {
-            // Try to find the resource in the directory where the current binary located.
-            NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
-            NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+    NSFileManager * fm = [NSFileManager defaultManager];
+    for (NSString * line in [src componentsSeparatedByString:@"\n"]) {
+        NSString * trimmed = [line stringByTrimmingCharactersInSet:[NSCharacterSet whitespaceCharacterSet]];
+        if ([trimmed isEqualToString:@"#pragma once"]) {
+            continue;
+        }

-            NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
-            if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
-
-                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
-                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
-                    // Optionally, if this is a symlink, try to resolve it.
-                    path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
-                    if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
-                        // It is a relative path, adding the binary directory as directory prefix.
-                        path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
-                    }
-                    if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
-                        // Link to the resource could not be resolved.
-                        path_lib_default = nil;
-                    } else {
-                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
-                    }
+        NSString * include_name = nil;
+        if (ggml_metal_library_parse_quoted_include(line, &include_name)) {
+            NSString * resolved = nil;
+            for (NSString * dir in search_paths) {
+                NSString * candidate = [dir stringByAppendingPathComponent:include_name];
+                if ([fm isReadableFileAtPath:candidate]) {
+                    resolved = candidate;
+                    break;
                }
-            } else {
-                // The resource couldn't be found in the binary's directory.
-                path_lib_default = nil;
            }
-
-            path_lib = path_lib_default;
+            if (!resolved) {
+                if (error) {
+                    NSString * msg = [NSString stringWithFormat:@"could not resolve include \"%@\" from '%@'", include_name, path];
+                    *error = [NSError errorWithDomain:@"ggml-metal-source-flatten" code:1
+                                             userInfo:@{NSLocalizedDescriptionKey: msg}];
+                }
+                return false;
+            }
+            if (!ggml_metal_library_flatten_file(dst, resolved, search_paths, seen, error)) {
+                return false;
+            }
+            continue;
        }

-        if (path_lib != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+        [dst appendString:line];
+        [dst appendString:@"\n"];
+    }

-            library = [device newLibraryWithURL:libURL error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        } else {
-            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+    return true;
+}

-            NSString * path_source;
-            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+static NSString * ggml_metal_library_flatten_source(NSString * path_source, NSError ** error) {
+    // Search paths cover both runtime layout (build/bin/kernels + build/bin)
+    // and source-tree layout (ggml/src/ggml-metal/kernels + ggml/src/ggml-metal + ggml/src).
+    NSString * path_kernels = [path_source stringByDeletingLastPathComponent];
+    NSString * path_base    = [path_kernels stringByDeletingLastPathComponent];
+    NSArray<NSString *> * search_paths = @[
+        path_kernels,
+        path_base,
+        [path_base stringByDeletingLastPathComponent],
+    ];

-            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+    NSMutableString * src = [[NSMutableString alloc] init];
+    NSMutableSet<NSString *> * seen = [NSMutableSet set];

-            if (path_resource) {
-                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+    if (!ggml_metal_library_flatten_file(src, path_source, search_paths, seen, error)) {
+        [src release];
+        return nil;
+    }
+    return src;
+}
+
+// Compile all per-kind libraries in parallel. `source_for_kind` returns the MSL
+// source for a kind (the helper takes ownership and releases it), or nil with
+// *err set on failure. On success the objs[] slots are populated and the routing
+// index is built; on any failure every error is logged and false is returned
+// (the caller is responsible for freeing `res`).
+static bool ggml_metal_library_compile_all(
+        ggml_metal_library_t res,
+        id<MTLDevice> device,
+        NSDictionary * prep,
+        NSString * (^source_for_kind)(int kind, NSError ** err),
+        const char * origin) {
+    const int64_t t_start = ggml_time_us();
+
+    int64_t  * t_per_lib   = calloc(GGML_METAL_LIB_COUNT, sizeof(int64_t));
+    NSError ** err_per_lib = calloc(GGML_METAL_LIB_COUNT, sizeof(NSError *));
+    __block atomic_bool any_failure = false;
+
+    dispatch_group_t group = dispatch_group_create();
+    dispatch_queue_t queue = dispatch_get_global_queue(QOS_CLASS_USER_INITIATED, 0);
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        dispatch_group_async(group, queue, ^{
+
+            const int64_t t0 = ggml_time_us();
+
+            NSError * error = nil;
+
+            NSString * src = source_for_kind(kind, &error);
+            if (!src) {
+                err_per_lib[kind] = [error retain];
+                atomic_store(&any_failure, true);
+                return;
            }

-            if (path_source == nil) {
-                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                path_source = @"ggml-metal.metal";
-            }
+            id<MTLLibrary> lib = nil;

-            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
-
-            src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return nil;
-            }
-        }
-#endif
-
-        if (!library) {
            @autoreleasepool {
-                // dictionary of preprocessor macros
-                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
-
-                if (ggml_metal_device_get_props(dev)->has_bfloat) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
-                }
-
-                if (ggml_metal_device_get_props(dev)->has_tensor) {
-                    [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
-                }
-
-#if GGML_METAL_EMBED_LIBRARY
-                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
-#endif
-
                MTLCompileOptions * options = [MTLCompileOptions new];
                options.preprocessorMacros = prep;

-                //[options setFastMathEnabled:false];
+                lib = [device newLibraryWithSource:src options:options error:&error];

-                library = [device newLibraryWithSource:src options:options error:&error];
-                if (error) {
-                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                    return nil;
-                }
-
-#if !__has_feature(objc_arc)
                [options release];
-#endif
+
+                // retain the error before the autorelease pool drains it
+                if (!lib) {
+                    err_per_lib[kind] = [error retain];
+                }
+            }
+
+            [src release];
+
+            t_per_lib[kind] = ggml_time_us() - t0;
+
+            if (!lib) {
+                atomic_store(&any_failure, true);
+                return;
+            }
+
+            res->objs[kind] = lib;
+        });
+    }
+    dispatch_group_wait(group, DISPATCH_TIME_FOREVER);
+    dispatch_release(group);
+
+    const bool ok = !atomic_load(&any_failure);
+
+    if (ok) {
+        const int64_t t_total = ggml_time_us() - t_start;
+        int64_t t_max = 0;
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            GGML_LOG_DEBUG("%s: compiled '%s' library in %.3f sec\n",
+                           __func__, k_lib_names[kind], t_per_lib[kind] / 1e6);
+            if (t_per_lib[kind] > t_max) t_max = t_per_lib[kind];
+        }
+        GGML_LOG_INFO("%s: loaded %d libraries from %s in %.3f sec (max single = %.3f sec)\n",
+                      __func__, GGML_METAL_LIB_COUNT, origin, t_total / 1e6, t_max / 1e6);
+
+        ggml_metal_library_build_index(res);
+    } else {
+        for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+            if (err_per_lib[kind]) {
+                GGML_LOG_ERROR("%s: failed to build '%s' library: %s\n", __func__,
+                               k_lib_names[kind], [[err_per_lib[kind] description] UTF8String]);
+                [err_per_lib[kind] release];
            }
        }
-
-#if GGML_METAL_EMBED_LIBRARY
-        [src release];
-#endif // GGML_METAL_EMBED_LIBRARY
-
-        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
    }

-    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
+    free(err_per_lib);
+    free(t_per_lib);

-    res->obj       = library;
+    return ok;
+}
+
+ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
+
+    ggml_metal_library_t res = calloc(1, sizeof(struct ggml_metal_library));
    res->dev       = dev;
    res->pipelines = ggml_metal_pipelines_init();
    res->lock      = [NSLock new];

+    // shared MTLCompileOptions preprocessor macros (matches the build-time defines)
+    NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+    if (ggml_metal_device_get_props(dev)->has_bfloat) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_BF16"];
+    }
+    if (ggml_metal_device_get_props(dev)->has_tensor) {
+        [prep setObject:@"1" forKey:@"GGML_METAL_HAS_TENSOR"];
+    }
+#if GGML_METAL_EMBED_LIBRARY
+    [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
+#if GGML_METAL_EMBED_LIBRARY
+    GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+    // start/end symbols emitted by CMake (see CMakeLists.txt), one pair per kind
+#define X(e, s) extern const char ggml_metallib_##s##_start[]; extern const char ggml_metallib_##s##_end[];
+    GGML_METAL_LIBS
+#undef X
+
+    static const char * const lib_start[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_start,
+    GGML_METAL_LIBS
+#undef X
+    };
+    static const char * const lib_end[GGML_METAL_LIB_COUNT] = {
+#define X(e, s) [GGML_METAL_LIB_##e] = ggml_metallib_##s##_end,
+    GGML_METAL_LIBS
+#undef X
+    };
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            (void) err;
+            return [[NSString alloc] initWithBytes:lib_start[kind]
+                                            length:(lib_end[kind] - lib_start[kind])
+                                          encoding:NSUTF8StringEncoding];
+        }, "embedded data");
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
    return res;
+#else
+#ifdef SWIFT_PACKAGE
+    NSBundle * bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+    NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+    const int64_t t_start = ggml_time_us();
+
+    NSError * error = nil;
+    NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+    if (path_lib == nil) {
+        // Try to find the resource in the directory where the current binary located.
+        NSString * bin_cur = [[NSProcessInfo processInfo] arguments][0];
+        NSString * bin_dir = [bin_cur stringByDeletingLastPathComponent];
+
+        NSString * path_lib_default = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+        if ([[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+            GGML_LOG_INFO("%s: found '%s'\n", __func__, [path_lib_default UTF8String]);
+
+            NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:path_lib_default error:&error];
+            if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                // Optionally, if this is a symlink, try to resolve it.
+                path_lib_default = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:path_lib_default error:&error];
+                if (path_lib_default && [path_lib_default length] > 0 && ![[path_lib_default substringToIndex:1] isEqualToString:@"/"]) {
+                    // It is a relative path, adding the binary directory as directory prefix.
+                    path_lib_default = [NSString pathWithComponents:@[bin_dir, path_lib_default]];
+                }
+                if (!path_lib_default || ![[NSFileManager defaultManager] isReadableFileAtPath:path_lib_default]) {
+                    // Link to the resource could not be resolved.
+                    path_lib_default = nil;
+                } else {
+                    GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [path_lib_default UTF8String]);
+                }
+            }
+        } else {
+            // The resource couldn't be found in the binary's directory.
+            path_lib_default = nil;
+        }
+
+        path_lib = path_lib_default;
+    }
+
+    if (path_lib != nil) {
+        // pre-compiled library found: a single combined default.metallib
+        NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+        GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+        res->objs[0]        = [device newLibraryWithURL:libURL error:&error];
+        res->single_library = true;
+        if (!res->objs[0]) {
+            GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+            ggml_metal_library_free(res);
+            return NULL;
+        }
+
+        GGML_LOG_INFO("%s: loaded in %.3f sec\n", __func__, (ggml_time_us() - t_start) / 1e6);
+        return res;
+    }
+
+    // no pre-compiled metallib: fall back to compiling each kernel source separately
+    GGML_LOG_INFO("%s: default.metallib not found, loading kernel sources\n", __func__);
+
+    NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+    if (path_resource) {
+        GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, [path_resource UTF8String]);
+    }
+
+    // resolve each kind's source path up front (file lookup/logging stays on the calling thread)
+    NSString ** path_per_kind = calloc(GGML_METAL_LIB_COUNT, sizeof(NSString *));
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        NSString * rel = [NSString stringWithFormat:@"kernels/%s.metal", k_lib_names[kind]];
+
+        NSString * path_source = nil;
+        if (path_resource) {
+            path_source = [path_resource stringByAppendingPathComponent:rel];
+        } else {
+            NSString * stem = [NSString stringWithFormat:@"kernels/%s", k_lib_names[kind]];
+            path_source = [bundle pathForResource:stem ofType:@"metal"];
+        }
+
+        if (path_source == nil || ![[NSFileManager defaultManager] isReadableFileAtPath:path_source]) {
+            GGML_LOG_WARN("%s: could not locate %s in bundle, falling back to cwd\n", __func__, [rel UTF8String]);
+            path_source = rel;
+        }
+
+        GGML_LOG_DEBUG("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+        path_per_kind[kind] = [path_source retain];
+    }
+
+    const bool ok = ggml_metal_library_compile_all(res, device, prep,
+        ^NSString * (int kind, NSError ** err) {
+            return ggml_metal_library_flatten_source(path_per_kind[kind], err);
+        }, "source");
+
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        [path_per_kind[kind] release];
+    }
+    free(path_per_kind);
+
+    if (!ok) {
+        ggml_metal_library_free(res);
+        return NULL;
+    }
+
+    return res;
+#endif
 }

 ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev, const char * source, bool verbose) {
@@ -319,10 +590,11 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
        return NULL;
    }

-    res->obj       = library;
-    res->dev       = dev;
-    res->pipelines = ggml_metal_pipelines_init();
-    res->lock      = [NSLock new];
+    res->objs[0]        = library;
+    res->single_library = true;
+    res->dev            = dev;
+    res->pipelines      = ggml_metal_pipelines_init();
+    res->lock           = [NSLock new];

    return res;
 }
@@ -332,8 +604,14 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
        return;
    }

-    if (lib->obj) {
-        [lib->obj release];
+    for (int kind = 0; kind < GGML_METAL_LIB_COUNT; ++kind) {
+        if (lib->objs[kind]) {
+            [lib->objs[kind] release];
+        }
+    }
+
+    if (lib->fn_to_lib) {
+        [lib->fn_to_lib release];
    }

    ggml_metal_pipelines_free(lib->pipelines);
@@ -394,11 +672,28 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_

        GGML_LOG_DEBUG("%s: compiling pipeline: base = '%s', name = '%s'\n", __func__, base, name);

+        // route to the library that actually defines this kernel; fn_to_lib is
+        // built from -[MTLLibrary functionNames] so it's always in sync
+        int lib_idx = 0;
+        if (!lib->single_library) {
+            NSNumber * idx = lib->fn_to_lib[base_func];
+            if (!idx) {
+                [lib->lock unlock];
+
+                GGML_LOG_ERROR("%s: kernel not found in any metal library: base = '%s', name = '%s'\n", __func__, base, name);
+
+                return res;
+            }
+            lib_idx = [idx intValue];
+        }
+
+        id<MTLLibrary> mtl_lib = lib->objs[lib_idx];
+
        id<MTLFunction> mtl_function;
        if (!cv) {
-            mtl_function = [lib->obj newFunctionWithName:base_func];
+            mtl_function = [mtl_lib newFunctionWithName:base_func];
        } else {
-            mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
+            mtl_function = [mtl_lib newFunctionWithName:base_func constantValues:cv->obj error:&error];
        }
        if (!mtl_function) {
            [lib->lock unlock];
@@ -667,6 +962,34 @@ void ggml_metal_rsets_free(ggml_metal_rsets_t rsets) {
    free(rsets);
 }

+static const struct {
+    const char *              name;
+    const char *              token;
+    enum ggml_metal_device_id id;
+} k_metal_devices[] = {
+#define DEV(name, id) { name, #id, id }
+    DEV("M1",       GGML_METAL_DEVICE_M1),
+    DEV("M1 Pro",   GGML_METAL_DEVICE_M1_PRO),
+    DEV("M1 Max",   GGML_METAL_DEVICE_M1_MAX),
+    DEV("M1 Ultra", GGML_METAL_DEVICE_M1_ULTRA),
+    DEV("M2",       GGML_METAL_DEVICE_M2),
+    DEV("M2 Pro",   GGML_METAL_DEVICE_M2_PRO),
+    DEV("M2 Max",   GGML_METAL_DEVICE_M2_MAX),
+    DEV("M2 Ultra", GGML_METAL_DEVICE_M2_ULTRA),
+    DEV("M3",       GGML_METAL_DEVICE_M3),
+    DEV("M3 Pro",   GGML_METAL_DEVICE_M3_PRO),
+    DEV("M3 Max",   GGML_METAL_DEVICE_M3_MAX),
+    DEV("M3 Ultra", GGML_METAL_DEVICE_M3_ULTRA),
+    DEV("M4",       GGML_METAL_DEVICE_M4),
+    DEV("M4 Pro",   GGML_METAL_DEVICE_M4_PRO),
+    DEV("M4 Max",   GGML_METAL_DEVICE_M4_MAX),
+    DEV("M5",       GGML_METAL_DEVICE_M5),
+    DEV("M5 Pro",   GGML_METAL_DEVICE_M5_PRO),
+    DEV("M5 Max",   GGML_METAL_DEVICE_M5_MAX),
+    DEV("M5 Ultra", GGML_METAL_DEVICE_M5_ULTRA),
+#undef DEV
+};
+
 static enum ggml_metal_device_id ggml_metal_device_id_parse(const char * name) {
    if (!name) {
        return GGML_METAL_DEVICE_GENERIC;
@@ -678,40 +1001,24 @@ static enum ggml_metal_device_id ggml_metal_device_id_parse(const char * name) {
    }
    const char * suffix = name + sizeof(prefix) - 1;

-    static const struct {
-        const char * name;
-        enum ggml_metal_device_id id;
-    } table[] = {
-        {"M1",       GGML_METAL_DEVICE_M1},
-        {"M1 Pro",   GGML_METAL_DEVICE_M1_PRO},
-        {"M1 Max",   GGML_METAL_DEVICE_M1_MAX},
-        {"M1 Ultra", GGML_METAL_DEVICE_M1_ULTRA},
-        {"M2",       GGML_METAL_DEVICE_M2},
-        {"M2 Pro",   GGML_METAL_DEVICE_M2_PRO},
-        {"M2 Max",   GGML_METAL_DEVICE_M2_MAX},
-        {"M2 Ultra", GGML_METAL_DEVICE_M2_ULTRA},
-        {"M3",       GGML_METAL_DEVICE_M3},
-        {"M3 Pro",   GGML_METAL_DEVICE_M3_PRO},
-        {"M3 Max",   GGML_METAL_DEVICE_M3_MAX},
-        {"M3 Ultra", GGML_METAL_DEVICE_M3_ULTRA},
-        {"M4",       GGML_METAL_DEVICE_M4},
-        {"M4 Pro",   GGML_METAL_DEVICE_M4_PRO},
-        {"M4 Max",   GGML_METAL_DEVICE_M4_MAX},
-        {"M5",       GGML_METAL_DEVICE_M5},
-        {"M5 Pro",   GGML_METAL_DEVICE_M5_PRO},
-        {"M5 Max",   GGML_METAL_DEVICE_M5_MAX},
-        {"M5 Ultra", GGML_METAL_DEVICE_M5_ULTRA},
-    };
-
-    for (size_t i = 0; i < sizeof(table)/sizeof(table[0]); ++i) {
-        if (strcmp(suffix, table[i].name) == 0) {
-            return table[i].id;
+    for (size_t i = 0; i < sizeof(k_metal_devices)/sizeof(k_metal_devices[0]); ++i) {
+        if (strcmp(suffix, k_metal_devices[i].name) == 0) {
+            return k_metal_devices[i].id;
        }
    }
    return GGML_METAL_DEVICE_GENERIC;
 }

-ggml_metal_device_t ggml_metal_device_init(int device) {
+const char * ggml_metal_device_id_token(enum ggml_metal_device_id id) {
+    for (size_t i = 0; i < sizeof(k_metal_devices)/sizeof(k_metal_devices[0]); ++i) {
+        if (k_metal_devices[i].id == id) {
+            return k_metal_devices[i].token;
+        }
+    }
+    return "GGML_METAL_DEVICE_GENERIC";
+}
+
+ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) {
    ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));

    assert(dev != NULL);
@@ -728,6 +1035,12 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
            dev->addr_virt = 0x000000400ULL;

            dev->props.device = device;
+
+            // the Metal backend uses the system default device as the single physical device;
+            // additional (virtual) devices are emulated on top of it via GGML_METAL_DEVICES
+            dev->props.device_phys = 0;
+            dev->props.device_virt = device;
+
            dev->props.has_simdgroup_reduction  = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
            dev->props.has_simdgroup_reduction |= [dev->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];

@@ -891,7 +1204,13 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
            }

            snprintf(dev->props.name, sizeof(dev->props.name), "%s%d", "MTL", device);
-            snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", [[dev->mtl_device name] UTF8String]);
+            const char * gpu_name = [[dev->mtl_device name] UTF8String];
+            if (n_devices > 1) {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s (dev p%d/v%d)",
+                         gpu_name, dev->props.device_phys, dev->props.device_virt);
+            } else {
+                snprintf(dev->props.desc, sizeof(dev->props.desc), "%s", gpu_name);
+            }

            dev->library = ggml_metal_library_init(dev);
            if (!dev->library) {
@@ -913,7 +1232,8 @@ ggml_metal_device_t ggml_metal_device_init(int device) {
            {
                for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
                    if ([dev->mtl_device supportsFamily:i]) {
-                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+                        dev->props.gpu_family = i - (int) MTLGPUFamilyApple1 + 1;
+                        GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, dev->props.gpu_family, i);
                        break;
                    }
                }
@@ -158,6 +158,10 @@
 #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
@@ -893,6 +897,8 @@ 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;
@@ -7,6 +7,7 @@
 #include "ggml-metal-impl.h"
 #include "ggml-metal-common.h"
 #include "ggml-metal-device.h"
+#include "ggml-metal-tuning.h"

 #include <cassert>
 #include <algorithm>
@@ -1676,6 +1677,7 @@ 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);
@@ -1721,6 +1723,8 @@ 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),
@@ -1750,26 +1754,53 @@ int ggml_metal_op_ssm_scan(ggml_metal_op_t ctx, int idx) {
        /*.nb0          =*/ nb0,
    };

-    auto pipeline = ggml_metal_library_get_pipeline_ssm_scan(lib, op);
+    constexpr int64_t CHUNK = OP_SSM_SCAN_SSD_CS;

-    GGML_ASSERT(d_state <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
+    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

-    const size_t smem = pipeline.smem;
+    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);

-    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_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_threadgroup_memory_size(enc, smem, 0);
+    if (!use_mma) {
+        dispatch(ggml_metal_library_get_pipeline_ssm_scan(lib, op, false), d_state, d_inner);
+        return 1;
+    }

-    ggml_metal_encoder_dispatch_threadgroups(enc, d_inner, n_head, n_seqs, d_state, 1, 1);
+    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);
+    }

    return 1;
 }
@@ -3346,12 +3377,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
    } else {
        // half4x4 kernel
-        const int nqptg = OP_FLASH_ATTN_EXT_VEC_NQPSG; // queries per threadgroup
+        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 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  == 0);
+        GGML_ASSERT(nqptg == 1 || nqptg == 2 || nqptg == 4);  // only instantiated Q values
        GGML_ASSERT(ncpsg  % 32 == 0);

        bool need_sync = false;
@@ -3410,7 +3447,7 @@ 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))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD(((GGML_PAD(ne00, 128) + 4*ncpsg + 2*GGML_PAD(ne20, 128))*(nsg)*nqptg)*(sizeof(float)/2), 16))

        int64_t nsg = 1;

@@ -3430,6 +3467,12 @@ 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;

@@ -3468,7 +3511,7 @@ 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, 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, nqptg, cfg.NE, nsg, nwg, use_kv_f16, ns10, ns20);

        GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

@@ -0,0 +1,77 @@
+#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
@@ -6,6 +6,7 @@
 #include "ggml-metal-device.h"
 #include "ggml-metal-context.h"
 #include "ggml-metal-ops.h"
+#include "ggml-metal-tuning.h"

 #include <mutex>
 #include <string>
@@ -203,6 +204,11 @@ 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;
@@ -870,10 +876,55 @@ 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;

@@ -891,7 +942,7 @@ 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),
+        /* .context = */ ggml_metal_device_get(device, g_devices),
    };
 }

@@ -0,0 +1,232 @@
+#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>;
@@ -0,0 +1,228 @@
+#include "common.h"
+
+// OP: 0 - add, 1 - sub, 2 - mul, 3 - div
+constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
+constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
+constant bool  FC_bin_rb [[function_constant(FC_BIN + 2)]];
+constant bool  FC_bin_cb [[function_constant(FC_BIN + 3)]];
+
+template <typename T0, typename T1, typename T>
+kernel void kernel_bin_fuse_impl(
+        constant ggml_metal_kargs_bin & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+#define FC_OP FC_bin_op
+#define FC_F  FC_bin_f
+#define FC_RB FC_bin_rb
+#define FC_CB FC_bin_cb
+
+    if (FC_RB) {
+        // row broadcast
+        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
+        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+
+        device const T0 * src0_row = (device const T0 *) (src0);
+        device       T  * dst_row  = (device       T  *) (dst);
+
+        if (FC_F == 1) {
+            device const T1 * src1_row = (device const T1 *) (src1 + args.o1[0]);
+
+            if (FC_OP == 0) {
+                dst_row[i0] = src0_row[i0] + src1_row[i1];
+            }
+
+            if (FC_OP == 1) {
+                dst_row[i0] = src0_row[i0] - src1_row[i1];
+            }
+
+            if (FC_OP == 2) {
+                dst_row[i0] = src0_row[i0] * src1_row[i1];
+            }
+
+            if (FC_OP == 3) {
+                dst_row[i0] = src0_row[i0] / src1_row[i1];
+            }
+        } else {
+            T0 res = src0_row[i0];
+
+            if (FC_OP == 0) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res += ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 1) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res -= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 2) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res *= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            if (FC_OP == 3) {
+                FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                    res /= ((device const T1 *) (src1 + args.o1[j]))[i1];
+                }
+            }
+
+            dst_row[i0] = res;
+        }
+    } else {
+        const int i03 = tgpig.z;
+        const int i02 = tgpig.y;
+        const int i01 = tgpig.x;
+
+        if (i01 >= args.ne01) {
+            return;
+        }
+
+        const int i13 = i03%args.ne13;
+        const int i12 = i02%args.ne12;
+        const int i11 = i01%args.ne11;
+
+        device const T0 * src0_ptr = (device const T0 *) (src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs);
+        device       T  * dst_ptr  = (device       T  *) (dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs);
+
+        if (FC_F == 1) {
+            device const T1 * src1_ptr = (device const T1 *) (src1 + args.o1[0] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                if (FC_OP == 0) {
+                    dst_ptr[i0] = src0_ptr[i0] + src1_ptr[i10];
+                }
+
+                if (FC_OP == 1) {
+                    dst_ptr[i0] = src0_ptr[i0] - src1_ptr[i10];
+                }
+
+                if (FC_OP == 2) {
+                    dst_ptr[i0] = src0_ptr[i0] * src1_ptr[i10];
+                }
+
+                if (FC_OP == 3) {
+                    dst_ptr[i0] = src0_ptr[i0] / src1_ptr[i10];
+                }
+            }
+        } else {
+            device const T1 * src1_ptr[8];
+            FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                src1_ptr[j] = (device const T1 *) (src1 + args.o1[j] + i13*args.nb13 + i12*args.nb12 + i11*args.nb11);
+            }
+
+            for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+                const int i10 = FC_CB ? i0%args.ne10 : i0;
+
+                T res = src0_ptr[i0];
+
+                if (FC_OP == 0) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res += src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 1) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res -= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 2) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res *= src1_ptr[j][i10];
+                    }
+                }
+
+                if (FC_OP == 3) {
+                    FOR_UNROLL (short j = 0; j < FC_F; ++j) {
+                        res /= src1_ptr[j][i10];
+                    }
+                }
+
+                dst_ptr[i0] = res;
+            }
+        }
+    }
+
+#undef FC_OP
+#undef FC_F
+#undef FC_RB
+#undef FC_CB
+}
+
+typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
+
+template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
+template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half,   half,   half>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half4,  half4,  half4>;
+
+kernel void kernel_add_id(
+        constant ggml_metal_kargs_add_id & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i1 = tgpig.x;
+    const int i2 = tgpig.y;
+
+    const int i11 = *((device const int32_t *) (src2 + i1*sizeof(int32_t) + i2*args.nb21));
+
+    const size_t nb1 = args.ne0 * sizeof(float);
+    const size_t nb2 = args.ne1 * nb1;
+
+    device       float * dst_row  = (device       float *)((device char *)dst  +  i1*nb1       + i2*nb2);
+    device const float * src0_row = (device const float *)((device char *)src0 +  i1*args.nb01 + i2*args.nb02);
+    device const float * src1_row = (device const float *)((device char *)src1 + i11*args.nb11);
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
+
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_repeat_bf16")]] kernel kernel_repeat_t kernel_repeat<bfloat>;
+#endif
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
@@ -0,0 +1,126 @@
+#pragma once
+
+#include "ggml-metal-impl.h"
+
+#include <metal_stdlib>
+
+#ifdef GGML_METAL_HAS_TENSOR
+#include <metal_tensor>
+
+#include <MetalPerformancePrimitives/MetalPerformancePrimitives.h>
+#endif
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define PAD2(x, n) (((x) + (n) - 1) & ~((n) - 1))
+
+#define FOR_UNROLL(x) _Pragma("clang loop unroll(full)") for (x)
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+// ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+//
+// cmd:
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/kernels/<src>.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/kernels/<src>.metal
+//
+#if __METAL_VERSION__ < 310 && defined(GGML_METAL_HAS_BF16)
+#undef GGML_METAL_HAS_BF16
+#endif
+
+#if defined(GGML_METAL_HAS_BF16)
+typedef matrix<bfloat, 4, 4> bfloat4x4;
+typedef matrix<bfloat, 2, 4> bfloat2x4;
+#endif
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+constexpr constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static inline float e8m0_to_fp32(uint8_t x) {
+    uint32_t bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint32_t) x << 23;
+    }
+
+    return as_type<float>(bits);
+}
+
+static inline float dot(float x, float y) {
+    return x*y;
+}
+
+static inline float sum(float x) {
+    return x;
+}
+
+static inline float sum(float4 x) {
+    return x[0] + x[1] + x[2] + x[3];
+}
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+constant float SQRT_2_INV      = 0.70710678118654752440084436210484f;
+
+// based on Abramowitz and Stegun formula 7.1.26 or similar Hastings' approximation
+// ref: https://www.johndcook.com/blog/python_erf/
+constant float p_erf  = 0.3275911f;
+constant float a1_erf = 0.254829592f;
+constant float a2_erf = -0.284496736f;
+constant float a3_erf = 1.421413741f;
+constant float a4_erf = -1.453152027f;
+constant float a5_erf = 1.061405429f;
+
+template<typename T>
+inline T erf_approx(T x) {
+    T sign_x = sign(x);
+    x = fabs(x);
+    T t = 1.0f / (1.0f + p_erf * x);
+    T y = 1.0f - (((((a5_erf * t + a4_erf) * t) + a3_erf) * t + a2_erf) * t + a1_erf) * t * exp(-x * x);
+    return sign_x * y;
+}
+
+template<typename T> T elu_approx(T x);
+
+template<> inline float elu_approx<float>(float x) {
+    return (x > 0.f) ? x : (exp(x) - 1);
+}
+
+template<> inline float4 elu_approx<float4>(float4 x) {
+    float4 res;
+
+    res[0] = (x[0] > 0.0f) ? x[0] : (exp(x[0]) - 1.0f);
+    res[1] = (x[1] > 0.0f) ? x[1] : (exp(x[1]) - 1.0f);
+    res[2] = (x[2] > 0.0f) ? x[2] : (exp(x[2]) - 1.0f);
+    res[3] = (x[3] > 0.0f) ? x[3] : (exp(x[3]) - 1.0f);
+
+    return res;
+}
@@ -0,0 +1,723 @@
+#include "common.h"
+
+typedef void (im2col_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
+
+          int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*args.s0 + ikw*args.d0 - args.p0;
+    const int64_t iih = ioh*args.s1 + ikh*args.d1 - args.p1;
+
+    int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*args.CHW + (iic*(KH*KW) + ikh*KW + ikw);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        while (in < args.N) {
+            pdst[offset_dst] = 0.0f;
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+
+            in += ntg[0];
+        }
+    } else {
+        int64_t offset_src = in*args.ofs0 + iic*args.ofs1 + iih*args.IW + iiw;
+
+        while (in < args.N) {
+            pdst[offset_dst] = x[offset_src];
+
+            offset_dst += ntg[0]*args.CHW*OH*OW;
+            offset_src += ntg[0]*args.ofs0;
+
+            in += ntg[0];
+        }
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+// TODO: optimize
+typedef void (im2col_ext_t)(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col_ext(
+        constant ggml_metal_kargs_im2col & args,
+        device const float * x,
+        device        char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
+    const int64_t KHW = (int64_t)args.KHW;
+
+    const int64_t d   = tgpig[0] / args.CHW;
+    const int64_t chw = tgpig[0] % args.CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
+
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    if (tpitg_0 >= args.N) {
+        return;
+    }
+
+    const int64_t tpitg_1 = HW / args.KW;
+    const int64_t tpitg_2 = HW % args.KW;
+
+    const int64_t iiw = tgpig[2] * args.s0 + tpitg_2 * args.d0 - args.p0;
+    const int64_t iih = tgpig[1] * args.s1 + tpitg_1 * args.d1 - args.p1;
+
+    const int64_t offset_dst =
+        (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * args.CHW +
+        (tgpig_0 * KHW + tpitg_1 * args.KW + tpitg_2);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= args.IH || iiw < 0 || iiw >= args.IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = tpitg_0 * args.ofs0 + tgpig_0 * args.ofs1;
+        pdst[offset_dst] = x[offset_src + iih * args.IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
+template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
+
+template <typename T>
+kernel void kernel_col2im_1d(
+        constant ggml_metal_kargs_col2im_1d & args,
+        device const T * col,
+        device       T * dst,
+        uint         tgpig [[threadgroup_position_in_grid]],
+        uint         tpitg [[thread_position_in_threadgroup]],
+        uint         ntg   [[threads_per_threadgroup]]) {
+
+    const int idx = tgpig * ntg + tpitg;
+    if (idx >= args.T_out * args.OC) {
+        return;
+    }
+
+    const int t_out = idx % args.T_out;
+    const int oc    = idx / args.T_out;
+    const int t_abs = t_out + args.p0;  // absolute position in uncropped signal
+
+    int t_in_min = (t_abs - args.K + args.s0) / args.s0;  // ceil((t_abs - K + 1) / s0)
+    if (t_in_min < 0) {
+        t_in_min = 0;
+    }
+    int t_in_max = t_abs / args.s0;
+    if (t_in_max >= args.T_in) {
+        t_in_max = args.T_in - 1;
+    }
+
+    float sum = 0.0f;
+    for (int t_in = t_in_min; t_in <= t_in_max; t_in++) {
+        const int k = t_abs - t_in * args.s0;
+        sum += float(col[(oc * args.K + k) + t_in * args.K_OC]);
+    }
+
+    dst[t_out + oc * args.T_out] = T(sum);
+}
+
+template [[host_name("kernel_col2im_1d_f32")]]  kernel void kernel_col2im_1d<float>(constant ggml_metal_kargs_col2im_1d &, device const float *, device float *, uint, uint, uint);
+template [[host_name("kernel_col2im_1d_f16")]]  kernel void kernel_col2im_1d<half>(constant ggml_metal_kargs_col2im_1d &, device const half *, device half *, uint, uint, uint);
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_col2im_1d_bf16")]] kernel void kernel_col2im_1d<bfloat>(constant ggml_metal_kargs_col2im_1d &, device const bfloat *, device bfloat *, uint, uint, uint);
+#endif
+
+template <typename TK>
+kernel void kernel_conv_2d(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const uint threads_per_tg = ntg.x * ntg.y * ntg.z;
+    const uint tg_index = (tgpig.z * tgpg.y + tgpig.y) * tgpg.x + tgpig.x;
+    const uint local_thread = tpitg.z * (ntg.x * ntg.y) + tpitg.y * ntg.x + tpitg.x;
+    const uint thread_index = tg_index * threads_per_tg + local_thread;
+    const uint64_t total_threads = (uint64_t) threads_per_tg * tgpg.x * tgpg.y * tgpg.z;
+    const uint64_t total_outputs = (uint64_t) args.N * args.OC * args.OH * args.OW;
+
+    for (uint64_t index = thread_index; index < total_outputs; index += total_threads) {
+        uint64_t tmp = index;
+
+        const int32_t ow = tmp % args.OW; tmp /= args.OW;
+        const int32_t oh = tmp % args.OH; tmp /= args.OH;
+        const int32_t oc = tmp % args.OC; tmp /= args.OC;
+        const int32_t  n = tmp;
+
+        float acc = 0.0f;
+
+        const int32_t base_x = ow*args.s0 - args.p0;
+        const int32_t base_y = oh*args.s1 - args.p1;
+
+        int32_t ky_start = 0;
+        if (base_y < 0) {
+            ky_start = (-base_y + args.d1 - 1)/args.d1;
+        }
+        int32_t ky_end = args.KH;
+        const int32_t y_max = args.IH - 1 - base_y;
+        if (y_max < 0) {
+            ky_end = ky_start;
+        } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+            ky_end = min(ky_end, y_max/args.d1 + 1);
+        }
+
+        int32_t kx_start = 0;
+        if (base_x < 0) {
+            kx_start = (-base_x + args.d0 - 1)/args.d0;
+        }
+        int32_t kx_end = args.KW;
+        const int32_t x_max = args.IW - 1 - base_x;
+        if (x_max < 0) {
+            kx_end = kx_start;
+        } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+            kx_end = min(kx_end, x_max/args.d0 + 1);
+        }
+
+        if (ky_start < ky_end && kx_start < kx_end) {
+            const uint64_t src_base_n = (uint64_t) n  * args.nb13;
+            const uint64_t w_base_oc  = (uint64_t) oc * args.nb03;
+
+            for (int32_t ic = 0; ic < args.IC; ++ic) {
+                const uint64_t src_base_nc = src_base_n + (uint64_t) ic * args.nb12;
+                const uint64_t w_base_ocic = w_base_oc  + (uint64_t) ic * args.nb02;
+
+                for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+                    const int32_t iy = base_y + ky*args.d1;
+                    const uint64_t src_base_row = src_base_nc + (uint64_t) iy * args.nb11;
+                    const uint64_t w_base_row   = w_base_ocic + (uint64_t) ky * args.nb01;
+
+                    for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                        const int32_t ix = base_x + kx*args.d0;
+                        const uint64_t src_offs = src_base_row + (uint64_t) ix * args.nb10;
+                        const uint64_t w_offs   = w_base_row   + (uint64_t) kx * args.nb00;
+
+                        const float x = *(device const float *)(src + src_offs);
+                        const float w = (float) (*(device const TK *)(weights + w_offs));
+
+                        acc += x * w;
+                    }
+                }
+            }
+        }
+
+        const uint64_t dst_offs =
+            (uint64_t) n  * args.nb3 +
+            (uint64_t) oc * args.nb2 +
+            (uint64_t) oh * args.nb1 +
+            (uint64_t) ow * args.nb0;
+
+        *(device float *)(dst + dst_offs) = acc;
+    }
+}
+
+template [[host_name("kernel_conv_2d_f32_f32")]]
+kernel void kernel_conv_2d<float>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_f16_f32")]]
+kernel void kernel_conv_2d<half>(
+        constant ggml_metal_kargs_conv_2d & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+typedef void (conv_transpose_1d_t)(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        constant ggml_metal_kargs_conv_transpose_1d & args,
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    // For output position j on the time axis, only input positions
+    //   i such that i*s0 <= j < i*s0 + K
+    // contribute -- i.e. i in [ceil((j - K + 1)/s0), floor(j/s0)]
+    // intersected with [0, IL-1]. That's at most ceil(K/s0) values
+    // (typically 2 for stride==K/2 transposed convs).
+    const int32_t j  = tgpig[0];
+    const int32_t s0 = args.s0;
+    const int32_t K  = args.K;
+    const int32_t IL = args.IL;
+
+    int32_t i_min;
+    {
+        int32_t a = j - K + 1;
+        i_min = a <= 0 ? 0 : (a + s0 - 1) / s0; // ceil(a/s0) for a>0
+    }
+    int32_t i_max = j / s0;
+    if (i_max > IL - 1) i_max = IL - 1;
+
+    float v = 0.0f;
+    if (i_min <= i_max) {
+        for (int64_t c = 0; c < args.IC; c++) {
+            const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+            const int32_t input_offset  = c * IL;
+
+            for (int32_t i = i_min; i <= i_max; i++) {
+                v += float(src0[kernel_offset + j - i * s0]) * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * args.nb0 + tgpig[1] * args.nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    constant ggml_metal_kargs_conv_transpose_1d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+
+typedef void (conv_transpose_2d_t)(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_2d(
+        constant ggml_metal_kargs_conv_transpose_2d & args,
+        device const T * src0,
+        device const float * src1,
+        device        char * dst,
+        threadgroup float * shared_sum [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int64_t out_x = tgpig[0];
+    const int64_t out_y = tgpig[1];
+    const int64_t out_c = tgpig[2];
+
+    const int64_t kw = tpitg[0];
+    const int64_t kh = tpitg[1];
+
+    float v = 0.0f;
+
+    for (int64_t in_c = 0; in_c < args.IC; in_c++) {
+        int64_t in_y = out_y - kh;
+
+        if (in_y < 0 || in_y % args.s0) continue;
+
+        in_y /= args.s0;
+
+        if (in_y >= args.IH) continue;
+
+        int64_t in_x = out_x - kw;
+
+        if (in_x < 0 || in_x % args.s0) continue;
+
+        in_x /= args.s0;
+
+        if (in_x >= args.IW) continue;
+
+        const int64_t input_idx = (args.IW * args.IH) * in_c + (args.IW) * in_y + in_x;
+        const int64_t kernel_idx = (args.KH * args.KW * args.OC) * in_c + (args.KH * args.KW) * out_c + (args.KW) * kh + kw;
+
+        v += (float)src0[kernel_idx] * src1[input_idx];
+    }
+
+    const uint tid = tpitg.y * ntg.x + tpitg.x;
+    shared_sum[tid] = v;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tid == 0) {
+        float total = 0.0f;
+        const uint num_threads = ntg.x * ntg.y;
+        for (uint i = 0; i < num_threads; i++) {
+            total += shared_sum[i];
+        }
+
+        device float * dst_ptr = (device float *) (dst + out_x*args.nb0 + out_y * args.nb1 + out_c*args.nb2);
+        dst_ptr[0] = total;
+    }
+}
+
+template [[host_name("kernel_conv_transpose_2d_f32_f32")]]
+kernel void kernel_conv_transpose_2d<float>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_transpose_2d_f16_f32")]]
+kernel void kernel_conv_transpose_2d<half>(
+    constant ggml_metal_kargs_conv_transpose_2d & args,
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    threadgroup float * shared_sum [[threadgroup(0)]],
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3   tpitg[[thread_position_in_threadgroup]],
+    uint3     ntg[[threads_per_threadgroup]]);
+
+// grid: x = C tile, y = OH, z = OW * N (for channel-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw_tiled(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t c = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (c >= args.C) {
+        return;
+    }
+
+    const int32_t oh = tgpig.y;
+    const int32_t own = tgpig.z;
+    const int32_t ow = own % args.OW;
+    const int32_t n  = own / args.OW;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    float acc = 0.0f;
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        const uint64_t w_base   = (uint64_t) c * args.nb02;
+        const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+// grid: x = OW tile, y = OH, z = C * N (for spatially-contiguous layouts)
+template <typename TK>
+kernel void kernel_conv_2d_dw(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]) {
+
+    const int32_t oh = tgpig.y;
+    const int32_t cn = tgpig.z;
+    const int32_t c  = cn % args.C;
+    const int32_t n  = cn / args.C;
+
+    const int32_t base_y = oh*args.s1 - args.p1;
+
+    int32_t ky_start = 0;
+    if (base_y < 0) {
+        ky_start = (-base_y + args.d1 - 1)/args.d1;
+    }
+    int32_t ky_end = args.KH;
+    const int32_t y_max = args.IH - 1 - base_y;
+    if (y_max < 0) {
+        ky_end = ky_start;
+    } else if (base_y + (args.KH - 1)*args.d1 >= args.IH) {
+        ky_end = min(ky_end, y_max/args.d1 + 1);
+    }
+
+    const uint64_t w_base   = (uint64_t) c * args.nb02;
+    const uint64_t src_base = (uint64_t) n * args.nb13 + (uint64_t) c * args.nb12;
+
+    const int32_t ow = (int32_t)(tgpig.x * ntg.x + tpitg.x);
+    if (ow >= args.OW) {
+        return;
+    }
+
+    float acc = 0.0f;
+
+    const int32_t base_x = ow*args.s0 - args.p0;
+
+    int32_t kx_start = 0;
+    if (base_x < 0) {
+        kx_start = (-base_x + args.d0 - 1)/args.d0;
+    }
+    int32_t kx_end = args.KW;
+    const int32_t x_max = args.IW - 1 - base_x;
+    if (x_max < 0) {
+        kx_end = kx_start;
+    } else if (base_x + (args.KW - 1)*args.d0 >= args.IW) {
+        kx_end = min(kx_end, x_max/args.d0 + 1);
+    }
+
+    if (ky_start < ky_end && kx_start < kx_end) {
+        for (int32_t ky = ky_start; ky < ky_end; ++ky) {
+            const int32_t iy = base_y + ky*args.d1;
+            const uint64_t src_row = src_base + (uint64_t) iy * args.nb11;
+            const uint64_t w_row = w_base + (uint64_t) ky * args.nb01;
+
+            for (int32_t kx = kx_start; kx < kx_end; ++kx) {
+                const int32_t ix = base_x + kx*args.d0;
+                const float x = *(device const float *)(src + src_row + (uint64_t) ix * args.nb10);
+                const float w = (float)(*(device const TK *)(weights + w_row + (uint64_t) kx * args.nb00));
+                acc += x * w;
+            }
+        }
+    }
+
+    const uint64_t dst_offs =
+        (uint64_t) n  * args.nb3 +
+        (uint64_t) c  * args.nb2 +
+        (uint64_t) oh * args.nb1 +
+        (uint64_t) ow * args.nb0;
+
+    *(device float *)(dst + dst_offs) = acc;
+}
+
+template [[host_name("kernel_conv_2d_dw_f32_f32")]]
+kernel void kernel_conv_2d_dw<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_f16_f32")]]
+kernel void kernel_conv_2d_dw<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f32_f32")]]
+kernel void kernel_conv_2d_dw_tiled<float>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template [[host_name("kernel_conv_2d_dw_tiled_f16_f32")]]
+kernel void kernel_conv_2d_dw_tiled<half>(
+        constant ggml_metal_kargs_conv_2d_dw & args,
+        device const char * weights,
+        device const char * src,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tpitg[[thread_position_in_threadgroup]],
+        uint3     ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_conv_3d(
+        constant ggml_metal_kargs_conv_3d & args,
+        device const  char * src0, // Weights [IC * OC, KD, KH, KW]
+        device const  char * src1, // Inputs  [IC * N,  ID, IH, IW]
+        device       char  * dst,  // Outputs [OC * N,  OD, OH, OW]
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+
+    // 1. Un-flatten the spatial dimension from Grid X
+    int64_t spatial_idx = tgpig.x * 32 + tpitg.x;
+
+    if (spatial_idx >= args.OW * args.OH * args.OD) {
+        return; // Thread falls outside the spatial volume
+    }
+
+    int64_t od = spatial_idx / (args.OW * args.OH);
+    int64_t oh = (spatial_idx / args.OW) % args.OH;
+    int64_t ow = spatial_idx % args.OW;
+
+    // 2. Map Y to Channels, Z to Batch
+    int64_t oc = tgpig.y;
+    int64_t batch_idx = tgpig.z;
+
+    // 3. Calculate anchor coordinates in the Input volume
+    int64_t i_w_base = ow * args.s0 - args.p0;
+    int64_t i_h_base = oh * args.s1 - args.p1;
+    int64_t i_d_base = od * args.s2 - args.p2;
+
+    float sum = 0.0f;
+
+    // 4. Gather Loop (Iterate over Input Channels -> Depth -> Height -> Width)
+    for (int64_t ic = 0; ic < args.IC; ++ic) {
+
+        // ggml packs batch and channel together in the 4th dimension
+        int64_t src_cn_idx = batch_idx * args.IC + ic;
+        int64_t w_cn_idx   = oc * args.IC + ic;
+
+        for (int64_t kz = 0; kz < args.KD; ++kz) {
+            int64_t id = i_d_base + kz * args.d2;
+            if (id < 0 || id >= args.ID) continue; // Boundary check (Padding)
+
+            for (int64_t ky = 0; ky < args.KH; ++ky) {
+                int64_t ih = i_h_base + ky * args.d1;
+                if (ih < 0 || ih >= args.IH) continue;
+
+                for (int64_t kx = 0; kx < args.KW; ++kx) {
+                    int64_t iw = i_w_base + kx * args.d0;
+                    if (iw < 0 || iw >= args.IW) continue;
+
+                    // Convert multi-dimensional coordinates to flat byte offsets
+                    int64_t w_idx = kx*args.nb00 + ky*args.nb01 + kz*args.nb02 + w_cn_idx*args.nb03;
+                    int64_t i_idx = iw*args.nb10 + ih*args.nb11 + id*args.nb12 + src_cn_idx*args.nb13;
+
+                    // Dereference memory and cast weights to f32 if they were f16
+                    float w_val = (float)*(device const T*)((device const char*)src0 + w_idx);
+                    float i_val = *(device const float*)((device const char*)src1 + i_idx);
+
+                    sum += w_val * i_val;
+                }
+            }
+        }
+    }
+
+    // 5. Write the accumulated value out to RAM
+    int64_t dst_cn_idx = batch_idx * args.OC + oc;
+    int64_t d_idx = ow*args.nb0 + oh*args.nb1 + od*args.nb2 + dst_cn_idx*args.nb3;
+
+    *(device float*)(dst + d_idx) = sum;
+}
+
+// Explicit instantiations so the JIT compiler can find them by name
+template [[host_name("kernel_conv_3d_f32_f32")]]
+kernel void kernel_conv_3d<float>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
+
+// Explicit instantiation for f16 weights
+template [[host_name("kernel_conv_3d_f16_f32")]]
+kernel void kernel_conv_3d<half>(
+    constant ggml_metal_kargs_conv_3d & args,
+    device const char  * src0,
+    device const char * src1,
+    device       char  * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]]);
@@ -0,0 +1,735 @@
+#pragma once
+
+#include "common.h"
+
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+#include "ggml-common.h"
+#endif
+
+#define QK_NL 16 // shared by mul_mm and get_rows_q instantiations
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f32_t4(device const float4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*src);
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+
+#if defined(GGML_METAL_HAS_BF16)
+template <typename type4x4>
+void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
+    reg = (type4x4)(*src);
+}
+
+template <typename type4>
+void dequantize_bf16_t4(device const bfloat4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src));
+}
+#endif
+
+template <typename type4x4>
+void dequantize_q1_0(device const block_q1_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+    const float neg_d = -d;
+
+    const int byte_offset = il * 2;  // il*16 bits = il*2 bytes
+    const uint8_t b0 = qs[byte_offset];
+    const uint8_t b1 = qs[byte_offset + 1];
+
+    float4x4 reg_f;
+
+    reg_f[0][0] = select(neg_d, d, bool(b0 & 0x01));
+    reg_f[0][1] = select(neg_d, d, bool(b0 & 0x02));
+    reg_f[0][2] = select(neg_d, d, bool(b0 & 0x04));
+    reg_f[0][3] = select(neg_d, d, bool(b0 & 0x08));
+    reg_f[1][0] = select(neg_d, d, bool(b0 & 0x10));
+    reg_f[1][1] = select(neg_d, d, bool(b0 & 0x20));
+    reg_f[1][2] = select(neg_d, d, bool(b0 & 0x40));
+    reg_f[1][3] = select(neg_d, d, bool(b0 & 0x80));
+
+    reg_f[2][0] = select(neg_d, d, bool(b1 & 0x01));
+    reg_f[2][1] = select(neg_d, d, bool(b1 & 0x02));
+    reg_f[2][2] = select(neg_d, d, bool(b1 & 0x04));
+    reg_f[2][3] = select(neg_d, d, bool(b1 & 0x08));
+    reg_f[3][0] = select(neg_d, d, bool(b1 & 0x10));
+    reg_f[3][1] = select(neg_d, d, bool(b1 & 0x20));
+    reg_f[3][2] = select(neg_d, d, bool(b1 & 0x40));
+    reg_f[3][3] = select(neg_d, d, bool(b1 & 0x80));
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q1_0_t4(device const block_q1_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const float neg_d = -d;
+    const int base = il * 4;
+    const uint8_t byte = xb->qs[base / 8];
+    const int s = base % 8;
+
+    float4 reg_f;
+    reg_f[0] = select(neg_d, d, bool((byte >> (s    )) & 1));
+    reg_f[1] = select(neg_d, d, bool((byte >> (s + 1)) & 1));
+    reg_f[2] = select(neg_d, d, bool((byte >> (s + 2)) & 1));
+    reg_f[3] = select(neg_d, d, bool((byte >> (s + 3)) & 1));
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q2_0(device const block_q2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    const int byte_offset = il * 4;  // il*16 elements = il*4 bytes (4 elements per byte)
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; i++) {
+        const uint8_t b = qs[byte_offset + i];
+        reg_f[i][0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+        reg_f[i][1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+        reg_f[i][2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+        reg_f[i][3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q2_0_t4(device const block_q2_0 * xb, short il, thread type4 & reg) {
+    const float d = xb->d;
+    const uint8_t b = xb->qs[il];
+
+    float4 reg_f;
+    reg_f[0] = ((float)((b >> 0) & 3) - 1.0f) * d;
+    reg_f[1] = ((float)((b >> 2) & 3) - 1.0f) * d;
+    reg_f[2] = ((float)((b >> 4) & 3) - 1.0f) * d;
+    reg_f[3] = ((float)((b >> 6) & 3) - 1.0f) * d;
+
+    reg = (type4) reg_f;
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = d1 * (qs[i] & mask0) + md;
+        reg_f[i/2][2*(i%2) + 1] = d2 * (qs[i] & mask1) + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
+
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        reg_f[i/2][2*(i%2) + 0] = ((qs[i] & mask0) * d1) + m;
+        reg_f[i/2][2*(i%2) + 1] = ((qs[i] & mask1) * d2) + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + md;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + md;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg_f[i/2][2*(i%2) + 0] = d * x0 + m;
+        reg_f[i/2][2*(i%2) + 1] = d * x1 + m;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    for (int i = 0; i < 4; ++i) {
+        reg_f[i] = float4(qs[4*il + i]) * d;
+    }
+
+    reg = (type4x4) reg_f;
+}
+
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
+    const float d = xb->d;
+
+    reg = (type4) (float4(qs[il]) * d);
+}
+
+template <typename type4x4>
+void dequantize_mxfp4(device const block_mxfp4 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const uint8_t shr = il >= 1 ? 4 : 0;
+
+    for (int i = 0; i < 4; ++i) {
+        reg[i][0] = d * kvalues_mxfp4_f[(q2[4*i + 0] >> shr) & 0x0F];
+        reg[i][1] = d * kvalues_mxfp4_f[(q2[4*i + 1] >> shr) & 0x0F];
+        reg[i][2] = d * kvalues_mxfp4_f[(q2[4*i + 2] >> shr) & 0x0F];
+        reg[i][3] = d * kvalues_mxfp4_f[(q2[4*i + 3] >> shr) & 0x0F];
+    }
+}
+
+template <typename type4>
+void dequantize_mxfp4_t4(device const block_mxfp4 * xb, short il, thread type4 & reg) {
+    device const uint8_t * q2 = (device const uint8_t *)xb->qs;
+
+    const float d = e8m0_to_fp32(xb->e);
+    const short il4 = il%4;
+
+    const uint8_t shr = il >= 4 ? 4 : 0;
+
+    reg[0] = d * kvalues_mxfp4_f[(q2[4*il4 + 0] >> shr) & 0x0F];
+    reg[1] = d * kvalues_mxfp4_f[(q2[4*il4 + 1] >> shr) & 0x0F];
+    reg[2] = d * kvalues_mxfp4_f[(q2[4*il4 + 2] >> shr) & 0x0F];
+    reg[3] = d * kvalues_mxfp4_f[(q2[4*il4 + 3] >> shr) & 0x0F];
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint16_t * ql = (device const uint16_t *)xb->ql;
+    device const uint16_t * qh = (device const uint16_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 32*(il/8) + 16*((il/2)&1) + 8*(il&1);
+    qh = qh + 16*(il/8) + 8*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint32_t kmask1 = il>1 ? (il>2 ? 0xC0C0C0C0 : 0x30303030) : (il>0 ? 0x0C0C0C0C : 0x03030303);
+    const uint32_t kmask2 = il>1 ? 0xF0F0F0F0                       : 0x0F0F0F0F;
+    const float ml = d_all * sc * 32.f;
+    const float dl0 = d_all * sc;
+    const float dl1 = dl0 / 256.f;
+    const float dl2 = dl0 / (256.f * 256.f);
+    const float dl3 = dl0 / (256.f * 256.f * 256.f);
+    const uint8_t shr_h = il>2 ? 2 : 0;
+    const uint8_t shl_h = il>1 ? 0 : (il>0 ? 2 : 4);
+    const uint8_t shr_l = il>1 ? 4 : 0;
+    for (int i = 0; i < 4; ++i) {
+        const uint32_t  low = (ql[2*i] | (uint32_t)(ql[2*i+1] << 16)) & kmask2;
+        const uint32_t high = (qh[2*i] | (uint32_t)(qh[2*i+1] << 16)) & kmask1;
+        const uint32_t q = ((high << shl_h) >> shr_h) | (low >> shr_l);
+        reg[i][0] = dl0 *  ((half)(q & 0xFF))       - ml;
+        reg[i][1] = dl1 * ((float)(q & 0xFF00))     - ml;
+        reg[i][2] = dl2 * ((float)(q & 0xFF0000))   - ml;
+        reg[i][3] = dl3 * ((float)(q & 0xFF000000)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4x4>
+void dequantize_tq2_0(device const block_tq2_0 * xb, short il, thread type4x4 & reg) {
+    device const uint8_t * qs = xb->qs;
+    const float d = xb->d;
+
+    float4x4 reg_f;
+
+    // 2 bits per element, 4 elements per byte, 128 elements per 32-byte group
+    const short base = il * 16;
+    for (int k = 0; k < 16; k++) {
+        const int i = base + k;
+        const int byte = ((i >> 7) & 1) * 32 + (i & 31);
+        const int l = (i >> 5) & 3;
+        reg_f[k/4][k%4] = d * (float)(((qs[byte] >> (2*l)) & 3) - 1);
+    }
+
+    reg = (type4x4) reg_f;
+}
--- a/Show More
+++ b/Show More