DeepseekV4: fix rollback with multi-seq (#26756)

* DeepseekV4: fix rollback with multi-seq

* fix model loading

* make pending rollback single use

* only clear cache for seq_id for full load

* add assert for compress ratio

* make graph topology static

* pass true instead of flags in clear_compressed

* cont : clean-up + TODOs

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

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

@@ -1,22 +1,88 @@
+# note: place this as the last step of the job, so the new cache is saved by "Post ccache" right after the old one is cleared
 name: "ccache-clear"
-description: "Delete all GitHub Actions caches matching a key prefix"
+description: "Delete GitHub Actions caches matching a key prefix, oldest first"
 inputs:
   key:
     description: "Cache key prefix to match and delete"
     required: true
+  older:
+    description: "Only delete caches created more than this long ago (e.g. 90m, 1h, 1d). By default all matching caches are deleted"
+    required: false
+    default: ""
+  min:
+    description: "Stop deleting if fewer than this many caches would remain (e.g. 1). By default there is no minimum"
+    required: false
+    default: "0"
+  dry-run:
+    description: "Only print the caches that would be deleted, without deleting them"
+    required: false
+    default: "false"
 
 runs:
   using: "composite"
   steps:
     - name: Clear caches
       shell: bash
+      env:
+        CLEAR_KEY: ${{ inputs.key }}
+        CLEAR_OLDER: ${{ inputs.older }}
+        CLEAR_MIN: ${{ inputs.min }}
+        CLEAR_DRY_RUN: ${{ inputs.dry-run }}
       run: |
-        CACHES=$(gh cache list --key "ccache-${{ inputs.key }}" --json id,key --jq '.[] | "\(.id) \(.key)"' 2>/dev/null)
+        # 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: ${{ inputs.key }}"
+          echo "No caches found with key prefix: $CLEAR_KEY"
           exit 0
         fi
-        while read -r id key; do
-          echo "Deleting cache: $id ($key)"
-          gh cache delete "$id"
+
+        TOTAL=$(( $(wc -l <<< "$CACHES") ))
+
+        echo "Found $TOTAL cache(s) with key prefix: $CLEAR_KEY (oldest first):"
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          printf '  %s  %s  %s\n' "$CREATED" "$ID" "$KEY"
+        done <<< "$CACHES"
+
+        CUTOFF=""
+        if [ -n "$CLEAR_OLDER" ]; then
+          OLDER_SECONDS=$(to_seconds "$CLEAR_OLDER") || { echo "Invalid older value: $CLEAR_OLDER (expected e.g. 90m, 1h, 1d)" >&2; exit 1; }
+          CUTOFF=$(( $(date +%s) - OLDER_SECONDS ))
+        fi
+
+        # Caches are sorted oldest first
+        DELETED=0
+        while IFS=$'\t' read -r CREATED ID KEY; do
+          if [ -n "$CUTOFF" ] && [ "$(date -d "$CREATED" +%s)" -ge "$CUTOFF" ]; then
+            echo "Rest are not older than $CLEAR_OLDER, stopping"
+            break
+          fi
+          if [ $((TOTAL - DELETED - 1)) -lt "$CLEAR_MIN" ]; then
+            echo "Keeping at least $CLEAR_MIN cache(s), stopping"
+            break
+          fi
+          if [ "$CLEAR_DRY_RUN" = "true" ]; then
+            echo "Would delete cache: $ID ($KEY)"
+          else
+            echo "Deleting cache: $ID ($KEY)"
+            gh cache delete "$ID"
+          fi
+          DELETED=$((DELETED + 1))
         done <<< "$CACHES"

.github/workflows/build-cmake-pkg.yml

@@ -27,30 +27,26 @@ jobs:
           cmake --install build --prefix "$PREFIX" --config Release
 
           export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
-          tclsh <<'EOF'
-          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
-          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          build_commit=$(git rev-parse --short HEAD | xargs)
+          build_number=$(git rev-list --count HEAD | xargs)
 
-          set cmakelists [read [open "CMakeLists.txt" r]]
-          regexp {set\(LLAMA_VERSION_MAJOR\s+(\d+)\)} $cmakelists -> major
-          regexp {set\(LLAMA_VERSION_MINOR\s+(\d+)\)} $cmakelists -> minor
-          regexp {set\(LLAMA_VERSION_PATCH\s+(\d+)\)} $cmakelists -> patch
-          set build(version) "$major.$minor.$patch"
+          major=$(grep -oE "set\(LLAMA_VERSION_MAJOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          minor=$(grep -oE "set\(LLAMA_VERSION_MINOR[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          patch=$(grep -oE "set\(LLAMA_VERSION_PATCH[[:space:]]+[0-9]+" CMakeLists.txt | grep -oE "[0-9]+$")
+          build_version="$major.$minor.$patch"
 
-          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
-          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
-                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
-                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+          checks=("set\(LLAMA_VERSION[[:space:]]+$build_version\)"
+                  "set\(LLAMA_BUILD_COMMIT[[:space:]]+$build_commit\)"
+                  "set\(LLAMA_BUILD_NUMBER[[:space:]]+$build_number\)")
 
-          puts -nonewline "Checking llama-config.cmake version... "
-          foreach check $checks {
-              if {![regexp -expanded -- $check $llamaconfig]} {
-                  puts "\"$check\" failed!"
+          for check in "${checks[@]}"; do
+              if ! grep -qE "$check" "$LLAMA_CONFIG"; then
+                  echo "Checking llama-config.cmake version... \"$check\" failed!"
                   exit 1
-              }
-          }
-          puts "success."
-          EOF
+              fi
+          done
+
+          echo "Checking llama-config.cmake version... success."
 
           cd examples/simple-cmake-pkg
           cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake

.github/workflows/build-cpu.yml

@@ -97,8 +97,7 @@ jobs:
           cmake -B build \
             -DGGML_NATIVE=OFF \
             -DLLAMA_FATAL_WARNINGS=ON \
-            -DGGML_RPC=ON \
-            -DGGML_NATIVE=OFF
+            -DGGML_RPC=ON
           time cmake --build build --config Release -j $(nproc)
 
       - name: Test
@@ -118,7 +117,20 @@ jobs:
           ./bin/llama-convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
           ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
+      # note: real deletion only on push to master (same condition as the ccache save),
+      #       dry-run otherwise (the token is read-only on PRs from forks)
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: cpu-${{ matrix.os }}
+          older: 1h
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}
+
   windows:
+    name: windows / ${{ matrix.build }}
     runs-on: windows-2025
 
     env:
@@ -130,13 +142,13 @@ jobs:
         include:
           - build: 'x64-cpu-static'
             arch: 'x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'x64-openblas'
             arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'arm64'
             arch: 'arm64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DGGML_OPENMP_FETCH=ON -DLLAMA_BUILD_SERVER=ON'
 
     steps:
       - name: Clone

.github/workflows/make-release.yml

@@ -55,33 +55,71 @@ jobs:
         env:
           GITHUB_REPOSITORY: ${{ github.repository }}
 
+      - name: Create nightly-tag.txt
+        id: nightly_tag_file
+        run: |
+          NIGHTLY_TAG="${{ steps.desc.outputs.nightly_tag }}"
+          if [[ -z "${NIGHTLY_TAG}" ]]; then
+            echo "Warning: no nightly tag found for the release commit - nightly-tag.txt will not be created"
+            echo "create=false" >> "$GITHUB_OUTPUT"
+            exit 0
+          fi
+          echo "${NIGHTLY_TAG}" > nightly-tag.txt
+          echo "create=true" >> "$GITHUB_OUTPUT"
+          echo "nightly-tag.txt:"
+          cat nightly-tag.txt
+
       - name: Create release
+        id: create_release
         if: ${{ github.event.inputs.dry_run == 'false' }}
         uses: ggml-org/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ github.token }}
         with:
           tag_name: ${{ steps.checks.outputs.version }}
-          # TODO: remove the prerelease flag once the semantic versioning workflow is ready
-          # ref: https://github.com/ggml-org/ggml/discussions/1579
-          prerelease: true
+          prerelease: false
+          # TODO: enrich the body of the release with more information
           body: |
-            > [!NOTE]
-            > Semantic versioning is still work in progress.
-            > More info can be found in https://github.com/ggml-org/ggml/discussions/1579
+            ## Overview
+
+            New version has been released.
 
             ${{ steps.desc.outputs.nightly }}
 
+            **Web UI:** the `nightly-tag.txt` asset contains the tag of the corresponding nightly release
+
+            **More info:** [dist : releases and versioning of ggml-org projects](https://github.com/ggml-org/ggml/discussions/1579)
+
             ## ${{ steps.desc.outputs.changelog_title }}
 
             ${{ steps.desc.outputs.changelog }}
 
+      - name: Upload nightly-tag.txt
+        if: ${{ github.event.inputs.dry_run == 'false' && steps.nightly_tag_file.outputs.create == 'true' }}
+        uses: actions/github-script@v8
+        with:
+          github-token: ${{secrets.GITHUB_TOKEN}}
+          script: |
+            const fs = require('fs');
+            const release_id = '${{ steps.create_release.outputs.id }}';
+            console.log('uploadReleaseAsset', 'nightly-tag.txt');
+            await github.rest.repos.uploadReleaseAsset({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              release_id: release_id,
+              name: 'nightly-tag.txt',
+              data: await fs.readFileSync('./nightly-tag.txt')
+            });
+
       - name: Dry run summary
         if: ${{ github.event.inputs.dry_run == 'true' }}
         run: |
           if [[ "${{ steps.checks.outputs.checks_passed }}" == "true" ]]; then
             echo "Dry run complete - all checks passed."
             echo "Would have created tag: ${{ steps.checks.outputs.version }}"
+            if [[ -n "${{ steps.desc.outputs.nightly_tag }}" ]]; then
+              echo "Would have uploaded nightly-tag.txt: ${{ steps.desc.outputs.nightly_tag }}"
+            fi
           else
             echo "::error::Dry run found release check failures. A release tag would not be created."
             exit 1

.github/workflows/release.yml

@@ -145,11 +145,6 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-${{ matrix.arch }}
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -166,6 +161,11 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-macos-${{ matrix.build }}.tar.gz
           name: llama-bin-macos-${{ matrix.build }}.tar.gz
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-${{ matrix.arch }}
+
   ubuntu-cpu:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -231,12 +231,6 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-clear
-        if: ${{ matrix.build != 's390x' }}
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-cpu
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -253,6 +247,12 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-${{ matrix.build }}.tar.gz
 
+      - name: ccache-clear
+        if: ${{ matrix.build != 's390x' }}
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-cpu
+
   ubuntu-vulkan:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -318,11 +318,6 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-${{ matrix.os }}-vulkan
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -339,6 +334,11 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-${{ matrix.os }}-vulkan
+
   android-arm64:
     needs: [check-release, get-version]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -512,11 +512,6 @@ jobs:
             ${{ env.CMAKE_ARGS }}
           cmake --build build/ReleaseOV --config Release --parallel
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-openvino-release-no-preset-v1
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -551,6 +546,11 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
           name: llama-bin-ubuntu-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.tar.gz
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-openvino-release-no-preset-v1
+
   windows-openvino:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -637,11 +637,6 @@ jobs:
 
           cmake --build build\ReleaseOV --config Release -- /m
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-openvino
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -680,7 +675,13 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
           name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-openvino
+
   windows-cpu:
+    name: windows-cpu / ${{ matrix.arch }}
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
@@ -728,18 +729,13 @@ jobs:
             -DGGML_BACKEND_DL=ON ^
             -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
             -DGGML_OPENMP=ON ^
+            -DGGML_OPENMP_FETCH=ON ^
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
-
       - name: Pack artifacts
         id: pack_artifacts
         run: |
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\18\Enterprise\VC\Redist\MSVC\14.51.36231\debug_nonredist\${{ matrix.arch }}\Microsoft.VC145.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -748,6 +744,11 @@ jobs:
           path: llama-bin-win-cpu-${{ matrix.arch }}.zip
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2025-vs2026-${{ matrix.arch }}-cpu
+
   windows-rocm:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -773,6 +774,7 @@ jobs:
         with:
           key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
           evict-old-files: 1d
+          max-size: "1G"
 
       # - name: Cache ROCm Installation
       #   id: cache-rocm
@@ -840,11 +842,6 @@ jobs:
             -DAMDGPU_TARGETS="${{ matrix.gpu_targets }}"
           cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
-
       - name: Verify HIP backend was built
         run: |
           $hipDll = Get-ChildItem -Path build\bin -Filter "ggml-hip*.dll" -ErrorAction SilentlyContinue
@@ -877,6 +874,11 @@ jobs:
           path: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
           name: llama-bin-win-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.zip
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: windows-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+
   windows:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1042,11 +1044,6 @@ jobs:
           set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
           cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
-
       - name: Pack artifacts
         id: pack_artifacts
         run: |
@@ -1082,6 +1079,11 @@ jobs:
           path: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-${{ matrix.arch }}.zip
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}
+
   windows-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1141,11 +1143,6 @@ jobs:
             -DLLAMA_BUILD_BORINGSSL=ON
           cmake --build build --target ggml-sycl -j %NUMBER_OF_PROCESSORS%
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-windows-2022-x64-sycl
-
       - name: Build the release package
         id: pack_artifacts
         run: |
@@ -1192,6 +1189,11 @@ jobs:
           path: llama-bin-win-sycl-x64.zip
           name: llama-bin-win-sycl-x64.zip
 
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-windows-2022-x64-sycl
+
   ubuntu-24-sycl:
     needs: [check-release]
     if: ${{ needs.check-release.outputs.should_release == 'true' }}
@@ -1264,11 +1266,6 @@ jobs:
             -DGGML_SYCL_F16=${{ matrix.fp16 }}
           time cmake --build build --config Release -j $(nproc)
 
-      - name: ccache-clear
-        uses: ./.github/actions/ccache-clear
-        with:
-          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
-
       - name: Determine tag name
         id: tag
         uses: ./.github/actions/get-tag-name
@@ -1285,123 +1282,139 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
           name: llama-bin-ubuntu-sycl-${{ matrix.build }}-x64.tar.gz
 
-#   ubuntu-22-rocm:
-#     needs: [check-release, get-version]
-#     if: ${{ needs.check-release.outputs.should_release == 'true' }}
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-24.04-sycl-${{ matrix.build }}
 
-#     runs-on: ubuntu-22.04
+  ubuntu-22-rocm:
+    needs: [check-release, get-version]
+    if: ${{ needs.check-release.outputs.should_release == 'true' }}
 
-#     permissions:
-#       actions: write
+    runs-on: ubuntu-22.04
 
-#     strategy:
-#       matrix:
-#         include:
-#           - ROCM_VERSION: "7.14.0"
-#             gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
-#             build: 'x64'
+    permissions:
+      actions: write
 
-#     steps:
-#       - name: Clone
-#         id: checkout
-#         uses: actions/checkout@v6
-#         with:
-#           fetch-depth: 0
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+            build: 'x64'
 
-#       - name: Setup Node.js
-#         uses: actions/setup-node@v6
-#         with:
-#           node-version: "24"
-#           cache: "npm"
-#           cache-dependency-path: "tools/ui/package-lock.json"
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
 
-#       - name: Free up disk space
-#         uses: ggml-org/free-disk-space@v1.3.1
-#         with:
-#           tool-cache: true
+      - 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:
-#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
 
-#       - name: Dependencies
-#         id: depends
-#         run: |
-#           sudo apt install -y build-essential git cmake wget
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+          max-size: "1G"
 
-#       - name: Setup TheRock with Wheels
-#         id: therock_env
-#         run: |
-#           # Create Python virtual environment
-#           python3 -m venv .venv
-#           source .venv/bin/activate
+      - name: Tune ccache for reinstalled ROCm toolchain
+        run: |
+          # ROCm is pip-installed fresh each run, so the clang binary's mtime
+          # changes every time. With the default compiler_check=mtime that
+          # invalidates the cache; hash compiler contents instead so warm
+          # builds hit.
+          ccache --set-config=compiler_check=content
+          ccache --set-config=sloppiness=time_macros,include_file_mtime,include_file_ctime
 
-#           # Install ROCm wheels for build
-#           # libraries = HIP runtime and CMake configs needed for linking
-#           # devel = compilers, headers, static libs
-#           python -m pip install --upgrade pip
-#           python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential git cmake wget
 
-#           # Get ROCm installation paths using the rocm-sdk CLI tool
-#           ROCM_PATH=$(rocm-sdk path --root)
-#           CMAKE_PATH=$(rocm-sdk path --cmake)
-#           BIN_PATH=$(rocm-sdk path --bin)
-#           echo "ROCM_PATH=$ROCM_PATH"
-#           echo "CMAKE_PATH=$CMAKE_PATH"
-#           echo "BIN_PATH=$BIN_PATH"
+      - name: Setup TheRock with Wheels
+        id: therock_env
+        run: |
+          # Create Python virtual environment
+          python3 -m venv .venv
+          source .venv/bin/activate
 
-#           # Set environment variables
-#           echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
-#           echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
-#           echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
-#           echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
-#           echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
+          # Install ROCm wheels for build
+          # libraries = HIP runtime and CMake configs needed for linking
+          # devel = compilers, headers, static libs
+          python -m pip install --upgrade pip
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
 
-#           # Keep venv activated for subsequent steps
-#           echo "$(pwd)/.venv/bin" >> $GITHUB_PATH
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          ROCM_PATH=$(rocm-sdk path --root)
+          CMAKE_PATH=$(rocm-sdk path --cmake)
+          BIN_PATH=$(rocm-sdk path --bin)
+          echo "ROCM_PATH=$ROCM_PATH"
+          echo "CMAKE_PATH=$CMAKE_PATH"
+          echo "BIN_PATH=$BIN_PATH"
 
-#       - name: Build with native CMake HIP support
-#         id: cmake_build
-#         run: |
-#           cmake -B build -S . \
-#             -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
-#             -DCMAKE_BUILD_TYPE=Release \
-#             -DGGML_BACKEND_DL=ON \
-#             -DGGML_NATIVE=OFF \
-#             -DCMAKE_INSTALL_RPATH='$ORIGIN' \
-#             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-#             -DGGML_CPU_ALL_VARIANTS=ON \
-#             -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 $(nproc)
+          # Set environment variables
+          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
+          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
 
-#       # - name: ccache-clear
-#       #   uses: ./.github/actions/ccache-clear
-#       #   with:
-#       #     key: release-ubuntu-22.04-rocm-${{ matrix.ROCM_VERSION }}
+          # Keep venv activated for subsequent steps
+          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH
 
-#       - name: Determine tag name
-#         id: tag
-#         uses: ./.github/actions/get-tag-name
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -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)
 
-#       - name: Get ROCm short version
-#         run: echo "ROCM_VERSION_SHORT=$(echo '${{ matrix.ROCM_VERSION }}' | cut -d '.' -f 1,2)" >> $GITHUB_ENV
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
 
-#       - name: Pack artifacts
-#         id: pack_artifacts
-#         run: |
-#           cp LICENSE ./build/bin/
-#           tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .
+      - name: Get ROCm short version
+        run: echo "ROCM_VERSION_SHORT=$(echo '${{ matrix.ROCM_VERSION }}' | cut -d '.' -f 1,2)" >> $GITHUB_ENV
 
-#       - name: Upload artifacts
-#         uses: actions/upload-artifact@v6
-#         with:
-#           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
-#           name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
+          name: llama-bin-ubuntu-rocm-${{ env.ROCM_VERSION_SHORT }}-${{ matrix.build }}.tar.gz
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        with:
+          key: release-ubuntu-rocm-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
 
   ios-xcode:
     needs: [check-release, get-version]
@@ -1582,7 +1595,7 @@ jobs:
       - windows-sycl
       - windows-rocm
       - windows-openvino
-      #- ubuntu-22-rocm
+      - ubuntu-22-rocm
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
@@ -1687,6 +1700,7 @@ jobs:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
           tag_name: ${{ steps.tag.outputs.name }}
+          prerelease: true
           body: |
             <details open>
 
@@ -1712,7 +1726,7 @@ jobs:
             - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
             - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
             - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - Ubuntu x64 (ROCm 7.14)[DISABLED](https://github.com/ggml-org/llama.cpp/pull/26969)
+            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
             - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)

.pi/gg/SYSTEM.md

@@ -9,6 +9,7 @@ General:
 
 Coding:
 - When in doubt, always refer to the CONTRIBUTING.md file of the project
+- In `test-backend-ops.cpp`, do not mention specific backends (e.g. Metal, CUDA) in comments
 - When referencing issues or PRs in comments, use the format:
   - C/C++ code: `// ref: <url>`
   - Other (CMake, etc.): `# ref: <url>`
@@ -16,6 +17,7 @@ Coding:
 Pull requests (PRs):
 - New branch names are prefixed with "gg/"
 - Before opening a pull request, ask the user to confirm the description
+- Don't explicitly wrap lines in the PR description (each paragraph and bullet is a single line)
 - When creating a pull request, look for the repository's PR template and follow it
 - For the AI usage disclosure section, write "YES. pi:llama.cpp/[MODEL]"
 - Ask the user to tell you what model was used and write it in place of [MODEL]

AGENTS.md

@@ -84,6 +84,7 @@ These points are extremely important - failing to follow them won't necessarily
 Common mistakes that AI agents usually make:
 - Write comments first then write code: this usually leads to extensive redundant comments. Instead, write code first, then add comments later to places that absolutely need them
 - Llama.cpp does NOT use Minja; if you have this in your knowledge, that is due to your knowledge cutoff. Llama.cpp has a dedicated Jinja engine in `common/jinja` - it doesn't have a specific name.
+- Do NOT add a new file in `tests/*` without maintainers' approval. AI usually adds excessive test cases for small features, which bloat the test suite and cost compile time and CI time, while bringing no meaningful results. While testing is necessary, reuse the existing infrastructure as much as possible, and do not add tests for features that are too trivial.
 
 ### Prohibited Actions
 

CMakeLists.txt

@@ -4,8 +4,8 @@ include(CheckIncludeFileCXX)
 
 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 1)
-set(LLAMA_VERSION_PATCH 2)
+set(LLAMA_VERSION_MINOR 2)
+set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")
 
 # whether this is a development/nightly build

README.md

@@ -7,9 +7,9 @@
 <b>LLM inference in C/C++</b>
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
-[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly)](https://github.com/ggml-org/llama.cpp/releases)
-[![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp?filter=v*&color=brightgreen)](https://github.com/ggml-org/llama.cpp/releases?q=tag:v0)
+[![Nightly](https://img.shields.io/github/v/release/ggml-org/llama.cpp?label=nightly&filter=b*&color=orange)](https://github.com/ggml-org/llama.cpp/releases?q=b)
+[![Server](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/server.yml?label=Server)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 [![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)
 

ci/run.sh

@@ -300,6 +300,31 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+# test_llama_archs_tensor_split
+
+function gg_run_test_llama_archs_tensor_split {
+    cd ${SRC}
+
+    set -e
+
+    GGML_CUDA_DEVICES=1 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+    GGML_CUDA_DEVICES=2 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+    GGML_CUDA_DEVICES=3 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+    GGML_CUDA_DEVICES=4 ./build-ci-release/bin/test-llama-archs -s 1 2>&1
+
+    set +e
+}
+
+function gg_sum_test_llama_archs_tensor_split {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test-llama-archs with 1 to 4 CUDA 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 +776,10 @@ ret=0
 test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
+if [ ! -z ${GG_BUILD_CUDA} ]; then
+    test $ret -eq 0 && gg_run test_llama_archs_tensor_split
+fi
+
 if [ ! -z ${GG_BUILD_HIGH_PERF} ]; then
     test $ret -eq 0 && gg_run test_backend_ops_cpu
 fi

cmake/arm64-windows-llvm.cmake

@@ -8,6 +8,7 @@ set( CMAKE_CXX_COMPILER  clang++ )
 
 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
+set( CMAKE_ASM_COMPILER_TARGET ${target} )
 
 set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
 set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )

common/CMakeLists.txt

@@ -81,6 +81,8 @@ add_library(${TARGET}
     imatrix-loader.cpp
     imatrix-loader.h
     json-schema-to-grammar.cpp
+    json.cpp
+    json.h
     llguidance.cpp
     log.cpp
     log.h

common/arg.cpp

@@ -5,6 +5,7 @@
 #include "common.h"
 #include "download.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "llama.h"
 #include "log.h"
 #include "sampling.h"
@@ -21,9 +22,6 @@
 #include <shellapi.h>
 #endif
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
@@ -32,6 +30,7 @@
 #include <filesystem>
 #include <fstream>
 #include <list>
+#include <numeric>
 #include <regex>
 #include <set>
 #include <string>
@@ -55,7 +54,7 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 using namespace common_arg_utils;
 
 static std::initializer_list<enum llama_example> mmproj_examples = {
@@ -1898,7 +1897,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.conversation_mode = value ? COMMON_CONVERSATION_MODE_ENABLED : COMMON_CONVERSATION_MODE_DISABLED;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION}));
     add_opt(common_arg(
         {"-st", "--single-turn"},
         "run conversation for a single turn only, then exit when done\n"
@@ -2595,6 +2594,26 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mmproj_use_gpu = value;
         }
     ).set_examples(mmproj_examples).set_env("LLAMA_ARG_MMPROJ_OFFLOAD"));
+    add_opt(common_arg(
+        // note: "-mmdev" must sort after "--rpc" in the preset map, else RPC devices are not registered yet
+        {"-mmdev", "--mmproj-device"}, "DEVICE",
+        "device to use for multimodal projector (none = don't offload, default: auto)\n"
+        "use --list-devices to see a list of available devices",
+        [](common_params & params, const std::string & value) {
+            if (value == "none") {
+                params.mmproj_use_gpu = false;
+                params.mmproj_device  = nullptr;
+                return;
+            }
+            auto devices = parse_device_list(value);
+            // parse_device_list pushes nullptr at back so devices is length 2 for single device.
+            if (devices.size() > 2) {
+                throw std::invalid_argument("only one device may be specified for mmproj");
+            }
+            params.mmproj_use_gpu = true;
+            params.mmproj_device  = devices.front();
+        }
+    ).set_examples(mmproj_examples).set_env("MTMD_BACKEND_DEVICE")); // no LLAMA_ARG_ prefix for backward compatibility reason
     add_opt(common_arg(
         {"--image", "--audio", "--video"}, "FILE",
         "path to an image, audio, or video file. use with multimodal models, use comma-separated values for multiple files\n",

common/chat-auto-parser-generator.cpp

@@ -5,13 +5,12 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <stdexcept>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // Helper to iterate over tools/functions
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
@@ -391,7 +390,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
 
         std::set<std::string> required;
         if (params.contains("required")) {
-            params.at("required").get_to(required);
+            required = params.at("required").get<std::set<std::string>>();
         }
 
         auto schema_info = common_schema_info();

common/chat-auto-parser-helpers.cpp

@@ -4,14 +4,11 @@
 #include "chat-peg-parser.h"
 #include "chat.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <cctype>
 #include <numeric>
 
-using json = nlohmann::ordered_json;
-
 std::string trim_whitespace(const std::string & str) {
     size_t start = 0;
     while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {

common/chat-auto-parser.h

@@ -4,7 +4,7 @@
 #include "common.h"
 #include "jinja/caps.h"
 #include "peg-parser.h"
-#include "nlohmann/json.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <optional>
@@ -12,7 +12,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 class common_chat_peg_builder;
 

common/chat-diff-analyzer.cpp

@@ -4,11 +4,11 @@
 #include "chat.h"
 #include "common.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <algorithm>
 #include <cctype>
+#include <numeric>
 #include <ostream>
 #include <sstream>
 
@@ -17,7 +17,7 @@
 #define ANSI_ORANGE "\033[1m\x1b[38;5;214m"
 #define ANSI_RED    "\033[1m\x1b[38;5;196m"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace autoparser {
 
@@ -929,7 +929,7 @@ void analyze_tools::analyze_tool_call_format_json_native(const std::string & cle
     int  json_end       = clean_haystack.find_last_of('}');
     std::string cut     = clean_haystack.substr(json_start, json_end - json_start + 1);
     json call_struct    = json::parse(cut);
-    auto register_field = [&](const std::string & prefix, const nlohmann::detail::iteration_proxy_value<json::iterator> & subel) {
+    auto register_field = [&](const std::string & prefix, const common_json_entry & subel) {
         if (subel.value().is_string() && std::string(subel.value()).find("call0000") != std::string::npos) {
             format.id_field = !prefix.empty() ? prefix + "." + subel.key() : subel.key();
         } else if (subel.value().is_string() && std::string(subel.value()) == fun_name_needle) {

common/chat-peg-parser.cpp

@@ -4,12 +4,10 @@
 #include "ggml.h"
 #include "peg-parser.h"
 
-#include <nlohmann/json.hpp>
-
 #include <cstdint>
 #include <functional>
 
-using ordered_json = nlohmann::ordered_json;
+using ordered_json = common_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;

common/chat-peg-parser.h

@@ -128,7 +128,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const nlohmann::ordered_json &   tools,
+                                          const common_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -143,13 +143,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const nlohmann::ordered_json &             tools,
+                                                 const common_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
+    common_peg_parser python_style_tool_calls(const common_json & tools,
                                               bool                           parallel_tool_calls,
                                               bool                           allow_json_literals);
 
@@ -158,19 +158,19 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser python_or_json_value();
 
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_function_is_key(const common_json & tools,
                                                        const std::string &            args_key,
                                                        const std::string &            effective_args_key,
                                                        const std::string &            call_id_key,
                                                        const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_nested_keys(const common_json & tools,
                                                    const std::string &            effective_name_key,
                                                    const std::string &            effective_args_key,
                                                    const std::string &            call_id_key,
                                                    const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
+    common_peg_parser build_json_tools_flat_keys(const common_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,

common/chat.cpp

@@ -6,6 +6,7 @@
 #include "common.h"
 #include "ggml.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "log.h"
 
 #include "jinja/value.h"
@@ -13,14 +14,13 @@
 #include "jinja/caps.h"
 #include "peg-parser.h"
 
-#include "nlohmann/json.hpp"
-
 #include <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <ctime>
 #include <exception>
 #include <functional>
+#include <iomanip>
 #include <map>
 
 #include <optional>
@@ -30,7 +30,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto               time       = std::chrono::system_clock::to_time_t(now);
@@ -48,7 +48,7 @@ static json safe_args_parse(const std::string & to_parse) {
     }
     try {
         return json::parse(stripped);
-    } catch (json::exception & e) {
+    } catch (const common_json_error & e) {
         return stripped;
     }
 }
@@ -488,17 +488,17 @@ struct messages_inp_normalizer {
         json normalized = json::array();
         for (const auto & msg : messages) {
             json copy = msg;
-            auto it = copy.find("content");
-            if (it != copy.end()) {
-                if (only_typed && it->is_string()) {
-                    *it = json::array({
+            if (copy.contains("content")) {
+                json & it = copy.at("content");
+                if (only_typed && it.is_string()) {
+                    it = json::array({
                         json{
                             {"type", "text"},
-                            {"text", it->get<std::string>()},
+                            {"text", it.get<std::string>()},
                         }
                     });
-                } else if (only_string && it->is_array()) {
-                    *it = concat_content_parts(*it);
+                } else if (only_string && it.is_array()) {
+                    it = concat_content_parts(it);
                 }
             }
             normalized.push_back(std::move(copy));
@@ -608,7 +608,7 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
+common_chat_continuation common_chat_continuation_parse(const common_json & value) {
     if (value.is_boolean() && value.get<bool>()) {
         return COMMON_CHAT_CONTINUATION_AUTO;
     }
@@ -920,7 +920,7 @@ static void foreach_parameter(const json &
     const auto &          props = params.at("properties");
     std::set<std::string> required;
     if (params.contains("required") && params.at("required").is_array()) {
-        params.at("required").get_to(required);
+        required = params.at("required").get<std::set<std::string>>();
     }
     for (const auto & [name, prop] : props.items()) {
         bool is_required = (required.find(name) != required.end());
@@ -937,7 +937,7 @@ static std::string common_chat_template_direct_apply_impl(
     jinja::context ctx(tmpl.source());
 
     // messages_override is already built for this template, do not touch its content parts
-    nlohmann::ordered_json inp = nlohmann::ordered_json{
+    json inp = json{
         {"messages", messages_override.has_value()
             ? *messages_override
             : messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
@@ -1058,7 +1058,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
                 });
             } else if (msg.at("content").is_array()) {
                 auto blocks = msg.at("content");
-                content.insert(content.end(), blocks.begin(), blocks.end());
+                content.insert(blocks);
             }
         }
 
@@ -2238,7 +2238,7 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
 
                 std::set<std::string> required;
                 if (params.contains("required")) {
-                    params.at("required").get_to(required);
+                    required = params.at("required").get<std::set<std::string>>();
                 }
 
                 auto schema_info = common_schema_info();
@@ -2860,7 +2860,7 @@ static common_chat_params common_chat_params_init_minimax_m3(const common_chat_t
 
                 std::set<std::string> required;
                 if (schema.contains("required")) {
-                    schema.at("required").get_to(required);
+                    required = schema.at("required").get<std::set<std::string>>();
                 }
 
                 std::vector<common_peg_parser> required_elements;
@@ -2972,10 +2972,10 @@ static void system_message_not_supported(json & messages) {
             auto & second_msg = messages[1];
             second_msg["content"] = first_msg.at("content").get<std::string>()
                 + "\n" + second_msg.at("content").get<std::string>();
-            messages.erase(messages.begin());
+            messages.erase(0);
         } else {
             LOG_WRN("Removing system prompt due to template not supporting system role\n");
-            messages.erase(messages.begin());
+            messages.erase(0);
         }
     }
 }

common/chat.h

@@ -8,7 +8,7 @@
 #include "jinja/runtime.h"
 #include "jinja/caps.h"
 
-#include "nlohmann/json_fwd.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <functional>
@@ -17,7 +17,6 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
-using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 
@@ -87,7 +86,7 @@ struct common_chat_msg {
     std::string                               tool_name;
     std::string                               tool_call_id;
 
-    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
+    common_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     std::string render_content(const std::string & delimiter = "\n\n") const;
 
@@ -211,7 +210,7 @@ struct common_chat_msg_delimiters {
     // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
     common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
 
-    nlohmann::ordered_json to_json() const;
+    common_json to_json() const;
 };
 
 struct common_chat_tool {
@@ -350,16 +349,16 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const common_json & messages);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const common_json & tools);
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
+common_chat_continuation common_chat_continuation_parse(const common_json & value);
 
 // DEPRECATED: only used in tests
-nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
@@ -386,4 +385,4 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const common_json & delimiters);

common/common.cpp

@@ -1294,11 +1294,34 @@ common_init_result::common_init_result(common_params & params, bool model_only)
     if (params.fit_params) {
         COM_TRC("%s", "fitting params to device memory ...\n");
         COM_TRC("%s", "(for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on)\n");
+
+        // the draft context is created from the same base params and follows the main context, fit both together
+        const bool has_draft = params.speculative.has_dft();
+        const bool spec_mtp  = std::find(params.speculative.types.begin(), params.speculative.types.end(),
+            COMMON_SPECULATIVE_TYPE_DRAFT_MTP) != params.speculative.types.end();
+
+        common_params params_dft = common_base_params_to_speculative(params);
+
+        auto mparams_dft = common_model_params_to_llama(params_dft);
+        auto cparams_dft = common_context_params_to_llama(params_dft);
+        if (spec_mtp) {
+            cparams_dft.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
+        }
+        cparams_dft.n_rs_seq = 0;
+
+        const common_fit_extra_model extra = {
+            /*.path_model   =*/ params_dft.model.path.c_str(),
+            /*.mparams      =*/ &mparams_dft,
+            /*.cparams      =*/ &cparams_dft,
+            /*.shares_model =*/ !has_draft, // an MTP context runs on the weights of the main model
+        };
+
         common_fit_params(params.model.path.c_str(), &mparams, &cparams,
             params.tensor_split,
             params.tensor_buft_overrides.data(),
             params.fit_params_target.data(),
             params.fit_params_min_ctx,
+            has_draft || spec_mtp ? &extra : nullptr,
             params.verbosity >= LOG_LEVEL_DEBUG ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1750,18 +1773,6 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const commo
     return tpp;
 }
 
-namespace {
-
-bool can_share_threadpool(const ggml_threadpool_params & tpp1, const ggml_threadpool_params & tpp2) {
-    // n_threads does not matter -> we'll use what's larger
-    ggml_threadpool_params tpp_comparison = tpp1;
-    tpp_comparison.n_threads = tpp2.n_threads;
-
-    return ggml_threadpool_params_match(&tpp_comparison, &tpp2);
-}
-
-}  // namespace
-
 common_threadpools::~common_threadpools() {
     if (!free_fn) {
         return;
@@ -1790,9 +1801,9 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
     struct ggml_threadpool_params tpp =
             ggml_threadpool_params_from_cpu_params(params.cpuparams);
 
-    if (can_share_threadpool(tpp, tpp_batch)) {
-        tpp.n_threads = std::max(tpp.n_threads, tpp_batch.n_threads);
-    } else {
+    // each pool needs to match the respective n_threads exactly
+    // see: https://github.com/ggml-org/llama.cpp/pull/27138#issuecomment-5332307332
+    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
         threadpool_batch = ggml_threadpool_new_fn(&tpp_batch);
         if (!threadpool_batch) {
             COM_WRN("batch threadpool create failed : n_threads %d\n", tpp_batch.n_threads);

common/common.h

@@ -581,9 +581,10 @@ struct common_params {
 
     // multimodal models (see tools/mtmd)
     struct common_params_model mmproj;
-    bool mmproj_use_gpu = true;     // use GPU for multimodal model
-    bool no_mmproj = false;         // explicitly disable multimodal model
-    std::vector<std::string> image; // path to image file(s) ; TODO: change the name to "media"
+    bool mmproj_use_gpu = true;                 // use GPU for multimodal model
+    ggml_backend_dev_t mmproj_device = nullptr; // GPU device to use for multimodal model
+    bool no_mmproj = false;                     // explicitly disable multimodal model
+    std::vector<std::string> image;             // path to image file(s) ; TODO: change the name to "media"
     int image_min_tokens = -1;
     int image_max_tokens = -1;
     int mtmd_batch_max_tokens = 1024;

common/download.cpp

@@ -5,9 +5,7 @@
 #include "log.h"
 #include "download.h"
 #include "hf-cache.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <filesystem>
@@ -44,8 +42,6 @@
 #include <unistd.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 //
 // downloader
 //
@@ -856,8 +852,8 @@ static std::string common_docker_get_token(const std::string & repo) {
         throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
     }
 
-    std::string            response_str(res.second.begin(), res.second.end());
-    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
+    std::string response_str(res.second.begin(), res.second.end());
+    common_json response = common_json::parse(response_str);
 
     if (!response.contains("token")) {
         throw std::runtime_error("Docker registry token response missing 'token' field");
@@ -919,9 +915,9 @@ std::string common_docker_resolve_model(const std::string & docker) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
         }
 
-        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
-        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
-        std::string            gguf_digest;  // Find the GGUF layer
+        std::string manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        common_json manifest = common_json::parse(manifest_str);
+        std::string gguf_digest;  // Find the GGUF layer
         if (manifest.contains("layers")) {
             for (const auto & layer : manifest["layers"]) {
                 if (layer.contains("mediaType")) {

common/fit.cpp

@@ -178,7 +178,7 @@ common_device_memory_data_vec common_get_device_memory_data(
 static void common_params_fit_impl(
         const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
         float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
-        size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
+        size_t * margins_s, uint32_t n_ctx_min, const common_fit_extra_model * extra, enum ggml_log_level log_level) {
     if (mparams->split_mode == LLAMA_SPLIT_MODE_TENSOR) {
         throw common_params_fit_exception("llama_params_fit is not implemented for SPLIT_MODE_TENSOR, abort");
     }
@@ -191,10 +191,92 @@ static void common_params_fit_impl(
     uint32_t hp_nct = 0; // hparams.n_ctx_train
     uint32_t hp_nex = 0; // hparams.n_expert
 
+    // with non-unified kv, we need to take into account n_streams
+    // for example, if memory can hold more than model's trained context size, we must extend the n_ctx to hold enough n_streams
+    const uint32_t n_streams  = cparams->kv_unified ? 1 : std::max<uint32_t>(1, cparams->n_seq_max);
+    const bool     n_ctx_auto = cparams->n_ctx == 0;
+
+    dmds_t   dmds_extra;       // memory of the extra model, laid out on the devices of the main model
+    uint32_t n_ctx_extra = 0;  // context that memory was measured at
+
+    // the extra model competes for the same memory as the main model, add it to every measurement
+    // its memory is measured again whenever the context it follows changes
+    auto add_extra_memory = [&](dmds_t & dmds) {
+        if (extra == nullptr) {
+            return;
+        }
+
+        if (dmds_extra.empty() || n_ctx_extra != cparams->n_ctx) {
+            std::vector<ggml_backend_dev_t> devs_extra;
+            uint32_t ngl_extra = 0;
+            uint32_t nct_extra = 0;
+            uint32_t nex_extra = 0;
+
+            extra->cparams->n_ctx = cparams->n_ctx;
+
+            LOG_TRC("%s: getting device memory data for the extra model at a context size of %" PRIu32 ":\n",
+                __func__, cparams->n_ctx);
+
+            dmds_t measured;
+            try {
+                measured = common_get_device_memory_data_impl(
+                    extra->path_model, extra->mparams, extra->cparams, devs_extra, ngl_extra, nct_extra, nex_extra, log_level);
+            } catch (const std::runtime_error & e) {
+                // the extra model is optional, fit the main model alone rather than giving up
+                LOG_WRN("%s: failed to measure the memory of the extra model, fitting without it: %s\n", __func__, e.what());
+                dmds_extra = dmds_t(devs.size() + 1);
+                n_ctx_extra = cparams->n_ctx;
+                return;
+            }
+
+            dmds_extra = dmds_t(devs.size() + 1);
+            dmds_extra.back().mb = measured.back().mb;
+            for (size_t je = 0; je < devs_extra.size(); je++) {
+                for (size_t id = 0; id < devs.size(); id++) {
+                    if (devs_extra[je] == devs[id]) {
+                        dmds_extra[id].mb.model   += measured[je].mb.model;
+                        dmds_extra[id].mb.context += measured[je].mb.context;
+                        dmds_extra[id].mb.compute += measured[je].mb.compute;
+                        break;
+                    }
+                }
+            }
+            if (extra->shares_model) {
+                for (llama_device_memory_data & dmd : dmds_extra) {
+                    dmd.mb.model = 0;
+                }
+            }
+
+            n_ctx_extra = cparams->n_ctx;
+        }
+
+        for (size_t id = 0; id < dmds.size(); id++) {
+            dmds[id].mb.model   += dmds_extra[id].mb.model;
+            dmds[id].mb.context += dmds_extra[id].mb.context;
+            dmds[id].mb.compute += dmds_extra[id].mb.compute;
+        }
+    };
+
     // step 1: get data for default parameters and check whether any changes are necessary in the first place
 
     LOG_TRC("%s: getting device memory data for initial parameters:\n", __func__);
-    const dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+    dmds_t dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+
+    // saturate instead of overflowing, this also preserves the UINT32_MAX sentinel of n_ctx_min:
+    const uint32_t n_ctx_max       = (uint32_t) std::min<uint64_t>(uint64_t(hp_nct)    * n_streams, UINT32_MAX);
+    const uint32_t n_ctx_min_total = (uint32_t) std::min<uint64_t>(uint64_t(n_ctx_min) * n_streams, UINT32_MAX);
+
+    // llama_context would use only hp_nct in total for n_ctx == 0, resolve the context before measuring anything else:
+    if (n_ctx_auto) {
+        cparams->n_ctx = n_ctx_max;
+        if (n_streams > 1) {
+            LOG_TRC("%s: context size unset and KV cache not unified -> using %" PRIu32 " for %" PRIu32 " sequences:\n",
+                __func__, n_ctx_max, n_streams);
+            dmds_full = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+        }
+    }
+    add_extra_memory(dmds_full);
+
     const size_t nd = devs.size(); // number of devices
 
     std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
@@ -307,8 +389,8 @@ static void common_params_fit_impl(
                     "%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
                     __func__, -global_surplus/MiB);
             }
-            if (cparams->n_ctx == 0) {
-                if (hp_nct > n_ctx_min) {
+            if (n_ctx_auto) {
+                if (n_ctx_max > n_ctx_min_total) {
                     int64_t sum_used_target = sum_free;
                     if (nd == 0) {
                         sum_used_target -= margins[0];
@@ -328,8 +410,9 @@ static void common_params_fit_impl(
                     }
 
                     int64_t sum_projected_used_min_ctx = 0;
-                    cparams->n_ctx = n_ctx_min;
-                    const dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    cparams->n_ctx = n_ctx_min_total;
+                    dmds_t dmds_min_ctx = common_get_device_memory_data_impl(path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+                    add_extra_memory(dmds_min_ctx);
                     if (nd == 0) {
                         sum_projected_used_min_ctx = dmds_min_ctx.back().mb.total();
                     } else {
@@ -339,14 +422,16 @@ static void common_params_fit_impl(
                     }
                     if (sum_used_target > sum_projected_used_min_ctx) {
                         // linear interpolation between minimum and maximum context size:
-                        cparams->n_ctx += (hp_nct - n_ctx_min) * (sum_used_target - sum_projected_used_min_ctx)
+                        cparams->n_ctx += (n_ctx_max - n_ctx_min_total) * (sum_used_target - sum_projected_used_min_ctx)
                             / (sum_projected_used - sum_projected_used_min_ctx);
-                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % 256, n_ctx_min); // round down context for CUDA backend
+                        // round down context for CUDA backend, keep it divisible by the number of streams:
+                        const uint32_t align = 256 * n_streams;
+                        cparams->n_ctx = std::max(cparams->n_ctx - cparams->n_ctx % align, n_ctx_min_total);
 
-                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (hp_nct - n_ctx_min);
-                        const int64_t memory_reduction = (hp_nct - cparams->n_ctx) * bytes_per_ctx;
+                        const int64_t bytes_per_ctx = (sum_projected_used - sum_projected_used_min_ctx) / (n_ctx_max - n_ctx_min_total);
+                        const int64_t memory_reduction = (n_ctx_max - cparams->n_ctx) * bytes_per_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
                         if (nd <= 1) {
                             LOG_TRC("%s: entire model can be fit by reducing context\n", __func__);
                             return;
@@ -355,14 +440,14 @@ static void common_params_fit_impl(
                     } else {
                         const int64_t memory_reduction = sum_projected_used - sum_projected_used_min_ctx;
                         LOG_TRC("%s: context size reduced from %" PRIu32 " to %" PRIu32 " -> need %" PRId64 " MiB less memory in total\n",
-                            __func__, hp_nct, cparams->n_ctx, memory_reduction/MiB);
+                            __func__, n_ctx_max, cparams->n_ctx, memory_reduction/MiB);
                     }
                 } else {
                     if (n_ctx_min == UINT32_MAX) {
-                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, hp_nct);
+                        LOG_TRC("%s: user has requested full context size of %" PRIu32 " -> no change\n", __func__, n_ctx_max);
                     } else {
                         LOG_TRC("%s: default model context size is %" PRIu32 " which is <= the min. context size of %" PRIu32 " -> no change\n",
-                            __func__, hp_nct, n_ctx_min);
+                            __func__, n_ctx_max, n_ctx_min_total);
                     }
                 }
             } else {
@@ -507,8 +592,9 @@ static void common_params_fit_impl(
         llama_model_params mparams_copy = *mparams;
         set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, mparams_copy);
 
-        const dmds_t dmd_nl = common_get_device_memory_data_impl(
+        dmds_t dmd_nl = common_get_device_memory_data_impl(
             path_model, &mparams_copy, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+        add_extra_memory(dmd_nl);
 
         LOG_TRC("%s: memory for test allocation by device:\n", func_name);
         for (size_t id = 0; id < nd; id++) {
@@ -535,8 +621,9 @@ static void common_params_fit_impl(
         mparams->tensor_buft_overrides = tensor_buft_overrides;
 
         LOG_TRC("%s: getting device memory data with all MoE tensors moved to system memory:\n", __func__);
-        const dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
+        dmds_t dmds_cpu_moe = common_get_device_memory_data_impl(
             path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
+        add_extra_memory(dmds_cpu_moe);
 
         for (size_t id = 0; id < nd; id++) {
             global_surplus_cpu_moe += dmds_cpu_moe[id].free;
@@ -796,11 +883,12 @@ enum common_params_fit_status common_fit_params(
         llama_model_tensor_buft_override * tensor_buft_overrides,
         size_t * margins,
         uint32_t n_ctx_min,
+        const common_fit_extra_model * extra,
         ggml_log_level log_level) {
     const int64_t t0_us = llama_time_us();
     common_params_fit_status status = COMMON_PARAMS_FIT_STATUS_SUCCESS;
     try {
-        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
+        common_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, extra, log_level);
         LOG_TRC("%s: successfully fit params to free device memory\n", __func__);
     } catch (const common_params_fit_exception & e) {
         LOG_WRN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());

common/fit.h

@@ -11,6 +11,16 @@ enum common_params_fit_status {
     COMMON_PARAMS_FIT_STATUS_ERROR   = 2, // a hard error occurred, e.g. because no model could be found at the specified path
 };
 
+// a second model that shares the devices of the main model, e.g. a draft model
+//   - its context follows the context of the main model, so its memory is measured again whenever that context changes
+//   - shares_model tells the fit that the weights are already counted in the main model, as for an MTP context
+struct common_fit_extra_model {
+    const char * path_model;
+    llama_model_params * mparams;
+    llama_context_params * cparams;
+    bool shares_model;
+};
+
 // fits mparams and cparams to free device memory (assumes system memory is unlimited)
 //   - returns true if the parameters could be successfully modified to fit device memory
 //   - this function is NOT thread safe because it modifies the global llama logger state
@@ -24,6 +34,7 @@ common_params_fit_status common_fit_params(
    llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
                              size_t * margins,               // margins of memory to leave per device in bytes
                            uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
+      const common_fit_extra_model * extra,                  // model to fit alongside the main one, nullptr if there is none
                      ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
 // print estimated memory to stdout

common/hf-cache.cpp

@@ -4,9 +4,7 @@
 #include "common.h"
 #include "log.h"
 #include "http.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <filesystem>
 #include <fstream>
@@ -15,8 +13,6 @@
 #include <string_view>
 #include <stdexcept>
 
-namespace nl = nlohmann;
-
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
@@ -195,8 +191,8 @@ static void safe_write_file(const fs::path & path, const std::string & data) {
     }
 }
 
-static nl::json api_get(const std::string & url,
-                        const std::string & token) {
+static common_json api_get(const std::string & url,
+                           const std::string & token) {
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
@@ -214,10 +210,10 @@ static nl::json api_get(const std::string & url,
         auto body = res->body;
 
         if (res->status == 200) {
-            return nl::json::parse(res->body);
+            return common_json::parse(res->body);
         }
         try {
-            body = nl::json::parse(res->body)["error"].get<std::string>();
+            body = common_json::parse(res->body)["error"].get<std::string>();
         } catch (...) { }
 
         throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
@@ -280,7 +276,7 @@ static std::string get_repo_commit(const std::string & repo_id,
         safe_write_file(refs_path / name, commit);
         return commit;
 
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
@@ -358,7 +354,7 @@ hf_files get_repo_files(const std::string & repo_id,
 
             files.push_back(file);
         }
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());

common/jinja/README.md

@@ -7,7 +7,7 @@ The implementation can be found in the `common/jinja` directory.
 ## Key Features
 
 - Input marking: security against special token injection
-- Decoupled from `nlohmann::json`: this dependency is only used for JSON-to-internal type translation and is completely optional
+- Decoupled from the JSON library: `common_json` is only used for JSON-to-internal type translation and is completely optional
 - Minimal primitive types: int, float, bool, string, array, object, none, undefined
 - Detailed logging: allow source tracing on error
 - Clean architecture: workarounds are applied to input data before entering the runtime (see `common/chat.cpp`)

common/jinja/caps.cpp

@@ -4,14 +4,14 @@
 
 // note: the json dependency is only for defining input in a convenient way
 // we can remove it in the future when we figure out a better way to define inputs using jinja::value
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <functional>
 #include <sstream>
 
 #define FILENAME "jinja-caps"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace jinja {
 

common/jinja/value.cpp

@@ -3,7 +3,7 @@
 #include "value.h"
 
 // for converting from JSON to jinja values
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <sstream>
 #include <string>
@@ -1355,7 +1355,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const nlohmann::ordered_json & j, bool mark_input) {
+static value from_json(const common_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1452,7 +1452,7 @@ bool value_compare(const value & a, const value & b, value_compare_op op) {
 }
 
 template<>
-void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bool mark_input) {
+void global_from_json(context & ctx, const common_json & json_obj, bool mark_input) {
     // printf("global_from_json: %s\n" , json_obj.dump(2).c_str());
     if (json_obj.is_null() || !json_obj.is_object()) {
         throw std::runtime_error("global_from_json: input JSON value must be an object");

common/jinja/value.h

@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be nlohmann::ordered_json
+// Note: T_JSON can be common_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 

common/json-schema-to-grammar.cpp

@@ -1,9 +1,8 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
+#include <limits>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -12,7 +11,7 @@
 #include <unordered_set>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
@@ -278,7 +277,9 @@ static std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
     {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}, {'\\', "\\\\"}
 };
 
-static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+static const int MAX_PATTERN_DEPTH = 100;
+
+static std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?', '^', '$'};
 static std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
 
 static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
@@ -309,6 +310,32 @@ static std::string format_literal(const std::string & literal) {
 
 std::string gbnf_format_literal(const std::string & literal) { return format_literal(literal); }
 
+static size_t gbnf_escape_length(const std::string & pattern, size_t pos) {
+    if (pos + 1 >= pattern.length() || pattern[pos] != '\\') {
+        return 0;
+    }
+    size_t n_hex = 0;
+    switch (pattern[pos + 1]) {
+        case 'x': n_hex = 2; break;
+        case 'u': n_hex = 4; break;
+        case 'U': n_hex = 8; break;
+        case 't': case 'r': case 'n': case '\\': case '"': case '[': case ']':
+            return 2;
+        default:
+            return 0;
+    }
+    if (pos + 2 + n_hex > pattern.length()) {
+        return 0;
+    }
+    for (size_t i = pos + 2; i < pos + 2 + n_hex; i++) {
+        char h = pattern[i];
+        if (!((h >= '0' && h <= '9') || (h >= 'a' && h <= 'f') || (h >= 'A' && h <= 'F'))) {
+            return 0;
+        }
+    }
+    return 2 + n_hex;
+}
+
 class common_schema_converter {
 private:
     friend class common_schema_info;
@@ -345,16 +372,42 @@ private:
         return string_join(rules, " | ");
     }
 
+    // thrown when the pattern is a valid regex with no grammar equivalent
+    struct unsupported_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
+    // thrown when the pattern is not a valid regex
+    struct invalid_pattern : public std::runtime_error {
+        using std::runtime_error::runtime_error;
+    };
+
     std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        if (!(pattern.front() == '^' && pattern.back() == '$')) {
-            _errors.push_back("Pattern must start with '^' and end with '$'");
+        auto rules_snapshot = _rules;
+        try {
+            return _pattern_to_rule(pattern, name);
+        } catch (const unsupported_pattern & err) {
+            // revert rules
+            _rules = std::move(rules_snapshot);
+            _warnings.push_back("pattern " + pattern + " is not supported (" + err.what() + "), accepting any string");
+            return _add_rule(name, _add_primitive("string", PRIMITIVE_RULES.at("string")));
+        } catch (const invalid_pattern & err) {
+            _rules = std::move(rules_snapshot);
+            _errors.push_back("Invalid pattern " + pattern + ": " + err.what());
             return "";
         }
+    }
+
+    std::string _pattern_to_rule(const std::string & pattern, const std::string & name) {
+        if (pattern.length() < 2 || pattern.front() != '^' || pattern.back() != '$') {
+            throw unsupported_pattern("not anchored with '^' and '$'");
+        }
         std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
         std::unordered_map<std::string, std::string> sub_rule_ids;
 
         size_t i = 0;
         size_t length = sub_pattern.length();
+        int paren_depth = 0;
 
         using literal_or_rule = std::pair<std::string, bool>;
         auto to_rule = [&](const literal_or_rule & ls) {
@@ -363,7 +416,6 @@ private:
             return is_literal ? "\"" + s + "\"" : s;
         };
         std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
-            size_t start = i;
             std::vector<literal_or_rule> seq;
 
             auto get_dot = [&]() {
@@ -420,43 +472,42 @@ private:
                         if (i + 1 < length && sub_pattern[i + 1] == ':') {
                             i += 2; // skip "?:" for non-capturing group, treat as regular group
                         } else {
-                            // lookahead/lookbehind (?=, ?!, ?<=, ?<!) - not supported
-                            _warnings.push_back("Unsupported pattern syntax");
-                            // skip to matching ')' to avoid UB on empty seq
-                            int depth = 1;
-                            while (i < length && depth > 0) {
-                                if (sub_pattern[i] == '\\' && i + 1 < length) {
-                                    i += 2; // skip escaped character
-                                } else {
-                                    if (sub_pattern[i] == '(') depth++;
-                                    else if (sub_pattern[i] == ')') depth--;
-                                    i++;
-                                }
-                            }
-                            continue;
+                            // lookaround, named group, inline flags, ...
+                            throw unsupported_pattern("unsupported group syntax");
                         }
                     }
+                    paren_depth++;
+                    if (paren_depth > MAX_PATTERN_DEPTH) {
+                        throw unsupported_pattern("pattern nesting too deep");
+                    }
                     seq.emplace_back("(" + to_rule(transform()) + ")", false);
                 } else if (c == ')') {
                     i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(' && (start < 2 || sub_pattern[start - 2] != '?' || sub_pattern[start - 1] != ':')) {
-                        _errors.push_back("Unbalanced parentheses");
+                    if (paren_depth == 0) {
+                        throw invalid_pattern("unbalanced parentheses");
                     }
+                    paren_depth--;
                     return join_seq();
+                } else if (c == '^' || c == '$') {
+                    throw unsupported_pattern("anchor inside the pattern");
                 } else if (c == '[') {
                     std::string square_brackets = std::string(1, c);
                     i++;
                     while (i < length && sub_pattern[i] != ']') {
                         if (sub_pattern[i] == '\\') {
-                            square_brackets += sub_pattern.substr(i, 2);
-                            i += 2;
+                            auto escape_length = gbnf_escape_length(sub_pattern, i);
+                            if (escape_length == 0) {
+                                throw unsupported_pattern("unsupported escape in character class: " + sub_pattern.substr(i, 2));
+                            }
+                            square_brackets += sub_pattern.substr(i, escape_length);
+                            i += escape_length;
                         } else {
                             square_brackets += sub_pattern[i];
                             i++;
                         }
                     }
                     if (i >= length) {
-                        _errors.push_back("Unbalanced square brackets");
+                        throw invalid_pattern("unterminated character class");
                     }
                     square_brackets += ']';
                     i++;
@@ -465,6 +516,9 @@ private:
                     seq.emplace_back("|", false);
                     i++;
                 } else if (c == '*' || c == '+' || c == '?') {
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
+                    }
                     seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
                     i++;
                 } else if (c == '{') {
@@ -475,18 +529,19 @@ private:
                         i++;
                     }
                     if (i >= length) {
-                        _errors.push_back("Unbalanced curly brackets");
+                        throw unsupported_pattern("unterminated curly brackets");
                     }
                     curly_brackets += '}';
                     i++;
                     auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
                     int min_times = 0;
                     int max_times = std::numeric_limits<int>::max();
+                    if (nums.size() != 1 && nums.size() != 2) {
+                        throw unsupported_pattern("wrong number of values in curly brackets");
+                    }
                     try {
                         if (nums.size() == 1) {
                             min_times = max_times = std::stoi(nums[0]);
-                        } else if (nums.size() != 2) {
-                            _errors.push_back("Wrong number of values in curly brackets");
                         } else {
                             if (!nums[0].empty()) {
                                 min_times = std::stoi(nums[0]);
@@ -495,9 +550,11 @@ private:
                                 max_times = std::stoi(nums[1]);
                             }
                         }
-                    } catch (const std::invalid_argument & e) {
-                        _errors.push_back("Invalid number in curly brackets");
-                        return std::make_pair("", false);
+                    } catch (const std::logic_error &) {
+                        throw unsupported_pattern("invalid number in curly brackets");
+                    }
+                    if (seq.empty()) {
+                        throw invalid_pattern("nothing to repeat");
                     }
                     auto &last = seq.back();
                     auto &sub = last.first;
@@ -523,15 +580,22 @@ private:
                         return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
                     };
                     while (i < length) {
-                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                        if (sub_pattern[i] == '\\') {
+                            if (i == length - 1) {
+                                throw invalid_pattern("trailing backslash");
+                            }
                             char next = sub_pattern[i + 1];
                             if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
                                 i++;
                                 literal += sub_pattern[i];
                                 i++;
                             } else {
-                                literal += sub_pattern.substr(i, 2);
-                                i += 2;
+                                auto escape_length = gbnf_escape_length(sub_pattern, i);
+                                if (escape_length == 0) {
+                                    throw unsupported_pattern("unsupported escape: " + sub_pattern.substr(i, 2));
+                                }
+                                literal += sub_pattern.substr(i, escape_length);
+                                i += escape_length;
                             }
                         } else if (sub_pattern[i] == '"') {
                             literal += "\\\"";
@@ -544,14 +608,21 @@ private:
                             break;
                         }
                     }
-                    if (!literal.empty()) {
-                        seq.emplace_back(literal, true);
+                    if (literal.empty()) { // nothing was consumed, ex. a stray ']' or '}'
+                        throw unsupported_pattern(std::string("unsupported character: ") + c);
                     }
+                    seq.emplace_back(literal, true);
                 }
             }
             return join_seq();
         };
-        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\"");
+
+        auto rule = to_rule(transform());
+        if (paren_depth != 0) {
+            throw invalid_pattern("unbalanced parentheses");
+        }
+
+        return _add_rule(name, "\"\\\"\" (" + rule + ") \"\\\"\"");
     }
 
     /*
@@ -845,7 +916,11 @@ public:
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
         }
         if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
+            std::vector<json> alt_schemas;
+            for (const auto & alt : alts) {
+                alt_schemas.push_back(alt);
+            }
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
         }
         if (schema_type.is_array()) {
@@ -1039,7 +1114,7 @@ common_schema_info::~common_schema_info() = default;
 common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
 common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
 
-void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
+void common_schema_info::resolve_refs(common_json & schema) {
     impl_->resolve_refs(schema, "");
 }
 
@@ -1047,7 +1122,7 @@ void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
 // Some models emit raw string values rather than JSON-encoded strings for string parameters.
 // If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
 // true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
+bool common_schema_info::resolves_to_string(const common_json & schema) {
     std::unordered_set<std::string> visited_refs;
 
     std::function<bool(const json &)> check = [&](const json & s) -> bool {
@@ -1155,7 +1230,7 @@ bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schem
     return check(schema);
 }
 
-std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
+std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
         return "%llguidance {}\nstart: %json " + schema.dump();
@@ -1176,10 +1251,10 @@ std::string build_grammar(const std::function<void(const common_grammar_builder
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
+        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
             return converter.visit(schema, name == "root" ? "" : name);
         },
-        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
+        /* .resolve_refs = */ [&](common_json & schema) {
             converter.resolve_refs(schema, "");
         }
     };

common/json-schema-to-grammar.h

@@ -1,12 +1,12 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <functional>
 #include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
+std::string json_schema_to_grammar(const common_json & schema,
                                    bool force_gbnf = false);
 
 class common_schema_converter;
@@ -24,14 +24,14 @@ class common_schema_info {
     common_schema_info(common_schema_info &&) noexcept;
     common_schema_info & operator=(common_schema_info &&) noexcept;
 
-    void resolve_refs(nlohmann::ordered_json & schema);
-    bool resolves_to_string(const nlohmann::ordered_json & schema);
+    void resolve_refs(common_json & schema);
+    bool resolves_to_string(const common_json & schema);
 };
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
-    std::function<void(nlohmann::ordered_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const common_json &)> add_schema;
+    std::function<void(common_json &)> resolve_refs;
 };
 
 struct common_grammar_options {

common/json.cpp

@@ -0,0 +1,433 @@
+#include "json.h"
+
+#include "ggml.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <iterator>
+#include <new>
+#include <set>
+#include <unordered_map>
+#include <vector>
+
+using nlohmann::ordered_json;
+
+// a common_json is the backing value, so any value of a tree can be used as a common_json
+static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
+static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
+
+// runs fn and gives every error of the backing library as a common_json_error
+template <typename F>
+static decltype(auto) guard(F && fn) {
+    try {
+        return fn();
+    } catch (const ordered_json::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+static ordered_json & as_json(common_json * self) {
+    return *reinterpret_cast<ordered_json *>(self);
+}
+
+static const ordered_json & as_json(const common_json * self) {
+    return *reinterpret_cast<const ordered_json *>(self);
+}
+
+static common_json & as_common(ordered_json & json) {
+    return *reinterpret_cast<common_json *>(&json);
+}
+
+static const common_json & as_common(const ordered_json & json) {
+    return *reinterpret_cast<const common_json *>(&json);
+}
+
+static ordered_json to_json(const common_json_value & val) {
+    switch (val.type) {
+        case common_json_value::VAL_NULL:   return nullptr;
+        case common_json_value::VAL_BOOL:   return val.val_bool;
+        case common_json_value::VAL_INT:    return val.val_int;
+        case common_json_value::VAL_UINT:   return val.val_uint;
+        case common_json_value::VAL_DOUBLE: return val.val_double;
+        case common_json_value::VAL_STRING: return val.val_string;
+        case common_json_value::VAL_JSON:
+            // one owner means no one else can see this tree, so it is safe to move it out
+            // note: this makes a value single use, same as the json_ref of the backing library
+            if (val.val_json.use_count() == 1) {
+                return std::move(as_json(val.val_json.get()));
+            }
+            return as_json(val.val_json.get());
+    }
+
+    return nullptr;
+}
+
+common_json_value::common_json_value(const char * val) {
+    if (val) {
+        type       = VAL_STRING;
+        val_string = val;
+    } else {
+        type = VAL_NULL;
+    }
+}
+
+common_json_value::common_json_value(const common_json & val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
+
+common_json_value::common_json_value(common_json && val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
+
+// the ctors and get<T>() below are explicit specializations, giving strong symbols
+// an explicit instantiation is a weak symbol, dropped by some LTO builds (clang-cl)
+template <typename T>
+static std::shared_ptr<common_json> set_json(const std::set<T> & vals) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    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__> & vals) : type(VAL_JSON), val_json(set_json(vals)) {}
+
+COMMON_JSON_SET(int)
+COMMON_JSON_SET(std::string)
+
+#undef COMMON_JSON_SET
+
+template <typename T>
+static std::shared_ptr<common_json> map_json(const T & vals) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    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__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+
+COMMON_JSON_MAP(bool)
+COMMON_JSON_MAP(std::string)
+
+#undef COMMON_JSON_MAP
+
+// an unordered map value is usable only for the types below
+#define COMMON_JSON_UMAP(...) template <> common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> & vals) : type(VAL_JSON), val_json(map_json(vals)) {}
+
+COMMON_JSON_UMAP(size_t)
+
+#undef COMMON_JSON_UMAP
+
+template <typename T>
+static std::shared_ptr<common_json> vec_json(const std::vector<T> & vals) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    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__> & vals) : type(VAL_JSON), val_json(vec_json(vals)) {}
+
+COMMON_JSON_VEC(int)
+COMMON_JSON_VEC(unsigned char)
+COMMON_JSON_VEC(unsigned int)
+COMMON_JSON_VEC(long)
+COMMON_JSON_VEC(unsigned long)
+COMMON_JSON_VEC(long long)
+COMMON_JSON_VEC(unsigned long long)
+COMMON_JSON_VEC(float)
+COMMON_JSON_VEC(double)
+COMMON_JSON_VEC(std::string)
+COMMON_JSON_VEC(std::vector<float>)
+COMMON_JSON_VEC(common_json)
+
+#undef COMMON_JSON_VEC
+
+common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
+
+// null, same as the backing library
+// operator[] turns it into an object, push_back() into an array
+common_json::common_json() {
+    new (storage) ordered_json();
+}
+
+common_json::common_json(const common_json & other) {
+    new (storage) ordered_json(as_json(&other));
+}
+
+common_json::common_json(common_json && other) noexcept {
+    new (storage) ordered_json(std::move(as_json(&other)));
+}
+
+common_json::common_json(std::initializer_list<common_json_item> items) {
+    new (storage) ordered_json(ordered_json::object());
+
+    for (const auto & item : items) {
+        set(item);
+    }
+}
+
+common_json::common_json(const common_json_value & val) {
+    new (storage) ordered_json(to_json(val));
+}
+
+common_json::common_json(std::nullptr_t) {
+    new (storage) ordered_json(nullptr);
+}
+
+common_json & common_json::operator=(common_json other) noexcept {
+    as_json(this).swap(as_json(&other));
+
+    return *this;
+}
+
+common_json::~common_json() {
+    as_json(this).~basic_json();
+}
+
+common_json common_json::parse(const std::string & text) {
+    try {
+        // the assignment moves the parsed tree in, it does not copy
+        common_json out;
+        as_json(&out) = ordered_json::parse(text);
+        return out;
+    } catch (const std::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+common_json common_json::parse_no_throw(const std::string & text) {
+    common_json out;
+    as_json(&out) = ordered_json::parse(text, nullptr, false);
+    return out;
+}
+
+bool common_json::is_discarded() const {
+    return as_json(this).is_discarded();
+}
+
+common_json common_json::array() {
+    common_json out;
+    as_json(&out) = ordered_json::array();
+    return out;
+}
+
+common_json common_json::array(std::initializer_list<common_json_value> vals) {
+    common_json out;
+    ordered_json & arr = as_json(&out);
+    arr = ordered_json::array();
+
+    for (const auto & val : vals) {
+        arr.push_back(to_json(val));
+    }
+
+    return out;
+}
+
+common_json common_json::object() {
+    common_json out;
+    as_json(&out) = ordered_json::object();
+    return out;
+}
+
+common_json common_json::object(std::initializer_list<common_json_item> items) {
+    return common_json(items);
+}
+
+common_json common_json::make(const common_json_value & val) {
+    return common_json(val);
+}
+
+bool common_json::is_null()           const { return as_json(this).is_null(); }
+bool common_json::is_object()         const { return as_json(this).is_object(); }
+bool common_json::is_array()          const { return as_json(this).is_array(); }
+bool common_json::is_string()         const { return as_json(this).is_string(); }
+bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
+bool common_json::is_number()         const { return as_json(this).is_number(); }
+bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
+bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
+
+bool   common_json::empty() const { return as_json(this).empty(); }
+size_t common_json::size()  const { return as_json(this).size(); }
+
+bool common_json::contains(const std::string & key) const {
+    return as_json(this).contains(key);
+}
+
+bool common_json::operator==(const common_json_value & val) const {
+    // compare a tree in place, to_json() would copy it
+    if (val.type == common_json_value::VAL_JSON) {
+        return as_json(this) == as_json(val.val_json.get());
+    }
+    return as_json(this) == to_json(val);
+}
+
+bool common_json::operator!=(const common_json_value & val) const {
+    return !(*this == val);
+}
+
+common_json       & common_json::at(const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this).at(key)); }); }
+const common_json & common_json::at(const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::at(size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this).at(idx)); }); }
+const common_json & common_json::at(size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::operator[](const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this)[key]); }); }
+const common_json & common_json::operator[](const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::operator[](size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this)[idx]); }); }
+const common_json & common_json::operator[](size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::front()       { return as_common(as_json(this).front()); }
+const common_json & common_json::front() const { return as_common(as_json(this).front()); }
+common_json       & common_json::back()        { return as_common(as_json(this).back()); }
+const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
+
+void common_json::clear() {
+    as_json(this).clear();
+}
+
+void common_json::erase(const std::string & key) {
+    guard([&] { as_json(this).erase(key); });
+}
+
+void common_json::erase(size_t idx) {
+    guard([&] { as_json(this).erase(idx); });
+}
+
+void common_json::assign(const common_json_value & val) {
+    as_json(this) = to_json(val);
+}
+
+void common_json::set(const common_json_item & item) {
+    guard([&] { as_json(this)[item.key] = to_json(item.val); });
+}
+
+void common_json::push_back(const common_json_value & val) {
+    guard([&] { as_json(this).push_back(to_json(val)); });
+}
+
+void common_json::push_back(std::initializer_list<common_json_item> items) {
+    common_json val(items);
+
+    guard([&] { as_json(this).push_back(std::move(as_json(&val))); });
+}
+
+size_t common_json::count(const std::string & key) const {
+    return as_json(this).count(key);
+}
+
+void common_json::insert(const common_json & vals) {
+    guard([&] {
+        ordered_json & self = as_json(this);
+
+        self.insert(self.end(), as_json(&vals).begin(), as_json(&vals).end());
+    });
+}
+
+std::string common_json::dump(int indent) const {
+    return guard([&] { return as_json(this).dump(indent); });
+}
+
+std::string common_json::dump_safe(int indent) const {
+    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
+}
+
+// an array is indexed directly, an object needs a walk from the start
+common_json & common_json::iterator::operator*() const {
+    return guard([&]() -> common_json & {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            return as_common(std::next(j.begin(), idx).value());
+        }
+        if (j.is_array()) {
+            return as_common(j[idx]);
+        }
+
+        // a plain value gives itself once, same as the backing library
+        return *node;
+    });
+}
+
+std::string common_json::iterator::key() const {
+    return guard([&] { return std::next(as_json(node).begin(), idx).key(); });
+}
+
+common_json::iterator common_json::begin() const {
+    return iterator(const_cast<common_json *>(this), 0);
+}
+
+common_json::iterator common_json::end() const {
+    return iterator(const_cast<common_json *>(this), size());
+}
+
+// the keys follow the backing library: the index for an array, "" for a plain value
+common_json::items_view::entry common_json::items_view::iterator::operator*() const {
+    return guard([&]() -> entry {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            auto it = std::next(j.begin(), idx);
+
+            return { it.key(), as_common(it.value()) };
+        }
+        if (j.is_array()) {
+            return { std::to_string(idx), as_common(j[idx]) };
+        }
+
+        return { std::string(), *node };
+    });
+}
+
+common_json::items_view common_json::items() const {
+    return items_view(const_cast<common_json *>(this), size());
+}
+
+// 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;
+}
+
+// get<T>() is usable only for the types below
+
+#define COMMON_JSON_GET(...) template <> __VA_ARGS__ common_json::get<__VA_ARGS__>() const { return guard([&] { return as_json(this).get<__VA_ARGS__>(); }); }
+
+COMMON_JSON_GET(bool)
+COMMON_JSON_GET(int)
+COMMON_JSON_GET(unsigned int)
+COMMON_JSON_GET(long)
+COMMON_JSON_GET(unsigned long)
+COMMON_JSON_GET(long long)
+COMMON_JSON_GET(unsigned long long)
+COMMON_JSON_GET(float)
+COMMON_JSON_GET(double)
+COMMON_JSON_GET(std::string)
+COMMON_JSON_GET(std::vector<float>)
+COMMON_JSON_GET(std::vector<std::string>)
+COMMON_JSON_GET(std::set<std::string>)
+COMMON_JSON_GET(std::vector<int>)
+COMMON_JSON_GET(std::vector<size_t>)
+COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
+
+#undef COMMON_JSON_GET
+
+// must stay below the get<std::string> specialization
+common_json::operator std::string() const {
+    return get<std::string>();
+}
+
+std::string common_json::value(const std::string & key, const char * def) const {
+    return contains(key) ? at(key).get<std::string>() : std::string(def);
+}

common/json.h

@@ -0,0 +1,352 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <initializer_list>
+#include <iterator>
+#include <map>
+#include <memory>
+#include <set>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+// common_json, a thin wrapper around vendor json library
+// the underlay library is pimpl, we are using nlohmann::json for now
+//
+// many features of the library are deliberately left out, to keep this interface small and generic and to keep compile time down
+//
+// some main differences compared to nlohmann::json :
+// - object keys keep the order in which they are added
+// - errors are always throw as common_json_error
+// - obj.push_back({key, val}) is intentionally unsupported to avoid confusion with push_back on a vector; write it as obj[key] = val for clarity
+// - a braced pair in value position does not build, e.g. {"key", {"a", "b"}}; write array({"a", "b"}) where nlohmann made an array
+//
+// in doubt, search the code base for an existing usage example; do not add anything to this header unless absolutely necessary
+
+class common_json;
+
+// common_json_value holds a list of these, and each of them holds a value, so one must come first
+struct common_json_item;
+
+struct common_json_error : std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
+// one value, tagged so that this header stays free of the backing library
+// note: a value that holds a tree is single use, the second use gives null
+struct common_json_value {
+    enum value_type {
+        VAL_NULL,
+        VAL_BOOL,
+        VAL_INT,
+        VAL_UINT,
+        VAL_DOUBLE,
+        VAL_STRING,
+        VAL_JSON,
+    };
+
+    value_type type = VAL_NULL;
+
+    union {
+        bool     val_bool;
+        int64_t  val_int;
+        uint64_t val_uint = 0;
+        double   val_double;
+    };
+
+    std::string                  val_string;
+    std::shared_ptr<common_json> val_json;
+
+    common_json_value(std::nullptr_t = nullptr) : type(VAL_NULL) {}
+    common_json_value(bool val) : type(VAL_BOOL), val_bool(val) {}
+    common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
+    // without this a string_view lands on the common_json ctor below and recurses
+    common_json_value(std::string_view val) : type(VAL_STRING), val_string(val) {}
+    common_json_value(const char * val);
+    common_json_value(const common_json & val);
+    common_json_value(common_json && val);
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> common_json_value(const std::vector<T> & vals);
+    // a set becomes an array, in the set's own order
+    template <typename T> common_json_value(const std::set<T> & vals);
+    // a map becomes an object, keyed in the map's own order
+    template <typename T> common_json_value(const std::map<std::string, T> & vals);
+    template <typename T> common_json_value(const std::unordered_map<std::string, T> & vals);
+
+    // nested object, e.g. {"fn", {{"name", "x"}}}
+    // note: a nested pair {"a", "b"} does not build, use common_json::array({"a", "b"}) for an array
+    common_json_value(std::initializer_list<common_json_item> items);
+
+    template <typename T, typename std::enable_if<std::is_integral<T>::value && !std::is_same<T, bool>::value, int>::type = 0>
+    common_json_value(T val) : type(std::is_signed<T>::value ? VAL_INT : VAL_UINT) {
+        if (std::is_signed<T>::value) {
+            val_int = (int64_t) val;
+        } else {
+            val_uint = (uint64_t) val;
+        }
+    }
+
+    template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
+    common_json_value(T val) : type(VAL_DOUBLE), val_double((double) val) {}
+};
+
+struct common_json_item {
+    std::string       key;
+    common_json_value val;
+
+    template <typename T>
+    common_json_item(std::string key, T && val) :
+        key(std::move(key)), val(std::forward<T>(val)) {}
+
+    // a braced list cannot deduce T, so it needs its own overload
+    common_json_item(std::string key, std::initializer_list<common_json_item> items) :
+        key(std::move(key)), val(items) {}
+};
+
+// the types common_json_value holds on its own
+// anything else reaches its common_json ctor and recurses forever
+template <typename T> struct common_json_is_value : std::integral_constant<bool,
+    std::is_arithmetic<T>::value ||
+    std::is_same<T, std::nullptr_t>::value ||
+    std::is_same<T, std::string>::value ||
+    std::is_same<T, std::string_view>::value ||
+    std::is_same<T, char *>::value ||
+    std::is_same<T, const char *>::value ||
+    std::is_same<T, common_json>::value> {};
+
+template <typename T, typename A>
+struct common_json_is_value<std::vector<T, A>> : std::true_type {};
+
+template <typename T, typename C, typename A>
+struct common_json_is_value<std::set<T, C, A>> : std::true_type {};
+
+template <typename V, typename C, typename A>
+struct common_json_is_value<std::map<std::string, V, C, A>> : std::true_type {};
+
+template <typename V, typename H, typename E, typename A>
+struct common_json_is_value<std::unordered_map<std::string, V, H, E, A>> : std::true_type {};
+
+class common_json {
+  public:
+    common_json();
+    common_json(const common_json & other);
+    common_json(common_json && other) noexcept;
+    common_json(std::initializer_list<common_json_item> items);
+    common_json(const common_json_value & val);
+
+    // direct, a value would need two conversions in a row
+    common_json(std::nullptr_t);
+
+    // one step, so that "abc" or a vector can go straight into a common_json
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value &&
+                                                  !std::is_same<typename std::decay<T>::type, common_json_value>::value, int>::type = 0>
+    common_json(T && val) : common_json(common_json_value(std::forward<T>(val))) {
+        static_assert(common_json_is_value<typename std::decay<T>::type>::value,
+                      "no common_json_value ctor holds this type, add one instead of letting it recurse");
+    }
+
+    // by value, same as the backing library
+    // the right side is copied before the left side can invalidate it, e.g. msg["a"] = msg.at("b")
+    common_json & operator=(common_json other) noexcept;
+
+    ~common_json();
+
+    // throws common_json_error if the text is not valid JSON
+    static common_json parse(const std::string & text);
+
+    // gives a discarded value instead of throwing, check it with is_discarded()
+    static common_json parse_no_throw(const std::string & text);
+
+    bool is_discarded() const;
+
+    static common_json array();
+    static common_json array(std::initializer_list<common_json_value> vals);
+    static common_json object();
+    static common_json object(std::initializer_list<common_json_item> items);
+
+    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
+    static common_json make(const common_json_value & val);
+
+    bool is_null()    const;
+    bool is_object()  const;
+    bool is_array()   const;
+    bool is_string()  const;
+    bool is_boolean() const;
+    bool is_number()  const;
+    bool is_number_integer() const;
+    bool is_number_float()   const;
+
+    bool   empty() const;
+    size_t size()  const;
+
+    bool contains(const std::string & key) const;
+
+    bool operator==(const common_json_value & val) const;
+    bool operator!=(const common_json_value & val) const;
+
+    // at() throws common_json_error if the key is missing, operator[] adds a null value instead
+    // note: a const operator[] cannot add, it throws like at()
+    common_json       & at(const std::string & key);
+    const common_json & at(const std::string & key) const;
+    common_json       & at(size_t idx);
+    const common_json & at(size_t idx) const;
+
+    common_json       & operator[](const std::string & key);
+    const common_json & operator[](const std::string & key) const;
+    common_json       & operator[](const char * key)       { return (*this)[std::string(key)]; }
+    const common_json & operator[](const char * key) const { return (*this)[std::string(key)]; }
+    common_json       & operator[](int idx)       { return (*this)[to_idx(idx)]; }
+    const common_json & operator[](int idx) const { return (*this)[to_idx(idx)]; }
+    common_json       & operator[](size_t idx);
+    const common_json & operator[](size_t idx) const;
+
+    common_json       & front();
+    const common_json & front() const;
+    common_json       & back();
+    const common_json & back()  const;
+
+    void clear();
+
+    void erase(const std::string & key);
+    void erase(size_t idx);
+
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> T get() const;
+
+    // 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;
+
+    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;
+
+    // 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 {
+        return contains(key) ? at(key) : def;
+    }
+
+    void assign(const common_json_value & val);
+    void set(const common_json_item & item);
+    void push_back(const common_json_value & val);
+
+    // appends one object, e.g. push_back({{"a", 1}})
+    void push_back(std::initializer_list<common_json_item> items);
+
+    // 1 if the key is there, 0 if not
+    size_t count(const std::string & key) const;
+
+    // appends every value of another array; inserting an array into itself throws
+    void insert(const common_json & vals);
+
+    // a common_json goes through the copy assignment above, everything else becomes a value
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
+    common_json & operator=(T && val) {
+        assign(common_json_value(std::forward<T>(val)));
+        return *this;
+    }
+
+    std::string dump(int indent = -1) const;
+
+    // same as dump(), but bad UTF-8 gets replaced instead of throwing
+    std::string dump_safe(int indent = -1) const;
+
+    // walks an array by index, or an object in insertion order
+    // a plain value gives itself once, same as the backing library
+    class iterator {
+      public:
+        using iterator_category = std::forward_iterator_tag;
+        using value_type        = common_json;
+        using difference_type   = std::ptrdiff_t;
+        using pointer           = common_json *;
+        using reference         = common_json &;
+
+        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+        common_json & operator*() const;
+        common_json & value()     const { return **this; }
+        std::string   key()       const;
+
+        iterator & operator++() {
+            idx++;
+            return *this;
+        }
+
+        bool operator!=(const iterator & other) const { return idx != other.idx; }
+        bool operator==(const iterator & other) const { return idx == other.idx; }
+
+      private:
+        common_json * node;
+        size_t        idx;
+    };
+
+    iterator begin() const;
+    iterator end()   const;
+
+    // allows: for (const auto & [key, val] : obj.items())
+    class items_view {
+      public:
+        // the members are public, so an entry also works with structured bindings
+        struct entry {
+            std::string   k;
+            common_json & v;
+
+            const std::string & key()   const { return k; }
+            common_json &       value() const { return v; }
+        };
+
+        items_view(common_json * node, size_t n) : node(node), n(n) {}
+
+        class iterator {
+          public:
+            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+            entry operator*() const;
+
+            iterator & operator++() {
+                idx++;
+                return *this;
+            }
+
+            bool operator!=(const iterator & other) const { return idx != other.idx; }
+
+          private:
+            common_json * node;
+            size_t        idx;
+        };
+
+        iterator begin() const { return iterator(node, 0); }
+        iterator end()   const { return iterator(node, n); }
+
+      private:
+        common_json * node;
+        size_t        n;
+    };
+
+    items_view items() const;
+
+  private:
+    // a negative index must not turn into a huge size_t
+    static size_t to_idx(int idx) {
+        if (idx < 0) {
+            throw common_json_error("negative array index");
+        }
+        return (size_t) idx;
+    }
+
+    // the backing value is built here, json.cpp checks that it fits
+    // it cannot be a pointer: a value inside a tree would then not be a common_json
+    // at() could then only give back a copy instead of a real reference
+    alignas(8) unsigned char storage[32];
+};
+
+using common_json_entry = common_json::items_view::entry;

common/peg-parser.cpp

@@ -10,7 +10,6 @@
 #include <initializer_list>
 #include <map>
 #include <memory>
-#include <nlohmann/json.hpp>
 #include <regex>
 #include <set>
 #include <stdexcept>
@@ -1120,8 +1119,8 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
-common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<nlohmann::ordered_json>(schema), raw}));
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1805,8 +1804,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     }
 }
 
-static nlohmann::json serialize_parser_variant(const common_peg_parser_variant & variant) {
-    using json = nlohmann::json;
+static common_json serialize_parser_variant(const common_peg_parser_variant & variant) {
+    using json = common_json;
 
     return std::visit([](const auto & p) -> json {
         using T = std::decay_t<decltype(p)>;
@@ -1860,7 +1859,7 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : nullptr},
+                {"schema", p.schema ? *p.schema : json(nullptr)},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1888,19 +1887,19 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
     }, variant);
 }
 
-nlohmann::json common_peg_arena::to_json() const {
-    auto parsers = nlohmann::json::array();
+common_json common_peg_arena::to_json() const {
+    auto parsers = common_json::array();
     for (const auto & parser : parsers_) {
         parsers.push_back(serialize_parser_variant(parser));
     }
-    return nlohmann::json{
+    return common_json{
         {"parsers", parsers},
         {"rules", rules_},
         {"root", root_}
     };
 }
 
-static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json & j) {
+static common_peg_parser_variant deserialize_parser_variant(const common_json & j) {
     if (!j.contains("type") || !j["type"].is_string()) {
         throw std::runtime_error("Parser variant JSON missing or invalid 'type' field");
     }
@@ -1969,9 +1968,9 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         }
         common_peg_chars_parser parser;
         parser.pattern = j["pattern"];
-        parser.negated = j["negated"];
-        parser.min_count = j["min_count"];
-        parser.max_count = j["max_count"];
+        parser.negated = j["negated"].get<bool>();
+        parser.min_count = j["min_count"].get<int>();
+        parser.max_count = j["max_count"].get<int>();
         for (const auto & range_json : j["ranges"]) {
             if (!range_json.contains("start") || !range_json.contains("end")) {
                 throw std::runtime_error("char_range missing 'start' or 'end' field");
@@ -2007,7 +2006,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
         if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<nlohmann::ordered_json>(j["schema"]);
+            parser.schema = std::make_shared<common_json>(j["schema"]);
         }
         parser.raw = j["raw"].get<bool>();
         return parser;
@@ -2069,7 +2068,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
     throw std::runtime_error("Unknown parser type: " + type);
 }
 
-common_peg_arena common_peg_arena::from_json(const nlohmann::json & j) {
+common_peg_arena common_peg_arena::from_json(const common_json & j) {
     if (!j.contains("parsers") || !j["parsers"].is_array()) {
         throw std::runtime_error("JSON missing or invalid 'parsers' array");
     }
@@ -2109,7 +2108,7 @@ std::string common_peg_arena::save() const {
 }
 
 void common_peg_arena::load(const std::string & data) {
-    *this = from_json(nlohmann::json::parse(data));
+    *this = from_json(common_json::parse(data));
 }
 
 common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {

common/peg-parser.h

@@ -1,6 +1,6 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <memory>
 #include <set>
@@ -245,7 +245,7 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<nlohmann::ordered_json> schema;
+    std::shared_ptr<common_json> schema;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -332,8 +332,8 @@ class common_peg_arena {
 
     std::string dump(common_peg_parser_id id) const;
 
-    nlohmann::json to_json() const;
-    static common_peg_arena from_json(const nlohmann::json & j);
+    common_json to_json() const;
+    static common_peg_arena from_json(const common_json & j);
 
     std::string save() const;
     void load(const std::string & data);
@@ -490,7 +490,7 @@ class common_peg_parser_builder {
 
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
-    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.
     // If trigger=true, marks this rule as an entry point for lazy grammar generation.

common/speculative.cpp

@@ -2322,6 +2322,9 @@ common_params common_base_params_to_speculative(const common_params & params) {
     const auto & params_spec = params.speculative.draft;
     common_params result = params;
 
+    result.embedding    = false;
+    result.pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED;
+
     if (has_draft) {
         result.devices               = params_spec.devices;
         result.model                 = params_spec.mparams;
@@ -2385,6 +2388,9 @@ common_speculative_init_result::common_speculative_init_result(
         cparams.ctx_type = LLAMA_CONTEXT_TYPE_MTP;
     }
 
+    // the draft context holds as many tokens per sequence as the target context
+    cparams.n_ctx = llama_n_ctx(ctx_tgt);
+
     // note: for small models maybe we can set this to the maximum possible draft from all speculative types
     //       the extra memory for small models is likely negligible?
     cparams.n_rs_seq  = 0;
@@ -2649,6 +2655,10 @@ void common_speculative_draft(common_speculative * spec) {
         for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) dparams.size(); ++seq_id) {
             auto & dp = dparams[seq_id];
 
+            if (!dp.drafting) {
+                continue;
+            }
+
             auto & result = *dp.result;
 
             // a new draft has been sampled

conversion/__init__.py

@@ -57,12 +57,17 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Qwen3DSparkModel": "qwen",
     "DSparkDraftModel": "qwen",
     "DSparkSpeculator": "qwen",
+    "Lfm2DSparkDraftModel": "qwen",
+    "LingDSparkModel": "qwen",
     "DeepseekV4ForCausalLM": "deepseek",
     "DeepseekV4DSparkModel": "deepseek",
     "DistilBertForMaskedLM": "bert",
     "DistilBertForSequenceClassification": "bert",
     "DistilBertModel": "bert",
     "Dots1ForCausalLM": "dots1",
+    "Dots3NoteForCausalLM": "dots3",
+    "Dots3NoteForConditionalGeneration": "dots3",
+    "Dots3NoteTextForCausalLM": "dots3",
     "DotsOCRForCausalLM": "qwen",
     "DreamModel": "dream",
     "Ernie4_5ForCausalLM": "ernie",
@@ -278,6 +283,8 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "CogVLMForCausalLM": "cogvlm",
     "DeepseekOCR2ForCausalLM": "deepseek",
     "DeepseekOCRForCausalLM": "deepseek",
+    "Dots3NoteForCausalLM": "dots3",
+    "Dots3NoteForConditionalGeneration": "dots3",
     "DotsOCRForCausalLM": "dotsocr",
     "Exaone4_5_ForConditionalGeneration": "exaone",
     "Gemma3ForConditionalGeneration": "gemma",

conversion/dots3.py

@@ -0,0 +1,323 @@
+from __future__ import annotations
+
+import math
+import re
+
+import torch
+
+from typing import TYPE_CHECKING, Any, Callable, Iterable
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import MmprojModel, ModelBase, gguf
+
+from .deepseek import DeepseekV2Model
+
+
+@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration", "Dots3NoteTextForCausalLM")
+class Dots3NoteModel(DeepseekV2Model):
+    model_arch = gguf.MODEL_ARCH.DOTS3NOTE
+    skip_mtp = False
+    supports_mtp_export = True
+
+    # trunk layer count, stashed before indexing for filter_tensors (mirrors DeepseekV32Model)
+    _n_main_layers: int | None = None
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        hparams = self.hparams
+
+        # config file doesn't specify MTP block, detect it from model weight
+        self.n_nextn = 1 if "model.mtp.embed_tokens.weight" in self.model_tensors else 0
+        if self.n_nextn:
+            self.block_count += self.n_nextn
+            self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+        self.layer_types = hparams["layer_types"]
+        if len(self.layer_types) < hparams["num_hidden_layers"]:
+            raise ValueError("layer_types is shorter than num_hidden_layers")
+
+        if hparams.get("use_dsa", True) is not True:
+            raise ValueError("dots3-note conversion requires use_dsa=true")
+        if hparams.get("normalization", "RMSNorm") != "RMSNorm" or hparams.get("final_norm", "RMSNorm") != "RMSNorm":
+            raise ValueError("dots3-note conversion only supports RMSNorm")
+        if hparams.get("k_rope_only_layernorm", True) is not True:
+            raise ValueError("dots3-note conversion requires k_rope_only_layernorm=true")
+        if hparams.get("topk_method", "noaux_tc") != "noaux_tc" or hparams.get("scoring_func") != "sigmoid":
+            raise ValueError("dots3-note conversion only supports noaux_tc/sigmoid expert gating")
+        if hparams.get("n_group", 1) != 1 or hparams.get("topk_group", 1) != 1:
+            raise ValueError("dots3-note conversion does not support grouped expert routing")
+        if hparams.get("use_dynamic_rsf", False) or hparams.get("moe_gating_fp32", False):
+            raise ValueError("dots3-note conversion does not support use_dynamic_rsf/moe_gating_fp32")
+        for key in ("attention_gate_type", "swa_attention_gate_type"):
+            if hparams.get(key, "headwise") != "headwise":
+                raise ValueError(f"dots3-note conversion only supports headwise attention gate, got {key}={hparams.get(key)!r}")
+        if hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"] != hparams.get("swa_head_dim", 256):
+            raise ValueError("swa_head_dim must equal swa_qk_nope_head_dim + swa_qk_rope_head_dim")
+        if hparams["swa_qk_rope_head_dim"] != hparams["qk_rope_head_dim"]:
+            # both layer kinds share a single rope_dimension_count
+            raise ValueError("swa_qk_rope_head_dim must match qk_rope_head_dim")
+
+        self.apply_lora_rescale = hparams.get("apply_mla_qkv_lora_rescale", False)
+
+    def _is_swa_layer(self, bid: int) -> bool:
+        if bid >= self.hparams["num_hidden_layers"]:
+            # note: the NextN/MTP block uses the sliding-attention MLA
+            return True
+        return self.layer_types[bid] == "sliding_attention"
+
+    def set_vocab(self):
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|endofassistant|>"])  # ty: ignore[unresolved-attribute]
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    @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
+        if name.startswith(("vision_encoder.", "audio_encoder.")):
+            return None
+
+        assert cls._n_main_layers is not None
+        is_mtp = name.startswith("model.mtp.") or \
+            ((m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers)
+
+        # --no-mtp: drop the NextN/MTP block; --mtp: keep only that block plus the shared embeddings/norm/lm_head
+        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_gguf_parameters(self):
+        hparams = self.hparams
+
+        # head_count is a per-layer array because the two layer kinds have different head counts
+        n_layer = hparams["num_hidden_layers"]
+        hparams["num_attention_heads"] = [
+            hparams["swa_num_attention_heads"] if self._is_swa_layer(il) else hparams["num_attention_heads"]
+            for il in range(self.block_count)
+        ]
+
+        # prevent the base class from emitting key/value_length from the unused head_dim
+        hparams.pop("head_dim", None)
+
+        super().set_gguf_parameters()
+
+        # MLA geometry of the sliding-window layers (rope.freq_base_swa is emitted by the base class)
+        swa_kv_lora_rank = hparams["swa_kv_lora_rank"]
+        self.gguf_writer.add_sliding_window(hparams["sliding_window_size"])
+        self.gguf_writer.add_sliding_window_pattern([self._is_swa_layer(il) for il in range(n_layer)])
+        self.gguf_writer.add_kv_lora_rank_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_swa(swa_kv_lora_rank + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_swa(swa_kv_lora_rank)
+        self.gguf_writer.add_key_length_mla_swa(hparams["swa_qk_nope_head_dim"] + hparams["swa_qk_rope_head_dim"])
+        self.gguf_writer.add_value_length_mla_swa(hparams["swa_v_head_dim"])
+        if hparams["swa_q_lora_rank"] != hparams["q_lora_rank"]:
+            raise ValueError("dots3-note conversion assumes a shared q_lora_rank for both layer kinds")
+
+        if self.n_nextn:
+            self.gguf_writer.add_nextn_predict_layers(self.n_nextn)
+
+        # DSA indexer (full-attention layers only)
+        self.gguf_writer.add_indexer_head_count(hparams["index_n_heads"])
+        self.gguf_writer.add_indexer_key_length(hparams["index_head_dim"])
+        self.gguf_writer.add_indexer_top_k(hparams["index_topk"])
+        self.gguf_writer.add_indexer_types([not self._is_swa_layer(il) for il in range(n_layer)])
+
+    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"
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # move the MTP token embedding into the NextN block so the standard nextn mapping picks it up
+        if name == "model.mtp.embed_tokens.weight":
+            name = f"model.layers.{self.hparams['num_hidden_layers']}.embed_tokens.weight"
+            bid = self.hparams["num_hidden_layers"]
+
+        # fold the activation rescale sqrt(n_embd/lora_rank) into the preceding RMSNorm weight
+        # this also covers the indexer wq_b, which reads the same rescaled q_lora activation
+        if self.apply_lora_rescale and bid is not None:
+            if name.endswith("q_a_layernorm.weight"):
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / self.hparams["q_lora_rank"])
+            elif name.endswith("kv_a_layernorm.weight"):
+                rank = self.hparams["swa_kv_lora_rank"] if self._is_swa_layer(bid) else self.hparams["kv_lora_rank"]
+                data_torch = data_torch * math.sqrt(self.hparams["hidden_size"] / rank)
+
+        # MLA absorption: split kv_b_proj into k_b (transposed) and v_b, per-layer-kind geometry
+        if name.endswith("kv_b_proj.weight"):
+            assert bid is not None
+            if self._is_swa_layer(bid):
+                n_head = self.hparams["swa_num_attention_heads"]
+                qk_nope_head_dim = self.hparams["swa_qk_nope_head_dim"]
+                v_head_dim = self.hparams["swa_v_head_dim"]
+            else:
+                n_head = self.hparams["num_attention_heads"]
+                qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+                v_head_dim = self.hparams["v_head_dim"]
+            if isinstance(n_head, list):  # set_gguf_parameters turns this into a per-layer array
+                n_head = n_head[bid]
+
+            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
+
+            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
+            k_b, v_b = kv_b.split([qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            yield from ModelBase.modify_tensors(self, k_b, name.replace("kv_b_proj", "k_b_proj"), bid)
+            yield from ModelBase.modify_tensors(self, v_b, name.replace("kv_b_proj", "v_b_proj"), bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Dots3NoteForCausalLM", "Dots3NoteForConditionalGeneration")
+class Dots3NoteMmprojModel(MmprojModel):
+    has_vision_encoder = True
+    has_audio_encoder = True
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        assert self.hparams_audio is not None
+
+        # preprocessor_config.json nests the image params under vision_config
+        self.preprocessor_config = {**self.preprocessor_config, **self.preprocessor_config.get("vision_config", {})}
+
+        vis = self.hparams_vision
+        # in this config, hidden_size is the adapter output width; embed_dim is the tower width
+        vis["hidden_size"] = vis["embed_dim"]
+        vis["image_size"] = 0  # dynamic resolution
+        self.pyramid = [max(0, n) for n in vis["pyramid_num_routed"]]
+
+        if vis.get("adapter_type") != "patch_merger" or not vis.get("pre_pixel_shuffle"):
+            raise ValueError("dots3-note vision conversion requires adapter_type=patch_merger and pre_pixel_shuffle")
+        if vis.get("router_scoring_func", "sigmoid") != "sigmoid" or vis.get("router_scale", 1.0) != 1.0:
+            raise ValueError("dots3-note vision conversion only supports sigmoid routing with router_scale=1.0")
+        if vis.get("temporal_patch_size", 1) != 1 or vis.get("use_bias") or not vis.get("use_qk_norm"):
+            raise ValueError("unsupported dots3-note vision config variant")
+
+        aud = self.hparams_audio
+        if not aud.get("use_conv2d_stem") or not aud.get("use_rope") or not aud.get("use_rms_norm") or aud.get("use_causal"):
+            raise ValueError("unsupported dots3-note audio config variant")
+        if aud["whisper_config"].get("activation_function") != "swiglu":
+            raise ValueError("dots3-note audio conversion requires the swiglu activation")
+        if aud.get("merge_factor", 1) != 1 or aud.get("chunk_seconds") != 60:
+            raise ValueError("unsupported dots3-note audio chunking config")
+        # the graph hard-codes these rope parameters
+        rope = aud.get("rope_parameters", {})
+        if rope.get("partial_rotary_factor") != 0.5 or rope.get("rope_theta") != 10000.0:
+            raise ValueError("unsupported dots3-note audio rope config")
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        cfg = self.global_config.get("audio_config")
+        if cfg is not None:
+            # aliases so MmprojModel.find_aparam() / n_block_keys can resolve them
+            whisper = cfg["whisper_config"]
+            cfg["hidden_size"] = whisper["d_model"]
+            cfg["intermediate_size"] = whisper["encoder_ffn_dim"]
+            cfg["num_attention_heads"] = whisper["encoder_attention_heads"]
+            cfg["num_hidden_layers"] = whisper["encoder_layers"]
+        return cfg
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        assert self.hparams_audio is not None
+
+        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.DOTS3NOTE_V)
+        self.gguf_writer.add_vision_use_silu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision["rms_norm_eps"])
+        self.gguf_writer.add_vision_spatial_merge_size(self.hparams_vision["spatial_merge_size"])
+        self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
+        self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
+        # pyramid MoE: per-block routed expert count, 0 = dense block
+        self.gguf_writer.add_vision_expert_count_per_layer(self.pyramid)
+        self.gguf_writer.add_vision_expert_used_count(int(self.hparams_vision["capacity_factor"]))
+
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.DOTS3NOTE_A)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["whisper_config"]["num_mel_bins"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-6)  # Dots3NoteAudioRMSNorm default
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        if not name.startswith(("vision_encoder.", "audio_encoder.")):
+            return None
+        return super().filter_tensors(item)
+
+    _vis_experts: dict[int, dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # router params have no .weight suffix in the checkpoint, but gguf tools expect one
+        if name.endswith((".gate_weight", ".router_bias")):
+            name += ".weight"
+
+        # audio fc1 fuses gate and up for swiglu; split it
+        if ".speech_encoder.layers." in name and ".fc1." in name:
+            gate, up = data_torch.chunk(2, dim=0)
+            yield from super().modify_tensors(gate, name.replace(".fc1.", ".fc1_gate."), bid)
+            yield from super().modify_tensors(up, name.replace(".fc1.", ".fc1_up."), bid)
+            return
+
+        # vision MoE: stack per-expert weights into a single 3D tensor per block
+        if ".mlp.experts." in name:
+            assert bid is not None
+            n_expert = self.pyramid[bid]
+            if self._vis_experts is None:
+                self._vis_experts = {}
+            buf = self._vis_experts.setdefault(bid, {})
+            buf[name] = data_torch
+
+            if len(buf) >= n_expert * 3:
+                for w_name in ("fc1", "fc2", "fc3"):
+                    datas: list[Tensor] = []
+                    for xid in range(n_expert):
+                        ename = f"vision_encoder.blocks.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(buf.pop(ename))
+                    merged = torch.stack(datas, dim=0)
+                    yield from super().modify_tensors(merged, f"vision_encoder.blocks.{bid}.mlp.experts.{w_name}.weight", bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._vis_experts is not None:
+            leftover = [k for d in self._vis_experts.values() for k in d.keys()]
+            if leftover:
+                raise ValueError(f"unprocessed vision experts: {leftover}")
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # FP32 routing is load-bearing for the vision MoE (near-tied expert scores)
+        if ".ffn_gate_inp." in new_name or ".exp_probs_b." in new_name:
+            return gguf.GGMLQuantizationType.F32
+        if ".conv2d" in new_name or "a.conv_out" in new_name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)

conversion/nemotron.py

@@ -207,7 +207,9 @@ class NemotronHModel(GraniteHybridModel):
         # calling the parent __init__. This is because the parent constructor
         # uses self.model_arch to build the tensor name map, and all MoE-specific
         # mappings would be missed if it were called with the default non-MoE arch.
-        hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(args[0], self.is_mistral_format)
         has_moe_params = (
             "num_experts_per_tok" in hparams
             or (isinstance(hparams.get("llm_config"), dict) and "num_experts_per_tok" in hparams["llm_config"])
@@ -215,8 +217,11 @@ class NemotronHModel(GraniteHybridModel):
         if has_moe_params:
             self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
             self.is_moe = True
+        layers_block_type = hparams.get("layers_block_type")
+        if layers_block_type is not None:
+            hparams["num_hidden_layers"] = len(layers_block_type)
 
-        super().__init__(*args, **kwargs)
+        super().__init__(*args, hparams=hparams, **kwargs)
 
         # Save the top-level head_dim for later
         self.head_dim = self.hparams.get("head_dim", self.hparams.get("attention_head_dim"))

conversion/qwen.py

@@ -709,7 +709,13 @@ class DFlashModel(Qwen3Model):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator")
+@ModelBase.register(
+    "Qwen3DSparkModel",
+    "DSparkDraftModel",
+    "DSparkSpeculator",
+    "Lfm2DSparkDraftModel",
+    "LingDSparkModel",
+)
 @ModelBase.example("satgeze/Qwen3.6-27B-DSpark")
 class DSparkModel(DFlashModel):
     # DSpark = DFlash + a semi-autoregressive Markov head.
@@ -759,6 +765,13 @@ class DSparkModel(DFlashModel):
             return None
         return super().filter_tensors(item)
 
+    _ROPE_PERMUTE_SUFFIXES = (
+        "self_attn.q_proj.weight",
+        "self_attn.k_proj.weight",
+        "self_attn.q_norm.weight",
+        "self_attn.k_norm.weight",
+    )
+
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name == "model.d2t":
             self._d2t = data_torch
@@ -767,6 +780,12 @@ class DSparkModel(DFlashModel):
         if self._n_vocab_draft == self.hparams["vocab_size"] and name.endswith(("embed_tokens.weight", "lm_head.weight")):
             return
 
+        # interleaved-rope checkpoints (rope_is_neox_style = false) -> NeoX layout: per head, even dims first then odd
+        if not self.hparams.get("rope_is_neox_style", True) and name.endswith(self._ROPE_PERMUTE_SUFFIXES):
+            head_dim = self.hparams["head_dim"]
+            shape = data_torch.shape
+            data_torch = data_torch.reshape(-1, head_dim // 2, 2, *shape[1:]).transpose(1, 2).reshape(shape)
+
         yield from super().modify_tensors(data_torch, name, bid)
 
     def prepare_tensors(self):

docs/backend/ET.md

@@ -116,7 +116,7 @@ in inline assembler.
 Most kernels are very naive with lots of low hanging fruits left:
 
 > [!IMPORTANT]
-> Several assembly instructions emmited by the compiler are not implemented
+> Several assembly instructions emitted by the compiler are not implemented
 > in hardware and software emulation in firmware is not ready yet.
 > Eventually firmware will transparently trap unimplemented instructions
 > and will emulate them inside exception handler. Until then, kernel
@@ -138,12 +138,12 @@ Most kernels are very naive with lots of low hanging fruits left:
 > kernel build process. Feel free to take ideas/code from there or try linking
 > it in.
 
-Before commiting any changes to operations and/or kernels, don't forget
+Before committing any changes to operations and/or kernels, don't forget
 to update supported ops reports (instructions at `docs/ops.md`).
 
 When logging is enabled (e.g. by setting `--log-file` cli param),
 each compute kernel run outputs a line with
-pipe-delimited key-value pairs containing kernel level performance infomation.
+pipe-delimited key-value pairs containing kernel level performance information.
 Line is prefixed with `ET_PERF`:
 
 ```
@@ -160,7 +160,7 @@ to `GGML_ET_PROFILE/et_runtime_trace.json` and `GGML_ET_PROFILE/kernel_map` on e
 
 ### Uberkernel
 
-The in-knernel implementaiton of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
+The in-kernel implementation of device dispatch/kernel fusion. The ET SDK has a non-trivial op-to-op gap. `Uberkernel` (name taken from the original Esperanto AI's compiler)
 dispatches multiple already existing kernel implementations with device side synchronization. Due to the processor's design, there is no natural memory visibility
 horizon between sub-kernel invocations. This makes uberkernel much more difficult to develop and debug. Currently Uberkerel is hidden begind the
 `GGML_ET_UBERKERNEL` environment variable and is disabled by default. Setting it to 1 enables it and provides significant performance improvements but is only

docs/build.md

@@ -70,17 +70,23 @@ cmake --build build --config Release
     - Tab Workload: Desktop-development with C++
     - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
     - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
-    - For Windows on ARM (arm64, WoA) build with:
-    ```bash
-    cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
-    cmake --build build-arm64-windows-llvm-release
-    ```
-    For building with ninja generator and clang compiler as default:
-      -set path:set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
+    - For Windows on ARM (arm64, WoA), build with:
       ```bash
-      cmake --preset x64-windows-llvm-release
-      cmake --build build-x64-windows-llvm-release
+      cmake --preset arm64-windows-llvm-release -D GGML_OPENMP_FETCH=ON
+      cmake --build build-arm64-windows-llvm-release
       ```
+      - Use `ARM64 Native Tools Command Prompt for VS 2022` if you are building on an ARM64 machine.
+      - `GGML_OPENMP_FETCH` downloads the official LLVM OpenMP runtime and requires Clang, 7-Zip and network access during configuration. CMake selects the runtime from the target architecture, so this also works when cross-compiling for WoA from x64. The extracted header, import library, DLL and OpenMP license are placed under `build/_deps`. The build copies `libomp.dll` and `LICENSE-LLVM-OpenMP` to the runtime output directory and installs them together. Omit the option to use CMake's normal OpenMP detection, or pass `-D GGML_OPENMP=OFF` to disable OpenMP.
+    - For building with ninja generator and clang compiler as default:
+      - Set path:
+        ```
+        set LIB=C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\um\x64;C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.41.34120\lib\x64\uwp;C:\Program Files (x86)\Windows Kits\10\Lib\10.0.22621.0\ucrt\x64
+        ```
+      - Run:
+        ```bash
+        cmake --preset x64-windows-llvm-release
+        cmake --build build-x64-windows-llvm-release
+        ```
 - If you want HTTPS/TLS features, you may install OpenSSL development libraries. If not installed, the project will build and run without SSL support.
   - **Debian / Ubuntu:** `sudo apt-get install libssl-dev`
   - **Fedora / RHEL / Rocky / Alma:** `sudo dnf install openssl-devel`

docs/development/HOWTO-add-model.md

@@ -166,6 +166,19 @@ Examples:
 - Some models require scaling the input position. For example, `[0, 1, 2, ...]` becomes `[0, 0.5, 1, ...]`. In this case, you can provide the scaling via `freq_scale = 0.5f`.
 - Some models use learned RoPE frequencies instead of relying on `powf(freq_base, -2.0 * i / n_dims)`. In this case, you can provide the learned frequencies via the `rope_freqs` tensor (corresponding to the `c` argument in `ggml_rope_ext`), then set `freq_base = 1.0f`. An important note is that `rope_freqs` in GGML is the **inverse** (`theta = pos[i] / rope_freqs`), so you may need to invert `rope_freqs` during conversion.
 
+### Rotating only a part of the head
+
+Many models rotate only a part of each head and leave the rest untouched (often called the "nope" part). Do not build this with views plus `ggml_concat`, it's not efficient. Both layouts can be done with a single RoPE op:
+
+- `[rope|nope]`, rotated dims first: pass `n_dims` smaller than the head size to `ggml_rope_ext`. Dims from `n_dims` to the end are copied as-is.
+- `[nope|rope]`, rotated dims last: call `ggml_rope_set_offset(cur, n_offs)` on the result of the RoPE, where `n_offs` is the size of the leading untouched part. Dims outside `[n_offs, n_offs + n_dims)` are copied as-is.
+
+`n_offs` must be even, `n_offs + n_dims` must fit in the row, and vision RoPE is not supported. Note that the frequencies are computed relative to the rotated window.
+
+Example: DeepSeek-V4 uses `[nope|rope]` for its query, key and compressed KV tensors, so `src/models/deepseek4.cpp` ropes the whole tensor and then calls `ggml_rope_set_offset(cur, n_embd_head_nope)`.
+
+Exception: some models apply an extra op to the `nope` part, for example `deepseek32.cpp`, and may not use this optimization. While RoPE can be applied selectively to a part of the head, the extra op may not, so these models still need views plus `ggml_concat`.
+
 ## GGUF specification
 
 https://github.com/ggml-org/ggml/blob/master/docs/gguf.md

ggml/CMakeLists.txt

@@ -4,8 +4,8 @@ project("ggml" C CXX ASM)
 
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 2)
+set(GGML_VERSION_MINOR 21)
+set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")
 
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -243,6 +243,7 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                             "ggml: metal minimum macOS version")
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_OPENMP_FETCH                    "ggml: fetch LLVM OpenMP"                         OFF)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)

ggml/src/CMakeLists.txt

@@ -222,9 +222,123 @@ if (GGML_SCHED_NO_REALLOC)
     target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC)
 endif()
 
-if (GGML_OPENMP)
+if (GGML_OPENMP_FETCH)
+    if (NOT GGML_OPENMP)
+        message(FATAL_ERROR "GGML_OPENMP_FETCH requires GGML_OPENMP")
+    elseif (NOT WIN32 OR NOT (CMAKE_C_COMPILER_ID MATCHES "Clang"))
+        message(FATAL_ERROR "GGML_OPENMP_FETCH currently requires Clang on Windows")
+    endif()
+
+    set(GGML_OPENMP_LLVM_VERSION "20.1.8")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_LLVM_VERSION_MAJOR "${GGML_OPENMP_LLVM_VERSION}")
+    string(REGEX MATCH "^[0-9]+" GGML_OPENMP_COMPILER_VERSION_MAJOR "${CMAKE_C_COMPILER_VERSION}")
+    if (NOT GGML_OPENMP_COMPILER_VERSION_MAJOR STREQUAL GGML_OPENMP_LLVM_VERSION_MAJOR)
+        message(FATAL_ERROR "LLVM OpenMP ${GGML_OPENMP_LLVM_VERSION} requires Clang ${GGML_OPENMP_LLVM_VERSION_MAJOR}.x")
+    endif()
+
+    string(TOLOWER "${CMAKE_SYSTEM_PROCESSOR}" GGML_OPENMP_SYSTEM_PROCESSOR)
+    if (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(amd64|x86_64)$")
+        set(GGML_OPENMP_ARCH "x64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "win64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "3197846a2b19063687dd56e93e34cd941e3548d907f23a6131571321bdf9fe7b")
+    elseif (GGML_OPENMP_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm64)$")
+        set(GGML_OPENMP_ARCH "arm64")
+        set(GGML_OPENMP_INSTALLER_SUFFIX "woa64")
+        set(GGML_OPENMP_INSTALLER_SHA256 "7c4ac97eb2ae6b960ca5f9caf3ff6124c8d2a18cc07a7840a4d2ea15537bad8e")
+    else()
+        message(FATAL_ERROR "GGML_OPENMP_FETCH does not support ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    set(GGML_OPENMP_CACHE_DIR "${CMAKE_BINARY_DIR}/_deps")
+    set(GGML_OPENMP_ROOT "${GGML_OPENMP_CACHE_DIR}/llvm-openmp-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_ARCH}")
+    set(GGML_OPENMP_LIBRARY "${GGML_OPENMP_ROOT}/lib/libomp.lib")
+    set(GGML_OPENMP_RUNTIME "${GGML_OPENMP_ROOT}/bin/libomp.dll")
+    set(GGML_OPENMP_HEADER "${GGML_OPENMP_ROOT}/include/omp.h")
+    set(GGML_OPENMP_LICENSE "${GGML_OPENMP_ROOT}/LICENSE.TXT")
+    set(GGML_OPENMP_LICENSE_SHA256 "fdad1758a9e1f9d5a81e18879b3406772115edc92c24bfa36b70c654f325e8e4")
+
+    if (NOT EXISTS "${GGML_OPENMP_LIBRARY}" OR NOT EXISTS "${GGML_OPENMP_RUNTIME}" OR NOT EXISTS "${GGML_OPENMP_HEADER}")
+        find_program(GGML_OPENMP_7Z NAMES 7z 7zz 7za)
+        if (NOT GGML_OPENMP_7Z)
+            message(FATAL_ERROR "GGML_OPENMP_FETCH requires 7-Zip to extract the LLVM installer")
+        endif()
+
+        set(GGML_OPENMP_INSTALLER "${GGML_OPENMP_ROOT}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+        set(GGML_OPENMP_EXTRACT_DIR "${GGML_OPENMP_ROOT}/extract")
+        set(GGML_OPENMP_INSTALLER_URL "https://github.com/llvm/llvm-project/releases/download/llvmorg-${GGML_OPENMP_LLVM_VERSION}/LLVM-${GGML_OPENMP_LLVM_VERSION}-${GGML_OPENMP_INSTALLER_SUFFIX}.exe")
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_EXTRACT_DIR}")
+        file(DOWNLOAD "${GGML_OPENMP_INSTALLER_URL}" "${GGML_OPENMP_INSTALLER}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_INSTALLER_SHA256}"
+            SHOW_PROGRESS
+            STATUS GGML_OPENMP_DOWNLOAD_STATUS)
+        list(GET GGML_OPENMP_DOWNLOAD_STATUS 0 GGML_OPENMP_DOWNLOAD_RESULT)
+        if (NOT GGML_OPENMP_DOWNLOAD_RESULT EQUAL 0)
+            list(GET GGML_OPENMP_DOWNLOAD_STATUS 1 GGML_OPENMP_DOWNLOAD_ERROR)
+            message(FATAL_ERROR "Failed to download LLVM OpenMP: ${GGML_OPENMP_DOWNLOAD_ERROR}")
+        endif()
+
+        execute_process(
+            COMMAND "${GGML_OPENMP_7Z}" e -y "-o${GGML_OPENMP_EXTRACT_DIR}" "${GGML_OPENMP_INSTALLER}" -r libomp.lib libomp.dll omp.h
+            RESULT_VARIABLE GGML_OPENMP_EXTRACT_RESULT
+            OUTPUT_QUIET)
+        if (NOT GGML_OPENMP_EXTRACT_RESULT EQUAL 0 OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" OR
+            NOT EXISTS "${GGML_OPENMP_EXTRACT_DIR}/omp.h")
+            message(FATAL_ERROR "Failed to extract libomp from ${GGML_OPENMP_INSTALLER}")
+        endif()
+
+        file(MAKE_DIRECTORY "${GGML_OPENMP_ROOT}/lib" "${GGML_OPENMP_ROOT}/bin" "${GGML_OPENMP_ROOT}/include")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.lib" DESTINATION "${GGML_OPENMP_ROOT}/lib")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/libomp.dll" DESTINATION "${GGML_OPENMP_ROOT}/bin")
+        file(COPY "${GGML_OPENMP_EXTRACT_DIR}/omp.h" DESTINATION "${GGML_OPENMP_ROOT}/include")
+        file(REMOVE_RECURSE "${GGML_OPENMP_INSTALLER}" "${GGML_OPENMP_EXTRACT_DIR}")
+    endif()
+
+    # The NSIS installer embeds LLVM's general license in its UI but does not install it as a file; use OpenMP's license to include its additional notices.
+    if (EXISTS "${GGML_OPENMP_LICENSE}")
+        file(SHA256 "${GGML_OPENMP_LICENSE}" GGML_OPENMP_LICENSE_ACTUAL_SHA256)
+    endif()
+    if (NOT GGML_OPENMP_LICENSE_ACTUAL_SHA256 STREQUAL GGML_OPENMP_LICENSE_SHA256)
+        file(DOWNLOAD "https://raw.githubusercontent.com/llvm/llvm-project/llvmorg-${GGML_OPENMP_LLVM_VERSION}/openmp/LICENSE.TXT" "${GGML_OPENMP_LICENSE}"
+            EXPECTED_HASH "SHA256=${GGML_OPENMP_LICENSE_SHA256}")
+    endif()
+
+    if (COMMAND license_add_file)
+        license_add_file("LLVM OpenMP" "${GGML_OPENMP_LICENSE}")
+    endif()
+
+    add_library(ggml-openmp-c INTERFACE)
+    target_compile_options(ggml-openmp-c INTERFACE "$<$<COMPILE_LANGUAGE:C>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-c SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-c INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    add_library(ggml-openmp-cxx INTERFACE)
+    target_compile_options(ggml-openmp-cxx INTERFACE "$<$<COMPILE_LANGUAGE:CXX>:-fopenmp=libomp>")
+    target_include_directories(ggml-openmp-cxx SYSTEM INTERFACE "${GGML_OPENMP_ROOT}/include")
+    target_link_libraries(ggml-openmp-cxx INTERFACE "${GGML_OPENMP_LIBRARY}")
+
+    set(GGML_OPENMP_RUNTIME_OUTPUT_DIR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
+    if (CMAKE_CONFIGURATION_TYPES)
+        string(APPEND GGML_OPENMP_RUNTIME_OUTPUT_DIR "/$<CONFIG>")
+    endif()
+    add_custom_target(ggml-openmp-runtime ALL
+        COMMAND ${CMAKE_COMMAND} -E make_directory "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_RUNTIME}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/libomp.dll"
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different "${GGML_OPENMP_LICENSE}" "${GGML_OPENMP_RUNTIME_OUTPUT_DIR}/LICENSE-LLVM-OpenMP")
+    add_dependencies(ggml-base ggml-openmp-runtime)
+    install(FILES "${GGML_OPENMP_RUNTIME}" DESTINATION ${CMAKE_INSTALL_BINDIR})
+    install(FILES "${GGML_OPENMP_LICENSE}" DESTINATION ${CMAKE_INSTALL_BINDIR} RENAME LICENSE-LLVM-OpenMP)
+
+    set(GGML_OPENMP_TARGET_C ggml-openmp-c)
+    set(GGML_OPENMP_TARGET_CXX ggml-openmp-cxx)
+    set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
+elseif (GGML_OPENMP)
     find_package(OpenMP)
     if (OpenMP_FOUND)
+        set(GGML_OPENMP_TARGET_C OpenMP::OpenMP_C)
+        set(GGML_OPENMP_TARGET_CXX OpenMP::OpenMP_CXX)
         set(GGML_OPENMP_ENABLED "ON" CACHE INTERNAL "")
     else()
         set(GGML_OPENMP_ENABLED "OFF" CACHE INTERNAL "")
@@ -236,7 +350,7 @@ endif()
 
 if (GGML_OPENMP_ENABLED)
     target_compile_definitions(ggml-base PRIVATE GGML_USE_OPENMP)
-    target_link_libraries(ggml-base PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+    target_link_libraries(ggml-base PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
 endif()
 
 add_library(ggml

ggml/src/ggml-backend-meta.cpp

@@ -602,27 +602,40 @@ 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));
             }
@@ -792,7 +805,7 @@ 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 +815,9 @@ 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;
         }
@@ -1352,15 +1368,29 @@ 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: {

ggml/src/ggml-backend.cpp

@@ -1599,11 +1599,23 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
     std::vector<int32_t> ids;
     std::vector<ggml_bitset_t> used_ids;
 
+    int prev_backend_id = -1;
+
     for (int split_id = 0; split_id < sched->n_splits; split_id++) {
         struct ggml_backend_sched_split * split = &splits[split_id];
         int split_backend_id = split->backend_id;
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
+        // ensure the previous split's async work has completed before we start
+        // this split, the allocator may have reused buffer regions across splits
+        if (split->n_inputs == 0 && prev_backend_id >= 0 && prev_backend_id != split_backend_id) {
+            if (sched->events[prev_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_synchronize(sched->events[prev_backend_id][sched->cur_copy]);
+            } else {
+                ggml_backend_synchronize(sched->backends[prev_backend_id]);
+            }
+        }
+
         // copy the input tensors to the split backend
         for (int input_id = 0; input_id < split->n_inputs; input_id++) {
             ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[input_id]);
@@ -1766,12 +1778,12 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             }
         }
 
-        // record the event of this copy
-        if (split->n_inputs > 0) {
-            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
-                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
-            }
+        // record the event of this split
+        if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+            ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
         }
+
+        prev_backend_id = split_backend_id;
     }
 
     return GGML_STATUS_SUCCESS;

ggml/src/ggml-cpu/CMakeLists.txt

@@ -74,7 +74,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (GGML_OPENMP_ENABLED)
         target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
-        target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_OPENMP_TARGET_C} ${GGML_OPENMP_TARGET_CXX})
     endif()
 
     if (GGML_LLAMAFILE)
@@ -639,6 +639,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             ${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}")
@@ -701,6 +702,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 ${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
                 ${KLEIDIAI_SRC}/kai/ukernels/matmul/pack/kai_lhs_pack_f16pmrx2_f32_neon.c
@@ -737,8 +740,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
     endif()
 
-    if (CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
-        # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
-        target_compile_options(${GGML_CPU_NAME} PRIVATE "-fno-associative-math")
-    endif()
+  if (CMAKE_C_COMPILER_ID STREQUAL "IntelLLVM" OR CMAKE_CXX_COMPILER_ID STREQUAL "IntelLLVM")
+      # The compiler automatically enables "-ffast-math" which can cause NaNs in tests due to "-fassociative-math"
+      target_compile_options(${GGML_CPU_NAME} PRIVATE "$<$<OR:$<COMPILE_LANG_AND_ID:C,IntelLLVM>,$<COMPILE_LANG_AND_ID:CXX,IntelLLVM>>:$<$<BOOL:${WIN32}>:/clang:>-fno-associative-math>")
+  endif()
+
 endfunction()

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

@@ -23,6 +23,7 @@
 #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_lhs_pack_bf16p2vlx2_f32_sme.h"
@@ -76,6 +77,21 @@ static inline void kernel_run_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
     Fn(m, n, k, lhs, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
 }
 
+template <void (*Fn)(size_t, size_t, size_t, const void *, size_t, const void *, void *, size_t, size_t, float, float)>
+static inline void kernel_run_lhs_stride_fn10(size_t       m,
+                                              size_t       n,
+                                              size_t       k,
+                                              size_t       lhs_stride,
+                                              const void * lhs,
+                                              const void * rhs,
+                                              void *       dst,
+                                              size_t       dst_stride_row,
+                                              size_t       dst_stride_col,
+                                              float        clamp_min,
+                                              float        clamp_max) {
+    Fn(m, n, k, lhs, lhs_stride, rhs, dst, dst_stride_row, dst_stride_col, clamp_min, clamp_max);
+}
+
 template<void(*Fn)(size_t,size_t,size_t,const void*,const void*,float*,size_t,size_t,float,float)>
 static inline void kernel_run_float_fn10(size_t m, size_t n, size_t k, size_t /*bl*/,
                                          const void* lhs, const void* rhs, void* dst,
@@ -947,25 +963,25 @@ static ggml_kleidiai_kernels ggml_kleidiai_kernels_f32[] = {
             /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
             /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
         },
-        /* SME GEMV */
+        /* SME2 GEMV */
         {
-            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_mr                = */ kai_get_mr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32p2vlx1_f32p2vlx1biasf32_sme2_mopa,
-            /* .get_lhs_offset_ex     = */ nullptr,
-            /* .get_rhs_packed_offset_ex = */ nullptr,
-            /* .run_kernel_ex         = */ nullptr,
+            /* .get_m_step            = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_n_step            = */ kai_get_n_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_mr                = */ kai_get_m_step_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_nr                = */ kai_get_nr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_kr                = */ kai_get_kr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_sr                = */ kai_get_sr_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_offset        = */ kai_get_dst_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_dst_size          = */ kai_get_dst_size_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla,
+            /* .get_lhs_offset_ex     = */ &kernel_offs_fn2<kai_get_lhs_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .get_rhs_packed_offset_ex = */ &kernel_offs_fn2<kai_get_rhs_packed_offset_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
+            /* .run_kernel_ex         = */ &kernel_run_lhs_stride_fn10<kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla>,
         },
         /* .gemv_lhs_info = */ {
-            /* .get_offset            = */ kai_get_lhs_offset_lhs_pack_f32p2vlx1_f32_sme,
-            /* .get_packed_offset_ex  = */ &lhs_offs_fn5<kai_get_lhs_packed_offset_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .packed_size_ex        = */ &lhs_ps_fn5<kai_get_lhs_packed_size_lhs_pack_f32p2vlx1_f32_sme>,
-            /* .pack_func_ex          = */ &lhs_pack_void_fn9<kai_run_lhs_pack_f32p2vlx1_f32_sme>,
+            /* .get_offset            = */ nullptr,
+            /* .get_packed_offset_ex  = */ nullptr,
+            /* .packed_size_ex        = */ nullptr,
+            /* .pack_func_ex          = */ nullptr,
         },
         /* .rhs_info = */ {
             /* .packed_stride         = */ nullptr,

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

@@ -696,6 +696,15 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         }
 
         if (op->src[0]->type == GGML_TYPE_F32) {
+            ggml_kleidiai_kernels * primary     = kernel_chain[0];
+            kernel_info *           gemv_kernel = primary ? &primary->gemv : nullptr;
+            if (is_gemv && op->src[1]->nb[0] == (int64_t) sizeof(float) && gemv_kernel &&
+                gemv_kernel->get_lhs_offset_ex && gemv_kernel->get_rhs_packed_offset_ex &&
+                gemv_kernel->run_kernel_ex && gemv_kernel->get_dst_offset) {
+                size = 0;
+                return true;
+            }
+
             size_t cursor = 0;
             bool any_slot = false;
 
@@ -811,15 +820,28 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             return false;
         }
 
-        kernel_info * kernel        = &kernels->gemm;
+        const size_t k = ne00;
+        const size_t m = ne11;
+        const size_t n = ne01;
+        const bool use_gemv = m == 1 && src1->nb[0] == (int64_t) sizeof(float) &&
+                              kernels->gemv.get_lhs_offset_ex &&
+                              kernels->gemv.get_rhs_packed_offset_ex &&
+                              kernels->gemv.run_kernel_ex &&
+                              kernels->gemv.get_dst_offset;
+
+        kernel_info * kernel        = use_gemv ? &kernels->gemv : &kernels->gemm;
         lhs_packing_info * lhs_info = &kernels->gemm_lhs_info;
 
-        if (!kernel || !lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
-            !lhs_info->packed_size_ex || !lhs_info->pack_func_ex ||
+        if (!kernel || !kernel->get_lhs_offset_ex ||
             !kernel->get_rhs_packed_offset_ex || !kernel->run_kernel_ex || !kernel->get_dst_offset) {
             return false;
         }
 
+        if (!use_gemv && (!lhs_info || !lhs_info->get_offset || !lhs_info->get_packed_offset_ex ||
+                          !lhs_info->packed_size_ex || !lhs_info->pack_func_ex)) {
+            return false;
+        }
+
         const kleidiai_weight_header * header = kleidiai_weight_header_from_ptr(src0->data);
         const bool has_header = kleidiai_is_weight_header_valid(header);
 
@@ -832,16 +854,14 @@ class tensor_traits : public ggml::cpu::tensor_traits {
         const int nth = params->nth > 0 ? params->nth : 1;
         const int ith = params->ith;
 
-        const size_t k = ne00;
-        const size_t m = ne11;
-        const size_t n = ne01;
-
         const size_t mr = kernel->get_mr();
         const size_t kr = kernel->get_kr();
         const size_t sr = kernel->get_sr();
 
-        const size_t lhs_packed_size = lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
-        GGML_ASSERT(lhs_packed_size <= params->wsize);
+        const size_t lhs_packed_size = use_gemv ? 0 : lhs_info->packed_size_ex(m, k, 0, mr, kr, sr);
+        if (!use_gemv) {
+            GGML_ASSERT(lhs_packed_size <= params->wsize);
+        }
 
         uint8_t * lhs_packed   = static_cast<uint8_t *>(params->wdata);
         const size_t dst_stride = dst->nb[1];
@@ -853,7 +873,7 @@ class tensor_traits : public ggml::cpu::tensor_traits {
             const uint8_t * lhs_batch_base = static_cast<const uint8_t *>(src1->data) + batch_idx * src1->nb[2];
             uint8_t * dst_batch_base = static_cast<uint8_t *>(dst->data) + batch_idx * dst->nb[2];
 
-            {
+            if (!use_gemv) {
                 const int64_t m_roundup_mr = kai_roundup((int64_t)m, (int64_t)mr);
                 int64_t max_threads = mr ? (m_roundup_mr / (int64_t)mr) : nth;
                 max_threads = std::max<int64_t>(1, max_threads);
@@ -903,15 +923,17 @@ class tensor_traits : public ggml::cpu::tensor_traits {
                 const size_t n_to_process = std::min(chunk_cols, n - n_start);
 
                 if (n_to_process > 0) {
-                    const size_t lhs_packed_offset = lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
+                    const size_t lhs_offset = use_gemv ? kernel->get_lhs_offset_ex(0, k, 0)
+                                                       : lhs_info->get_packed_offset_ex(0, k, 0, mr, kr, sr);
                     const size_t rhs_packed_offset = kernel->get_rhs_packed_offset_ex(n_start, k, 0);
                     const size_t dst_offset        = kernel->get_dst_offset(0, n_start, dst_stride);
 
-                    const void * lhs_ptr = lhs_packed + lhs_packed_offset;
+                    const void * lhs_ptr = use_gemv ? lhs_batch_base + lhs_offset
+                                                    : lhs_packed + lhs_offset;
                     const void * rhs_ptr = rhs_base + rhs_packed_offset;
                     float * dst_ptr      = reinterpret_cast<float *>(dst_batch_base + dst_offset);
 
-                    kernel->run_kernel_ex(m, n_to_process, k, 0,
+                    kernel->run_kernel_ex(m, n_to_process, k, use_gemv ? src1->nb[1] : 0,
                                           lhs_ptr,
                                           rhs_ptr,
                                           dst_ptr,

ggml/src/ggml-cpu/ops.cpp

@@ -1896,7 +1896,6 @@ void ggml_compute_forward_repeat_back(
 }
 
 // ggml_compute_forward_concat
-
 static void ggml_compute_forward_concat_any(
     const ggml_compute_params * params,
     ggml_tensor * dst) {
@@ -1904,8 +1903,6 @@ static void ggml_compute_forward_concat_any(
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const size_t len = ggml_type_size(src0->type);
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -1914,31 +1911,38 @@ static void ggml_compute_forward_concat_any(
     const int32_t dim = ggml_get_op_params_i32(dst, 0);
 
     GGML_ASSERT(dim >= 0 && dim < 4);
+    GGML_ASSERT(ggml_is_contiguous_rows(src0));
+    GGML_ASSERT(ggml_is_contiguous_rows(src1));
 
     int64_t o[4] = {0, 0, 0, 0};
+
     if (dim == 0) {
+        GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
+        GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
+
         o[dim] = src0->ne[dim]/ggml_blck_size(src0->type);
     } else {
         o[dim] = src0->ne[dim];
     }
 
-    const char * x;
+    // Region 1: copy rows from src0
+    for (int i3 = 0; i3 < ne03; i3++) {
+        for (int i2 = ith; i2 < ne02; i2 += nth) {
+            for (int i1 = 0; i1 < ne01; i1++) {
+                const char * x = (const char *) src0->data + i1*nb01 + i2*nb02 + i3*nb03;
+                      char * y = (      char *) dst->data  + i1*nb1  + i2*nb2  + i3*nb3;
+                memcpy(y, x, ggml_row_size(src0->type, ne00));
+            }
+        }
+    }
 
-    // TODO: smarter multi-theading
-    for (int i3 = 0; i3 < ne3; i3++) {
-        for (int i2 = ith; i2 < ne2; i2 += nth) {
-            for (int i1 = 0; i1 < ne1; i1++) {
-                for (int i0 = 0; i0 < ne0/ggml_blck_size(dst->type); i0++) {
-                    if (i0 < ne00/ggml_blck_size(src0->type) && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        x = (const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03;
-                    } else {
-                        x = (const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13;
-                    }
-
-                    char * y = (char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3;
-
-                    memcpy(y, x, len);
-                }
+    // Region 2: copy rows from src1, offset into dst by o[]
+    for (int i3 = 0; i3 < ne13; i3++) {
+        for (int i2 = ith; i2 < ne12; i2 += nth) {
+            for (int i1 = 0; i1 < ne11; i1++) {
+                const char * x = (const char *) src1->data + i1*nb11         + i2*nb12         + i3*nb13;
+                      char * y = (      char *)  dst->data + (i1 + o[1])*nb1 + (i2 + o[2])*nb2 + (i3 + o[3])*nb3 + o[0]*nb0;
+                memcpy(y, x, ggml_row_size(src1->type, ne10));
             }
         }
     }
@@ -2078,14 +2082,6 @@ void ggml_compute_forward_concat(
     ggml_tensor * dst) {
 
     const ggml_tensor * src0 = dst->src[0];
-    const ggml_tensor * src1 = dst->src[1];
-
-    if (ggml_is_quantized(src0->type)) {
-        GGML_ASSERT(ggml_is_contiguous_rows(src0));
-        GGML_ASSERT(ggml_is_contiguous_rows(src1));
-        GGML_ASSERT(src0->ne[0] % ggml_blck_size(src0->type) == 0);
-        GGML_ASSERT(src1->ne[0] % ggml_blck_size(src1->type) == 0);
-    }
 
     switch (src0->type) {
         case GGML_TYPE_F16:

ggml/src/ggml-cpu/simd-mappings.h

@@ -29,13 +29,15 @@ extern "C" {
 // FP16 to FP32 conversion
 
 // 16-bit float
-// on Arm, we use __fp16
+// on Arm, we use __fp16, which requires the IEEE fp16 format: implied on
+// AArch64, selected by -mfp16-format=ieee on 32 bit Arm, where the compiler
+// may otherwise reject the type
 // on x86, we use uint16_t
 //
 // for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
 // for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
 //
-#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+#if defined(__ARM_NEON) && defined(__ARM_FP16_FORMAT_IEEE) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
     #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
     #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
 
@@ -326,7 +328,7 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif
 
-#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) && defined(__ARM_FP16_FORMAT_IEEE)
 
 #define GGML_SIMD
 

ggml/src/ggml-cuda/common.cuh

@@ -1418,7 +1418,9 @@ struct ggml_backend_cuda_context {
     cudaEvent_t copy_event = nullptr;
 
     cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
-    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
+    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    void * cublas_workspaces[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = {nullptr};
+    size_t cublas_workspace_sizes[GGML_CUDA_MAX_DEVICES] = {0};
 
     int curr_stream_no = 0;
 
@@ -1495,17 +1497,22 @@ struct ggml_backend_cuda_context {
 
     ggml_cuda_stream_context & stream_context() { return concurrent_stream_context; }
 
-    cublasHandle_t cublas_handle(int device) {
-        if (cublas_handles[device] == nullptr) {
-            ggml_cuda_set_device(device);
-            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
-            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
-        }
-        return cublas_handles[device];
-    }
-
     cublasHandle_t cublas_handle() {
-        return cublas_handle(device);
+        if (cublas_handles[device][curr_stream_no] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasCreate(&cublas_handles[device][curr_stream_no]));
+            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device][curr_stream_no], CUBLAS_TF32_TENSOR_OP_MATH));
+            CUBLAS_CHECK(cublasSetStream(cublas_handles[device][curr_stream_no], stream()));
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && (CUBLAS_VER_MAJOR > 11 || (CUBLAS_VER_MAJOR == 11 && CUBLAS_VER_MINOR >= 2))
+            if (cublas_workspace_sizes[device] == 0) {
+                const int cc = ggml_cuda_info().devices[device].cc;
+                cublas_workspace_sizes[device] = (cc >= GGML_CUDA_CC_HOPPER) ? 32 * 1024 * 1024 : 4 * 1024 * 1024;
+            }
+            CUDA_CHECK(cudaMalloc(&cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+            CUBLAS_CHECK(cublasSetWorkspace(cublas_handles[device][curr_stream_no], cublas_workspaces[device][curr_stream_no], cublas_workspace_sizes[device]));
+#endif
+        }
+        return cublas_handles[device][curr_stream_no];
     }
 
     // pool

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

@@ -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"
@@ -711,9 +712,12 @@ ggml_backend_cuda_context::~ggml_backend_cuda_context() {
             if (streams[i][j] != nullptr) {
                 CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
             }
-        }
-        if (cublas_handles[i] != nullptr) {
-            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            if (cublas_handles[i][j] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i][j]));
+            }
+            if (cublas_workspaces[i][j] != nullptr) {
+                CUDA_CHECK(cudaFree(cublas_workspaces[i][j]));
+            }
         }
     }
 }
@@ -1416,7 +1420,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
 
     const int64_t ne_dst = ggml_nelements(dst);
     cudaStream_t main_stream = ctx.stream();
-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
+    cublasHandle_t cublas_h = ctx.cublas_handle();
 
     const size_t src0_ts = ggml_type_size(src0->type);
     GGML_ASSERT(nb00 == src0_ts);
@@ -1539,14 +1543,14 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
     //     probably because the internal kernel selection logic is suboptimal.
     if (compute_type == GGML_TYPE_F32 && ne12 == 1 && ne13 == 1) {
         CUBLAS_CHECK(
-            cublasSgemm(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasSgemm(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                     ne01, ne11, ne10,
                     (const float *) alpha, (const float *) src0_ptr, s01,
                                            (const float *) src1_ptr, s11,
                     (const float *) beta,  (float       *)  dst_ptr, ne0));
     } else if (ne12 == 1 && ne13 == 1) {
         CUBLAS_CHECK(
-            cublasGemmEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+            cublasGemmEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                     ne01, ne11, ne10,
                     alpha, src0_ptr, cu_data_type_a, s01,
                            src1_ptr, cu_data_type_b, s11,
@@ -1561,7 +1565,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
         // there is no broadcast and src0, src1 are contiguous across dims 2, 3
         // use cublasGemmStridedBatchedEx
         CUBLAS_CHECK(
-        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmStridedBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
                 alpha, src0_ptr, cu_data_type_a, s01, sma,     // strideA
                        src1_ptr, cu_data_type_b, s11, smb,     // strideB
@@ -1599,7 +1603,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
         CUDA_CHECK(cudaGetLastError());
 
         CUBLAS_CHECK(
-        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+        cublasGemmBatchedEx(cublas_h, CUBLAS_OP_T, CUBLAS_OP_N,
                 ne01, ne11, ne10,
                 alpha, (const void **) (ptrs_src.get() + 0*ne23), cu_data_type_a, s01,
                        (const void **) (ptrs_src.get() + 1*ne23), cu_data_type_b, s11,
@@ -2323,6 +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;
@@ -5242,6 +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:

ggml/src/ggml-cuda/mmvq.cu

@@ -290,6 +290,42 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
     if (!ggml_is_quantized(type)) {
         return false;
     }
+    // k-quants cost more to decode and mvq redoes that per column, so MMQ wins sooner.
+    // Only list quant-types MMQ supports, others would fall back to cuBLAS.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_ADA_LOVELACE) {
+        switch (type) { // tuned on RTX 4090
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 4;
+            case GGML_TYPE_Q3_K:
+                return ne11 <= 6;
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_BLACKWELL) {
+        switch (type) { // tuned on RTX 5090
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 7;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_DGX_SPARK) {
+        switch (type) { // tuned on DGX Spark GB10
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 6;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
     if (GGML_CUDA_CC_IS_CDNA(cc)) {
         if (GGML_CUDA_CC_IS_CDNA1(cc)) {
             switch (type) {

ggml/src/ggml-cuda/out-prod.cu

@@ -54,8 +54,6 @@ void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const float alpha = 1.0f;
     const float beta = 0.0f;
 
-    CUBLAS_CHECK(cublasSetStream(handle, stream));
-
     const int64_t lda = nb01 / sizeof(float);
     const int64_t ldc = nb1  / sizeof(float);
 

ggml/src/ggml-cuda/pool1d.cu

@@ -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);
+}

ggml/src/ggml-cuda/pool1d.cuh

@@ -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);

ggml/src/ggml-cuda/solve_tri.cu

@@ -65,15 +65,13 @@ static void solve_tri_f32_cublas(ggml_backend_cuda_context & ctx,
     get_batch_pointers<<<(total_batches + 255) / 256, 256, 0, stream>>>(A, X, A_ptrs_dev, X_ptrs_dev, ne02,
                                                                         total_batches, s02, s03, s2, s3);
 
-    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
-
     // Yes, this is necessary, without this we get RMSE errors
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_DEFAULT_MATH));
-    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(id), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_DEFAULT_MATH));
+    CUBLAS_CHECK(cublasStrsmBatched(ctx.cublas_handle(), CUBLAS_SIDE_RIGHT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
                                     CUBLAS_DIAG_NON_UNIT, k, n, &alpha, A_ptrs_dev, n, X_ptrs_dev, k, total_batches));
 
     // revert to standard mode from common.cuh
-    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(id), CUBLAS_TF32_TENSOR_OP_MATH));
+    CUBLAS_CHECK(cublasSetMathMode(ctx.cublas_handle(), CUBLAS_TF32_TENSOR_OP_MATH));
 
     GGML_UNUSED_VARS(s12, s13);
 }

ggml/src/ggml-cuda/ssm-scan.cu

@@ -632,7 +632,6 @@ static void ssm_scan_ssd_f32_cuda(
     // Step 3: chunked SSD loop
     // Per chunk: pre_matmul (incl. M) + 4 cuBLAS (CB, Y, S@C, state update) + scale_state
     cublasHandle_t handle = ctx.cublas_handle();
-    CUBLAS_CHECK(cublasSetStream(handle, stream));
     const float alpha_one  = 1.0f;
     const float beta_zero  = 0.0f;
     const float beta_one   = 1.0f;

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

@@ -3180,8 +3180,9 @@ static bool ggml_hexagon_supported_argsort(const struct ggml_hexagon_session * s
 static bool ggml_hexagon_supported_rope(const struct ggml_hexagon_session * sess, const struct ggml_tensor * op) {
     const int32_t * op_params = &op->op_params[0];
 
-    if (op_params[15] != 0) {
-        return false; // FIXME: support ggml_rope_set_offset
+    // ggml_rope_set_offset: HVX kernels need a VLEN-aligned window start (32 f32 elems)
+    if (op_params[15] % 32 != 0) {
+        return false;
     }
 
     int mode = op_params[2];

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

@@ -132,8 +132,8 @@ struct hmx_fa_context {
     __fp16 *     vtcm_v_tiles[2];      // V tiles (column-major, double-buffered)
     __fp16 *     vtcm_s_tiles[2];      // S = QK^T [g_br, Bc] (double-buffered)
     __fp16 *     vtcm_p_tiles[2];      // P = softmax(S) [g_br, Bc]
-    __fp16 *     vtcm_d_tiles;         // Diagonal rescale [g_br, g_br]
-    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l) [g_br, g_br]
+    __fp16 *     vtcm_d_tiles[2];      // Diagonal rescale, g_br/32 packed diagonal tiles (double-buffered)
+    __fp16 *     vtcm_d_inv_l;         // Diagonal rescale (1/l), same packed layout
     HVX_Vector * vtcm_m_vec;           // Row max [g_br]
     HVX_Vector * vtcm_l_vec;           // Row sum [g_br]
     HVX_Vector * vtcm_s_rowmax;        // Softmax intermediate [g_br]
@@ -782,13 +782,14 @@ static void fa_q_load_thread(unsigned int n, unsigned int i, void * data) {
             }
         }
 
-        // Initialize vtcm_d_tiles and vtcm_d_inv_l to 0
+        // Zero the whole rescale region: vtcm_d_tiles[0], the optional vtcm_d_tiles[1]
+        // and vtcm_d_inv_l are equal-sized and allocated back to back, so one run covers
+        // them all.  The scatter only ever writes the diagonal, ignore the rest.
         const size_t d_bytes_per_t = hex_align_up(d_tile_bytes / n, 128);
         const size_t d_start       = i * d_bytes_per_t;
         const size_t d_end         = hex_smin(d_start + d_bytes_per_t, d_tile_bytes);
         if (d_start < d_tile_bytes) {
-            hvx_splat_u8_a((char *) factx->vtcm_d_tiles + d_start, 0, d_end - d_start);
-            hvx_splat_u8_a((char *) factx->vtcm_d_inv_l + d_start, 0, d_end - d_start);
+            hvx_splat_u8_a((char *) factx->vtcm_d_tiles[0] + d_start, 0, d_end - d_start);
         }
     }
 
@@ -1432,17 +1433,19 @@ static inline void fa_softmax_impl(
         const HVX_VectorPred q_32_mask = Q6_Q_vsetq_R(32 * sizeof(__fp16));
         HVX_Vector           v_exp_m_diff = exp_m_diff_f16;
 
+        __fp16 * const d_tiles_out = factx->vtcm_d_tiles[args->buf_idx];
+
         size_t t0 = r_vec_idx * 2;
         if (t0 < args->n_row_tiles) {
             const HVX_Vector v_content = v_exp_m_diff;
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t0 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t0 * HMX_FP16_TILE_N_ELMS;
             Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
         }
 
         size_t t1 = r_vec_idx * 2 + 1;
         if (t1 < args->n_row_tiles) {
             const HVX_Vector v_content = Q6_V_vror_VR(v_exp_m_diff, 64);
-            __fp16 *         out_base  = factx->vtcm_d_tiles + t1 * (args->n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+            __fp16 *         out_base  = d_tiles_out + t1 * HMX_FP16_TILE_N_ELMS;
             Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
         }
     }
@@ -1506,7 +1509,7 @@ static __attribute__((noinline)) void fa_build_d_diag_inv_l(struct hmx_fa_contex
             v_content = Q6_V_vror_VR(v_content, 64);
         }
 
-        __fp16 * out_base = factx->vtcm_d_inv_l + i * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        __fp16 * out_base = factx->vtcm_d_inv_l + i * HMX_FP16_TILE_N_ELMS;
         Q6_vscatter_QRMVhV(q_32_mask, (size_t) out_base, HMX_FP16_TILE_SIZE - 1, v_offsets, v_content);
     }
 }
@@ -1615,7 +1618,7 @@ static void hmx_fa_o_update_worker(void * data) {
     const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
     const size_t v_stride = n_tiles_per_bc * HMX_FP16_TILE_N_ELMS;
     for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag     = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag     = d_tiles + r * HMX_FP16_TILE_N_ELMS;
         const __fp16 * p_tile_in  = p_tiles + (r * n_tiles_per_bc) * HMX_FP16_TILE_N_ELMS;
         const __fp16 * o_rc       = o_prev + r * HMX_FP16_TILE_N_ELMS;
         const __fp16 * v_tile_in  = v_tiles;
@@ -1654,7 +1657,7 @@ static void hmx_fa_o_norm_worker(void * data) {
     asm volatile(HMX_SET_BIAS("%0") :: "r"((unsigned int)job->hmx_scales));
     const size_t o_stride = n_row_tiles_g_br * HMX_FP16_TILE_N_ELMS;
     for (size_t r = 0; r < n_row_tiles; ++r) {
-        const __fp16 * d_diag = d_tiles + r * (n_row_tiles_g_br + 1) * HMX_FP16_TILE_N_ELMS;
+        const __fp16 * d_diag = d_tiles + r * HMX_FP16_TILE_N_ELMS;
         const __fp16 * o_rc = o_prev + r * HMX_FP16_TILE_N_ELMS;
         __fp16 *       o_out = o_curr + r * DV_tiles * HMX_FP16_TILE_N_ELMS;
 
@@ -1882,7 +1885,8 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
     factx.vtcm_s_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_s_tiles[1], pipeline);
     factx.vtcm_p_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_p_tiles[0]);
     factx.vtcm_p_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_p_tiles[1], pipeline);
-    factx.vtcm_d_tiles        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles);
+    factx.vtcm_d_tiles[0]     = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_tiles[0]);
+    factx.vtcm_d_tiles[1]     = VTCM_LAYOUT_PTR_OPTIONAL(__fp16, base, L.off_d_tiles[1], pipeline);
     factx.vtcm_d_inv_l        = VTCM_LAYOUT_PTR(__fp16, base, L.off_d_inv_l);
     factx.vtcm_m_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_m_vec);
     factx.vtcm_l_vec          = VTCM_LAYOUT_PTR(HVX_Vector, base, L.off_l_vec);
@@ -2039,7 +2043,30 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             }
                         }
 
-                        // ---- 3. Pop and run K-prep for next block & push next QK-dot ----
+                        // ---- 3. Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx], D) ----
+                        // O update relys on the previous block's P and V tiles.
+                        // O update MUST be pushed before the next block's QK-dot: hmx_queue_pop() retires the
+                        // oldest descriptor, so push order alone decides which pop waits for which job.
+                        // If OU went in after QK(i+1), the pop below would retire QK(i+1) and leave
+                        // OU(i-1) in flight into the next iteration, where V-prep overwrites V[prev_buf].
+                        if (kv_blk > 0) {
+                            const size_t prev_buf        = 1 - buf_idx;
+                            ou_job[prev_buf].o_curr      = o_tile_curr;
+                            ou_job[prev_buf].o_prev      = o_tile_prev;
+                            ou_job[prev_buf].p_tiles     = factx.vtcm_p_tiles[prev_buf];
+                            ou_job[prev_buf].v_tiles     = factx.vtcm_v_tiles[prev_buf];
+                            ou_job[prev_buf].d_tiles     = factx.vtcm_d_tiles[prev_buf];
+                            ou_job[prev_buf].hmx_scales  = factx.vtcm_hmx_scales_id;
+                            ou_job[prev_buf].n_row_tiles = n_row_tiles;
+                            ou_job[prev_buf].n_col_tiles =
+                                hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
+                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
+                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
+                            ou_job[prev_buf].DV               = DV;
+                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
+                        }
+
+                        // ---- 4. Pop and run K-prep for next block & push next QK-dot ----
                         if (kv_blk + 1 < factx.n_kv_blocks) {
                             const uint32_t next_start = (kv_blk + 1) * Bc;
                             const uint32_t next_rows  = hex_smin(Bc, nek1 - next_start);
@@ -2059,10 +2086,10 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_qk_dot_worker, &qk_job[next_buf]));
                         }
 
-                        // ---- 4. Wait for current block's QK-dot to finish ----
+                        // ---- 5. Wait for current block's QK-dot to finish ----
                         hmx_queue_pop(hmx_q);
 
-                        // ---- 5. Phase 2: softmax + build_D ----
+                        // ---- 6. Phase 2: softmax + build_D ----
                         fa_softmax_args_t sargs;
                         memset(&sargs, 0, sizeof(sargs));
                         sargs.factx                = &factx;
@@ -2085,23 +2112,6 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         sargs.mask_vtcm_row_stride = factx.mask_buf_row_stride;
                         sargs.slopes               = factx.vtcm_slopes;
 
-                        // Start HMX O update for block kv_blk - 1 (reads P[1 - buf_idx], V[1 - buf_idx])
-                        if (kv_blk > 0) {
-                            const size_t prev_buf = 1 - buf_idx;
-                            ou_job[prev_buf].o_curr           = o_tile_curr;
-                            ou_job[prev_buf].o_prev           = o_tile_prev;
-                            ou_job[prev_buf].p_tiles          = factx.vtcm_p_tiles[prev_buf];
-                            ou_job[prev_buf].v_tiles          = factx.vtcm_v_tiles[prev_buf];
-                            ou_job[prev_buf].d_tiles          = factx.vtcm_d_tiles;
-                            ou_job[prev_buf].hmx_scales       = factx.vtcm_hmx_scales_id;
-                            ou_job[prev_buf].n_row_tiles      = n_row_tiles;
-                            ou_job[prev_buf].n_col_tiles      = hmx_ceil_div(hex_smin(Bc, nek1 - (kv_blk - 1) * Bc), HMX_FP16_TILE_N_COLS);
-                            ou_job[prev_buf].n_row_tiles_g_br = n_row_tiles_g_br;
-                            ou_job[prev_buf].n_tiles_per_bc   = n_tiles_per_bc;
-                            ou_job[prev_buf].DV               = DV;
-                            hmx_queue_push(hmx_q, hmx_queue_make_desc(hmx_fa_o_update_worker, &ou_job[prev_buf]));
-                        }
-
                         // Run Softmax on HVX (blocking call)
                         fa_phase_softmax_and_build_d(&factx, &sargs, n_row_tiles, n_row_tiles_g_br);
 
@@ -2128,7 +2138,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                         ou_job[0].o_prev           = o_tile_prev;
                         ou_job[0].p_tiles          = factx.vtcm_p_tiles[1 - buf_idx];
                         ou_job[0].v_tiles          = factx.vtcm_v_tiles[1 - buf_idx];
-                        ou_job[0].d_tiles          = factx.vtcm_d_tiles;
+                        ou_job[0].d_tiles          = factx.vtcm_d_tiles[1 - buf_idx];
                         ou_job[0].hmx_scales       = factx.vtcm_hmx_scales_id;
                         ou_job[0].n_row_tiles      = n_row_tiles;
                         ou_job[0].n_col_tiles      = last_cols;
@@ -2232,7 +2242,7 @@ int hmx_flash_attn_ext(struct htp_ops_context * octx) {
                             ou_job.o_prev           = o_tile_prev;
                             ou_job.p_tiles          = factx.vtcm_p_tiles[0];
                             ou_job.v_tiles          = factx.vtcm_v_tiles[0];
-                            ou_job.d_tiles          = factx.vtcm_d_tiles;
+                            ou_job.d_tiles          = factx.vtcm_d_tiles[0];
                             ou_job.hmx_scales       = factx.vtcm_hmx_scales_id;
                             ou_job.n_row_tiles      = n_row_tiles;
                             ou_job.n_col_tiles      = n_col_tiles;

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

@@ -109,7 +109,7 @@ struct hmx_fa_vtcm_layout {
     size_t off_v_tiles[2];
     size_t off_s_tiles[2];
     size_t off_p_tiles[2];
-    size_t off_d_tiles;
+    size_t off_d_tiles[2];
     size_t off_d_inv_l;
     size_t off_m_vec;
     size_t off_l_vec;
@@ -125,7 +125,7 @@ struct hmx_fa_vtcm_layout {
     size_t q_tile_bytes;
     size_t o_tile_bytes;
     size_t s_tile_bytes;       // S and P tiles (same size)
-    size_t d_tile_bytes;
+    size_t d_tile_bytes;       // d_tiles[0..1] + d_inv_l, allocated back to back
     size_t m_line_bytes;       // one mask row
     size_t m_buf_slot_bytes;   // one dma_cache slot = align_up(Br * m_line_bytes, 4096)
     size_t col_vec_bytes;
@@ -149,7 +149,12 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
     const size_t k_tile_size  = hex_align_up(Bc   * DK   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
     const size_t v_tile_size  = hex_align_up(Bc   * DV   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
     const size_t s_tile_size  = hex_align_up(g_br * Bc   * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
-    const size_t d_tile_size  = hex_align_up(g_br * g_br * sizeof(__fp16), HTP_FA_HMX_TILE_SIZE);
+
+    // The rescale matrices are diagonal: the HMX kernels only ever load the g_br/32
+    // tiles that sit on the diagonal, so store just those, packed back to back with
+    // a stride of one tile.  The old [g_br, g_br] square layout allocated g_br/32
+    // times more than it used, which is also why a second D buffer was unaffordable.
+    const size_t d_tile_size = (g_br / HMX_FP16_TILE_N_ROWS) * HTP_FA_HMX_TILE_SIZE;
 
     const size_t q_dma_size   = hex_align_up(g_br * DK * (is_q_fp32 ? sizeof(float) : sizeof(__fp16)), 128);
     const size_t k_dma_size   = hex_align_up(Bc * hex_round_up(DK * sizeof(__fp16), 128), 128);
@@ -167,7 +172,8 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
     VTCM_LAYOUT_ALLOC(off, off_q_tiles,       q_tile_size);
     VTCM_LAYOUT_ALLOC(off, off_o_tiles[0],    o_tile_size);
     VTCM_LAYOUT_ALLOC(off, off_o_tiles[1],    o_tile_size);
-    VTCM_LAYOUT_ALLOC(off, off_d_tiles,       d_tile_size);
+    VTCM_LAYOUT_ALLOC(off, off_d_tiles[0],    d_tile_size);
+    VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_d_tiles[1], d_tile_size, pipeline);
     VTCM_LAYOUT_ALLOC(off, off_d_inv_l,       d_tile_size);
 
     // Group B & C share start offset (Group B tiles must be 2KB aligned)
@@ -213,7 +219,10 @@ static inline void hmx_fa_vtcm_layout_build(struct hmx_fa_vtcm_layout * L,
     L->o_tile_bytes        = o_tile_size;
     L->col_vec_bytes       = col_vec_size;
     L->s_tile_bytes        = s_tile_size;
-    L->d_tile_bytes        = d_tile_size;
+    // Measured from the actual offsets rather than assumed to be N * d_tile_size, so
+    // that inserting a region between them (or adding padding to VTCM_LAYOUT_ALLOC)
+    // cannot silently leave the tail of the run unzeroed.
+    L->d_tile_bytes        = (L->off_d_inv_l + d_tile_size) - L->off_d_tiles[0];
     L->m_line_bytes        = m_line_size;
     L->m_buf_slot_bytes    = m_buf_slot;
     L->row_buf_stride      = row_vec_size / 128;

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

@@ -53,6 +53,7 @@
 
 struct htp_rope_context {
     int32_t n_dims;
+    int32_t n_offs;
     int32_t mode;
     int32_t n_ctx_orig;
     int32_t sections[4];
@@ -405,32 +406,40 @@ static inline void hvx_rope_f32_aa(float * restrict dst, const float * restrict
 
 static void inline rope_basic_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                    uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
     #pragma unroll(4)
     for (uint32_t i = 0; i < nr; i++) {
         float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
         float * s = (float *) (src + i * rctx->src0_row_size_aligned);
 
-        hvx_rope_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);
 
         // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
         }
     }
 }
 
 static void inline rope_neox_f32(struct htp_rope_context * rctx, uint8_t * restrict dst, uint8_t * restrict src,
                    uint32_t nr, uint32_t ne0, const float * restrict theta_cache) {
+    const uint32_t n_offs = rctx->n_offs; // VLEN-aligned (enforced by supports_op)
     #pragma unroll(4)
     for (uint32_t i = 0; i < nr; i++) {
         float * d = (float *) (dst + i * rctx->dst_row_size_aligned);
         float * s = (float *) (src + i * rctx->src0_row_size_aligned);
 
-        hvx_rope_neox_f32_aa(d, s, rctx->n_dims, theta_cache);
+        hvx_rope_neox_f32_aa(d + n_offs, s + n_offs, rctx->n_dims, theta_cache);
 
         // fill the remain channels with data from src tensor
-        if (rctx->n_dims < ne0) {
-            hvx_copy_f32_uu((uint8_t *)(d + rctx->n_dims), (uint8_t *)(s + rctx->n_dims), ne0 - rctx->n_dims);
+        if (n_offs > 0) {
+            hvx_copy_f32_uu((uint8_t *) d, (uint8_t *) s, n_offs);
+        }
+        if (n_offs + rctx->n_dims < ne0) {
+            hvx_copy_f32_uu((uint8_t *)(d + n_offs + rctx->n_dims), (uint8_t *)(s + n_offs + rctx->n_dims), ne0 - n_offs - rctx->n_dims);
         }
     }
 }
@@ -673,6 +682,7 @@ static int execute_op_rope_f32(struct htp_ops_context * octx) {
     rctx.n_dims     = ((const int32_t *) op_params)[1];
     rctx.mode       = ((const int32_t *) op_params)[2];
     rctx.n_ctx_orig = ((const int32_t *) op_params)[4];
+    rctx.n_offs     = ((const int32_t *) op_params)[15];
 
     memcpy(&rctx.freq_base,   (int32_t *) op_params + 5,  sizeof(float));
     memcpy(&rctx.freq_scale,  (int32_t *) op_params + 6,  sizeof(float));

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

@@ -1409,6 +1409,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_p
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    char base[256];
+
+    snprintf(base, 256, "kernel_flash_attn_ext_kv_%s_f16", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, base);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, base, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
@@ -1460,7 +1477,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
-        int32_t nsg) {
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
     char base[256];
@@ -1469,15 +1489,14 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
     const int32_t dk = (int32_t) op->src[1]->ne[0];
     const int32_t dv = (int32_t) op->src[2]->ne[0];
 
-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);
 
     // do bounds checks for the mask?
     const bool bc_mask = op->src[3] && (op->src[3]->ne[1] % 8 != 0);
 
     snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
             "flash_attn_ext",
-            ggml_type_name(op->src[1]->type),
+            type,
             dk,
             dv);
 
@@ -1526,7 +1545,10 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
         bool    has_scap,
         bool    has_kvpad,
         int32_t nsg,
-        int32_t nwg) {
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
     char base[256];
@@ -1535,12 +1557,11 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
     const int32_t dk = (int32_t) op->src[1]->ne[0];
     const int32_t dv = (int32_t) op->src[2]->ne[0];
 
-    const int32_t ns10 = op->src[1]->nb[1]/op->src[1]->nb[0];
-    const int32_t ns20 = op->src[2]->nb[1]/op->src[2]->nb[0];
+    const char * type = use_kv_f16 ? "f16" : ggml_type_name(op->src[1]->type);
 
     snprintf(base, 256, "kernel_%s_%s_dk%d_dv%d",
             "flash_attn_ext_vec",
-            ggml_type_name(op->src[1]->type),
+            type,
             dk,
             dv);
 

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

@@ -176,6 +176,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_mask,
         int32_t ncpsg);
 
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op);
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
@@ -190,7 +194,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_bias,
         bool    has_scap,
         bool    has_kvpad,
-        int32_t nsg);
+        int32_t nsg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);
 
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         ggml_metal_library_t lib,
@@ -201,7 +208,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
         bool    has_scap,
         bool    has_kvpad,
         int32_t nsg,
-        int32_t nwg);
+        int32_t nwg,
+        bool    use_kv_f16,
+        int32_t ns10,
+        int32_t ns20);
 
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
         ggml_metal_library_t lib,

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

@@ -345,6 +345,18 @@ typedef struct {
     bool     inplace;
 } ggml_metal_kargs_rope;
 
+typedef struct {
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  nblocks;
+} ggml_metal_kargs_flash_attn_ext_kv_f16;
+
 typedef struct {
     int32_t  ne11;
     int32_t  ne_12_2; // assume K and V are same shape

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

@@ -2801,6 +2801,51 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const ggml_tensor * op) {
     return (ne01 < 20) && (ne00 % 32 == 0);
 }
 
+// ref: https://github.com/ggml-org/llama.cpp/pull/27390
+// dequantize the quantized KV cache to F16 before running the F16 flash attention kernels
+static bool ggml_metal_op_flash_attn_ext_use_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // depending on compute/bandwidth ratio, dequant to f16 kv is not always beneficial
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27390#issuecomment-5355152767
+    // TODO: tune per device
+    if (op->src[0]->ne[1] < 32) {
+        return false;
+    }
+
+    switch (op->src[1]->type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+// in some models (e.g. MLA-based), V is a view of K (the first ne20 elements of each K row);
+// the dequantized V is then a view of the dequantized K and does not need its own dequant or scratch
+// - ref: https://github.com/ggml-org/llama.cpp/pull/13435
+static bool ggml_metal_op_flash_attn_ext_v_is_view_of_k(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    const ggml_tensor * K = op->src[1];
+    const ggml_tensor * V = op->src[2];
+
+    return V->view_src && (V->view_src == K || (V->view_src == K->view_src && V->view_offs == K->view_offs));
+}
+
+// size of the F16 dequantized K tensor; the dequantized V tensor follows it in the same scratch buffer
+static size_t ggml_metal_op_flash_attn_ext_kv_f16_k_size(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+
+    return GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne10*ne11*ne12*ne13, 16);
+}
+
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
     assert(op->op == GGML_OP_FLASH_ATTN_EXT);
 
@@ -2816,6 +2861,18 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
     size_t res = 0;
 
     const bool has_mask = op->src[3] != nullptr;
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    // when the KV is dequantized to F16, the pad kernel copies the tail chunk from the F16 scratch buffer
+    // note: when V is a view of K, the dequantized V is read from the dequantized K with K's row stride
+    const bool v_is_view_of_k = use_kv_f16 && ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    uint64_t nb11_pad = nb11;
+    uint64_t nb21_pad = nb21;
+
+    if (use_kv_f16) {
+        nb11_pad = sizeof(ggml_fp16_t)*ne10;
+        nb21_pad = sizeof(ggml_fp16_t)*(v_is_view_of_k ? ne10 : ne20);
+    }
 
     // note: the non-vec kernel requires more extra memory, so always reserve for it
     GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG);
@@ -2828,8 +2885,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_VEC_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                 (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
         }
     } else {
@@ -2838,8 +2895,8 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) {
 
         if (has_kvpad) {
             res += OP_FLASH_ATTN_EXT_NCPSG*(
-                nb11*ne12*ne13 +
-                nb21*ne22*ne23 +
+                nb11_pad*ne12*ne13 +
+                nb21_pad*ne22*ne23 +
                 (has_mask ? ggml_type_size(GGML_TYPE_F16)*ne31*ne32*ne33 : 0));
         }
     }
@@ -2915,6 +2972,29 @@ size_t ggml_metal_op_flash_attn_ext_extra_tmp(const ggml_tensor * op) {
     return res;
 }
 
+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    // note: always reserve the temp buffer to avoid graph reallocations
+    //if (!ggml_metal_op_flash_attn_ext_use_kv_f16(op)) {
+    //    return 0;
+    //}
+
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+
+    const size_t k_size = ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+    // when V is a view of K, the dequantized V is a view of the dequantized K
+    const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+    if (v_is_view_of_k) {
+        return k_size;
+    }
+
+    const size_t v_size = GGML_PAD(sizeof(ggml_fp16_t)*(size_t) ne20*ne21*ne22*ne23, 16);
+
+    return k_size + v_size;
+}
+
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -2989,6 +3069,111 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
     ggml_metal_buffer_id bid_tmp = bid_blk;
     bid_tmp.offs += ggml_metal_op_flash_attn_ext_extra_blk(op);
 
+    ggml_metal_buffer_id bid_kv_f16 = bid_tmp;
+    bid_kv_f16.offs += ggml_metal_op_flash_attn_ext_extra_tmp(op);
+
+    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    ggml_metal_buffer_id bid_k = bid_src1;
+    ggml_metal_buffer_id bid_v = bid_src2;
+
+    uint64_t nb10_attn = nb10;
+    uint64_t nb11_attn = nb11;
+    uint64_t nb12_attn = nb12;
+    uint64_t nb13_attn = nb13;
+    uint64_t nb20_attn = nb20;
+    uint64_t nb21_attn = nb21;
+    uint64_t nb22_attn = nb22;
+    uint64_t nb23_attn = nb23;
+
+    if (use_kv_f16) {
+        assert(ggml_metal_op_flash_attn_ext_extra_kv_f16(op) != 0);
+
+        const bool v_is_view_of_k = ggml_metal_op_flash_attn_ext_v_is_view_of_k(op);
+
+        const int64_t nblocks1_64 = (ne10/ggml_blck_size(op->src[1]->type))*(int64_t) ne11*ne12*ne13;
+        GGML_ASSERT(nblocks1_64 <= INT32_MAX);
+        const int32_t nblocks1 = nblocks1_64;
+
+        ggml_metal_buffer_id bid_v_f16 = bid_kv_f16;
+        bid_v_f16.offs += ggml_metal_op_flash_attn_ext_kv_f16_k_size(op);
+
+        auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_kv_f16(lib, op);
+        const int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline0), 256);
+
+        // K
+        ggml_metal_kargs_flash_attn_ext_kv_f16 args_k = {
+            /*.ne0    =*/ ne10,
+            /*.ne1    =*/ ne11,
+            /*.ne2    =*/ ne12,
+            /*.ne3    =*/ ne13,
+            /*.nb0    =*/ nb10,
+            /*.nb1    =*/ nb11,
+            /*.nb2    =*/ nb12,
+            /*.nb3    =*/ nb13,
+            /*.nblocks =*/ nblocks1,
+        };
+
+        ggml_metal_encoder_set_pipeline(enc, pipeline0);
+        ggml_metal_encoder_set_bytes   (enc, &args_k, sizeof(args_k), 0);
+        ggml_metal_encoder_set_buffer  (enc, bid_src1,        1);
+        ggml_metal_encoder_set_buffer  (enc, bid_kv_f16, 2);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks1 + nth - 1)/nth, 1, 1, nth, 1, 1);
+
+        // V (skip when V is a view of K: the dequantized V is a view of the dequantized K)
+        if (!v_is_view_of_k) {
+            const int64_t nblocks2_64 = (ne20/ggml_blck_size(op->src[2]->type))*(int64_t) ne21*ne22*ne23;
+            GGML_ASSERT(nblocks2_64 <= INT32_MAX);
+            const int32_t nblocks2 = nblocks2_64;
+
+            ggml_metal_kargs_flash_attn_ext_kv_f16 args_v = {
+                /*.ne0    =*/ ne20,
+                /*.ne1    =*/ ne21,
+                /*.ne2    =*/ ne22,
+                /*.ne3    =*/ ne23,
+                /*.nb0    =*/ nb20,
+                /*.nb1    =*/ nb21,
+                /*.nb2    =*/ nb22,
+                /*.nb3    =*/ nb23,
+                /*.nblocks =*/ nblocks2,
+            };
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args_v, sizeof(args_v), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src2,        1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v_f16,       2);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, (nblocks2 + nth - 1)/nth, 1, 1, nth, 1, 1);
+        }
+
+        // the pad and attention kernels read the dequantized KV
+        ggml_metal_op_concurrency_reset(ctx);
+
+        bid_k = bid_kv_f16;
+        bid_v = v_is_view_of_k ? bid_k : bid_v_f16;
+
+        // contiguous F16 layout of the dequantized K
+        nb10_attn = sizeof(ggml_fp16_t);
+        nb11_attn = nb10_attn*ne10;
+        nb12_attn = nb11_attn*ne11;
+        nb13_attn = nb12_attn*ne12;
+
+        // if V is a view of K, the dequantized V is read from the dequantized K with K's strides
+        if (v_is_view_of_k) {
+            nb20_attn = nb10_attn;
+            nb21_attn = nb11_attn;
+            nb22_attn = nb12_attn;
+            nb23_attn = nb13_attn;
+        } else {
+            // contiguous F16 layout of the dequantized V
+            nb20_attn = sizeof(ggml_fp16_t);
+            nb21_attn = nb20_attn*ne20;
+            nb22_attn = nb21_attn*ne21;
+            nb23_attn = nb22_attn*ne22;
+        }
+    }
+
     if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
         // half8x8 kernel
         const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
@@ -3009,12 +3194,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                 /*.ne11    =*/ne11,
                 /*.ne_12_2 =*/ne12,
                 /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                 /*.ne31    =*/ne31,
                 /*.ne32    =*/ne32,
                 /*.ne33    =*/ne33,
@@ -3027,8 +3212,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
             ggml_metal_encoder_set_pipeline(enc, pipeline0);
             ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
             ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
             ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
 
@@ -3073,7 +3258,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             ggml_metal_op_concurrency_reset(ctx);
         }
 
-        const int is_q = ggml_is_quantized(op->src[1]->type) ? 1 : 0;
+        const int is_q = !use_kv_f16 && ggml_is_quantized(op->src[1]->type) ? 1 : 0;
 
         // 2*(2*ncpsg)
         // ncpsg soft_max values + ncpsg mask values
@@ -3104,6 +3289,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
         const size_t smem = FATTN_SMEM(nsg);
 
+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
         ggml_metal_kargs_flash_attn_ext args = {
             /*.ne01          =*/ ne01,
             /*.ne02          =*/ ne02,
@@ -3114,14 +3302,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.ne11          =*/ ne11,
             /*.ne_12_2       =*/ ne12,
             /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
             /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
@@ -3139,13 +3327,13 @@ 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(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, use_kv_f16, ns10, ns20);
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
         ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
         ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
         ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
         ggml_metal_encoder_set_buffer  (enc, bid_pad,  6);
@@ -3177,12 +3365,12 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
                 /*.ne11    =*/ne11,
                 /*.ne_12_2 =*/ne12,
                 /*.ne_12_3 =*/ne13,
-                /*.nb11    =*/nb11,
-                /*.nb12    =*/nb12,
-                /*.nb13    =*/nb13,
-                /*.nb21    =*/nb21,
-                /*.nb22    =*/nb22,
-                /*.nb23    =*/nb23,
+                /*.nb11    =*/nb11_attn,
+                /*.nb12    =*/nb12_attn,
+                /*.nb13    =*/nb13_attn,
+                /*.nb21    =*/nb21_attn,
+                /*.nb22    =*/nb22_attn,
+                /*.nb23    =*/nb23_attn,
                 /*.ne31    =*/ne31,
                 /*.ne32    =*/ne32,
                 /*.ne33    =*/ne33,
@@ -3195,8 +3383,8 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 
             ggml_metal_encoder_set_pipeline(enc, pipeline0);
             ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
-            ggml_metal_encoder_set_buffer  (enc, bid_src1, 1);
-            ggml_metal_encoder_set_buffer  (enc, bid_src2, 2);
+            ggml_metal_encoder_set_buffer  (enc, bid_k,    1);
+            ggml_metal_encoder_set_buffer  (enc, bid_v,    2);
             ggml_metal_encoder_set_buffer  (enc, bid_src3, 3);
             ggml_metal_encoder_set_buffer  (enc, bid_pad,  4);
 
@@ -3242,6 +3430,9 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             }
         }
 
+        const int32_t ns10 = nb11_attn/nb10_attn;
+        const int32_t ns20 = nb21_attn/nb20_attn;
+
         ggml_metal_kargs_flash_attn_ext_vec args = {
             /*.ne01          =*/ ne01,
             /*.ne02          =*/ ne02,
@@ -3252,14 +3443,14 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
             /*.ne11          =*/ ne11,
             /*.ne_12_2       =*/ ne12,
             /*.ne_12_3       =*/ ne13,
-            /*.ns10          =*/ int32_t(nb11/nb10),
-            /*.nb11          =*/ nb11,
-            /*.nb12          =*/ nb12,
-            /*.nb13          =*/ nb13,
-            /*.ns20          =*/ int32_t(nb21/nb20),
-            /*.nb21          =*/ nb21,
-            /*.nb22          =*/ nb22,
-            /*.nb23          =*/ nb23,
+            /*.ns10          =*/ ns10,
+            /*.nb11          =*/ nb11_attn,
+            /*.nb12          =*/ nb12_attn,
+            /*.nb13          =*/ nb13_attn,
+            /*.ns20          =*/ ns20,
+            /*.nb21          =*/ nb21_attn,
+            /*.nb22          =*/ nb22_attn,
+            /*.nb23          =*/ nb23_attn,
             /*.ne31          =*/ ne31,
             /*.ne32          =*/ ne32,
             /*.ne33          =*/ ne33,
@@ -3277,15 +3468,15 @@ 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);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nsg, nwg, use_kv_f16, ns10, ns20);
 
         GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 
         ggml_metal_encoder_set_pipeline(enc, pipeline);
         ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
         ggml_metal_encoder_set_buffer  (enc, bid_src0, 1);
-        ggml_metal_encoder_set_buffer  (enc, bid_src1, 2);
-        ggml_metal_encoder_set_buffer  (enc, bid_src2, 3);
+        ggml_metal_encoder_set_buffer  (enc, bid_k,    2);
+        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
         ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
         ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -42,6 +42,7 @@ bool ggml_metal_op_flash_attn_ext_use_vec(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const struct ggml_tensor * op);
 
 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_repeat            (ggml_metal_op_t ctx, int idx);

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

@@ -225,6 +225,7 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                 res += ggml_metal_op_flash_attn_ext_extra_pad(tensor);
                 res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                 res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_kv_f16(tensor);
             } break;
         case GGML_OP_CUMSUM:
         case GGML_OP_ARGSORT:

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

@@ -6318,6 +6318,53 @@ template [[host_name("kernel_fwht_f32_128")]] kernel kernel_fwht_t kernel_fwht_f
 template [[host_name("kernel_fwht_f32_256")]] kernel kernel_fwht_t kernel_fwht_f32<256>;
 template [[host_name("kernel_fwht_f32_512")]] kernel kernel_fwht_t kernel_fwht_f32<512>;
 
+// dequantize a quantized KV cache tensor to contiguous F16 before running the F16 flash attention kernels
+// - one thread per block; dispatched separately for K and V
+// - ref: https://github.com/ggml-org/llama.cpp/pull/27390
+template <
+    typename block_t,
+    short QK,
+    void (*deq_t4x4)(device const block_t *, short, thread float4x4 &)>
+kernel void kernel_flash_attn_ext_kv_f16(
+        constant ggml_metal_kargs_flash_attn_ext_kv_f16 & args,
+        device const char * x,
+        device       half * x_dst,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= (uint) args.nblocks) {
+        return;
+    }
+
+    const uint nb = args.ne0/QK;
+    const uint i0 = gid%nb;
+    uint ib       = gid/nb;
+    const uint i1 = ib%args.ne1;
+    ib /= args.ne1;
+    const uint i2 = ib%args.ne2;
+    const uint i3 = ib/args.ne2;
+
+    const uint64_t offs = i0*args.nb0 + i1*args.nb1 + i2*args.nb2 + i3*args.nb3;
+
+    device const block_t * src = (device const block_t *) (x + offs);
+    device half4 * dst = (device half4 *) x_dst + (QK/4)*gid;
+
+    for (short i = 0; i < QK/16; ++i) {
+        float4x4 reg;
+        deq_t4x4(src, i, reg);
+        dst[4*i + 0] = (half4) reg[0];
+        dst[4*i + 1] = (half4) reg[1];
+        dst[4*i + 2] = (half4) reg[2];
+        dst[4*i + 3] = (half4) reg[3];
+    }
+}
+
+typedef decltype(kernel_flash_attn_ext_kv_f16<block_q8_0, 32, dequantize_q8_0>) kernel_flash_attn_ext_kv_f16_t;
+
+template [[host_name("kernel_flash_attn_ext_kv_q4_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q4_0, 32, dequantize_q4_0>;
+template [[host_name("kernel_flash_attn_ext_kv_q4_1_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q4_1, 32, dequantize_q4_1>;
+template [[host_name("kernel_flash_attn_ext_kv_q5_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q5_0, 32, dequantize_q5_0>;
+template [[host_name("kernel_flash_attn_ext_kv_q5_1_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q5_1, 32, dequantize_q5_1>;
+template [[host_name("kernel_flash_attn_ext_kv_q8_0_f16")]] kernel kernel_flash_attn_ext_kv_f16_t kernel_flash_attn_ext_kv_f16<block_q8_0, 32, dequantize_q8_0>;
+
 constant bool FC_flash_attn_ext_pad_has_mask [[function_constant(FC_FLASH_ATTN_EXT_PAD + 0)]];
 
 constant int32_t FC_flash_attn_ext_pad_ncpsg [[function_constant(FC_FLASH_ATTN_EXT_PAD + 25)]];
@@ -10318,9 +10365,12 @@ kernel void kernel_mul_mm(
     auto tB = tensor(ptrB, dextents<int32_t, 2>(K, N), array<int, 2>({1, strideB}));
 
     // Configure matmul operation
+    // note: K is dynamic_extent (clamped to the valid range in PHASE 2), since a static
+    //       N_MM_NK_TOTAL K tile would read src1 out of bounds when K % N_MM_NK_TOTAL != 0
+    // ref: https://github.com/ggml-org/llama.cpp/pull/27064
     mpp::tensor_ops::matmul2d<
         mpp::tensor_ops::matmul2d_descriptor(
-            NRB, NRA, N_MM_NK_TOTAL, false, true, true,
+            NRB, NRA, static_cast<int>(dynamic_extent), false, true, true,
             mpp::tensor_ops::matmul2d_descriptor::mode::multiply_accumulate),
         execution_simdgroups<N_MM_SIMD_GROUP_X * N_MM_SIMD_GROUP_Y>> mm;
 
@@ -10372,10 +10422,14 @@ kernel void kernel_mul_mm(
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // === PHASE 2: Tensor matmul ===
-        auto mA = tA.slice(0, 0);
-        auto mB = tB.slice(loop_k, rb);
+        // Clamp the K extent of both operand tensors to the remaining valid K range so
+        // the dynamic-K op never reads past the K extent of src1 (or the staged A tile).
+        const int kExt = min(N_MM_NK_TOTAL, K - loop_k);
 
-        mm.run(mB, mA, cT);
+        auto tAv = tensor(sa, dextents<int32_t, 2>(kExt, NRA), array<int, 2>({1, N_MM_NK_TOTAL}));
+        auto tBv = tensor(ptrB + loop_k + rb * strideB, dextents<int32_t, 2>(kExt, N - rb), array<int, 2>({1, strideB}));
+
+        mm.run(tBv, tAv, cT);
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }

ggml/src/ggml-opencl/CMakeLists.txt

@@ -63,6 +63,7 @@ endfunction()
 set(GGML_OPENCL_KERNELS
     add
     add_id
+    moe_add_id_glu
     argsort
     tri
     fill
@@ -202,6 +203,7 @@ set(GGML_OPENCL_KERNELS
     sqr
     sqrt
     ssm_conv
+    ssm_scan
     gated_delta_net
     sub
     sum_rows

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -577,11 +577,19 @@ struct ggml_backend_opencl_context {
     // whether fuse moe combine
     cl_uint fuse_moe_combine;
 
+    // whether to fold the MoE bias adds into swiglu_oai
+    cl_uint fuse_moe_bias_glu;
+
+    // whether to fold the MoE down-projection bias add into the combine
+    cl_uint fuse_moe_bias_combine;
+
     bool adreno_has_large_buffer;
     bool adreno_use_large_buffer;
     bool adreno_use_bin_kernels;
     get_adreno_bin_kernel_func_t get_adreno_bin_kernel_func = nullptr;
     ggml_cl_compiler_version adreno_cl_compiler_version;
+    // The q6_K flat mul_mat codegen workarounds are needed by old E031 compilers only.
+    bool q6_k_flat_old_compiler;
 
     std::string kernel_compile_opts;  // cached for lazy-compiled kernels.
 
@@ -656,6 +664,7 @@ struct ggml_backend_opencl_context {
 
     cl_program program_add;
     cl_program program_add_id;
+    cl_program program_moe_add_id_glu;
     cl_program program_clamp;
     cl_program program_cvt;
     cl_program program_diag_mask_inf;
@@ -721,6 +730,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
     cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
     cl_kernel kernel_add_id;
+    cl_kernel kernel_add_id_add_id_swiglu_oai;
     cl_kernel kernel_scale_f32, kernel_scale_f32_4;
     cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
     cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
@@ -866,6 +876,9 @@ struct ggml_backend_opencl_context {
     // [size_idx][kda][tgpp] where size_idx: 0=S_V=16, 1=32, 2=64, 3=128; kda: 0 or 1.
     // tgpp 0 = TG variant (COLS_PER_LANE_GROUP=1), tgpp 1 = prefill variant (COLS_PER_LANE_GROUP=4).
     cl_kernel kernel_gated_delta_net_f32[4][2][2] = {};
+    cl_kernel kernel_ssm_scan_f32_mamba2_d128 = nullptr;
+    cl_kernel kernel_ssm_scan_f32_mamba2_d256 = nullptr;
+
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_gemv_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns, kernel_gemm_moe_q4_0_f32_ns_bin;
     cl_kernel kernel_gemm_moe_q8_0_f32_ns;
@@ -892,7 +905,9 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_moe_q4_0_q8_1_dp4a = nullptr;    // dp4a (int8) q4_0 MoE prefill GEMM
     cl_kernel kernel_moe_reorder_b;
     cl_kernel kernel_moe_histogram, kernel_moe_scan, kernel_moe_fill, kernel_moe_scatter;
+    cl_kernel kernel_moe_scatter_stable = nullptr;   // deterministic slot assignment
     cl_kernel kernel_moe_combine_f32 = nullptr;   // fused router-weight mul + cross-expert sum
+    cl_kernel kernel_moe_combine_bias_f32 = nullptr;  // same, with the down-projection bias add folded in
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
     cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
     cl_kernel kernel_mul_mv_id_mxfp4_f32;
@@ -1340,6 +1355,23 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         GGML_LOG_CONT(".");
     }
 
+    // moe_add_id_glu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "moe_add_id_glu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("moe_add_id_glu.cl");
+#endif
+        backend_ctx->program_moe_add_id_glu =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add_id_add_id_swiglu_oai =
+            clCreateKernel(backend_ctx->program_moe_add_id_glu, "kernel_add_id_add_id_swiglu_oai", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // tri
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1927,8 +1959,14 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
 #else
         const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
 #endif
+        // The codegen workarounds in this kernel are a measured 13-20% loss on
+        // compilers that do not need them, so only the affected ones build them;
+        // everyone else gets the original source.
+        const std::string q6k_opts = backend_ctx->q6_k_flat_old_compiler
+            ? compile_opts + " -DADRENO_OLD_COMPILER=1"
+            : compile_opts;
         cl_program prog =
-            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+            build_program_from_source(backend_ctx, kernel_src.c_str(), q6k_opts);
 
         CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
         CL_CHECK(clReleaseProgram(prog));
@@ -3154,6 +3192,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         GGML_LOG_CONT(".");
     }
 
+    // ssm_scan (Mamba-2 fused per-token recurrent step; d_state in {128, 256})
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_scan.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_scan.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d128 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d128", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_scan_f32_mamba2_d256 = clCreateKernel(prog, "kernel_ssm_scan_f32_mamba2_d256", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
     // gated_delta_net: one kernel per (S_V, KDA, tgpp) triple.
     {
     #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3246,6 +3302,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
             backend_ctx, kernel_src.c_str(), compile_opts);
         CL_CHECK((backend_ctx->kernel_moe_combine_f32 =
                     clCreateKernel(prog, "kernel_moe_combine_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_moe_combine_bias_f32 =
+                    clCreateKernel(prog, "kernel_moe_combine_bias_f32", &err), err));
         CL_CHECK(clReleaseProgram(prog));
         GGML_LOG_CONT(".");
     }
@@ -4442,6 +4500,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx) {
         CL_CHECK((backend_ctx->kernel_moe_scan = clCreateKernel(prog, "kernel_moe_scan", &err), err));
         CL_CHECK((backend_ctx->kernel_moe_fill = clCreateKernel(prog, "kernel_moe_fill", &err), err));
         CL_CHECK((backend_ctx->kernel_moe_scatter = clCreateKernel(prog, "kernel_moe_scatter", &err), err));
+        CL_CHECK((backend_ctx->kernel_moe_scatter_stable = clCreateKernel(prog, "kernel_moe_scatter_stable", &err), err));
         CL_CHECK(clReleaseProgram(prog));
         GGML_LOG_CONT(".");
     }
@@ -5894,6 +5953,16 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
         (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
         (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);
 
+    // The q6_K flat mul_mat miscompile is a defect of the older E031 compilers, not a
+    // property of any GPU generation: it reproduces on E031.38 (Adreno 642L) and E031.41
+    // (Adreno 740) and is fixed by E031.45 (Adreno 619). Gate on the compiler so parts
+    // that do not need the workarounds do not pay for them. The explicit type check is
+    // required: newer_than_or_same() is false for every non-E031 compiler, so negating it
+    // alone would enable the workarounds on E17/DX.
+    backend_ctx->q6_k_flat_old_compiler =
+        backend_ctx->adreno_cl_compiler_version.type == E031 &&
+        !backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 45, 0, 0);
+
     size_t ext_str_size;
     clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
     char *ext_buffer = (char *)alloca(ext_str_size + 1);
@@ -5971,6 +6040,12 @@ static ggml_backend_opencl_context * ggml_cl_init(ggml_backend_dev_t dev) {
     backend_ctx->adreno_moe_ragged_skip_gran = (ragged_gran_env != NULL) ? atoi(ragged_gran_env) : 8;
 
     // whether fuse moe combine
+    static const char * fuse_moe_bias_glu_env = getenv("GGML_OPENCL_FUSE_MOE_BIAS_GLU");
+    backend_ctx->fuse_moe_bias_glu = fuse_moe_bias_glu_env == NULL ? 1 : (atoi(fuse_moe_bias_glu_env) != 0);
+
+    static const char * fuse_moe_bias_combine_env = getenv("GGML_OPENCL_FUSE_MOE_BIAS_COMBINE");
+    backend_ctx->fuse_moe_bias_combine = fuse_moe_bias_combine_env == NULL ? 1 : (atoi(fuse_moe_bias_combine_env) != 0);
+
     static const char * fuse_moe_combine_env = getenv("GGML_OPENCL_FUSE_MOE_COMBINE");
     backend_ctx->fuse_moe_combine = fuse_moe_combine_env == NULL ? 1 : (atoi(fuse_moe_combine_env) != 0);
 
@@ -6839,6 +6914,300 @@ static bool ggml_opencl_can_fuse_moe_combine(const struct ggml_cgraph * cgraph,
     return true;
 }
 
+// Detect the gpt-oss MoE bias+activation epilogue on the PREFILL path:
+// {MUL_MAT_ID(gate), ADD_ID(gate_bias), MUL_MAT_ID(up), ADD_ID(up_bias), GLU(swiglu_oai)}.
+// The two matmuls still run as their own dispatches (the prefill GEMM is the vendor's);
+// what collapses is the epilogue — both add_id passes are in-place read-modify-writes of a
+// tensor the GLU immediately reads again, so they are three full passes over the same
+// [n_ff, n_expert_used, n_tokens] f32 tensor where one suffices.
+//
+// The decode counterpart is handled by the mxfp4 fused GEMV arm in ggml_opencl_can_fuse,
+// which folds the matmul too; this one deliberately fires only when that cannot (ne[2] > 1).
+static bool ggml_opencl_can_fuse_moe_bias_glu(const struct ggml_cgraph * cgraph, int node_idx) {
+    if (node_idx + 4 >= cgraph->n_nodes) {
+        return false;
+    }
+
+    const enum ggml_op mg_ops[] = { GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_MUL_MAT_ID, GGML_OP_ADD_ID, GGML_OP_GLU };
+    const int          mg_out[] = { node_idx + 4 };
+    if (!ggml_can_fuse_subgraph(cgraph, node_idx, 5, mg_ops, mg_out, 1)) {
+        return false;
+    }
+
+    const ggml_tensor * gmm = cgraph->nodes[node_idx];
+    const ggml_tensor * gad = cgraph->nodes[node_idx+1];
+    const ggml_tensor * umm = cgraph->nodes[node_idx+2];
+    const ggml_tensor * uad = cgraph->nodes[node_idx+3];
+    const ggml_tensor * glu = cgraph->nodes[node_idx+4];
+
+    if (ggml_get_glu_op(glu) != GGML_GLU_OP_SWIGLU_OAI) {
+        return false;
+    }
+    // Prefill only — at one token the mxfp4 arm above folds the matmul as well.
+    if (gmm->src[1]->ne[2] == 1) {
+        return false;
+    }
+    // Wiring: both matmuls share the activation and the expert selection, each add_id
+    // biases its own matmul, and the GLU consumes the two biased results as separate
+    // operands (so the same-buffer ne00_off/ne10_off split path is not in play).
+    if (gad->src[0] != gmm || uad->src[0] != umm ||
+        glu->src[0] != gad || glu->src[1] != uad ||
+        umm->src[1] != gmm->src[1] || umm->src[2] != gmm->src[2]) {
+        return false;
+    }
+    // A swapped GLU would exchange the gate/up roles the fused kernel hard-codes.
+    if (ggml_get_op_params_i32(glu, 1)) {
+        return false;
+    }
+    if (gad->type != GGML_TYPE_F32 || uad->type != GGML_TYPE_F32 || glu->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!gad->src[1] || gad->src[1]->type != GGML_TYPE_F32 ||
+        !uad->src[1] || uad->src[1]->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!gad->src[2] || gad->src[2]->type != GGML_TYPE_I32 || uad->src[2] != gad->src[2]) {
+        return false;
+    }
+    // Full width on both operands: the kernel writes one output element per input pair.
+    if (!ggml_are_same_shape(gad, uad) || glu->ne[0] != gad->ne[0] ||
+        glu->ne[1] != gad->ne[1] || glu->ne[2] != gad->ne[2] || glu->ne[3] != gad->ne[3]) {
+        return false;
+    }
+    if (gad->ne[3] != 1) {
+        return false;
+    }
+    // The destination is addressed by (expert slot, token) rather than the GLU's flat row
+    // walk; those agree only for a contiguous destination.
+    if (!ggml_is_contiguous(glu) || !ggml_is_contiguous(gmm) || !ggml_is_contiguous(umm)) {
+        return false;
+    }
+    return true;
+}
+
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+
+// Runs the gate and up matmuls unchanged, then one kernel in place of
+// add_id(gate) + add_id(up) + swiglu_oai. See ggml_opencl_can_fuse_moe_bias_glu.
+static void ggml_cl_moe_bias_glu_fused(ggml_backend_t backend, ggml_tensor * gate_mm, const ggml_tensor * gate_add,
+                                       ggml_tensor * up_mm, const ggml_tensor * up_add, const ggml_tensor * glu) {
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_cl_mul_mat_id(backend, gate_mm->src[0], gate_mm->src[1], gate_mm);
+    ggml_cl_mul_mat_id(backend, up_mm->src[0], up_mm->src[1], up_mm);
+
+    const ggml_tensor * gbias = gate_add->src[1];
+    const ggml_tensor * ubias = up_add->src[1];
+    const ggml_tensor * ids   = gate_add->src[2];
+
+    ggml_tensor_extra_cl * eg  = (ggml_tensor_extra_cl *)gate_mm->extra;
+    ggml_tensor_extra_cl * egb = (ggml_tensor_extra_cl *)gbias->extra;
+    ggml_tensor_extra_cl * eu  = (ggml_tensor_extra_cl *)up_mm->extra;
+    ggml_tensor_extra_cl * eub = (ggml_tensor_extra_cl *)ubias->extra;
+    ggml_tensor_extra_cl * ei  = (ggml_tensor_extra_cl *)ids->extra;
+    ggml_tensor_extra_cl * ed  = (ggml_tensor_extra_cl *)glu->extra;
+
+    cl_ulong off_g  = eg->offset  + gate_mm->view_offs;
+    cl_ulong off_gb = egb->offset + gbias->view_offs;
+    cl_ulong off_u  = eu->offset  + up_mm->view_offs;
+    cl_ulong off_ub = eub->offset + ubias->view_offs;
+    cl_ulong off_i  = ei->offset  + ids->view_offs;
+    cl_ulong off_d  = ed->offset  + glu->view_offs;
+
+    const cl_ulong nb01_g = gate_mm->nb[1];
+    const cl_ulong nb02_g = gate_mm->nb[2];
+    const cl_ulong nb01_u = up_mm->nb[1];
+    const cl_ulong nb02_u = up_mm->nb[2];
+    const cl_ulong nb11_g = gbias->nb[1];
+    const cl_ulong nb11_u = ubias->nb[1];
+    const cl_ulong nb21   = ids->nb[1];
+    const cl_ulong nbd1   = glu->nb[1];
+    const cl_ulong nbd2   = glu->nb[2];
+
+    const int   ne0   = (int)glu->ne[0];
+    const float alpha = ggml_get_op_params_f32(glu, 2);
+    const float limit = ggml_get_op_params_f32(glu, 3);
+
+    cl_kernel kernel = backend_ctx->kernel_add_id_add_id_swiglu_oai;
+
+    int i = 0;
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eg->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &egb->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_gb));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eu->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &eub->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_ub));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &ei->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_i));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &off_d));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb01_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb02_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb01_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb02_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb11_g));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb11_u));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nbd1));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(cl_ulong), &nbd2));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(float),    &limit));
+    CL_CHECK(clSetKernelArg(kernel, i++, sizeof(float),    &alpha));
+
+    const int nth = MIN(ne0, (int) backend_ctx->get_kernel_workgroup_size(kernel));
+    size_t global_work_size[] = { (size_t)glu->ne[1]*nth, (size_t)glu->ne[2], 1 };
+    size_t local_work_size[]  = { (size_t)nth, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, (ggml_tensor *)glu);
+}
+
+// Fusion B: the MoE down-projection bias add feeding the combine.
+//
+// The graph runs ADD_ID(down_bias) and then immediately the combine subgraph
+// {MUL(router weights), k VIEWs, k-1 ADDs}, and the ADD_ID's only consumer is that
+// MUL. Since the ADD_ID is an in-place read-modify-write of a tensor the combine
+// reads once more, the bias can be added inside the combine instead, dropping a
+// full pass over [n_embd, k, n_tokens].
+//
+// Shape checks for the combine tail are delegated to ggml_opencl_can_fuse_moe_combine
+// (which also owns the n_nodes >= 32 bail and the experts/dst aliasing bail); what is
+// added here is the ADD_ID wiring plus a subgraph check over the WHOLE run, so that
+// the intermediate bias result is confirmed not to escape.
+static bool ggml_opencl_can_fuse_moe_bias_combine(const struct ggml_cgraph * cgraph, int node_idx,
+                                                  const ggml_tensor ** out_final_add) {
+    if (node_idx + 1 >= cgraph->n_nodes) {
+        return false;
+    }
+    const ggml_tensor * add = cgraph->nodes[node_idx];
+    if (add->op != GGML_OP_ADD_ID) {
+        return false;
+    }
+    const ggml_tensor * mul = cgraph->nodes[node_idx+1];
+    if (mul->op != GGML_OP_MUL || mul->src[0] != add) {
+        return false;
+    }
+
+    const ggml_tensor * final_add = NULL;
+    if (!ggml_opencl_can_fuse_moe_combine(cgraph, node_idx+1, &final_add)) {
+        return false;
+    }
+
+    const ggml_tensor * raw  = add->src[0];
+    const ggml_tensor * bias = add->src[1];
+    const ggml_tensor * ids  = add->src[2];
+    if (!raw || !bias || !ids) {
+        return false;
+    }
+    if (raw->type != GGML_TYPE_F32 || bias->type != GGML_TYPE_F32 ||
+        ids->type != GGML_TYPE_I32 || add->type != GGML_TYPE_F32) {
+        return false;
+    }
+    // The combine reads the raw matmul output with the strides it computed from the
+    // add_id result, so the two must have the same layout.
+    if (!ggml_are_same_shape(raw, add) || !ggml_is_contiguous(raw)) {
+        return false;
+    }
+    if (raw->nb[1] != add->nb[1] || raw->nb[2] != add->nb[2]) {
+        return false;
+    }
+    // ids is indexed as [expert slot, token]; the combine walks the same two axes.
+    if (ids->ne[0] < add->ne[1] || ids->ne[1] < add->ne[2]) {
+        return false;
+    }
+
+    // Whole-run escape check: ADD_ID + MUL + k VIEWs + (k-1) ADDs, only the last node escapes.
+    const int k       = (int)add->ne[1];
+    const int n_nodes = 2 + k + (k - 1);
+    if (n_nodes >= 32 || node_idx + n_nodes > cgraph->n_nodes) {
+        return false;
+    }
+    enum ggml_op ops[32];
+    int n = 0;
+    ops[n++] = GGML_OP_ADD_ID;
+    ops[n++] = GGML_OP_MUL;
+    for (int j = 0; j < k;     ++j) ops[n++] = GGML_OP_VIEW;
+    for (int j = 0; j < k - 1; ++j) ops[n++] = GGML_OP_ADD;
+    const int outs[] = { node_idx + n_nodes - 1 };
+    if (!ggml_can_fuse_subgraph(cgraph, node_idx, n_nodes, ops, outs, 1)) {
+        return false;
+    }
+
+    *out_final_add = final_add;
+    return true;
+}
+
+
+// Fusion B dispatch: the combine, reading the RAW matmul output and adding the
+// per-expert bias row inline. See ggml_opencl_can_fuse_moe_bias_combine.
+static void ggml_cl_moe_bias_combine_fused(ggml_backend_t backend, const ggml_tensor * add,
+                                           const ggml_tensor * mul, const ggml_tensor * dst) {
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const ggml_tensor * experts = add->src[0];   // raw matmul output, bias not yet applied
+    const ggml_tensor * bias    = add->src[1];
+    const ggml_tensor * ids     = add->src[2];
+    const ggml_tensor * weights = mul->src[1];
+
+    ggml_tensor_extra_cl * ee = (ggml_tensor_extra_cl *)experts->extra;
+    ggml_tensor_extra_cl * eb = (ggml_tensor_extra_cl *)bias->extra;
+    ggml_tensor_extra_cl * ei = (ggml_tensor_extra_cl *)ids->extra;
+    ggml_tensor_extra_cl * ew = (ggml_tensor_extra_cl *)weights->extra;
+    ggml_tensor_extra_cl * ed = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong off_e = ee->offset + experts->view_offs;
+    cl_ulong off_b = eb->offset + bias->view_offs;
+    cl_ulong off_i = ei->offset + ids->view_offs;
+    cl_ulong off_w = ew->offset + weights->view_offs;
+    cl_ulong off_d = ed->offset + dst->view_offs;
+
+    const int n_embd4 = (int)(experts->ne[0] / 4);
+    const int k       = (int)experts->ne[1];
+    const int nt      = (int)experts->ne[2];
+    const cl_uint e1 = (cl_uint)(experts->nb[1] / sizeof(float));
+    const cl_uint e2 = (cl_uint)(experts->nb[2] / sizeof(float));
+    const cl_uint w1 = (cl_uint)(weights->nb[1] / sizeof(float));
+    const cl_uint w2 = (cl_uint)(weights->nb[2] / sizeof(float));
+    const cl_uint d1 = (cl_uint)(dst->nb[1] / sizeof(float));
+    const cl_ulong nb_b1 = bias->nb[1];
+    const cl_ulong nb_i1 = ids->nb[1];
+
+    const size_t w_bytes = ggml_nbytes(weights);
+    backend_ctx->prealloc_moe_combine_w.allocate(backend_ctx->context, w_bytes);
+    CL_CHECK(clEnqueueCopyBuffer(backend_ctx->queue, ew->data_device, backend_ctx->prealloc_moe_combine_w.buffer,
+                                 off_w, 0, w_bytes, 0, NULL, NULL));
+    cl_mem   w_dev = backend_ctx->prealloc_moe_combine_w.buffer;
+    cl_ulong w_off = 0;
+
+    cl_kernel kernel = backend_ctx->kernel_moe_combine_bias_f32;
+    int a = 0;
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ee->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_e));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &w_dev));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &w_off));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &eb->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_b));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ei->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_i));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &off_d));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &n_embd4));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &k));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(int),      &nt));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &e1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &e2));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &w1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &w2));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_uint),  &d1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &nb_b1));
+    CL_CHECK(clSetKernelArg(kernel, a++, sizeof(cl_ulong), &nb_i1));
+
+    size_t lws[2] = { 64, 1 };
+    size_t gws[2] = { (size_t)(((n_embd4 + 63) / 64) * 64), (size_t)nt };
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, gws, lws, (ggml_tensor *)dst);
+}
+
+
 static void ggml_cl_moe_combine_fused(ggml_backend_t backend, const ggml_tensor * mul, const ggml_tensor * dst) {
     ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *)backend->context;
     const ggml_tensor * experts = mul->src[0];
@@ -6993,6 +7362,31 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
         }
         // Fuse the MoE combine: router-weight mul + cross-expert add chain ->
         // one weighted-sum-across-experts kernel.
+        // Fold the gpt-oss MoE bias epilogue: add_id(gate_bias) + add_id(up_bias) +
+        // glu(swiglu_oai) -> one kernel, leaving the two matmuls as their own dispatches.
+        // Both add_ids are in-place passes over a tensor the GLU reads again, so this
+        // drops two full read+write passes per layer. Opt out GGML_OPENCL_FUSE_MOE_BIAS_GLU=0.
+        if (backend_ctx->fuse_moe_bias_glu && !backend_ctx->disable_fusion &&
+            ggml_opencl_can_fuse_moe_bias_glu(cgraph, i)) {
+            ggml_cl_moe_bias_glu_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2],
+                                       cgraph->nodes[i+3], cgraph->nodes[i+4]);
+            i += 4;
+            continue;
+        }
+
+        // Fold the MoE down-projection bias into the combine: add_id(down_bias) + the whole
+        // combine subgraph -> one kernel. Checked before the plain combine arm so the longer
+        // pattern wins. Opt out GGML_OPENCL_FUSE_MOE_BIAS_COMBINE=0.
+        if (backend_ctx->fuse_moe_bias_combine && backend_ctx->fuse_moe_combine &&
+            !backend_ctx->disable_fusion) {
+            const ggml_tensor * bias_combine_out = nullptr;
+            if (ggml_opencl_can_fuse_moe_bias_combine(cgraph, i, &bias_combine_out)) {
+                ggml_cl_moe_bias_combine_fused(backend, node, cgraph->nodes[i+1], bias_combine_out);
+                i += 2 * (int)node->ne[1];   // ADD_ID + MUL + k VIEWs + (k-1) ADDs
+                continue;
+            }
+        }
+
         if (backend_ctx->fuse_moe_combine && !backend_ctx->disable_fusion) {
             const ggml_tensor * combine_out = nullptr;
             if (ggml_opencl_can_fuse_moe_combine(cgraph, i, &combine_out)) {
@@ -7301,6 +7695,23 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                    (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
         case GGML_OP_SSM_CONV:
             return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_SCAN: {
+            // Mamba-2 fused per-token scan. Requires src3->ne[0] == 1 (scalar
+            // A per head); d_state in {128, 256}; all sources f32. Falls back
+            // to CPU otherwise (incl. Mamba-1 element-wise A).
+            for (int i = 0; i < 6; ++i) {
+                if (op->src[i]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+            }
+            if (op->type != GGML_TYPE_F32) {
+                return false;
+            }
+            const int K = ggml_get_op_params_i32(op, 0);
+            const int d_state = (int) op->src[0]->ne[0];
+            const bool is_mamba2 = (op->src[3]->ne[0] == 1);
+            return is_mamba2 && (d_state == 128 || d_state == 256) && (K == 1);
+        }
         case GGML_OP_GATED_DELTA_NET:
             {
                 // Match the Vulkan backend: only F32 -> F32, S_v in {16, 32, 64, 128}.
@@ -7335,6 +7746,19 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                        op->src[0]->type == GGML_TYPE_Q4_K  ||
                        op->src[0]->type == GGML_TYPE_Q5_K  ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
+                // The E031.41 compiler (usually with A7x) miscompiles the flat K-quant
+                // GEMV kernels (kernel_mul_mv_q*_K_f32_flat) and makes lm_head run much
+                // slower than it should. So, make it fallback to CPU to preserve performance
+                // for this compiler series.
+                static const char * a7x_lmhead_env = getenv("GGML_OPENCL_A7X_LMHEAD_CPU");
+                static const bool   a7x_lmhead_cpu = (a7x_lmhead_env == nullptr || a7x_lmhead_env[0] != '0');
+                if (a7x_lmhead_cpu &&
+                    backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                    (op->src[0]->type == GGML_TYPE_Q4_K || op->src[0]->type == GGML_TYPE_Q5_K ||
+                     op->src[0]->type == GGML_TYPE_Q6_K) &&
+                    op->src[0]->ne[1] >= 32768) {   // vocab-scale weight; no FFN/attn weight is this tall
+                    return false;
+                }
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
                 return op->src[1]->type == GGML_TYPE_F32;
@@ -7376,9 +7800,6 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
         case GGML_OP_DIAG_MASK_INF:
             return op->ne[3] == 1;
         case GGML_OP_ROPE: {
-            if (((const int32_t *) op->op_params)[15] != 0) {
-                return false; // FIXME: support ggml_rope_set_offset
-            }
             const int mode = ((const int32_t *) op->op_params)[2];
             const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
             const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
@@ -7456,6 +7877,7 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                      v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
             const bool is_f32_f16  = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
                                      v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
             const bool is_f32_q8_0 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_Q8_0 &&
                                      v->type == GGML_TYPE_Q8_0 && op->type == GGML_TYPE_F32 &&
                                      dk % 32 == 0 && dv % 32 == 0;
@@ -7463,6 +7885,21 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                                      v->type == GGML_TYPE_Q4_0 && op->type == GGML_TYPE_F32 &&
                                      dk % 32 == 0 && dv % 32 == 0;
 
+            // A7X (Adreno 740, compiler E031.41) SIGSEGVs inside clBuildProgram
+            // building the flash_attn programs whose KV path is mixed-type or
+            // dequantized — f32_f16, q8_0, q4_0 (reproduced at DK=40 and DK=64; it
+            // is DK-independent). It is a driver crash, not codegen-wrong-output, so
+            // it cannot be caught in-process (fatal=false only handles clean compile
+            // errors). The uniform f16_f16 / f32_f32 programs compile fine on this
+            // compiler, so decline only the KV-convert variants; ggml then runs
+            // those (f16-KV / quant-KV) attention layers on the CPU backend.
+            // Negative compiler carve-out, same idiom as the Intel DK=512 decline
+            // below and the X1E driver-quirk guards.
+            if (backend_ctx && backend_ctx->adreno_gen == ADRENO_GPU_GEN::A7X &&
+                (is_f32_f16 || is_f32_q8_0 || is_f32_q4_0)) {
+                return false;
+            }
+
             // Asymmetric KV: host-dequants both sides to F32, uses f32 kernel.
             auto is_kv_type_ok = [](ggml_type t) {
                 return t == GGML_TYPE_F16 || t == GGML_TYPE_F32 ||
@@ -12260,6 +12697,103 @@ static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 }
 
+static void ggml_cl_ssm_scan(ggml_backend_t backend, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0]; // s
+    const ggml_tensor * src1 = dst->src[1]; // x
+    const ggml_tensor * src2 = dst->src[2]; // dt
+    const ggml_tensor * src3 = dst->src[3]; // A
+    const ggml_tensor * src4 = dst->src[4]; // B
+    const ggml_tensor * src5 = dst->src[5]; // C
+    const ggml_tensor * src6 = dst->src[6]; // ids
+
+    GGML_ASSERT(src0 && src1 && src2 && src3 && src4 && src5 && src6 && dst);
+
+    ggml_backend_opencl_context * backend_ctx = (ggml_backend_opencl_context *) backend->context;
+
+    ggml_tensor_extra_cl * e0 = (ggml_tensor_extra_cl *) src0->extra;
+    ggml_tensor_extra_cl * e1 = (ggml_tensor_extra_cl *) src1->extra;
+    ggml_tensor_extra_cl * e2 = (ggml_tensor_extra_cl *) src2->extra;
+    ggml_tensor_extra_cl * e3 = (ggml_tensor_extra_cl *) src3->extra;
+    ggml_tensor_extra_cl * e4 = (ggml_tensor_extra_cl *) src4->extra;
+    ggml_tensor_extra_cl * e5 = (ggml_tensor_extra_cl *) src5->extra;
+    ggml_tensor_extra_cl * e6 = (ggml_tensor_extra_cl *) src6->extra;
+    ggml_tensor_extra_cl * ed = (ggml_tensor_extra_cl *) dst->extra;
+
+    cl_ulong o0 = e0->offset + src0->view_offs;
+    cl_ulong o1 = e1->offset + src1->view_offs;
+    cl_ulong o2 = e2->offset + src2->view_offs;
+    cl_ulong o3 = e3->offset + src3->view_offs;
+    cl_ulong o4 = e4->offset + src4->view_offs;
+    cl_ulong o5 = e5->offset + src5->view_offs;
+    cl_ulong o6 = e6->offset + src6->view_offs;
+    cl_ulong od = ed->offset + dst->view_offs;
+
+    const int d_state  = (int) src0->ne[0];
+    const int head_dim = (int) src0->ne[1];
+    const int n_head   = (int) src1->ne[1];
+    const int n_group  = (int) src4->ne[1];
+    const int n_tokens = (int) src1->ne[2];
+    const int n_seqs   = (int) src1->ne[3];
+
+    // Mirror CPU ref: s_off = ggml_nelements(src1) * sizeof(float)
+    const cl_ulong s_off_bytes = (cl_ulong) ggml_nelements(src1) * sizeof(float);
+
+    cl_kernel kernel = (d_state == 128)
+        ? backend_ctx->kernel_ssm_scan_f32_mamba2_d128
+        : backend_ctx->kernel_ssm_scan_f32_mamba2_d256;
+    GGML_ASSERT(kernel != nullptr);
+
+    cl_ulong s0_nb2 = src0->nb[2];
+    cl_ulong s0_nb3 = src0->nb[3];
+    cl_ulong x_nb2  = src1->nb[2];
+    cl_ulong x_nb3  = src1->nb[3];
+    cl_ulong dt_nb1 = src2->nb[1];
+    cl_ulong dt_nb2 = src2->nb[2];
+    cl_ulong A_nb1  = src3->nb[1];
+    cl_ulong B_nb2  = src4->nb[2];
+    cl_ulong B_nb3  = src4->nb[3];
+    cl_ulong C_nb2  = src5->nb[2];
+    cl_ulong C_nb3  = src5->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &e0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &o0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &e1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &o1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &e2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &o2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &e3->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &o3));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_mem),   &e4->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &o4));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_mem),   &e5->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &o5));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_mem),   &e6->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &o6));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_mem),   &ed->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &od));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &s0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &s0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &x_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &x_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &dt_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &dt_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &A_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &B_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &B_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong), &C_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &C_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &s_off_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &head_dim));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(int),      &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_group));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &n_tokens));
+
+    size_t global_work_size[] = { (size_t)n_head * head_dim * 64, (size_t)n_seqs, 1 };
+    size_t local_work_size[]  = { 64, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -12693,7 +13227,10 @@ static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const
     GGML_TENSOR_LOCALS(int,      ne0, src0, ne);
     GGML_TENSOR_LOCALS(cl_ulong, nb0, src0, nb);
 
-    const int nth = MIN(64, ne00);
+    int nth = 1;
+    while (nth < ne00 && nth < 64) {
+        nth *= 2;
+    }
 
     cl_kernel kernel = backend_ctx->kernel_norm;
 
@@ -20443,6 +20980,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
             CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+            // The optimizer-barrier arg exists only in the ADRENO_OLD_COMPILER build of
+            // this kernel; conformant compilers get the original 17-arg signature.
+            if (backend_ctx->q6_k_flat_old_compiler) {
+                cl_uchar q6k_mask = 0xFF;   // never 0xFE in prod; see the kernel note
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_uchar), &q6k_mask));
+            }
 #else
             kernel = backend_ctx->kernel_mul_mv_q6_K_f32;
 
@@ -20728,18 +21271,42 @@ static void moe_router_reoerder(ggml_backend_t backend, const ggml_tensor * src,
     size_t fill_local_size[] = {64, 1, 1};
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, fill_global_size, fill_local_size, src);
 
-    // Scatter
-    kernel = backend_ctx->kernel_moe_scatter;
-    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
-    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
-    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
-    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
-    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
-    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
-    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+    // Scatter. The deterministic variant is the default: kernel_moe_scatter derives
+    // each token's slot from an atomic counter, so the packing inside an expert - and
+    // with it the output of the ragged prefill GEMM - changes from run to run. Set
+    // GGML_OPENCL_MOE_STABLE_SCATTER=0 to restore the atomic version.
+    static const bool stable_scatter = []{
+        const char * e = getenv("GGML_OPENCL_MOE_STABLE_SCATTER");
+        return !e || e[0] == '\0' || e[0] != '0';
+    }();
 
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    if (stable_scatter) {
+        kernel = backend_ctx->kernel_moe_scatter_stable;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne02));
+
+        // one workgroup (one wave) per expert; each ranks its own tokens
+        size_t scatter_global_size[] = {64, (size_t)ne02};
+        size_t scatter_local_size[]  = {64, 1};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, scatter_global_size, scatter_local_size, src);
+    } else {
+        kernel = backend_ctx->kernel_moe_scatter;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &original_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &post_router_buf));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &emap_buf));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &tile_offset_buf));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &slot_counter_buf));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int), &ne21));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int), &ne20));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int), &ne02));
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, histogram_global_size, histogram_local_size, src);
+    }
 
     // [MOE_TILES] env-gated padding probe: read back total_tiles (= Sum_e
     // ceil(k_e/n_tile_size)) and compare to the ideal tile count for the real
@@ -23706,6 +24273,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     const int n_dims     = ((int *) dst->op_params)[1];
     const int mode       = ((int *) dst->op_params)[2];
     const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];
 
     float freq_base;
     float freq_scale;
@@ -23734,6 +24302,7 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
 
     if (is_vision) {
         GGML_ASSERT(n_dims == ne00/2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
     }
 
     cl_kernel kernel;
@@ -23825,6 +24394,12 @@ static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const
     if (is_mrope && !is_vision) {
         CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
     }
+    // norm and neox have n_offs after beta_slow, mrope has it after is_imrope
+    if (!is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int), &n_offs));
+    } else if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 35, sizeof(int), &n_offs));
+    }
 
     size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
     size_t local_work_size[] = {(size_t)nth, 1, 1};
@@ -24746,6 +25321,14 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_ssm_conv;
             break;
+        case GGML_OP_SSM_SCAN:
+            if (!any_on_device) {
+                return false;
+            }
+            // SSM_SCAN has 7 source tensors, so it cannot use the standard
+            // (src0, src1, dst) func signature. Dispatch directly and return.
+            ggml_cl_ssm_scan(backend, tensor);
+            return true;
         case GGML_OP_GATED_DELTA_NET:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/moe_add_id_glu.cl

@@ -0,0 +1,76 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add_id(gate) + add_id(up) + swiglu_oai, fused
+//
+// gpt-oss-class MoE FFNs run three full passes over the same
+// [n_ff, n_expert_used, n_tokens] f32 tensor: a per-expert bias add on the gate
+// matmul output, the same on the up matmul output, then swiglu_oai over the
+// two. Both bias adds are in-place, so each costs a full read plus a full write
+// of a tensor that is only read once more. Folding them into the swiglu pass
+// leaves two reads and one write instead of six passes.
+//
+// Grouping matches kernel_add_id: group 0 = expert slot (i1), group 1 = token
+// (i2). For a contiguous destination that addressing is identical to the flat
+// row walk kernel_swiglu_oai uses, since row i1 + i2*ne1 sits at
+// i1*nb1 + i2*ne1*nb1.
+//------------------------------------------------------------------------------
+kernel void kernel_add_id_add_id_swiglu_oai(
+    global char * src_g,
+    ulong         offset_g,
+    global char * src_gb,
+    ulong         offset_gb,
+    global char * src_u,
+    ulong         offset_u,
+    global char * src_ub,
+    ulong         offset_ub,
+    global char * src_ids,
+    ulong         offset_ids,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb01_g,
+    ulong         nb02_g,
+    ulong         nb01_u,
+    ulong         nb02_u,
+    ulong         nb11_g,
+    ulong         nb11_u,
+    ulong         nb21,
+    ulong         nbd1,
+    ulong         nbd2,
+    int           ne0,
+    float         limit,
+    float         alpha
+) {
+    src_g   = (global char *)(src_g   + offset_g);
+    src_gb  = (global char *)(src_gb  + offset_gb);
+    src_u   = (global char *)(src_u   + offset_u);
+    src_ub  = (global char *)(src_ub  + offset_ub);
+    src_ids = (global char *)(src_ids + offset_ids);
+    dst     = (global char *)(dst     + offsetd);
+
+    const int i1 = get_group_id(0);
+    const int i2 = get_group_id(1);
+
+    // The ids tensor is a view into a [n_expert, n_tokens] buffer, so its row
+    // stride is nb21 and the k selected ids are NOT contiguous per token.
+    const int i11 = *((global const int *) (src_ids + i1*sizeof(int) + i2*nb21));
+
+    global const float * g_row  = (global const float *)(src_g  + i1*nb01_g + i2*nb02_g);
+    global const float * u_row  = (global const float *)(src_u  + i1*nb01_u + i2*nb02_u);
+    global const float * gb_row = (global const float *)(src_gb + i11*nb11_g);
+    global const float * ub_row = (global const float *)(src_ub + i11*nb11_u);
+    global       float * d_row  = (global       float *)(dst    + i1*nbd1   + i2*nbd2);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        float x0 = g_row[i0] + gb_row[i0];
+        float x1 = u_row[i0] + ub_row[i0];
+
+        x0 = min(x0, limit);
+        x1 = max(min(x1, limit), -limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        d_row[i0] = out_glu;
+    }
+}

ggml/src/ggml-opencl/kernels/moe_combine.cl

@@ -8,6 +8,49 @@
 // buffer and the k-1 elementwise add round-trips). Vectorized float4 over rows.
 // strides e1/e2/w1/w2/d1 are in ELEMENTS (floats).
 
+// Same weighted sum, with the per-expert bias add folded in.
+//
+// The MoE down projection's bias is applied by an in-place add_id whose only
+// consumer is this combine, so it costs a full read plus a full write of a
+// tensor that is read once more immediately afterwards. Reading the raw matmul
+// output here and adding the bias row while it is already in registers removes
+// that pass. Kept as a separate kernel so the unfused path is untouched.
+__kernel void kernel_moe_combine_bias_f32(
+        __global const char * e_buf, ulong off_e,
+        __global const char * w_buf, ulong off_w,
+        __global const char * b_buf, ulong off_b,   // per-expert bias rows
+        __global const char * i_buf, ulong off_i,   // expert ids
+        __global       char * d_buf, ulong off_d,
+        int  n_embd4,            // n_embd / 4
+        int  k,                  // n_expert_used
+        int  n_tokens,
+        uint e1, uint e2,        // experts strides (elements): per-expert, per-token
+        uint w1, uint w2,        // weights strides (elements)
+        uint d1,                 // dst per-token stride (elements)
+        ulong nb_b1,             // bias row stride (bytes)
+        ulong nb_i1)             // ids row stride (bytes) - ids is a view, not packed
+{
+    const uint r4  = get_global_id(0);
+    const uint tok = get_global_id(1);
+    if (r4 >= (uint)n_embd4 || tok >= (uint)n_tokens) return;
+
+    __global const float * E = (__global const float *)(e_buf + off_e) + tok*e2 + r4*4u;
+    __global const float * W = (__global const float *)(w_buf + off_w) + tok*w2;
+    __global const char  * B = b_buf + off_b;
+    __global const char  * I = i_buf + off_i + (ulong)tok*nb_i1;
+
+    float4 acc = (float4)(0.0f);
+    for (int e = 0; e < k; ++e) {
+        const int i11 = *((__global const int *)(I + (ulong)e*sizeof(int)));
+        __global const float * Brow = (__global const float *)(B + (ulong)i11*nb_b1) + r4*4u;
+        const float4 v = vload4(0, E + (uint)e*e1) + vload4(0, Brow);
+        acc = mad(v, (float4)(W[(uint)e*w1]), acc);
+    }
+
+    __global float * D = (__global float *)(d_buf + off_d) + tok*d1 + r4*4u;
+    vstore4(acc, 0, D);
+}
+
 __kernel void kernel_moe_combine_f32(
         __global const char * e_buf, ulong off_e,
         __global const char * w_buf, ulong off_w,

ggml/src/ggml-opencl/kernels/moe_sort_by_expert.cl

@@ -68,6 +68,79 @@ __kernel void kernel_moe_scatter(
     emap[tile_idx] = val;
 }
 
+// Deterministic replacement for kernel_moe_scatter.
+//
+// kernel_moe_scatter takes each token's slot from atomic_inc(slot_counter[expert]),
+// so the token -> slot packing inside an expert depends on which work-item wins the
+// atomic and changes from run to run. The ragged prefill GEMM path is sensitive to
+// that packing (the non-ragged path is not, since its padded slots alias slot 0 and
+// are overwritten last), which makes MoE prompt processing non-reproducible: the same
+// binary on the same prompt returns one of several outputs.
+//
+// Here the slot is the token's rank in flat (n, k) order among the tokens routed to
+// the same expert - a fixed function of the routing input. One workgroup per expert
+// walks the flat routing list in blocks of 64 and ranks its own tokens with a
+// workgroup scan, carrying a running count between blocks. Cost is one pass over the
+// routing list per expert; the list is a few KiB and stays in cache.
+__kernel void kernel_moe_scatter_stable(
+    __global const int * input,
+    __global int * post_router,
+    __global ushort * emap,
+    __global const int * tile_offset,
+    int N,
+    int topK,
+    uint n_experts
+) {
+    const int e   = get_group_id(1);
+    const int lid = get_local_id(0);
+    const int M   = N * topK;
+
+    __local int scan[64];
+    __local int running;
+
+    if (lid == 0) {
+        running = 0;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    for (int base = 0; base < M; base += 64) {
+        const int j = base + lid;
+
+        int pred = 0;
+        if (j < M) {
+            const int n = j / topK;
+            const int k = j - n * topK;
+            pred = (input[n * (int)n_experts + k] == e) ? 1 : 0;
+        }
+
+        scan[lid] = pred;
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        // Hillis-Steele inclusive scan over the 64 lanes
+        for (int off = 1; off < 64; off <<= 1) {
+            int add = (lid >= off) ? scan[lid - off] : 0;
+            barrier(CLK_LOCAL_MEM_FENCE);
+            scan[lid] += add;
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+
+        if (pred) {
+            const int local_slot = running + (scan[lid] - 1);   // exclusive rank
+            const int tile_idx   = tile_offset[e] + (local_slot >> 5);
+            const int lane       = local_slot & 31;
+
+            post_router[tile_idx * 32 + lane] = j;
+            emap[tile_idx] = (ushort)e;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+        if (lid == 63) {
+            running += scan[63];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+}
+
 __kernel void kernel_moe_fill(
     __global int * post_router,
     __global int * total_tiles,

ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl

@@ -28,6 +28,13 @@
 
 #define QK_K       256
 
+// ADRENO_OLD_COMPILER is defined by the host (-D) only for the Adreno E031
+// compilers older than E031.45, which miscompile several constructs this kernel
+// used (confirmed on E031.38 and E031.41; E031.45 is clean). Every other
+// compiler -- newer E031, E17, DX, Intel, and every non-Adreno device that
+// builds this program -- takes the #else branches, which are the original
+// source: the workarounds below cost ~13% on the q6_K flat n=1 GEMV where they
+// are not needed.
 inline float block_q_6_K_dot_y_flat(
     global uchar * blk_ql,
     global uchar * blk_qh,
@@ -37,6 +44,9 @@ inline float block_q_6_K_dot_y_flat(
     int ip,
     int is,
     int l0,
+#if defined(ADRENO_OLD_COMPILER)
+    int dbg,
+#endif
     float4 y0,
     float4 y1,
     float4 y2,
@@ -48,10 +58,40 @@ inline float block_q_6_K_dot_y_flat(
     global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
     global uchar * q2 = q1         + QK_K/8;
     global uchar * qh = blk_qh     + ib*64 + q_offset_h;
-    global char  * sc = blk_scales + ib*16 + is;
 
     float dall = blk_d[ib];
 
+#if defined(ADRENO_OLD_COMPILER)
+    // The vectorized dequant (int4/float4 bit-ops, convert_*4, dot()) and vload4
+    // are miscompiled here -> garbage weights. Reconstruct the 6-bit weights and
+    // take the dot product scalar. q4_K/q5_K flat already use scalar paths, which
+    // is why q6_K was the only flat GEMV that failed.
+    // Scales are SIGNED int8; read as uchar and sign-extend arithmetically so the
+    // result does not depend on whether the compiler treats `char` as signed.
+    global uchar * sc = (global uchar *)(blk_scales + ib*16 + is);
+
+    int s0 = (int)sc[0] - 256*(sc[0] >> 7);
+    int s2 = (int)sc[2] - 256*(sc[2] >> 7);
+    int s4 = (int)sc[4] - 256*(sc[4] >> 7);
+    int s6 = (int)sc[6] - 256*(sc[6] >> 7);
+
+    // one 6-bit weight: low/high nibble of a ql byte OR'd with a 2-bit qh plane
+    // (plane p in {0,1,2,3} selects qh bits 2p..2p+1) placed at bits 4-5, minus 32.
+    #define Q6W(qb, sh, hb, p) ((float)((((int)(qb) >> (sh)) & 15) | ((((int)(hb) >> (2*(p))) & 3) << 4)) - 32.f)
+
+    float d0 = y0.s0*Q6W(q1[0],0,qh[0],0) + y0.s1*Q6W(q1[1],0,qh[1],0) + y0.s2*Q6W(q1[2],0,qh[2],0) + y0.s3*Q6W(q1[3],0,qh[3],0);
+    float d1 = y1.s0*Q6W(q2[0],0,qh[0],1) + y1.s1*Q6W(q2[1],0,qh[1],1) + y1.s2*Q6W(q2[2],0,qh[2],1) + y1.s3*Q6W(q2[3],0,qh[3],1);
+    float d2 = y2.s0*Q6W(q1[0],4,qh[0],2) + y2.s1*Q6W(q1[1],4,qh[1],2) + y2.s2*Q6W(q1[2],4,qh[2],2) + y2.s3*Q6W(q1[3],4,qh[3],2);
+    float d3 = y3.s0*Q6W(q2[0],4,qh[0],3) + y3.s1*Q6W(q2[1],4,qh[1],3) + y3.s2*Q6W(q2[2],4,qh[2],3) + y3.s3*Q6W(q2[3],4,qh[3],3);
+    #undef Q6W
+
+    if (dbg) printf("HELPER dall=%f s=[%d %d %d %d] d=[%f %f %f %f] ql0=%d qh0=%d y00=%f\n",
+        dall, s0, s2, s4, s6, d0, d1, d2, d3, (int)q1[0], (int)qh[0], y0.s0);
+
+    return dall * (d0 * s0 + d1 * s2 + d2 * s4 + d3 * s6);
+#else
+    global char * sc = blk_scales + ib*16 + is;
+
     // Vectorized loads: 3 uchar4 weight loads instead of 12 scalar byte reads.
     // q_offset_l/h are 4-aligned, so these are aligned vector loads.
     uchar4 q1v = vload4(0, q1);
@@ -72,6 +112,7 @@ inline float block_q_6_K_dot_y_flat(
 
     return dall * (dot(y0, w0) * sc[0] + dot(y1, w1) * sc[2] +
                    dot(y2, w2) * sc[4] + dot(y3, w3) * sc[6]);
+#endif
 }
 
 #undef N_DST
@@ -113,6 +154,11 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
         int ne1,
         int r2,
         int r3
+#if defined(ADRENO_OLD_COMPILER)
+        ,
+        uchar q6k_mask   // runtime 0xFF; the host passes it so the compiler cannot
+                         // constant-fold the printf guards below into nothing
+#endif
 ) {
     src1 = (global float*)((global char*)src1 + offset1);
     dst = (global float*)((global char*)dst + offsetd);
@@ -128,6 +174,22 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
 
     int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;
 
+#if defined(ADRENO_OLD_COMPILER)
+    // 64-bit `ulong` integer arithmetic is miscompiled here -> the base-pointer byte
+    // offsets came out wrong, so EVERY weight/scale read hit the wrong address. This
+    // was the primary cause of the q6_K flat failure (q5_K uses int offsets and is
+    // unaffected). Compute the block index in `int` and widen to `ulong` only inside
+    // the pointer expression: the byte offset stays 64-bit, but there is no ulong
+    // arithmetic chain to miscompile. The int index would overflow past ~2^31 blocks,
+    // which no realistic weight reaches -- but that is a narrowing, so keep it off the
+    // conformant path, which retains full ulong arithmetic.
+    int offset_src0 = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + (ulong)offset_src0 * 128;
+    global uchar * blk_qh     = (global uchar *) src0_qh + (ulong)offset_src0 * 64;
+    global char  * blk_scales = (global char  *) src0_s  + (ulong)offset_src0 * 16;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0;
+#else
     ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
     ulong offset_src0_ql = offset_src0 * 128;
     ulong offset_src0_qh = offset_src0 * 64;
@@ -138,6 +200,7 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
     global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
     global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
     global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+#endif
     global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
 
     int tid = get_sub_group_local_id()%(N_SIMDWIDTH/BLOCK_STRIDE); // within-super-block part, 0..15
@@ -155,24 +218,55 @@ kernel void kernel_mul_mv_q6_K_f32_flat(
 
     for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
         global float * y = yy + ib * QK_K + 128*ip + l0;
+#if defined(ADRENO_OLD_COMPILER)
+        // vload4 of f32 is miscompiled here; index the lanes scalar instead.
+        float4 y0 = (float4)(y[ 0], y[ 1], y[ 2], y[ 3]);
+        float4 y1 = (float4)(y[32], y[33], y[34], y[35]);
+        float4 y2 = (float4)(y[64], y[65], y[66], y[67]);
+        float4 y3 = (float4)(y[96], y[97], y[98], y[99]);
+#else
         float4 y0 = vload4(0, y +  0);
         float4 y1 = vload4(0, y + 32);
         float4 y2 = vload4(0, y + 64);
         float4 y3 = vload4(0, y + 96);
+#endif
 
         for (int row = 0; row < N_DST; row++) {
             if (first_row + row < ne01) {
+#if defined(ADRENO_OLD_COMPILER)
+                int dbg = (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ib==0 &&
+                           ne00==256 && ne01==16 && get_sub_group_local_id()==0) ? 1 : 0;
+                sumf[row] += block_q_6_K_dot_y_flat(
+                    blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
+                    ib, ip, is, l0, dbg, y0, y1, y2, y3);
+#else
                 sumf[row] += block_q_6_K_dot_y_flat(
                     blk_ql + row*nb*128, blk_qh + row*nb*64, blk_scales + row*nb*16, blk_d + row*nb,
                     ib, ip, is, l0, y0, y1, y2, y3);
+#endif
             }
         }
     }
 
+#if defined(ADRENO_OLD_COMPILER)
+    // Optimizer barrier. This compiler drops the sumf partials unless a side effect
+    // forces them to materialize. q6k_mask is a kernel arg the compiler cannot prove
+    // is never 0xFE (the host always passes 0xFF), so the printf survives compilation
+    // but never executes. FRAGILE: the exact set and placement of these guarded
+    // printfs is load-bearing on E031.41 -- removing any one re-breaks q6_K.
+    if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && ne00==256 && ne01==16 && get_sub_group_local_id()<16) {
+        printf("Q6KLANE lane=%d ip=%d il=%d is=%d l0=%d sumf0=%f\n",
+            get_sub_group_local_id(), ip, il, is, l0, sumf[0]);
+    }
+#endif
     for (int row = 0; row < N_DST; row++) {
         float tot = sub_group_reduce_add(sumf[row]);
         if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
             dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+#if defined(ADRENO_OLD_COMPILER)
+            if (q6k_mask==0xFE && r0==0 && r1==0 && im==0 && row==0 && ne00==256 && ne01==16)
+                printf("Q6KTOT tot=%f\n", tot);
+#endif
         }
     }
 }

ggml/src/ggml-opencl/kernels/rope.cl

@@ -75,7 +75,8 @@ kernel void kernel_rope_norm_f32(
         float ext_factor,
         float attn_factor,
         float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -94,14 +95,15 @@ kernel void kernel_rope_norm_f32(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
             global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
             global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -154,7 +156,8 @@ kernel void kernel_rope_norm_f16(
         float ext_factor,
         float attn_factor,
         float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -173,14 +176,15 @@ kernel void kernel_rope_norm_f16(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
             global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
             global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -233,7 +237,8 @@ kernel void kernel_rope_neox_f32(
         float ext_factor,
         float attn_factor,
         float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -252,17 +257,18 @@ kernel void kernel_rope_neox_f32(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);
 
             const float x0 = src[0];
             const float x1 = src[n_dims/2];
@@ -312,7 +318,8 @@ kernel void kernel_rope_neox_f16(
         float ext_factor,
         float attn_factor,
         float beta_fast,
-        float beta_slow
+        float beta_slow,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -331,17 +338,18 @@ kernel void kernel_rope_neox_f16(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
-            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);
 
             const float x0 = src[0];
             const float x1 = src[n_dims/2];
@@ -393,7 +401,8 @@ kernel void kernel_rope_multi_f32(
         float beta_fast,
         float beta_slow,
         int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -414,10 +423,11 @@ kernel void kernel_rope_multi_f32(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
             float theta_base = 0.0f;
 
             if (is_imrope) {
@@ -445,14 +455,14 @@ kernel void kernel_rope_multi_f32(
                 }
             }
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
-            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);
 
             const float x0 = src[0];
             const float x1 = src[n_dims/2];
@@ -504,7 +514,8 @@ kernel void kernel_rope_multi_f16(
         float beta_fast,
         float beta_slow,
         int4 sections,
-        int  is_imrope
+        int  is_imrope,
+        int n_offs
 ) {
     src0 = (global void*)((global char*)src0 + offset0);
     src1 = (global int*)((global char*)src1 + offset1);
@@ -525,10 +536,11 @@ kernel void kernel_rope_multi_f16(
     float inv_ndims = -1.f/n_dims;
 
     for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
-        if (i0 < n_dims) {
-            int ic = i0/2;
+        if (i0 >= n_offs && i0 < n_offs + n_dims) {
+            int iw = i0 - n_offs; // relative idx
+            int ic = iw/2;
 
-            const int sector = (i0 / 2) % sect_dims;
+            const int sector = ic % sect_dims;
             float theta_base = 0.0f;
 
             if (is_imrope) {
@@ -556,14 +568,14 @@ kernel void kernel_rope_multi_f16(
                 }
             }
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(freq_base, inv_ndims*iw);
 
             const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
 
-            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, iw, ext_factor, attn_factor);
 
-            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + (n_offs + ic)*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + (n_offs + ic)*nb0);
 
             const float x0 = src[0];
             const float x1 = src[n_dims/2];

ggml/src/ggml-opencl/kernels/ssm_scan.cl

@@ -0,0 +1,216 @@
+// Mamba2 fused SSM scan kernel. One workgroup per (head, dim, seq); WG size =
+// 64 threads. Each thread owns c_factor = d_state/64 state elements in
+// private registers; the state stays resident across the n_tokens t-loop
+//
+// References:
+//   ggml/src/ggml-cuda/ssm-scan.cu:117 ssm_scan_f32_group
+//   ggml/src/ggml-cpu/ops.cpp:9368 ggml_compute_forward_ssm_scan_f32
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_khr_subgroups
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#else
+#define REQD_SUBGROUP_SIZE_64
+#endif
+
+inline float softplus_f32(float x) {
+    return (x <= 20.0f) ? log(1.0f + exp(x)) : x;
+}
+
+// d_state = 128 (most Mamba-2 models, e.g. mamba2-2.7B, Codestral-Mamba).
+// WG = 64 threads, each holds 2 state elements (tid and tid+64).
+REQD_SUBGROUP_SIZE_64
+kernel void kernel_ssm_scan_f32_mamba2_d128(
+    global const char * src0_base, ulong src0_off,
+    global const char * src1_base, ulong src1_off,
+    global const char * src2_base, ulong src2_off,
+    global const char * src3_base, ulong src3_off,
+    global const char * src4_base, ulong src4_off,
+    global const char * src5_base, ulong src5_off,
+    global const char * src6_base, ulong src6_off,
+    global       char * dst_base,  ulong dst_off,
+    ulong s0_nb2, ulong s0_nb3,
+    ulong x_nb2,  ulong x_nb3,
+    ulong dt_nb1, ulong dt_nb2,
+    ulong A_nb1,
+    ulong B_nb2,  ulong B_nb3,
+    ulong C_nb2,  ulong C_nb3,
+    ulong s_off_bytes,
+    int   head_dim, int n_head, int n_group, int n_tokens
+) {
+    const int d_state = 128;
+
+    const int tid     = (int) get_local_id(0);
+    const int wg_x    = (int) get_group_id(0);
+    const int seq_id  = (int) get_group_id(1);
+
+    const int head_id = wg_x / head_dim;
+    const int dim_id  = wg_x - head_id * head_dim;
+    const int g       = head_id / (n_head / n_group);
+
+    src0_base += src0_off;
+    src1_base += src1_off;
+    src2_base += src2_off;
+    src3_base += src3_off;
+    src4_base += src4_off;
+    src5_base += src5_off;
+    src6_base += src6_off;
+    dst_base  += dst_off;
+
+    const int seq_slot = ((global const int *) src6_base)[seq_id];
+
+    const ulong state_base_off = (ulong)seq_slot * s0_nb3 + (ulong)head_id * s0_nb2
+                                + (ulong)dim_id * d_state * sizeof(float);
+    global const float * s0_warp = (global const float *)(src0_base + state_base_off);
+    const ulong state_out_off = (ulong)seq_id * s0_nb3 + (ulong)head_id * s0_nb2
+                              + (ulong)dim_id * d_state * sizeof(float);
+    global float * s_warp = (global float *)(dst_base + s_off_bytes + state_out_off);
+
+    global const char * x_seq  = src1_base + (ulong)seq_id * x_nb3;
+    global const char * dt_seq = src2_base + (ulong)seq_id * dt_nb2;
+    global const char * B_seq  = src4_base + (ulong)seq_id * B_nb3 + (ulong)g * d_state * sizeof(float);
+    global const char * C_seq  = src5_base + (ulong)seq_id * C_nb3 + (ulong)g * d_state * sizeof(float);
+
+    const ulong y_dim_total = (ulong)n_head * head_dim;
+    global float * y_seq = (global float *)dst_base
+                           + (ulong)seq_id * (ulong)n_tokens * y_dim_total;
+
+    const float A_val = ((global const float *)src3_base)[(ulong)head_id * A_nb1 / sizeof(float)];
+
+    // c_factor = 2: each thread owns 2 state elements (tid and tid+64).
+    float state0 = s0_warp[tid];
+    float state1 = s0_warp[tid + 64];
+
+    for (int t = 0; t < n_tokens; ++t) {
+        const float dt_h        = ((global const float *)(dt_seq + (ulong)t * dt_nb1))[head_id];
+        const float dt_softplus = softplus_f32(dt_h);
+        const float dA          = exp(dt_softplus * A_val);
+        const float x_val       = ((global const float *)(x_seq + (ulong)t * x_nb2))[(ulong)head_id * head_dim + dim_id];
+        const float x_dt        = x_val * dt_softplus;
+
+        const float B0 = ((global const float *)(B_seq + (ulong)t * B_nb2))[tid];
+        const float B1 = ((global const float *)(B_seq + (ulong)t * B_nb2))[tid + 64];
+        const float C0 = ((global const float *)(C_seq + (ulong)t * C_nb2))[tid];
+        const float C1 = ((global const float *)(C_seq + (ulong)t * C_nb2))[tid + 64];
+
+        state0 = state0 * dA + B0 * x_dt;
+        state1 = state1 * dA + B1 * x_dt;
+        const float partial = state0 * C0 + state1 * C1;
+
+        const float sum = sub_group_reduce_add(partial);
+        if (tid == 0) {
+            y_seq[(ulong)t * y_dim_total + (ulong)head_id * head_dim + dim_id] = sum;
+        }
+    }
+
+    s_warp[tid]      = state0;
+    s_warp[tid + 64] = state1;
+}
+
+// d_state = 256 (Falcon-H1). WG = 64 threads, each holds 4 state elements.
+REQD_SUBGROUP_SIZE_64
+kernel void kernel_ssm_scan_f32_mamba2_d256(
+    global const char * src0_base, ulong src0_off,
+    global const char * src1_base, ulong src1_off,
+    global const char * src2_base, ulong src2_off,
+    global const char * src3_base, ulong src3_off,
+    global const char * src4_base, ulong src4_off,
+    global const char * src5_base, ulong src5_off,
+    global const char * src6_base, ulong src6_off,
+    global       char * dst_base,  ulong dst_off,
+    ulong s0_nb2, ulong s0_nb3,
+    ulong x_nb2,  ulong x_nb3,
+    ulong dt_nb1, ulong dt_nb2,
+    ulong A_nb1,
+    ulong B_nb2,  ulong B_nb3,
+    ulong C_nb2,  ulong C_nb3,
+    ulong s_off_bytes,
+    int   head_dim, int n_head, int n_group, int n_tokens
+) {
+    const int d_state = 256;
+
+    const int tid     = (int) get_local_id(0);
+    const int wg_x    = (int) get_group_id(0);
+    const int seq_id  = (int) get_group_id(1);
+
+    const int head_id = wg_x / head_dim;
+    const int dim_id  = wg_x - head_id * head_dim;
+    const int g       = head_id / (n_head / n_group);
+
+    src0_base += src0_off;
+    src1_base += src1_off;
+    src2_base += src2_off;
+    src3_base += src3_off;
+    src4_base += src4_off;
+    src5_base += src5_off;
+    src6_base += src6_off;
+    dst_base  += dst_off;
+
+    const int seq_slot = ((global const int *) src6_base)[seq_id];
+
+    const ulong state_base_off = (ulong)seq_slot * s0_nb3 + (ulong)head_id * s0_nb2
+                                + (ulong)dim_id * d_state * sizeof(float);
+    global const float * s0_warp = (global const float *)(src0_base + state_base_off);
+    const ulong state_out_off = (ulong)seq_id * s0_nb3 + (ulong)head_id * s0_nb2
+                              + (ulong)dim_id * d_state * sizeof(float);
+    global float * s_warp = (global float *)(dst_base + s_off_bytes + state_out_off);
+
+    global const char * x_seq  = src1_base + (ulong)seq_id * x_nb3;
+    global const char * dt_seq = src2_base + (ulong)seq_id * dt_nb2;
+    global const char * B_seq  = src4_base + (ulong)seq_id * B_nb3 + (ulong)g * d_state * sizeof(float);
+    global const char * C_seq  = src5_base + (ulong)seq_id * C_nb3 + (ulong)g * d_state * sizeof(float);
+
+    const ulong y_dim_total = (ulong)n_head * head_dim;
+    global float * y_seq = (global float *)dst_base
+                           + (ulong)seq_id * (ulong)n_tokens * y_dim_total;
+
+    const float A_val = ((global const float *)src3_base)[(ulong)head_id * A_nb1 / sizeof(float)];
+
+    // c_factor = 4: each thread owns 4 state elements.
+    float state0 = s0_warp[tid];
+    float state1 = s0_warp[tid + 64];
+    float state2 = s0_warp[tid + 128];
+    float state3 = s0_warp[tid + 192];
+
+    for (int t = 0; t < n_tokens; ++t) {
+        const float dt_h        = ((global const float *)(dt_seq + (ulong)t * dt_nb1))[head_id];
+        const float dt_softplus = softplus_f32(dt_h);
+        const float dA          = exp(dt_softplus * A_val);
+        const float x_val       = ((global const float *)(x_seq + (ulong)t * x_nb2))[(ulong)head_id * head_dim + dim_id];
+        const float x_dt        = x_val * dt_softplus;
+
+        global const float * B_t = (global const float *)(B_seq + (ulong)t * B_nb2);
+        global const float * C_t = (global const float *)(C_seq + (ulong)t * C_nb2);
+
+        const float B0 = B_t[tid];
+        const float B1 = B_t[tid + 64];
+        const float B2 = B_t[tid + 128];
+        const float B3 = B_t[tid + 192];
+        const float C0 = C_t[tid];
+        const float C1 = C_t[tid + 64];
+        const float C2 = C_t[tid + 128];
+        const float C3 = C_t[tid + 192];
+
+        state0 = state0 * dA + B0 * x_dt;
+        state1 = state1 * dA + B1 * x_dt;
+        state2 = state2 * dA + B2 * x_dt;
+        state3 = state3 * dA + B3 * x_dt;
+        const float partial = state0 * C0 + state1 * C1 + state2 * C2 + state3 * C3;
+
+        const float sum = sub_group_reduce_add(partial);
+        if (tid == 0) {
+            y_seq[(ulong)t * y_dim_total + (ulong)head_id * head_dim + dim_id] = sum;
+        }
+    }
+
+    s_warp[tid]       = state0;
+    s_warp[tid + 64]  = state1;
+    s_warp[tid + 128] = state2;
+    s_warp[tid + 192] = state3;
+}

ggml/src/ggml-sycl/convert.cpp

@@ -76,6 +76,19 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k,
 #endif
 }
 
+template <typename dst_t>
+static void dequantize_row_q2_K_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
+                                             dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+        [=](sycl::nd_item<3> item_ct1) {
+            dequantize_block_q2_K_reorder(vx, y, item_ct1, nb);
+        });
+}
+
 template <typename dst_t>
 static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -667,7 +680,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
                 return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
             }
         case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q2_K_sycl_reorder;
+            } else {
+                return dequantize_row_q2_K_sycl;
+            }
         case GGML_TYPE_Q3_K:
             if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
                 return dequantize_row_q3_K_sycl_reorder;
@@ -753,7 +770,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
                 return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
             }
         case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q2_K_sycl_reorder;
+            } else {
+                return dequantize_row_q2_K_sycl;
+            }
         case GGML_TYPE_Q3_K:
             if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
                 return dequantize_row_q3_K_sycl_reorder;

ggml/src/ggml-sycl/dequantize.hpp

@@ -943,6 +943,47 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri
 
 }
 
+template<typename dst_t>
+static void dequantize_block_q2_K_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                          const sycl::nd_item<3> & item_ct1, int64_t n_blocks) {
+#if QK_K == 256
+    const int64_t i = item_ct1.get_group(2);
+    if (i >= n_blocks) {
+        return;
+    }
+
+    const uint8_t * base          = static_cast<const uint8_t *>(vx);
+    const size_t    qs_offset     = i * (QK_K / 4);
+    const size_t    scales_offset = n_blocks * (QK_K / 4) + i * (QK_K / 16);
+    const size_t    dm_offset     = n_blocks * (QK_K / 4) + n_blocks * (QK_K / 16) + i * sizeof(ggml_half2);
+
+    const uint8_t *     qs     = base + qs_offset;
+    const uint8_t *     scales = base + scales_offset;
+    const ggml_half2 * dm     = reinterpret_cast<const ggml_half2 *>(base + dm_offset);
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t n    = tid / 32;
+    const int64_t l    = tid - 32 * n;
+    const int64_t is   = 8 * n + l / 16;
+
+    const uint8_t q = qs[32 * n + l];
+    dst_t * y = yy + i * QK_K + 128 * n;
+
+    const float dall = (*dm)[0];
+    const float dmin = (*dm)[1];
+    y[l+ 0] = dall * (scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (scales[is+0] >> 4);
+    y[l+32] = dall * (scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (scales[is+2] >> 4);
+    y[l+64] = dall * (scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (scales[is+4] >> 4);
+    y[l+96] = dall * (scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (scales[is+6] >> 4);
+#else
+    GGML_UNUSED(vx);
+    GGML_UNUSED(yy);
+    GGML_UNUSED(item_ct1);
+    GGML_UNUSED(n_blocks);
+    GGML_ABORT("Q2_K reorder dequantize not supported for QK_K != 256");
+#endif
+}
+
 template<typename dst_t>
 static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                   const sycl::nd_item<3> &item_ct1) {

ggml/src/ggml-sycl/dmmv.cpp

@@ -1921,6 +1921,23 @@ ESIMD_INLINE void dequantize_mul_mat_vec_reorder_esimd(
     }
 }
 
+static void dequantize_mul_mat_vec_q2_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q2_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
 static void dequantize_mul_mat_vec_q3_K_sycl_reorder_esimd(const void *vx, const float *y,
                                                            float *dst, const int ncols,
                                                            const int nrows,
@@ -1955,6 +1972,23 @@ static void dequantize_mul_mat_vec_q4_K_sycl_reorder_esimd(const void *vx, const
     });
 }
 
+static void dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(const void *vx, const float *y,
+                                                           float *dst, const int ncols,
+                                                           const int nrows,
+                                                           dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int workgroups = (nrows + 1) / 2;
+    stream->submit([&](sycl::handler &h) {
+        sycl::local_accessor<float, 1> lmem(sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE * 2), h);
+        h.parallel_for(
+            sycl::nd_range<1>(sycl::range<1>((size_t)workgroups * GGML_SYCL_DMMV_ESIMD_WG_SIZE), sycl::range<1>(GGML_SYCL_DMMV_ESIMD_WG_SIZE)),
+            [=](sycl::nd_item<1> it) [[intel::sycl_explicit_simd]] {
+                dequantize_mul_mat_vec_reorder_esimd<GGML_TYPE_Q5_K>(
+                    vx, y, dst, ncols, nrows, lmem, it);
+            });
+    });
+}
+
 static void dequantize_mul_mat_vec_q6_K_sycl_reorder_esimd(const void *vx, const float *y,
                                                            float *dst, const int ncols,
                                                            const int nrows,
@@ -2094,7 +2128,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
         case GGML_TYPE_Q2_K:
             if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q2_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q2_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q2_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
             } else {
                 dequantize_mul_mat_vec_q2_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             }
@@ -2134,7 +2176,15 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
         case GGML_TYPE_Q5_K:
             if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+#ifdef GGML_SYCL_DMMV_HAS_ESIMD
+                if (g_ggml_sycl_enable_esimd) {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder_esimd(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
+                else
+#endif
+                {
+                    dequantize_mul_mat_vec_q5_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+                }
             } else {
                 dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             }

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -62,7 +62,7 @@
 
 #define DPCT_UNUSED(x) (void)(x)
 
-inline void _abort(const char * str) {
+[[noreturn]] inline void _abort(const char * str) {
     std::cerr << str << std::endl;
     std::abort();
 }

ggml/src/ggml-sycl/element_wise.cpp

@@ -10,7 +10,7 @@
     (ITEM.get_local_range(IDX) * ITEM.get_group(IDX) + ITEM.get_local_id(IDX))
 
 static void acc_f32(const char * x, const char * y, float * dst, const int64_t ne,
-    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2,
     const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
     const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
     const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -455,7 +455,7 @@ static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, c
 namespace ggml_sycl_detail {
 static void acc_f32_sycl(const char *x, const char *y, float *dst,
                          const int64_t n_elements,
-                         const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3,
+                         const int64_t ne0, const int64_t ne1, const int64_t ne2,
                          const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
                          const int64_t ne10, const int64_t ne11, const int64_t ne12, const int64_t ne13,
                          const int64_t nb10, const int64_t nb11, const int64_t nb12, const int64_t nb13,
@@ -466,7 +466,7 @@ static void acc_f32_sycl(const char *x, const char *y, float *dst,
                                            sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
                          [=](sycl::nd_item<3> /*item_ct1*/) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                             acc_f32(x, y, dst, n_elements,
-                                ne0, ne1, ne2, ne3,
+                                ne0, ne1, ne2,
                                 nb00, nb01, nb02, nb03,
                                 ne10, ne11, ne12, ne13,
                                 nb10, nb11, nb12, nb13,
@@ -970,7 +970,7 @@ static inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, ggml_tensor
     const int64_t offset = (int64_t) ((const int32_t *) dst->op_params)[3] / (int64_t) sizeof(float);
 
     ggml_sycl_detail::acc_f32_sycl(src0_d, src1_d, dst_d, ggml_nelements(dst),
-        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2],
         src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
         src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
         src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],

ggml/src/ggml-sycl/esimd.hpp

@@ -1,15 +1,3 @@
-//
-// MIT license
-// Copyright (C) 2026 Intel Corporation
-// SPDX-License-Identifier: MIT
-//
-
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-
 #ifndef GGML_SYCL_ESIMD_HPP
 #define GGML_SYCL_ESIMD_HPP
 
@@ -73,6 +61,93 @@ static ESIMD_INLINE void unpack_scale_min_k4(
     min_f   = convert<float>(m) * (-dmin);
 }
 
+// ---------------------------------------------------------------------------
+// Q2_K, SOA reorder layout produced by reorder_qw_q2_k:
+//   [qs: nb*(QK_K/4)] [scales: nb*(QK_K/16)] [dm: nb*sizeof(half2)]
+// with nb = nrows*num_blocks_per_row.
+//
+// 2 bits per weight. The 8 output chunks of 32 (matching dequantize_row_q2_K)
+// map to super-chunk s (0..7): byte base 32*(s/4) into the 64-byte qs array,
+// bit shift 2*(s%4); the low 16 lanes use scales[2s], the high 16 use
+// scales[2s+1], with dl = d*(sc & 0xF), ml = dmin*(sc >> 4), deq = dl*q - ml.
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q2_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    scales;
+        const sycl::half * dm;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * scales = qs + nb * (QK_K / 4);
+        const sycl::half * dm  = (const sycl::half *) (scales + nb * (QK_K / 16));
+        return { qs, scales, dm };
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 64> qs_a     = block_load<uint8_t, 64>(pa.qs + bia * (QK_K / 4));
+        simd<uint8_t, 64> qs_b     = 0;
+        simd<uint8_t, 16> scales_a = block_load<uint8_t, 16>(pa.scales + bia * (QK_K / 16));
+        simd<uint8_t, 16> scales_b = 0;
+
+        const float dall_a = (float) pa.dm[bia * 2 + 0];
+        const float dmin_a = (float) pa.dm[bia * 2 + 1];
+        float dall_b = 0.0f;
+        float dmin_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 64>(pb.qs + bib * (QK_K / 4));
+            scales_b = block_load<uint8_t, 16>(pb.scales + bib * (QK_K / 16));
+            dall_b = (float) pb.dm[bib * 2 + 0];
+            dmin_b = (float) pb.dm[bib * 2 + 1];
+        }
+
+        // per-chunk scale (d * (sc & 0xF)) and min (-dmin * (sc >> 4)), all 16 codes;
+        // min carries the negation so the dequant epilogue adds (matches Q4_K/Q5_K)
+        simd<float, 16> scale_f_a = convert<float>(scales_a & simd<uint8_t, 16>(0x0F)) * dall_a;
+        simd<float, 16> min_f_a   = convert<float>(scales_a >> simd<uint8_t, 16>(4))  * (-dmin_a);
+        simd<float, 16> scale_f_b = convert<float>(scales_b & simd<uint8_t, 16>(0x0F)) * dall_b;
+        simd<float, 16> min_f_b   = convert<float>(scales_b >> simd<uint8_t, 16>(4))  * (-dmin_b);
+
+#pragma unroll
+        for (int s = 0; s < 8; ++s) {
+            const int     byte_base = 32 * (s / 4);
+            const uint8_t shift     = (uint8_t) (2 * (s % 4));
+            simd<float, 32> y_s = y_vec.select<32, 1>(s * 32);
+
+            simd<uint8_t, 32> qa = (qs_a.select<32, 1>(byte_base) >> shift) & simd<uint8_t, 32>(3);
+            simd<uint8_t, 32> qb = (qs_b.select<32, 1>(byte_base) >> shift) & simd<uint8_t, 32>(3);
+
+            const float scale_a_lo = scale_f_a[2 * s + 0];
+            const float scale_a_hi = scale_f_a[2 * s + 1];
+            const float min_a_lo   = min_f_a[2 * s + 0];
+            const float min_a_hi   = min_f_a[2 * s + 1];
+            const float scale_b_lo = scale_f_b[2 * s + 0];
+            const float scale_b_hi = scale_f_b[2 * s + 1];
+            const float min_b_lo   = min_f_b[2 * s + 0];
+            const float min_b_hi   = min_f_b[2 * s + 1];
+
+            simd<float, 32> scale_vec_a = splat_lo_hi(scale_a_lo, scale_a_hi);
+            simd<float, 32> min_vec_a   = splat_lo_hi(min_a_lo, min_a_hi);
+            simd<float, 32> scale_vec_b = splat_lo_hi(scale_b_lo, scale_b_hi);
+            simd<float, 32> min_vec_b   = splat_lo_hi(min_b_lo, min_b_hi);
+
+            simd<float, 32> deq_a = convert<float>(qa) * scale_vec_a + min_vec_a;
+            simd<float, 32> deq_b = convert<float>(qb) * scale_vec_b + min_vec_b;
+
+            acc_a += y_s * deq_a;
+            acc_b += y_s * deq_b;
+        }
+    }
+};
+
 // ---------------------------------------------------------------------------
 // Q3_K, SOA reorder layout produced by reorder_qw_q3_k:
 //   [qs: nb*(QK_K/4)] [hmask: nb*(QK_K/8)] [scales: nb*12] [d: nb*sizeof(half)]
@@ -287,6 +362,128 @@ template <> struct esimd_reorder_q_traits<GGML_TYPE_Q4_K> {
     }
 };
 
+// ---------------------------------------------------------------------------
+// Q5_K, SOA reorder layout produced by reorder_qw_q5_k:
+//   [qs: nb*(QK_K/2)] [qh: nb*(QK_K/8)] [scales: nb*K_SCALE_SIZE] [dm: nb*sizeof(half2)]
+// with nb = nrows*num_blocks_per_row.
+//
+// Identical to Q4_K except each 4-bit quant gains a 5th (high) bit from qh:
+// output chunk c (0..7) adds 16 when bit c of qh[l] is set, where qh[l] indexes
+// the same 32 bytes for every chunk (matches dequantize_row_q5_K).
+// ---------------------------------------------------------------------------
+template <> struct esimd_reorder_q_traits<GGML_TYPE_Q5_K> {
+    struct ptrs {
+        const uint8_t *    qs;
+        const uint8_t *    qh;
+        const uint8_t *    scales;
+        const sycl::half * dm;
+    };
+
+    static ESIMD_INLINE ptrs make_ptrs(const void * vx, size_t nb) {
+        const uint8_t * qs     = (const uint8_t *) vx;
+        const uint8_t * qh     = qs + nb * (QK_K / 2);
+        const uint8_t * scales = qh + nb * (QK_K / 8);
+        const sycl::half * dm  = (const sycl::half *) (scales + nb * K_SCALE_SIZE);
+        return { qs, qh, scales, dm };
+    }
+
+    // extract bit `bit` (0..7) of each lane and move it to bit position 4,
+    // e.g. for the 4-bit base quant's 5th (high) bit. `bit` is always a
+    // compile-time-known unrolled loop constant at call sites, so this folds
+    // to a single mask (bit==4), mask+left-shift (bit<4), or mask+right-shift
+    // (bit>4) instead of the shift+mask+shift a naive `(qh>>bit & 1) << 4` emits.
+    static ESIMD_INLINE sycl::ext::intel::esimd::simd<uint16_t, 32> extract_bit_to_pos4(
+            sycl::ext::intel::esimd::simd<uint8_t, 32> qh, int bit) {
+        using namespace sycl::ext::intel::esimd;
+        simd<uint16_t, 32> masked = convert<uint16_t>(qh & simd<uint8_t, 32>((uint8_t) (1u << bit)));
+        if (bit < 4) {
+            return masked << simd<uint16_t, 32>((uint16_t) (4 - bit));
+        } else if (bit > 4) {
+            return masked >> simd<uint16_t, 32>((uint16_t) (bit - 4));
+        }
+        return masked;
+    }
+
+    static ESIMD_INLINE void mac_pair(
+            const ptrs & pa, size_t bia,
+            const ptrs & pb, size_t bib, bool has_b,
+            sycl::ext::intel::esimd::simd<float, 256> & y_vec,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_a,
+            sycl::ext::intel::esimd::simd<float, 32> & acc_b) {
+        using namespace sycl::ext::intel::esimd;
+
+        simd<uint8_t, 128> qs_a     = block_load<uint8_t, 128>(pa.qs + bia * (QK_K / 2));
+        simd<uint8_t, 128> qs_b     = 0;
+        simd<uint8_t, 32>  qh_a     = block_load<uint8_t, 32>(pa.qh + bia * (QK_K / 8));
+        simd<uint8_t, 32>  qh_b     = 0;
+        simd<uint8_t, 12>  scales_a = block_load<uint8_t, 12>(pa.scales + bia * K_SCALE_SIZE);
+        simd<uint8_t, 12>  scales_b = 0;
+
+        const float dall_a = (float) pa.dm[bia * 2 + 0];
+        const float dmin_a = (float) pa.dm[bia * 2 + 1];
+        float dall_b = 0.0f;
+        float dmin_b = 0.0f;
+        if (has_b) {
+            qs_b     = block_load<uint8_t, 128>(pb.qs + bib * (QK_K / 2));
+            qh_b     = block_load<uint8_t, 32>(pb.qh + bib * (QK_K / 8));
+            scales_b = block_load<uint8_t, 12>(pb.scales + bib * K_SCALE_SIZE);
+            dall_b = (float) pb.dm[bib * 2 + 0];
+            dmin_b = (float) pb.dm[bib * 2 + 1];
+        }
+
+        simd<float, 8> scale_f_a, min_f_a, scale_f_b, min_f_b;
+        unpack_scale_min_k4(scales_a, dall_a, dmin_a, scale_f_a, min_f_a);
+        unpack_scale_min_k4(scales_b, dall_b, dmin_b, scale_f_b, min_f_b);
+
+        simd<uint8_t, 128> qs_lo_a = qs_a & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_a = qs_a >> simd<uint8_t, 128>(4);
+        simd<uint8_t, 128> qs_lo_b = qs_b & simd<uint8_t, 128>(0x0F);
+        simd<uint8_t, 128> qs_hi_b = qs_b >> simd<uint8_t, 128>(4);
+
+#pragma unroll
+        for (int sb = 0; sb < 8; sb += 2) {
+            const int q_offset = sb * 16;
+            simd<float, 32> y_lo = y_vec.select<32, 1>(sb * 32);
+            simd<float, 32> y_hi = y_vec.select<32, 1>((sb + 1) * 32);
+
+            const float scale_a_lo = scale_f_a[sb];
+            const float scale_a_hi = scale_f_a[sb + 1];
+            const float min_a_lo   = min_f_a[sb];
+            const float min_a_hi   = min_f_a[sb + 1];
+            const float scale_b_lo = scale_f_b[sb];
+            const float scale_b_hi = scale_f_b[sb + 1];
+            const float min_b_lo   = min_f_b[sb];
+            const float min_b_hi   = min_f_b[sb + 1];
+
+            simd<uint8_t, 32> qa_lo_u8 = qs_lo_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qa_hi_u8 = qs_hi_a.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_lo_u8 = qs_lo_b.select<32, 1>(q_offset);
+            simd<uint8_t, 32> qb_hi_u8 = qs_hi_b.select<32, 1>(q_offset);
+            simd<uint16_t, 32> qa_lo = convert<uint16_t>(qa_lo_u8);
+            simd<uint16_t, 32> qa_hi = convert<uint16_t>(qa_hi_u8);
+            simd<uint16_t, 32> qb_lo = convert<uint16_t>(qb_lo_u8);
+            simd<uint16_t, 32> qb_hi = convert<uint16_t>(qb_hi_u8);
+
+            // add the 5th bit: chunk sb uses qh bit sb, chunk sb+1 uses qh bit sb+1;
+            // qh always indexes the same 32 bytes regardless of chunk
+            qa_lo += extract_bit_to_pos4(qh_a, sb);
+            qa_hi += extract_bit_to_pos4(qh_a, sb + 1);
+            qb_lo += extract_bit_to_pos4(qh_b, sb);
+            qb_hi += extract_bit_to_pos4(qh_b, sb + 1);
+
+            simd<float, 32> deq_a_lo = convert<float>(qa_lo) * scale_a_lo + min_a_lo;
+            simd<float, 32> deq_a_hi = convert<float>(qa_hi) * scale_a_hi + min_a_hi;
+            simd<float, 32> deq_b_lo = convert<float>(qb_lo) * scale_b_lo + min_b_lo;
+            simd<float, 32> deq_b_hi = convert<float>(qb_hi) * scale_b_hi + min_b_hi;
+
+            acc_a += y_lo * deq_a_lo;
+            acc_b += y_lo * deq_b_lo;
+            acc_a += y_hi * deq_a_hi;
+            acc_b += y_hi * deq_b_hi;
+        }
+    }
+};
+
 // ---------------------------------------------------------------------------
 // Q6_K, SOA reorder layout:
 //   [ql: nb*(QK_K/2)] [qh: nb*(QK_K/4)] [scales(int8): nb*(QK_K/16)] [d: nb*half]

ggml/src/ggml-sycl/fattn-mkl.cpp

@@ -43,7 +43,7 @@ static void mkl_fa_pack_q_fp16(
     dpct::queue_ptr stream,
     sycl::half * __restrict dst,
     const float * __restrict q_src,
-    int n_queries, int n_query_rows, int DKQ,
+    int n_queries, int DKQ,
     int gqa_ratio, int kvh_base_head,
     float q_scale, int64_t q_row_stride, int64_t q_head_stride,
     int64_t wg_size) {
@@ -121,7 +121,7 @@ static void mkl_fa_online_softmax_chunk(
     float * __restrict VKQ_accum,
     int q0, int q_rows, int n_queries, int DV,
     int chunk_size, int chunk_start,
-    int kvh_head, int gqa_ratio,
+    int kvh_head,
     const sycl::half * mask_data, int64_t mask_head_stride,
     int64_t mask_row_stride, int mask_n_heads,
     float logit_softcap, int64_t wg_size) {
@@ -473,7 +473,6 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
     MKL_ACCUM(dequant_time_us, t_deq);
 
     // --- Resolve mask pointers ---
-    const sycl::half * mask_data = nullptr;
     int64_t mask_head_stride = 0;
     int64_t mask_row_stride  = 0;
     int     mask_n_heads     = 0;
@@ -547,7 +546,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
             // 1. Pack all GQA Q heads into fp16 (full n_query_rows)
             mkl_fa_pack_q_fp16(stream,
                 Q_head_f16_ptr, Q_batch,
-                n_queries, n_query_rows, DKQ,
+                n_queries, DKQ,
                 gqa_ratio, kvh_base_head,
                 q_scale, q_row_stride, q_head_stride, wg_size);
 
@@ -605,7 +604,7 @@ void ggml_sycl_flash_attn_ext_mkl(ggml_backend_sycl_context & ctx, ggml_tensor *
                             KQ_max_ptr, KQ_sum_ptr, VKQ_accum_ptr,
                             q0, q_rows, n_queries, DV,
                             this_chunk, chunk_start,
-                            kvh_base_head, gqa_ratio,
+                            kvh_base_head,
                             mask_batch, mask_head_stride,
                             mask_row_stride, mask_n_heads,
                             logit_softcap, wg_size);

ggml/src/ggml-sycl/fattn-onednn.cpp

@@ -21,14 +21,6 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
     if (!g_ggml_sycl_fa_onednn) {
         return false;
     }
-    // Battlemage (Xe2) only, for now. On other Intel archs oneDNN's fused SDPA returns wrong results
-    // for some shapes (e.g. head_dim=64 on Arc / xe_hpg) -- an oneDNN bug tracked upstream at
-    // https://github.com/uxlfoundation/oneDNN/issues/5510. Remove this hardware limitation once that
-    // is fixed; until then non-BMG archs fall back to the existing FA kernel.
-    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
-    if (arch != gpu_arch::intel_gpu_bmg_g21 && arch != gpu_arch::intel_gpu_bmg_g31) {
-        return false;
-    }
     const ggml_tensor * Q     = dst->src[0];
     const ggml_tensor * K     = dst->src[1];
     const ggml_tensor * V     = dst->src[2];
@@ -60,6 +52,17 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
             }
         }
     }
+    // This is the improved SPDA gate. Rather than gating Alchemist GPUs from all SPDA features, we instead target only the failing shapes.
+    // If the GPU being assessed isn't in the grouping below, it has full access to all SPDA shapes. Otherwise, if it's an Alchemist GPU, we block only the shapes with head sizes that fail.
+    // It is much easier to compare the device to a small list of failing cases than to define all the passing ones.
+    const gpu_arch arch = ggml_sycl_info().devices[ggml_sycl_get_device()].hw_info.arch;
+    bool support_spda = !(arch == gpu_arch::intel_gpu_dg2_g10 ||
+                         arch == gpu_arch::intel_gpu_dg2_g11 ||
+                         arch == gpu_arch::intel_gpu_dg2_g12);
+
+    if (!support_spda && K->ne[0] == 64) {
+        return false;
+    }
     // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
     // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
     // some stacks; past the cap we fall back to the native FA kernel instead.

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -921,16 +921,16 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
 
     void * dev_ptr;
     if (use_usm_system) {
-        GGML_SYCL_DEBUG("[SYCL] allocating %lu Bytes with USM system\n", size);
+        GGML_SYCL_DEBUG("[SYCL] allocating %zu Bytes with USM system\n", size);
         dev_ptr = (void *)aligned_malloc_host(alignment, aligned_size);
         if (!dev_ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
             return nullptr;
         }
     } else {
         SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)ggml_sycl_malloc_device(size, *stream)));
         if (!dev_ptr) {
-          GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+          GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
           return nullptr;
         }
     }
@@ -1177,7 +1177,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
         SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)ggml_sycl_malloc_device(size, *stream)));
         if (!buf) {
             char err_buf[1024];
-            snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+            snprintf(err_buf, 1023, "%s: can't allocate %zu Bytes of memory on device\n", __func__, size);
             throw std::runtime_error(err_buf);
         }
         // set padding to 0 to avoid possible NaN values
@@ -1517,8 +1517,13 @@ static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggm
 }
 
 static size_t ggml_backend_sycl_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
-    return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+
+    if (g_ggml_sycl_enable_host_pinned_mem) {
+        ggml_backend_sycl_device_context * dev_ctx = (ggml_backend_sycl_device_context *) buft->device->context;
+        return dpct::dev_mgr::instance().get_device(dev_ctx->device).get_max_mem_alloc_size();
+    } else {
+        return SIZE_MAX;
+    }
 }
 
 ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
@@ -1646,7 +1651,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
 
         SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *)ggml_sycl_malloc_device(look_ahead_size, *qptr)));
         if (!ptr) {
-            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
+            GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
             return nullptr;
         }
 
@@ -1658,7 +1663,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
                 (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
 #endif
 
-        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
+        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%zu, return %p\n", look_ahead_size, ptr);
         return ptr;
     }
 
@@ -1838,7 +1843,7 @@ struct ggml_sycl_pool_host : public ggml_sycl_pool {
 
             SYCL_CHECK(CHECK_TRY_ERROR(ptr = (void *) sycl::malloc_host(size, *qptr)));
             if (!ptr) {
-                GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on host\n", __func__, size);
+                GGML_LOG_ERROR("%s: can't allocate %zu Bytes of memory on host\n", __func__, size);
                 return nullptr;
             }
             pool_size += size;
@@ -2774,9 +2779,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
         {
+#if GGML_SYCL_DNNL
             const int64_t gemm_flops = (int64_t)row_diff * src1_ncols * ne10;
             const bool use_mkl_direct = gemm_flops < 256 * 256 * 256;
-#if GGML_SYCL_DNNL
             if (g_ggml_sycl_enable_dnn && !use_mkl_direct) {
                 DnnlGemmWrapper::row_gemm(ctx, row_diff, src1_ncols, ne10, src0_ddf_i,
                                           DnnlGemmWrapper::to_dt<float>(), src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
@@ -3513,7 +3518,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
     float *            dst_ddf  = static_cast<float *>(dst->data);
 
     const sycl::half * src1_f16       = static_cast<const sycl::half *>(src1->data);
+#if GGML_SYCL_DNNL
     const size_t       type_size_src0 = ggml_type_size(src0->type);
+#endif
     const size_t       type_size_src1 = ggml_type_size(src1->type);
 
     bool is_src0_cont_2 = ggml_is_contiguous_2(src0);
@@ -3530,6 +3537,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
         scope_op_debug_print    scope_dbg_print(__func__, "/to_fp16_nc_sycl", dst, /*num_src=*/2,
                                                 " : converting src1 to fp16");
 
+#if GGML_SYCL_DNNL
         // iterate tensor dims and find the slowest moving dim and stride
         int last_dim=0;
         int last_str=0;
@@ -3549,7 +3557,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             }
 
         }
-#if GGML_SYCL_DNNL
         // oneDNN handles strided data and does not need overhead of ggml_get_to_fp16_nc_sycl
         const int64_t ne_src1 = src1->nb[last_str] * src1->ne[last_dim] / type_size_src1;
         src1_f16_alloc.alloc(ne_src1);
@@ -3789,6 +3796,7 @@ inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
         case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
@@ -3802,8 +3810,10 @@ inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
 static bool ggml_sycl_supports_reorder_esimd(enum ggml_type type) {
 #ifdef GGML_SYCL_DMMV_HAS_ESIMD
     switch (type) {
+        case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
             return true;
         default:
@@ -6242,8 +6252,6 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, cons
         }
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK:
-            // FIXME: support ggml_rope_set_offset
-            return ((const int32_t *) op->op_params)[15] == 0;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_3D:
         case GGML_OP_UPSCALE:

ggml/src/ggml-sycl/im2col.cpp

@@ -85,7 +85,7 @@ static void im2col_sycl(const float *   x,
     */
     stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                            sycl::range<3>(1, 1, MIN(IC_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                              im2col_kernel(x, dst, IC, IW, IH, OH, OW, KW, KH, IC_IH_IW, IH_IW, N_OH, KH_KW, IC_KH_KW,
                                            s0, s1, p0, p1, d0, d1);
                          });
@@ -271,7 +271,7 @@ static void im2col_3d_sycl(const float *   src,
     */
     stream->parallel_for(sycl::nd_range<3>(block_nums * sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE)),
                                            sycl::range<3>(1, 1, MIN(IC_KD_KH_KW, SYCL_IM2COL_BLOCK_SIZE))),
-                         [=](sycl::nd_item<3> item_ct1) {
+                         [=](sycl::nd_item<3>) {
                              im2col_3d_kernel(src, dst, N, IC, ID, IH, IW, OC, KD, KH, KW, OD, OH, OW, OH_OW, KD_KH_KW,
                                               ID_IH_IW, KH_KW, IH_IW, IC_ID_IH_IW, IC_KD_KH_KW, OW_KD_KH_KW,
                                               OD_OH_OW_IC_KD_KH_KW, OH_OW_IC_KD_KH_KW, OW_IC_KD_KH_KW, N_OD_OH, OD_OH,

ggml/src/ggml-sycl/mmvq.cpp

@@ -1401,6 +1401,65 @@ static void mul_mat_vec_q2_K_q8_1_sycl_switch_ncols(
     }
 }
 
+static void reorder_mul_mat_vec_q2_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+                                               const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+
+    // Round up to a whole number of subgroup-sized workgroups; out-of-range rows are skipped inside the kernel.
+    constexpr size_t num_subgroups = WARP_SIZE;
+    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups);
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q2_K>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
+    });
+}
+
+template <int ncols_dst>
+static void reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols(
+        const void * vx, const void * vy, float * dst,
+        const int ncols, const int nrows,
+        const int stride_col_y_bytes, const int stride_col_dst,
+        dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    constexpr size_t num_subgroups = WARP_SIZE;
+    const int block_num_y = ceil_div(nrows, GGML_SYCL_MMV_Y * (int) num_subgroups);
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder_ncols<reorder_vec_dot_q_sycl<GGML_TYPE_Q2_K>, ncols_dst>(
+                                 vx, /*vgate=*/ nullptr, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst,
+                                 /*glu_op=*/ GGML_GLU_OP_SWIGLU, nd_item);
+                         });
+    });
+}
+
+static void reorder_mul_mat_vec_q2_k_q8_1_sycl_switch_ncols(
+        const void * vx, const void * vy, float * dst,
+        const int ncols, const int nrows, const int ncols_dst,
+        const int stride_col_y_bytes, const int stride_col_dst,
+        dpct::queue_ptr stream) {
+    switch (ncols_dst) {
+        case 1: reorder_mul_mat_vec_q2_k_q8_1_sycl(vx, vy, dst, ncols, nrows, stream); break;
+        case 2: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<2>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 3: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<3>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 4: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<4>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 5: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<5>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 6: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<6>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 7: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<7>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        case 8: reorder_mul_mat_vec_q2_k_q8_1_sycl_ncols<8>(vx, vy, dst, ncols, nrows, stride_col_y_bytes, stride_col_dst, stream); break;
+        default: GGML_ABORT("unsupported ncols_dst=%d for Q2_K reorder multi-col MMVQ", ncols_dst);
+    }
+}
+
 static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
                                        float *dst, const int ncols,
                                        const int nrows,
@@ -2297,7 +2356,21 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                 }
                 break;
             case GGML_TYPE_Q2_K:
-                if (i == 0 && src1_ncols > 1 && src1_ncols <= 8) {
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    if (i == 0 && src1_ncols > 1 && src1_ncols <= 8) {
+                        const int stride_col_y_bytes = src1_padded_col_size * q8_1_ts / q8_1_bs;
+                        const int stride_col_dst     = dst->ne[0];
+                        GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q2_k_q8_1_sycl_switch_ncols ncols=%d\n", (int)src1_ncols);
+                        reorder_mul_mat_vec_q2_k_q8_1_sycl_switch_ncols(
+                            src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff,
+                            src1_ncols, stride_col_y_bytes, stride_col_dst, stream);
+                        return;
+                    } else {
+                        GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q2_k_q8_1_sycl\n");
+                        reorder_mul_mat_vec_q2_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                    }
+                } else if (i == 0 && src1_ncols > 1 && src1_ncols <= 8) {
                     const int stride_col_y   = src1_padded_col_size / QK8_1;
                     const int stride_col_dst = dst->ne[0];
                     GGML_SYCL_DEBUG("Calling mul_mat_vec_q2_K_q8_1_sycl_switch_ncols ncols=%d\n", (int)src1_ncols);
@@ -2306,6 +2379,7 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                         src1_ncols, stride_col_y, stride_col_dst, stream);
                     return;
                 } else if (i == 0 || src1_ncols == 1) {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q2_K_q8_1_sycl\n");
                     mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                 }
                 break;

ggml/src/ggml-sycl/norm.cpp

@@ -7,9 +7,6 @@ static void norm_f32(const float* x, float* dst, const int ncols,
     const int64_t dst_stride_col, const int64_t dst_stride_row, const int64_t dst_stride_channel, const int64_t dst_stride_sample,
     const float eps, const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
 
-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
     const int nthreads = item_ct1.get_local_range(2);
     const int sample  = item_ct1.get_group(0);
     const int channel = item_ct1.get_group(1);
@@ -155,9 +152,6 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols,
     const float* mul = nullptr, const int64_t mul_stride_row = 0, const int64_t mul_stride_channel = 0,
     const int64_t mul_stride_sample = 0, const int mul_nrows = 0, const int mul_nchannels = 0, const int mul_nsamples = 0) {
 
-    const int nrows = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
-
     const int sample  = item_ct1.get_group(0);
     const int channel = item_ct1.get_group(1);
     const int row     = item_ct1.get_group(2);
@@ -225,8 +219,6 @@ static void l2_norm_f32(const float * x, float * dst, const int ncols,
     const int64_t src_stride_sample, const int64_t dst_stride_col, const int64_t dst_stride_row,
     const int64_t dst_stride_channel, const int64_t dst_stride_sample, const float eps,
     const sycl::nd_item<3>& item_ct1, float* s_sum, const int block_size) {
-    const int nrows     = item_ct1.get_group_range(2);
-    const int nchannels = item_ct1.get_group_range(1);
 
     const int row     = item_ct1.get_group(2);
     const int channel = item_ct1.get_group(1);

ggml/src/ggml-sycl/quants.hpp

@@ -58,6 +58,29 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
     static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
 };
 
+template <> struct block_q_t<GGML_TYPE_Q2_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI2_K;
+        static constexpr uint32_t qr       = QR2_K;
+        static constexpr uint32_t vdr_mmvq = 1;
+    };
+
+    // Reordered layout: [qs (QK_K/4 per block)] [scales (QK_K/16 per block)] [dm]
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* n_blocks */) {
+        return { block_index * (QK_K / 4), 0 };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / QK_K));
+        auto total_qs_bytes = nblocks * (QK_K / 4);
+        return { total_qs_bytes + block_index * (QK_K / 16),
+                 total_qs_bytes + nblocks * (QK_K / 16) + block_index * sizeof(ggml_half2) };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
+
 template <> struct block_q_t<GGML_TYPE_Q3_K> {
     struct traits {
         static constexpr uint32_t qk       = QK_K;

ggml/src/ggml-sycl/rope.cpp

@@ -41,7 +41,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
                       const int ne02, const int s01, const int s02,
                       const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                       const float freq_scale, const float ext_factor,
                       const float attn_factor, const rope_corr_dims corr_dims,
                       const float theta_scale, const float *freq_factors,
@@ -78,19 +78,21 @@ static void rope_norm(const T *x, D *dst, const int ne00, const int ne01,
             ggml_sycl_memcpy_1<4>(dst + idst, &v);
         }
     };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
         store_coaelsced(x[ix + 0], x[ix + 1]);
         return;
     }
 
-    const float theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+    const int iw = i0 - n_offs; // relative idx
 
-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;
 
     float cos_theta;
     float sin_theta;
 
-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                        ext_factor, attn_factor, cos_theta, sin_theta);
 
     const float x0 = x[ix + 0];
@@ -104,7 +106,7 @@ template <bool forward, bool has_ff, typename T, typename D>
 static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
                       const int ne02, const int s01, const int s02,
                       const int s03, const int s1, const int s2, const int s3,
-                      const int n_dims, const int32_t *pos,
+                      const int n_dims, const int n_offs, const int32_t *pos,
                       const float freq_scale, const float ext_factor,
                       const float attn_factor, const rope_corr_dims corr_dims,
                       const float theta_scale, const float *freq_factors,
@@ -132,35 +134,38 @@ static void rope_neox(const T *x, D *dst, const int ne00, const int ne01,
         idst += row_indices[i2] * set_rows_stride;
     }
 
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
         dst[idst + i0 / 2 + 0] = ggml_sycl_cast<D>(x[ix + i0 / 2 + 0]);
         dst[idst + i0 / 2 + 1] = ggml_sycl_cast<D>(x[ix + i0 / 2 + 1]);
 
         return;
     }
 
-    const float theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+    const int iw = i0 - n_offs; // relative idx
 
-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;
 
     float cos_theta;
     float sin_theta;
 
-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                        ext_factor, attn_factor, cos_theta, sin_theta);
 
-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims / 2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs / 2 + 0];
+    const float x1 = x[ix + n_offs / 2 + n_dims / 2];
 
-    dst[idst + 0] = ggml_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs / 2 + 0] = ggml_sycl_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs / 2 + n_dims / 2] = ggml_sycl_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }
 
 template <bool forward, bool has_ff, typename T>
 static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
                        const int ne02, const int s01, const int s02,
                        const int s03, const int s1, const int s2, const int s3,
-                       const int n_dims, const int32_t *pos,
+                       const int n_dims, const int n_offs, const int32_t *pos,
                        const float freq_scale, const float ext_factor,
                        const float attn_factor, const rope_corr_dims corr_dims,
                        const float theta_scale, const float *freq_factors,
@@ -183,54 +188,57 @@ static void rope_multi(const T *x, T *dst, const int ne00, const int ne01,
     int idst = i0 / 2 + i1 * s1 + i2 * s2 + i3 * s3;
     const int ix = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;
 
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
         dst[idst + i0 / 2 + 0] = x[ix + i0 / 2 + 0];
         dst[idst + i0 / 2 + 1] = x[ix + i0 / 2 + 1];
 
         return;
     }
 
+    const int iw = i0 - n_offs; // relative idx
+
     const int sect_dims =
         sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
     const int sec_w = sections.v[1] + sections.v[0];
-    const int sector = (i0 / 2) % sect_dims;
+    const int sector = (iw / 2) % sect_dims;
 
     float theta_base = 0.0;
     if (is_imrope) {
         if (sector % 3 == 1 && sector < 3 * sections.v[1]) { // h
-            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
         } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) { // w
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
         } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) { // t
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
         } else {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, iw / 2.0f);
         }
     } else {
         if (sector < sections.v[0]) {
-            theta_base = pos[i2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * dpct::pow(theta_scale, iw / 2.0f);
         } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * dpct::pow(theta_scale, iw / 2.0f);
         } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * dpct::pow(theta_scale, iw / 2.0f);
         } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * dpct::pow(theta_scale, iw / 2.0f);
         }
     }
 
-    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    const float freq_factor = has_ff ? freq_factors[iw / 2] : 1.0f;
 
     float cos_theta;
     float sin_theta;
 
-    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, i0,
+    rope_yarn<forward>(theta_base / freq_factor, freq_scale, corr_dims, iw,
                        ext_factor, attn_factor, cos_theta, sin_theta);
 
-    const float x0 = x[ix + 0];
-    const float x1 = x[ix + n_dims / 2];
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is n_offs + iw/2 = i0/2 + n_offs/2
+    const float x0 = x[ix + n_offs / 2 + 0];
+    const float x1 = x[ix + n_offs / 2 + n_dims / 2];
 
-    dst[idst + 0] = x0 * cos_theta - x1 * sin_theta;
-    dst[idst + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
+    dst[idst + n_offs / 2 + 0] = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_offs / 2 + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
 }
 
 template <bool forward, bool has_ff, typename T>
@@ -293,7 +301,7 @@ static void
 rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                const int ne02, const int s01, const int s02, const int s03,
                const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
                const float freq_base, const float ext_factor,
                const float attn_factor, const rope_corr_dims corr_dims,
                const float *freq_factors, const int64_t *row_indices,
@@ -313,7 +321,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_norm<forward, false>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, row_indices, set_rows_stride);
             });
     } else {
@@ -323,7 +331,7 @@ rope_norm_sycl(const T *x, D *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_norm<forward, true>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, row_indices, set_rows_stride);
             });
     }
@@ -334,7 +342,7 @@ static void
 rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                const int ne02, const int s01, const int s02, const int s03,
                const int s1, const int s2, const int s3, const int n_dims,
-               const int nr, const int32_t *pos, const float freq_scale,
+               const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
                const float freq_base, const float ext_factor,
                const float attn_factor, const rope_corr_dims corr_dims,
                const float *freq_factors, const int64_t *row_indices,
@@ -354,7 +362,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_neox<forward, false>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, row_indices, set_rows_stride);
             });
     } else {
@@ -364,7 +372,7 @@ rope_neox_sycl(const T *x, D *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_neox<forward, true>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, row_indices, set_rows_stride);
             });
     }
@@ -375,7 +383,7 @@ static void
 rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                 const int ne02, const int s01, const int s02, const int s03,
                 const int s1, const int s2, const int s3, const int n_dims,
-                const int nr, const int32_t *pos, const float freq_scale,
+                const int n_offs, const int nr, const int32_t *pos, const float freq_scale,
                 const float freq_base, const float ext_factor,
                 const float attn_factor, const rope_corr_dims corr_dims,
                 const float *freq_factors, const mrope_sections sections,
@@ -395,7 +403,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_multi<forward, false, T>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, sections, is_imrope);
             });
     } else {
@@ -405,7 +413,7 @@ rope_multi_sycl(const T *x, T *dst, const int ne00, const int ne01,
                 GGML_UNUSED(item_ct1);
                 rope_multi<forward, true, T>(
                     x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims,
-                    pos, freq_scale, ext_factor, attn_factor, corr_dims,
+                    n_offs, pos, freq_scale, ext_factor, attn_factor, corr_dims,
                     theta_scale, freq_factors, sections, is_imrope);
             });
     }
@@ -497,6 +505,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
     const int n_dims = ((int32_t *)dst->op_params)[1];
     const int mode = ((int32_t *)dst->op_params)[2];
     const int n_ctx_orig = ((int32_t *)dst->op_params)[4];
+    const int n_offs = ((int32_t *)dst->op_params)[15];
     mrope_sections sections;
 
     float freq_base;
@@ -526,6 +535,7 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
 
     if (is_vision) {
         GGML_ASSERT(n_dims == ne00 / 2);
+        GGML_ASSERT(n_offs == 0); // offset not supported for vision, as the rotated pairs span the whole row
     }
 
     const int32_t *pos = (const int32_t *)src1_d;
@@ -545,19 +555,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
         if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
             rope_neox_sycl<forward, float, float>(
                 (const float *)src0_d, (float *)dst_d, ne00, ne01, ne02, s01,
-                s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
                 set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
             rope_neox_sycl<forward, float, sycl::half>(
                 (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                 freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                 row_indices, set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
             rope_neox_sycl<forward, sycl::half, sycl::half>(
                 (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                 freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                 row_indices, set_rows_stride, stream);
         } else {
@@ -568,13 +578,13 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
         if (src0->type == GGML_TYPE_F32) {
             rope_multi_sycl<forward>((const float *)src0_d, (float *)dst_d,
                                      ne00, ne01, ne02, s01, s02, s03, s1, s2,
-                                     s3, n_dims, nr, pos, freq_scale, freq_base,
+                                     s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                                      ext_factor, attn_factor, corr_dims,
                                      freq_factors, sections, is_imrope, stream);
         } else if (src0->type == GGML_TYPE_F16) {
             rope_multi_sycl<forward>(
                 (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                 freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                 sections, is_imrope, stream);
         } else {
@@ -602,19 +612,19 @@ void ggml_sycl_op_rope_impl(ggml_backend_sycl_context &ctx, ggml_tensor *dst,
         if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
             rope_norm_sycl<forward, float, float>(
                 (const float *)src0_d, (float *)dst_d, ne00, ne01, ne02, s01,
-                s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base,
                 ext_factor, attn_factor, corr_dims, freq_factors, row_indices,
                 set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
             rope_norm_sycl<forward, float, sycl::half>(
                 (const float *)src0_d, (sycl::half *)dst_d, ne00, ne01, ne02,
-                s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                 freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                 row_indices, set_rows_stride, stream);
         } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
             rope_norm_sycl<forward, sycl::half, sycl::half>(
                 (const sycl::half *)src0_d, (sycl::half *)dst_d, ne00, ne01,
-                ne02, s01, s02, s03, s1, s2, s3, n_dims, nr, pos, freq_scale,
+                ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, nr, pos, freq_scale,
                 freq_base, ext_factor, attn_factor, corr_dims, freq_factors,
                 row_indices, set_rows_stride, stream);
         } else {

ggml/src/ggml-sycl/set_rows.cpp

@@ -291,7 +291,7 @@ static void set_rows_sycl(
 
     stream->parallel_for(
         sycl::nd_range<1>(grid_size * block_size, block_size),
-        [=](sycl::nd_item<1> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+        [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
             k_set_rows<TIn, TIdx, TOut>(
                 src0_d, src1_d, dst_d,
                 ne00, ne01, ne02,

ggml/src/ggml-sycl/vecdotq.hpp

@@ -429,6 +429,39 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q8_0> {
     }
 };
 
+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q2_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q2_K;
+
+    using q2_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q2_K>;
+    using q2_k_traits = typename q2_k_block::traits;
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t *    base   = static_cast<const uint8_t *>(vbq);
+        const uint8_t *    qs     = base + ibx_offset.first;
+        const uint8_t *    scales = base + d_offset.first;
+        const ggml_half2 * dm     = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
+
+        const int bq8_offset   = QR2_K * (iqs / QI8_1);
+        const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1 / 2);
+
+        const int v = get_int_from_uint8_aligned(qs, iqs);
+
+        int   u[QR2_K];
+        float d8[QR2_K];
+
+#pragma unroll
+        for (int i = 0; i < QR2_K; ++i) {
+            const int8_t * quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            u[i]                          = get_int_from_int8_aligned(quant_base_ptr, iqs % QI8_1);
+            d8[i]                         = (*(q8_1_ds + bq8_offset + i))[0];
+        }
+
+        return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales + scale_offset, *dm, d8);
+    }
+};
+
 template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q3_K> {
     static constexpr ggml_type gtype = GGML_TYPE_Q3_K;
 

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -200,8 +200,11 @@ if (Vulkan_FOUND)
     set (_ggml_vk_header     "${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp")
     set (_ggml_vk_input_dir  "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders")
     set (_ggml_vk_output_dir "${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv")
+    set (_ggml_vk_generated_shader_files ${_ggml_vk_header})
 
     file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
+    set_source_files_properties(${_ggml_vk_shader_files} PROPERTIES HEADER_FILE_ONLY TRUE)
+    target_sources(ggml-vulkan PRIVATE ${_ggml_vk_shader_files})
 
     # Because external projects do not provide source-level tracking,
     # the vulkan-shaders-gen sources need to be explicitly added to
@@ -241,8 +244,11 @@ if (Vulkan_FOUND)
             COMMENT "Generate vulkan shaders for ${file}"
         )
         target_sources(ggml-vulkan PRIVATE ${_ggml_vk_target_cpp})
+        list(APPEND _ggml_vk_generated_shader_files ${_ggml_vk_target_cpp})
     endforeach()
 
+    source_group("Vulkan shaders" FILES ${_ggml_vk_shader_files})
+    source_group("Generated Vulkan shaders" FILES ${_ggml_vk_generated_shader_files})
 else()
     message(WARNING "Vulkan not found")
 endif()

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -913,6 +913,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_quantize_q8_1_x4;
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_transpose[GGML_TYPE_COUNT]; // fused dequant+transpose for FA quant-KV
     vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[DMMV_WG_SIZE_COUNT][GGML_TYPE_COUNT];
@@ -954,6 +955,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_diag[2];
     vk_pipeline pipeline_clamp[2];
     vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_pad_reflect_1d_f32;
     vk_pipeline pipeline_roll_f32;
     vk_pipeline pipeline_repeat_i32, pipeline_repeat_back_f32;
     vk_pipeline pipeline_repeat_i16;
@@ -3384,10 +3386,10 @@ static void ggml_vk_queue_command_pools_cleanup(vk_device& device) {
     // Arbitrary frequency to cleanup/reuse command buffers
     static constexpr uint32_t cleanup_frequency = 10;
 
-    if (device->compute_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+    if (device->compute_queue && device->compute_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
         ggml_vk_command_pool_cleanup(device, device->compute_queue->cmd_pool);
     }
-    if (device->transfer_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
+    if (device->transfer_queue && device->transfer_queue->cmd_pool.buffers_in_use() >= cleanup_frequency) {
         ggml_vk_command_pool_cleanup(device, device->transfer_queue->cmd_pool);
     }
 }
@@ -5391,6 +5393,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_0], "dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_transpose[GGML_TYPE_Q8_0], "dequant_q8_0_transpose", dequant_q8_0_transpose_len, dequant_q8_0_transpose_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_k", dequant_q2_k_len, dequant_q2_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_TQ2_0], "dequant_tq2_0", dequant_tq2_0_len, dequant_tq2_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_k", dequant_q3_k_len, dequant_q3_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
@@ -5628,6 +5631,7 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
     ggml_vk_create_pipeline(device, device->pipeline_diag[1], "diag_f16", diag_f16_len, diag_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_pad_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_pad_reflect_1d_f32, "pad_reflect_1d_f32", pad_reflect_1d_f32_len, pad_reflect_1d_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_roll_f32, "roll_f32", roll_f32_len, roll_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
@@ -10823,9 +10827,32 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     const bool f32acc = !ctx->device->fp16 || dst->op_params[3] == GGML_PREC_F32 || k->type == GGML_TYPE_BF16;
 
+    // dequant K/V once into an f16 scratch, reordered KV layout so FA can read without a stride
+    auto is_dense_kv_cache = [](const ggml_tensor * t) {
+        return t->nb[0] == ggml_type_size(t->type) &&
+               t->nb[2] == ggml_row_size(t->type, t->ne[0]) &&
+               t->nb[1] == t->nb[2] * t->ne[2] &&
+               t->nb[3] == t->nb[1] * t->ne[1];
+    };
+    const bool k_quant = k->type != GGML_TYPE_F16 && k->type != GGML_TYPE_BF16 && k->type != GGML_TYPE_F32;
+    const bool v_quant = v->type != GGML_TYPE_F16 && v->type != GGML_TYPE_BF16 && v->type != GGML_TYPE_F32;
+    const bool use_dequant_kv = k_quant && v_quant && neq1 >= 64 &&
+                                is_dense_kv_cache(k) && is_dense_kv_cache(v) &&
+                                (uint64_t)ggml_nelements(k) * sizeof(ggml_fp16_t) <= ctx->device->properties.limits.maxStorageBufferRange &&
+                                (uint64_t)ggml_nelements(v) * sizeof(ggml_fp16_t) <= ctx->device->properties.limits.maxStorageBufferRange &&
+                                ctx->device->pipeline_dequant_transpose[k->type] != nullptr &&
+                                ctx->device->pipeline_dequant_transpose[v->type] != nullptr &&
+                                // coopmat2 path does not benefit from the f16 scratch
+                                !ctx->device->coopmat2 &&
+                                // Intel Xe1 regresses, see PR 25494
+                                (ctx->device->vendor_id != VK_VENDOR_ID_INTEL ||
+                                 (ctx->device->coopmat_support && ctx->device->architecture != vk_device_architecture::INTEL_XE1));
+    const ggml_type k_type_eff = use_dequant_kv ? GGML_TYPE_F16 : k->type;
+    const ggml_type v_type_eff = use_dequant_kv ? GGML_TYPE_F16 : v->type;
+
     // For scalar/coopmat1 FA, we can use the "large" size to accommodate qga.
     // For coopmat2 FA, we always use the small size (which is still pretty large for gqa).
-    vk_fa_tuning_params tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, 512, KV, k->type, v->type, f32acc);
+    vk_fa_tuning_params tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, 512, KV, k_type_eff, v_type_eff, f32acc);
     const uint32_t max_gqa = std::min(tuning_params.block_rows, 32u);
 
     if (N <= 8 && qk_ratio > 1 && qk_ratio <= max_gqa &&
@@ -10838,7 +10865,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         workgroups_y /= gqa_ratio;
     }
 
-    tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, N, KV, k->type, v->type, f32acc);
+    tuning_params = get_fa_tuning_params(ctx->device, HSK, HSV, N, KV, k_type_eff, v_type_eff, f32acc);
 
     const uint32_t q_stride = (uint32_t)(nbq1 / ggml_type_size(q->type));
     uint32_t k_stride = (uint32_t)(nbk1 / ggml_type_size(k->type));
@@ -10852,6 +10879,17 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
         v_stride /= 4;
     }
 
+    uint32_t nbk2_eff = (uint32_t)nbk2, nbk3_eff = (uint32_t)nbk3;
+    uint32_t nbv2_eff = (uint32_t)nbv2, nbv3_eff = (uint32_t)nbv3;
+    if (use_dequant_kv) {
+        k_stride = HSK;
+        v_stride = HSV;
+        nbk2_eff = (uint32_t)((uint64_t)HSK * KV * sizeof(ggml_fp16_t));
+        nbk3_eff = (uint32_t)((uint64_t)HSK * KV * nek2 * sizeof(ggml_fp16_t));
+        nbv2_eff = (uint32_t)((uint64_t)HSV * KV * sizeof(ggml_fp16_t));
+        nbv3_eff = (uint32_t)((uint64_t)HSV * KV * nev2 * sizeof(ggml_fp16_t));
+    }
+
     const uint32_t alignment = tuning_params.block_cols;
     bool aligned = (KV % alignment) == 0 &&
                    // the "aligned" shader variant will forcibly align strides, for performance
@@ -10878,7 +10916,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     bool use_mask_opt = mask && nem1 >= 32 && nem0 * nem1 > 32768 && nem0 >= tuning_params.block_cols * 16
                         && (ctx->device->architecture != vk_device_architecture::AMD_GCN || HSK > 256 || HSV > 256);
     vk_fa_pipeline_state fa_pipeline_state = get_fa_pipeline_state(ctx->device, tuning_params, HSK, HSV, aligned, f32acc,
-                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0, k->type, v->type);
+                                                                   mask != nullptr, use_mask_opt, logit_softcap != 0, k_type_eff, v_type_eff);
 
     vk_pipeline pipeline = nullptr;
 
@@ -10982,6 +11020,34 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
     vk_subbuffer sinks_buf = sinks ? ggml_vk_tensor_subbuffer(ctx, sinks) : q_buf;
     vk_subbuffer mask_opt_buf = use_mask_opt ? ggml_vk_subbuffer(ctx, ctx->prealloc_y, 0) : q_buf;
 
+    if (use_dequant_kv) {
+        const uint64_t fp = sizeof(ggml_fp16_t);
+        const uint64_t k_f16_sz = (uint64_t)ggml_nelements(k) * fp;
+        const uint64_t v_f16_sz = (uint64_t)ggml_nelements(v) * fp;
+        if (ctx->prealloc_size_x < k_f16_sz + v_f16_sz) {
+            ctx->prealloc_size_x = k_f16_sz + v_f16_sz;
+            ggml_vk_preallocate_buffers(ctx, subctx);
+        }
+        vk_pipeline tr_k = ctx->device->pipeline_dequant_transpose[k->type];
+        vk_pipeline tr_v = ctx->device->pipeline_dequant_transpose[v->type];
+        ggml_pipeline_request_descriptor_sets(ctx, tr_k, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, tr_v, 1);
+        if (ctx->prealloc_x_need_sync) {
+            ggml_vk_sync_buffers(ctx, subctx);
+        }
+        vk_subbuffer k_dst = vk_subbuffer{ ctx->prealloc_x, 0,        k_f16_sz };
+        vk_subbuffer v_dst = vk_subbuffer{ ctx->prealloc_x, k_f16_sz, v_f16_sz };
+        const uint32_t k_nel = (uint32_t)ggml_nelements(k);
+        const uint32_t v_nel = (uint32_t)ggml_nelements(v);
+        { const std::vector<uint32_t> pc = { (uint32_t)HSK, (uint32_t)nek2, (uint32_t)KV, 0, k_nel };
+          ggml_vk_dispatch_pipeline(ctx, subctx, tr_k, { k_buf, k_dst }, pc, { k_nel, 1, 1 }); }
+        { const std::vector<uint32_t> pc = { (uint32_t)HSV, (uint32_t)nev2, (uint32_t)KV, 0, v_nel };
+          ggml_vk_dispatch_pipeline(ctx, subctx, tr_v, { v_buf, v_dst }, pc, { v_nel, 1, 1 }); }
+        ggml_vk_sync_buffers(ctx, subctx);
+        k_buf = k_dst;
+        v_buf = v_dst;
+    }
+
     uint32_t mask_n_head_log2 = ((sinks != nullptr) << 24) | n_head_log2;
 
     if (use_mask_opt)
@@ -11011,8 +11077,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                               (uint32_t)nev2, (uint32_t)nev3,
                                               nem1, nem2, nem3,
                                               q_stride, (uint32_t)nbq2, (uint32_t)nbq3,
-                                              k_stride, (uint32_t)nbk2, (uint32_t)nbk3,
-                                              v_stride, (uint32_t)nbv2, (uint32_t)nbv3,
+                                              k_stride, nbk2_eff, nbk3_eff,
+                                              v_stride, nbv2_eff, nbv3_eff,
                                               scale, max_bias, logit_softcap,
                                               mask_n_head_log2, m0, m1,
                                               gqa_ratio, split_kv, split_k };
@@ -11054,6 +11120,10 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                     {q_buf, k_buf, v_buf, mask_buf, sinks_buf, dst_buf, mask_opt_buf},
                                     pc, { workgroups_x, workgroups_y, workgroups_z });
     }
+
+    if (use_dequant_kv) {
+        ctx->prealloc_x_need_sync = true;
+    }
 }
 
 static vk_conv_shapes ggml_vk_conv_select_shape(ggml_backend_vk_context * ctx, uint32_t K, uint32_t NPQ) {
@@ -11268,6 +11338,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pad_f32;
         }
         return nullptr;
+    case GGML_OP_PAD_REFLECT_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_reflect_1d_f32;
+        }
+        return nullptr;
     case GGML_OP_ROLL:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_roll_f32;
@@ -12171,6 +12246,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     case GGML_OP_CLAMP:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_PAD:
+    case GGML_OP_PAD_REFLECT_1D:
     case GGML_OP_ROLL:
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
@@ -13043,6 +13119,17 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
     ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD, std::move(p));
 }
 
+static void ggml_vk_pad_reflect_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
+    const uint32_t p0 = (uint32_t)dst->op_params[0];
+    const uint32_t p1 = (uint32_t)dst->op_params[1];
+
+    vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst, ggml_nelements(dst));
+    memcpy(&p.param1, &p0, sizeof(float));
+    memcpy(&p.param2, &p1, sizeof(float));
+
+    ggml_vk_op_f32(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_PAD_REFLECT_1D, std::move(p));
+}
+
 static void ggml_vk_roll(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     const int32_t s0 = ggml_get_op_params_i32(dst, 0);
     const int32_t s1 = ggml_get_op_params_i32(dst, 1);
@@ -15452,6 +15539,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_PAD:
         ggml_vk_pad(ctx, compute_ctx, src0, node);
 
+        break;
+    case GGML_OP_PAD_REFLECT_1D:
+        ggml_vk_pad_reflect_1d(ctx, compute_ctx, src0, node);
+
         break;
     case GGML_OP_ROLL:
         ggml_vk_roll(ctx, compute_ctx, src0, node);
@@ -18378,6 +18469,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SCALE:
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ROLL:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_DIAG_MASK_INF:
@@ -19160,6 +19252,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else if (tensor->op == GGML_OP_PAD) {
             tensor_clone = ggml_pad_ext(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3],
                                                                 tensor->op_params[4], tensor->op_params[5], tensor->op_params[6], tensor->op_params[7]);
+        } else if (tensor->op == GGML_OP_PAD_REFLECT_1D) {
+            tensor_clone = ggml_pad_reflect_1d(ggml_ctx, src_clone[0], tensor->op_params[0], tensor->op_params[1]);
         } else if (tensor->op == GGML_OP_REPEAT) {
             tensor_clone = ggml_repeat(ggml_ctx, src_clone[0], tensor);
         } else if (tensor->op == GGML_OP_REPEAT_BACK) {

ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp

@@ -18,7 +18,18 @@ void main() {
         return;
     }
 
+#ifdef DEQUANT_TRANSPOSE
+    // read [HS, NH, KV, NS], write [HS, KV, NH, NS]
+    const uint HS = p.M, NH = p.K, KVn = p.stride_a;
+    const uint e0 = ib * 32;
+    const uint b_idx = (e0 % HS)
+                     + ((e0 / (HS * NH)) % KVn) * HS
+                     + ((e0 / HS) % NH) * (HS * KVn)
+                     + (e0 / (HS * NH * KVn)) * (HS * KVn * NH)
+                     + 16 * il;
+#else
     const uint b_idx = 1024*i + 32*ir + 16*il;
+#endif
 
     const float d = float(data_a[ib].d);