opencl: fix local size for norm (#27339 )

ui: Stores split refactor (#27240 )
* ui: Extract server stream lifecycle from chatStore into ChatStreamManager Discovery, attach/replay, resume retry and the remote-running snapshot formed a cohesive cluster inside chatStore. It now lives in chat-streams.svelte.ts as ChatStreamManager, owned by chatStore, which keeps the public entry points as delegates so components are unchanged. chatStore: 2877 -> 2418 lines. * ui: Extract user interaction gates from agenticStore into AgenticGates Tool permission requests, turn-limit continue prompts and queued steering messages are the state the loop waits on between turns. They had no coupling to session state, so they now live in agentic-gates.svelte.ts; agenticStore keeps delegates so components are unchanged. agenticStore: 1196 -> 1073 lines. * ui: Compose MCP resources under mcpStore.resources Resource state was a second import scope next to mcpStore. Consumers now go through mcpStore.resources, so the MCP surface is one store; mcp-resources.svelte.ts stays a separate file owned by mcpStore. * ui: Reorganize stores into domain namespaces * fix: Update stale doc comments * ui: Consolidate conv running-state into a chat activity ledger Running-state was split across chatStore.chatLoadingStates (local pipes), ChatStreamManager.remoteRunningConvs (backend sessions) and attachingConvs (attach lifecycle), unioned by hand in getAllLoadingChats and cross-cleaned by setChatLoading calling streams.clearRemoteRunning - the 'spinner ghosts until tab toggle' workaround. chatActivityStore now owns both sets with one transition per event: markLocal / localEnded (local pipe end also drops the stale remote hint, no cross-owner call) / applyRemoteSnapshot (diffed). The sidebar reads chatStore.activity.loadingConvs through the unchanged getAllLoadingChats entry point. Consequences: - isStreamingActive and its five manual writers are gone; isStreaming() now reports whether the active conversation has a live streaming pipe, which is what all four consumers (assistant row, stop action, context gauge, chat screen) actually check - isLoading/isReasoning become derived from the per-conv maps plus the active conversation, dropping the manual resync in syncLoadingStateForChat and clearUIState - attachingConvs and the last-attach coordination disappear from ChatStreamManager - getAllStreamingChats (no consumers) is removed * ui: Give store collaborators narrow host interfaces Collaborators took 'host: typeof <store>', i.e. the store's entire public surface, which is how chatStore's streamChatCompletion, createAssistantMessage, getApiOptions and setStreamingActive got widened to public. Replace with per-collaborator interfaces carrying only the members each one drives: - ChatStreamHost (chat/streams) - activity, processing, streaming states, abort controller, loading/streaming setters - ChatFlowsHost (chat/flows) - streaming core, message creation, per-conv state setters - McpHealthHost (mcp/health) - connection registry + reconnection - ModelPropsHost / ModelStatusHost (models) - model rows, feed updates; the managers write modalities/status back onto the host's rows, so those members stay writable - ConversationsPreferencesHost (conversations) - the active row and the conversation list The store classes now declare 'implements <Host>' so the contract is visible at the class level, and the 'import type { <store> }' back references in the collaborators disappear entirely - the host contract is local to each collaborator file, and collaborators can no longer reach around their slice. Members stay public (structural typing), but the collaborator side is now compiler-enforced. * test: Chat Activity store test * refactor: Cleanup * chore: Remove legacy architecture docs * ui: Memoize findMessageIndex for the streaming hot path Streaming looks up the same message index on every chunk, a linear scan of activeMessages each time. Cache the last lookup and reuse it after validating the id still sits at the same position (O(1)); any structural change to the array fails validation and falls back to a full scan. * ui: Throttle per-chunk stream state writes to localStorage saveStreamState ran JSON.stringify + a synchronous localStorage.setItem on every decoded chunk of the stream. The read loop now goes through a new saveStreamStateThrottled (one write per conversation per 500ms, latest value held pending); the public saveStreamState keeps its immediate-write contract for stream start and pre-fetch, and also resets the throttle window. A pending offset is force-flushed at resume boundaries (resumeStream reads the offset back from localStorage), on visibilitychange->hidden and on pagehide, so a reload always finds a usable offset. The resume offset only needs to be roughly current since the server retransmits from a line boundary and the client discards its partial line. Adds unit tests for the throttled/flush/clear interplay. * ui: Compute context gauge timing stats in one pass currentRead/Fresh/Cache/Output were separate deriveds, each running a full reverse scan of activeMessages for the last assistant timings, and cumulative ran its own forward scan plus an agentic filter - 4-5 O(n) passes per chunk while streaming. Replace with a single summarizeAssistantTimings() pass (last assistant timings, last agentic llm totals and the cumulative sums) feeding a shared derived snapshot. Semantics unchanged, including the live-stats overrides and the agentic llm-totals branch. * agentic : clear session state when a conversation is deleted Every conversation that ran an agentic flow left an AgenticSession in the store forever; clearSession was never called. conversationsStore now notifies deletion listeners and agenticStore drops the matching sessions, avoiding a circular import back into conversationsStore. * chat : extract ChatService.normalizeMessagesForApi The DB->API message normalization (convert + drop empty system messages) was duplicated in sendMessage, preEncode and the agentic flow. Extract it into one shared method and call it from all three. * sse : share record splitting and data extraction splitSseRecords and extractSseDataPayload centralize the record-boundary splitting and data: line extraction used by parseSseJsonStream and the models status feed. chat.service keeps its own line-based parser for resume support. * api : delegate apiFetchWithParams to apiFetch apiFetchWithParams duplicated apiFetch's headers/fetch/error handling body-for-body; it only differs in URL construction. Build the URL and delegate. * chat flows : dedupe title, timings and cleanup handling - conversationsStore.applyTitleFromContent centralizes the title-from-first- message logic duplicated in 5 places - ChatProcessingStore.applyStreamTimings centralizes the onTimings handler shared by the chat and continue flows - host.cleanupStreaming centralizes the loading/streaming/processing reset repeated across the continue flow's exit paths * conversations : centralize conversation update mirroring rename, pin, mcp override, reasoning effort and cwd all repeated the same write-DB-then-mirror-into-list-and-active dance. A single applyConversationUpdate(id, updates) on the host collapses all five and removes the forgot-to-mirror-one-field bug class. Drops the redundant array reassignment in setCwd (deep field assignment is reactive). * mcp : dedupe tool execution, server parsing and tool indexing - executeTool delegates to executeToolByName (only diff was argument parsing) - drop the private #parseServerSettings copy; use parseMcpServerSettings - cache getServers() keyed on the raw config value (hot path) - indexServerTools() unifies the three identical toolsIndex rebuild loops Assisted-by: Claude * mcp : share cursor pagination and tool indexing - MCPService.paginate() collapses the identical do-while loops in listAllResources and listAllResourceTemplates - promoteHealthCheckToConnection now uses indexServerTools like the other connect paths Assisted-by: Claude * database : share message parent-child bookkeeping - addChildToParent() dedups the append-to-children update in createMessageBranch and createSystemMessage - removeChildFromParent() dedups the remove-from-children cleanup in deleteMessage and deleteMessageCascading - bulkAdd the cloned messages when forking a conversation instead of one add per message Assisted-by: Claude * chore: Lint/format * fix: `pagehide` event from `window` * refactor: Api Fetch util * docs : rewrite architecture sections in README Update the high-level diagram, routes, hooks, stores, services and data flow tables to match the current UI structure (mcp/settings/search routes, agentic/tools/mcp stores, MCPService/ToolsService/SandboxService, /tools API). Fix stale architectural patterns for per-conversation state and modality validation. * chore : add ESLint rule for blank lines between accessors Enforce a blank line between consecutive class accessors. The core padding-line-between-statements rule does not cover class members, so a local rule is needed. * refactor : reorder store members and unify naming Order store class members as public fields, private fields, constructor, getters, public methods, then private methods. Normalize private naming to the `private` keyword (drop `#` and the `_` prefix where there is no matching public getter). Rename conversationsStore.init() to initialize() to match the other stores. * refactor : prefix lookup methods with get in agentic and chat stores Unify bare-name lookup methods with the get* prefix used across the other stores (mcp, models, tools, settings). Renames currentTurn, totalToolCalls, lastError, streamingToolCall, executingToolCallId, pendingPermissionRequest, pendingContinueRequest, pendingSteeringMessageContent, pendingSteeringMessageExtras in the agentic store and pendingMessageContent, pendingMessageExtras in the chat store. Updates the two consuming components and a doc comment. * refactor: Clean up comments in stores' and services' code * chore : add ESLint rule for class member ordering Enforce structural order (public fields -> private fields -> constructor -> getters -> setters -> public methods -> private methods) with alphabetical sorting within each group via perfectionist/sort-classes. Dependency detection keeps Svelte $derived fields in a valid dependency order instead of alphabetizing them, since Svelte rejects forward references. Assisted-by: Claude * refactor : reorder class members to match new ESLint rule Apply the sort-classes rule across stores, services, hooks and utils. Pure reordering - verified no logic changes by comparing sorted line multisets before/after. All tests and svelte-check pass.
2026-08-20 19:52:43 +02:00 · 2026-08-20 10:52:07 -07:00 · 2026-08-20 19:02:04 +02:00 · 2026-08-20 18:45:37 +02:00 · 2026-08-20 16:36:57 +02:00 · 2026-08-20 09:18:11 -05:00
295 changed files with 16189 additions and 14361 deletions
@@ -1,18 +1,18 @@
-ARG OPENVINO_VERSION_MAJOR=2026.2.1
-ARG OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3
+ARG OPENVINO_VERSION_MAJOR=2026.3
+ARG OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c
 ARG UBUNTU_VERSION=24.04

 # Intel GPU driver versions. https://github.com/intel/compute-runtime/releases
-ARG IGC_VERSION=v2.36.3
-ARG IGC_VERSION_FULL=2_2.36.3+21719
-ARG COMPUTE_RUNTIME_VERSION=26.22.38646.4
-ARG COMPUTE_RUNTIME_VERSION_FULL=26.22.38646.4-0
+ARG IGC_VERSION=v2.38.2
+ARG IGC_VERSION_FULL=2_2.38.2+22051
+ARG COMPUTE_RUNTIME_VERSION=26.27.39122.11
+ARG COMPUTE_RUNTIME_VERSION_FULL=26.27.39122.11-0
 ARG IGDGMM_VERSION=22.10.0

 # Intel NPU driver versions. https://github.com/intel/linux-npu-driver/releases
-ARG NPU_DRIVER_VERSION=v1.33.0
-ARG NPU_DRIVER_FULL=v1.33.0.20260529-26625960453
-ARG LIBZE1_VERSION=1.27.0-1~24.04~ppa2
+ARG NPU_DRIVER_VERSION=v1.35.0
+ARG NPU_DRIVER_FULL=v1.35.0.20260722-29947505341
+ARG LIBZE1_VERSION=1.28.2-1~24.04~ppa1

 # Optional proxy build arguments
 ARG http_proxy=
@@ -170,7 +170,7 @@ RUN --mount=type=cache,target=/var/cache/intel-npu,sharing=locked \
    fi; \
    DEB=/var/cache/intel-npu/libze1_${LIBZE1_VERSION}_amd64.deb; \
    if [ ! -f "$DEB" ]; then \
-        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260324T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
+        wget -q -O "$DEB" https://snapshot.ppa.launchpadcontent.net/kobuk-team/intel-graphics/ubuntu/20260606T100000Z/pool/main/l/level-zero-loader/libze1_${LIBZE1_VERSION}_amd64.deb; \
    fi; \
    mkdir /tmp/npu/ && cd /tmp/npu/ && tar -xf "$TGZ" && cp "$DEB" .; \
    apt-get update; \
@@ -6,8 +6,7 @@ inputs:
    required: true
  cuda_arch:
    description: "CUDA target architecture"
-    required: false
-    default: "x64"
+    required: true

 runs:
  using: "composite"
@@ -40,9 +40,9 @@ jobs:
    runs-on: ubuntu-24.04

    env:
-      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -69,8 +69,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -119,6 +119,7 @@ jobs:
          ./bin/llama-completion -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256

  windows:
+    name: windows / ${{ matrix.build }}
    runs-on: windows-2025

    env:
@@ -130,13 +131,13 @@ 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
@@ -22,6 +22,7 @@ env:

 jobs:
  cuda:
+    name: windows-cuda (${{ matrix.cuda }}, ${{ matrix.arch }})
    runs-on: windows-2022

    permissions:
@@ -29,7 +30,16 @@ jobs:

    strategy:
      matrix:
-        cuda: ['12.4', '13.3']
+        include:
+          - cuda: '12.4'
+            arch: x64
+            defines: '-DGGML_CUDA_CUB_3DOT2=ON'
+          - cuda: '13.3'
+            arch: x64
+            defines: ''
+          - cuda: '13.4'
+            arch: arm64
+            defines: '-DCMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc-cuda.cmake'

    steps:
      - name: Clone
@@ -39,12 +49,13 @@ jobs:
      - name: ccache
        uses: ggml-org/ccache-action@v1.2.21
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

      - name: Install Cuda Toolkit
        uses: ./.github/actions/windows-setup-cuda
        with:
          cuda_version: ${{ matrix.cuda }}
+          cuda_arch: ${{ matrix.arch }}

      - name: Install Ninja
        id: install_ninja
@@ -54,26 +65,21 @@ jobs:
      - name: Build
        id: cmake_build
        shell: cmd
-        # TODO: Remove GGML_CUDA_CUB_3DOT2 flag once CCCL 3.2 is bundled within CTK and that CTK version is used in this project
        run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
          cmake -S . -B build -G "Ninja Multi-Config" ^
-            -DLLAMA_BUILD_SERVER=ON ^
-            -DLLAMA_BUILD_BORINGSSL=ON ^
-            -DGGML_NATIVE=OFF ^
            -DGGML_BACKEND_DL=ON ^
-            -DGGML_CPU_ALL_VARIANTS=ON ^
+            -DGGML_NATIVE=OFF ^
+            -DGGML_CPU=OFF ^
            -DGGML_CUDA=ON ^
-            -DGGML_RPC=ON ^
-            -DGGML_CUDA_CUB_3DOT2=ON
+            -DLLAMA_BUILD_BORINGSSL=ON ${{ matrix.defines }}
          set /A NINJA_JOBS=%NUMBER_OF_PROCESSORS%-1
-          cmake --build build --config Release -j %NINJA_JOBS% -t ggml
-          cmake --build build --config Release
+          cmake --build build --config Release -j %NINJA_JOBS% --target ggml-cuda

      - name: ccache-clear
        uses: ./.github/actions/ccache-clear
        with:
-          key: release-windows-2022-x64-cuda-${{ matrix.cuda }}
+          key: release-windows-2022-${{ matrix.arch }}-cuda-${{ matrix.cuda }}

  hip:
    runs-on: windows-2022
@@ -39,8 +39,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -81,7 +81,7 @@ jobs:
        # TODO: fix and re-enable the `test-llama-archs` test below
        run: |
          cd ${{ github.workspace }}
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 2000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 2000

      - name: Test (GPU)
        id: cmake_test_gpu
@@ -89,15 +89,15 @@ jobs:
        run: |
          cd ${{ github.workspace }}
          export GGML_OPENVINO_DEVICE=GPU
-          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs" --verbose --timeout 3000
+          ctest --test-dir build/ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" --verbose --timeout 3000

  openvino-windows-2022:
    runs-on: windows-2022

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -166,4 +166,4 @@ jobs:
          call "%OPENVINO_ROOT%\setupvars.bat"

          cd build
-          ctest --test-dir ReleaseOV -L main -E "test-llama-archs" -C Release --verbose --timeout 3000
+          ctest --test-dir ReleaseOV -L main -E "test-llama-archs|test-recurrent-state-rollback-nemotron-h" -C Release --verbose --timeout 3000
@@ -288,8 +288,8 @@ jobs:

    env:
      # Sync versions in build.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Clone
@@ -394,6 +394,11 @@ jobs:
    name: Create shared tags from digests
    needs: [prepare_matrices, push_to_registry, create_tag]
    runs-on: ubuntu-24.04
+    permissions:
+      contents: read
+      packages: write
+      id-token: write
+      attestations: write
    strategy:
      fail-fast: false
      matrix:
@@ -428,6 +433,7 @@ jobs:
          password: ${{ secrets.GITHUB_TOKEN }}

      - name: Create tags from digests
+        id: create_tags
        shell: bash
        run: |
          set -euo pipefail
@@ -439,6 +445,7 @@ jobs:
          SRC_TAG="${{ needs.create_tag.outputs.source_tag }}"
          BUILD_DATE="${{ steps.build_date.outputs.date }}"
          COMMIT_SHA="${{ steps.checkout.outputs.commit }}"
+          echo "image_repo=${IMAGE_REPO}" >> "$GITHUB_OUTPUT"
          TAGS="${{ matrix.config.tag }}"
          ARCHES="${{ matrix.config.arches }}"
          DIGEST_GLOB="/tmp/digests/*.tsv"
@@ -505,6 +512,16 @@ jobs:

              echo "Creating ${merged_versioned_tag} from ${refs[*]}"
              docker buildx imagetools create "${annotations[@]}" --tag "${merged_versioned_tag}" "${refs[@]}"
+
+              if [[ "$tag_name" == "${TAGS%% *}" ]]; then
+                  local digest
+                  digest="$(docker buildx imagetools inspect "${merged_versioned_tag}" --format '{{.Manifest.Digest}}')"
+                  if [[ ! "$digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
+                      echo "Invalid digest for ${merged_versioned_tag}: ${digest}" >&2
+                      exit 1
+                  fi
+                  echo "${image_type}_digest=${digest}" >> "$GITHUB_OUTPUT"
+              fi
          }

          for tag in $TAGS; do
@@ -528,3 +545,24 @@ jobs:
          done
        env:
          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - name: Attest full image
+        if: ${{ matrix.config.full }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.full_digest }}
+
+      - name: Attest light image
+        if: ${{ matrix.config.light }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.light_digest }}
+
+      - name: Attest server image
+        if: ${{ matrix.config.server }}
+        uses: actions/attest@v4
+        with:
+          subject-name: ${{ steps.create_tags.outputs.image_repo }}
+          subject-digest: ${{ steps.create_tags.outputs.server_digest }}
@@ -446,8 +446,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -562,8 +562,8 @@ jobs:

    env:
      # Sync versions in build-openvino.yml, build-self-hosted.yml, release.yml, build-cache.yml, .devops/openvino.Dockerfile
-      OPENVINO_VERSION_MAJOR: "2026.2.1"
-      OPENVINO_VERSION_FULL: "2026.2.1.21919.ede283a88e3"
+      OPENVINO_VERSION_MAJOR: "2026.3"
+      OPENVINO_VERSION_FULL: "2026.3.0.22451.bd8d6542e3c"

    steps:
      - name: Set OpenVINO version output
@@ -681,6 +681,7 @@ jobs:
          name: llama-bin-win-openvino-${{ env.OPENVINO_VERSION_MAJOR }}-x64.zip

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

@@ -728,6 +729,7 @@ 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

@@ -739,7 +741,6 @@ jobs:
      - 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
@@ -1569,6 +1570,8 @@ jobs:
    # https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
    permissions:
        contents: write # for creating release
+        id-token: write
+        attestations: write

    runs-on: ubuntu-slim

@@ -1577,14 +1580,14 @@ jobs:
      - windows
      - windows-cpu
      - windows-cuda
-      #- windows-sycl
+      - windows-sycl
      - windows-rocm
      - windows-openvino
      #- ubuntu-22-rocm
      - ubuntu-cpu
      - ubuntu-vulkan
      - ubuntu-24-openvino
-      #- ubuntu-24-sycl
+      - ubuntu-24-sycl
      - android-arm64
      - macos-cpu
      - ios-xcode
@@ -1662,6 +1665,12 @@ jobs:
        run: |
          tar -czvf release/llama-${{ steps.tag.outputs.name }}-ui.tar.gz --transform "s,^\.,llama-${{ steps.tag.outputs.name }}," -C ./ui-dist .

+      - name: Attest release artifacts
+        id: attest
+        uses: actions/attest@v4
+        with:
+          subject-path: 'release/*'
+
      - name: Create and push git tag
        run: |
          TAG="${{ steps.tag.outputs.name }}"
@@ -1689,6 +1698,9 @@ jobs:
            **Website:**
            - <https://llama.app>

+            **Attestations:**
+            - <${{ steps.attest.outputs.attestation-url }}>
+
            **macOS/iOS:**
            - [macOS Apple Silicon (arm64)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.tar.gz)
            - macOS Apple Silicon (arm64, KleidiAI enabled) [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23780)
@@ -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>`
@@ -7,10 +7,11 @@
 <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)](https://github.com/ggml-org/llama.cpp/releases)
+[![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)
-[![Docker](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
-[![Winget](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.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)

 [manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3A0cc4m%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [compile times](https://github.com/ggml-org/llama.cpp-dev/blob/master/README-compile-times.md) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

@@ -190,7 +190,7 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

    # TODO: fix and re-enable the `test-llama-archs` test below
-    CTEST_EXTRA="-E test-llama-archs"
+    CTEST_EXTRA="-E test-llama-archs|test-recurrent-state-rollback-nemotron-h"
 fi

 ## helpers
@@ -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" )
@@ -2595,6 +2595,26 @@ 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",
@@ -4658,6 +4678,12 @@ void common_params_add_preset_options(std::vector<common_arg> & args) {
        [](common_params &, int) { /* unused */ }
    ).set_env(COMMON_ARG_PRESET_STOP_TIMEOUT).set_preset_only());

+    args.push_back(common_arg(
+        {"dedup-cache-models"}, "0|1",
+        "in server router mode, hide a cached model from the model list when this preset resolves to the same model file",
+        [](common_params &, const std::string &) { /* unused */ }
+    ).set_env(COMMON_ARG_PRESET_DEDUP_CACHE_MODELS).set_preset_only());
+
    // args.push_back(common_arg(
    //     {"pin"},
    //     "in server router mode, do not unload this model if models_max is exceeded",
@@ -11,8 +11,9 @@
 #include <memory>

 // pseudo-env variable to identify preset-only arguments
-#define COMMON_ARG_PRESET_LOAD_ON_STARTUP "__PRESET_LOAD_ON_STARTUP"
-#define COMMON_ARG_PRESET_STOP_TIMEOUT    "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_LOAD_ON_STARTUP    "__PRESET_LOAD_ON_STARTUP"
+#define COMMON_ARG_PRESET_STOP_TIMEOUT       "__PRESET_STOP_TIMEOUT"
+#define COMMON_ARG_PRESET_DEDUP_CACHE_MODELS "__PRESET_DEDUP_CACHE_MODELS"

 //
 // CLI argument parsing
@@ -1778,6 +1778,8 @@ void common_threadpools::init(llama_context * ctx, const common_params & params)
    struct ggml_threadpool_params tpp =
            ggml_threadpool_params_from_cpu_params(params.cpuparams);

+    // 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) {
@@ -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;
@@ -989,6 +989,26 @@ std::vector<common_cached_model_info> common_list_cached_models() {
    return result;
 }

+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file) {
+    auto [repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
+
+    auto files = hf_cache::get_cached_files(repo);
+    if (files.empty()) {
+        return "";
+    }
+
+    if (!hf_file.empty()) {
+        for (const auto & f : files) {
+            if (f.path == hf_file) {
+                return f.local_path;
+            }
+        }
+        return "";
+    }
+
+    return find_best_model(files, tag).local_path;
+}
+
 bool common_download_remove(const std::string & hf_repo_with_tag) {
    namespace fs = std::filesystem;

@@ -85,6 +85,10 @@ std::vector<std::string> common_download_get_all_parts(const std::string & url);
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();

+// resolve the local cached file path for a HF repo without network access (hf_file, if given, must match exactly)
+// returns an empty string if the model is not present in the cache
+std::string common_download_resolve_path(const std::string & hf_repo_with_tag, const std::string & hf_file = "");
+
 // download single file from url to local path
 // returns status code or -1 on error
 // skip_etag: if true, don't read/write .etag files (for HF cache where filename is the hash)
@@ -278,7 +278,9 @@ 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 +311,32 @@ 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 +373,42 @@ 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 +417,6 @@ 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 +473,42 @@ 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 +517,9 @@ 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 +530,19 @@ 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 +551,11 @@ 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 +581,22 @@ 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 +609,21 @@ 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 + ") \"\\\"\"");
    }

    /*
@@ -2649,6 +2649,10 @@ 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
@@ -57,6 +57,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3DSparkModel": "qwen",
    "DSparkDraftModel": "qwen",
    "DSparkSpeculator": "qwen",
+    "Lfm2DSparkDraftModel": "qwen",
    "DeepseekV4ForCausalLM": "deepseek",
    "DeepseekV4DSparkModel": "deepseek",
    "DistilBertForMaskedLM": "bert",
@@ -109,6 +110,8 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "GraniteSwitchForCausalLM": "granite",
    "GraniteSpeechForConditionalGeneration": "granite",
    "GraniteSpeechPlusForConditionalGeneration": "granite",
+    "GraniteSWAForCausalLM": "granite",
+    "GraniteMoeSWAForCausalLM": "granite",
    "Grok1ForCausalLM": "grok",
    "GrokForCausalLM": "grok",
    "GroveMoeForCausalLM": "grovemoe",
@@ -74,6 +74,108 @@ class GraniteModel(LlamaModel):
        return super().filter_tensors(item)


+@ModelBase.register("GraniteSWAForCausalLM")
+class GraniteSWAModel(GraniteModel):
+    """Conversion for IBM's GraniteSWAForCausalLM (interleaved sliding window attention)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if name.endswith("sinks"):
+            name += ".weight"
+
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        """GraniteSWA uses Granite parameters plus sliding window configuration."""
+        super().set_gguf_parameters()
+
+        # Add sliding_window from config
+        sliding_window = self.hparams.get("sliding_window", 128)
+        self.gguf_writer.add_sliding_window(sliding_window)
+        logger.info("gguf: (granite_swa) sliding_window = %s", sliding_window)
+
+        # Derive sliding_window_pattern from layer_types
+        if layer_types := self.hparams.get("layer_types"):
+            is_swa = [t == "sliding_attention" for t in layer_types]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern = %d SWA layers / %d total",
+                        sum(is_swa), len(is_swa))
+        else:
+            # Fall back to period-based pattern: i % 4 != 0
+            # This matches the transformers default pattern
+            n_layers = self.block_count
+            is_swa = [i % 4 != 0 for i in range(n_layers)]
+            self.gguf_writer.add_sliding_window_pattern(is_swa)
+            logger.info("gguf: (granite_swa) sliding_window_pattern (inferred) = %d SWA layers / %d total",
+                        sum(is_swa), n_layers)
+
+        # Add rope_pattern from no_rope_layers
+        if no_rope_layers := self.hparams.get("no_rope_layers"):
+            # Convert 1/0 to bool (1 = use RoPE, 0 = NoPE)
+            rope_pattern = [bool(x) for x in no_rope_layers]
+            self.gguf_writer.add_rope_pattern(rope_pattern)
+            logger.info("gguf: (granite_swa) rope_pattern = %d RoPE layers / %d total",
+                        sum(rope_pattern), len(rope_pattern))
+
+
+@ModelBase.register("GraniteMoeSWAForCausalLM")
+class GraniteMoeSWAModel(GraniteSWAModel):
+    """Conversion for IBM's GraniteMoeSWAForCausalLM (unified dense + MoE with iSWA)"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_SWA
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        if shared_intermediate_size := self.hparams.get("shared_intermediate_size"):
+            self.gguf_writer.add_expert_shared_feed_forward_length(shared_intermediate_size)
+            logger.info("gguf: (granitemoewa) shared_intermediate_size = %s", shared_intermediate_size)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """Split merged MoE tensors (gate+up) following standard MoE pattern."""
+
+        # Handle expert FFN tensors (merged gate+up) - swash format: experts.gate_up_proj
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.experts.gate_up_proj"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, f"Merged FFN tensor size must be 2 * intermediate_size, got {data_torch.shape[-2]}"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle expert FFN down projection - swash format: experts.down_proj
+        if name.endswith("block_sparse_moe.experts.down_proj"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_EXP, bid), bid)
+            return
+
+        # Handle expert FFN tensors (merged gate+up) - standard granite format: input_linear.weight
+        # Kept fused since inference (build_moe_ffn) supports a single gate_up_exps
+        # tensor for the routed experts.
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_UP_EXP, bid), bid)
+            return
+
+        # Handle shared expert FFN tensors (if present) - kept fused since
+        # inference (build_ffn) supports a single ffn_up_shexp tensor with
+        # LLM_FFN_SWIGLU for the shared expert.
+        if name.endswith("shared_mlp.input_linear.weight"):
+            ffn_dim = self.hparams.get("shared_intermediate_size", self.hparams["intermediate_size"])
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * shared_intermediate_size"
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_SHEXP, bid), bid)
+            return
+
+        # Handle shared expert output (if present)
+        if name.endswith("shared_mlp.output_linear.weight"):
+            yield from ModelBase.modify_tensors(self, data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, bid), bid)
+            return
+
+        # Pass through to parent for all other tensors (including sinks)
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
@ModelBase.register("GraniteMoeForCausalLM", "GraniteMoeSharedForCausalLM")
@ModelBase.example("ibm-granite/granite-3.1-3b-a800m-instruct")
 class GraniteMoeModel(GraniteModel):
@@ -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"))
@@ -709,7 +709,7 @@ class DFlashModel(Qwen3Model):
        yield from super().modify_tensors(data_torch, name, bid)


-@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator")
+@ModelBase.register("Qwen3DSparkModel", "DSparkDraftModel", "DSparkSpeculator", "Lfm2DSparkDraftModel")
@ModelBase.example("satgeze/Qwen3.6-27B-DSpark")
 class DSparkModel(DFlashModel):
    # DSpark = DFlash + a semi-autoregressive Markov head.
@@ -759,6 +759,13 @@ 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 +774,12 @@ 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):
@@ -237,8 +237,8 @@ chmod +x ubuntu-llamacpp-ov-install.sh
 # ============================================
 set -euo pipefail

-OPENVINO_VERSION_MAJOR="2026.2.1"
-OPENVINO_VERSION_FULL="2026.2.1.21919.ede283a88e3"
+OPENVINO_VERSION_MAJOR="2026.3"
+OPENVINO_VERSION_FULL="2026.3.0.22451.bd8d6542e3c"

 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OPENVINO_INSTALL_DIR="/opt/intel/openvino_${OPENVINO_VERSION_MAJOR}"
@@ -334,7 +334,7 @@ echo "  ./build/ReleaseOV/bin/llama-cli -m model.gguf"
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release.

 </details>

@@ -364,8 +364,8 @@ REM ============================================
 REM llama.cpp OpenVINO Build Script (Ninja)
 REM ============================================

-set "OPENVINO_VERSION_MAJOR=2026.2.1"
-set "OPENVINO_VERSION_FULL=2026.2.1.21919.ede283a88e3"
+set "OPENVINO_VERSION_MAJOR=2026.3"
+set "OPENVINO_VERSION_FULL=2026.3.0.22451.bd8d6542e3c"

 set "SCRIPT_DIR=%~dp0"
 set "VCPKG_DIR=C:\vcpkg"
@@ -547,7 +547,7 @@ endlocal
 ```

 > [!NOTE]
-> The script pins OpenVINO `2026.2.1` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.
+> The script pins OpenVINO `2026.3` via the `OPENVINO_VERSION_MAJOR` / `OPENVINO_VERSION_FULL` variables at the top — edit them to track a different release. From any new shell, source the matching `setupvars` script via the junction — `call "C:\Intel\openvino\setupvars.bat"` from `cmd`, or `& "C:\Intel\openvino\setupvars.ps1"` from PowerShell. If `winget` cannot register Visual Studio Build Tools on first run, install them once manually and re-run the script from an elevated **Developer Command Prompt for VS 2022**.

 </details>

@@ -72,9 +72,10 @@ cmake --build build --config Release
    - 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 --preset arm64-windows-llvm-release -D GGML_OPENMP_FETCH=ON
    cmake --build build-arm64-windows-llvm-release
    ```
+    `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
      ```bash
@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 20)
-set(GGML_VERSION_PATCH 1)
+set(GGML_VERSION_PATCH 2)
 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)
@@ -7,7 +7,7 @@ extern "C" {
 #endif

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

 #ifdef  __cplusplus
@@ -1981,6 +1981,14 @@ extern "C" {
            float                 beta_fast,
            float                 beta_slow);

+    // set the offset dims for RoPE
+    // a must be GGML_OP_ROPE or GGML_OP_ROPE_BACK
+    // vision RoPE is not supported
+    // example: (marking: x = rotated, 0 = unrotated)
+    //     n_embd = 10, n_dims = 4, offset = 2 --> [00xxxx0000]
+    GGML_API struct ggml_tensor * ggml_rope_set_offset(
+            struct ggml_tensor  * a,
+            int                   n_offs);

    // clamp
    // in-place, returns view(a)
@@ -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
@@ -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;
@@ -2534,6 +2534,9 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
            }
        case GGML_OP_ROPE:
            {
+                if (((const int32_t *) op->op_params)[15] != 0) {
+                    return false; // FIXME: support ggml_rope_set_offset
+                }
                if (op->src[0]->ne[0] > 896) {
                    return false;
                }
@@ -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)
@@ -5979,6 +5979,8 @@ static void ggml_compute_forward_rope_flt(
    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);

+    const int n_offs = ((int32_t *) dst->op_params)[15];
+
    GGML_TENSOR_UNARY_OP_LOCALS

    //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3);
@@ -5995,6 +5997,10 @@ static void ggml_compute_forward_rope_flt(
    GGML_ASSERT(n_dims <= ne0);
    GGML_ASSERT(n_dims % 2 == 0);

+    GGML_ASSERT(n_offs >= 0);
+    GGML_ASSERT(n_offs % 2 == 0);
+    GGML_ASSERT(n_offs + n_dims <= ne0);
+
    // rows per thread
    const int dr = (nr + nth - 1)/nth;

@@ -6020,6 +6026,7 @@ static void ggml_compute_forward_rope_flt(

    if (is_vision) {
        GGML_ASSERT(n_dims == ne0/2);
+        GGML_ASSERT(n_offs == 0);
    }

    const float * freq_factors = NULL;
@@ -6068,12 +6075,12 @@ static void ggml_compute_forward_rope_flt(

                switch (mode) {
                    case GGML_ROPE_TYPE_NORMAL:
-                        rotate_pairs<T>(n_dims, 1, cache, src, dst_data, 1);
+                        rotate_pairs<T>(n_dims, 1, cache, src + n_offs, dst_data + n_offs, 1);
                        break;
                    case GGML_ROPE_TYPE_NEOX:
                    case GGML_ROPE_TYPE_MROPE:
                    case GGML_ROPE_TYPE_IMROPE:
-                        rotate_pairs<T>(n_dims, n_dims/2, cache, src, dst_data);
+                        rotate_pairs<T>(n_dims, n_dims/2, cache, src + n_offs, dst_data + n_offs);
                        break;
                    case GGML_ROPE_TYPE_VISION:
                        rotate_pairs<T>(ne0, n_dims, cache, src, dst_data);
@@ -6084,7 +6091,11 @@ static void ggml_compute_forward_rope_flt(

                if (!is_vision) {
                    // fill the remain channels with data from src tensor
-                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+                        if (i0 == n_offs) {
+                            i0 += n_dims - 2; // skip the rotated channels
+                            continue;
+                        }
                        const T * const src = (T *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                        T * dst_data  = (T *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);

@@ -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

@@ -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
@@ -711,9 +711,12 @@ 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 +1419,7 @@ 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 +1542,14 @@ 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 +1564,7 @@ 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 +1602,7 @@ 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,
@@ -2723,6 +2726,12 @@ static bool ggml_cuda_should_fuse_rms_norm_mul_rope(const ggml_tensor * rms_norm
        return false;
    }

+    // ggml_rope_set_offset is not yet supported in the fused kernel
+    const int n_offs = ((const int32_t *) rope->op_params)[15];
+    if (n_offs != 0) {
+        return false;
+    }
+
    return true;
 }

@@ -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) {
@@ -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);

@@ -53,6 +53,7 @@ static __global__ void rope_norm(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -61,7 +62,8 @@ static __global__ void rope_norm(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -92,19 +94,24 @@ static __global__ void rope_norm(const T *            x,
            ggml_cuda_memcpy_1<4>(dst + idst, &v);
        }
    };
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        store_coaelsced(x[ix + 0], x[ix + 1]);
        return;
    }

-    const float theta_base = pos[i2]*powf(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]*powf(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, ext_factor, attn_factor, cos_theta, sin_theta);
+    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 + 1];
@@ -125,6 +132,7 @@ static __global__ void rope_neox(const T *            x,
                                 const int            s2,
                                 const int            s3,
                                 const int            n_dims,
+                                 const int            n_offs,
                                 const int32_t *      pos,
                                 const float          freq_scale,
                                 const float          ext_factor,
@@ -133,7 +141,8 @@ static __global__ void rope_neox(const T *            x,
                                 const float          theta_scale,
                                 const float *        freq_factors,
                                 const int64_t *      row_indices,
-                                 const int            set_rows_stride) {
+                                 const int            set_rows_stride,
+                                 const bool           inplace) {
    ggml_cuda_pdl_lc();
    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);

@@ -158,27 +167,33 @@ static __global__ void rope_neox(const T *            x,
        idst += row_indices[i2] * set_rows_stride;
    }

-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        dst[idst + i0 / 2 + 0] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 0]);
        dst[idst + i0 / 2 + 1] = ggml_cuda_cast<D>(x[ix + i0 / 2 + 1]);

        return;
    }

-    const float theta_base = pos[i2]*powf(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]*powf(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, ext_factor, attn_factor, cos_theta, sin_theta);
+    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_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
-    dst[idst + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
+    dst[idst + n_offs/2 + 0]          = ggml_cuda_cast<D>(x0 * cos_theta - x1 * sin_theta);
+    dst[idst + n_offs/2 + n_dims / 2] = ggml_cuda_cast<D>(x0 * sin_theta + x1 * cos_theta);
 }

 template <bool forward, bool has_ff, typename T>
@@ -194,6 +209,7 @@ static __global__ void rope_multi(const T *            x,
                                  const int            s2,
                                  const int            s3,
                                  const int            n_dims,
+                                  const int            n_offs,
                                  const int32_t *      pos,
                                  const float          freq_scale,
                                  const float          ext_factor,
@@ -202,7 +218,8 @@ static __global__ void rope_multi(const T *            x,
                                  const float          theta_scale,
                                  const float *        freq_factors,
                                  const mrope_sections sections,
-                                  const bool           is_imrope) {
+                                  const bool           is_imrope,
+                                  const bool           inplace) {
    const int i0 = 2 * (blockDim.y * blockIdx.y + threadIdx.y);

    if (i0 >= ne00) {
@@ -219,52 +236,58 @@ static __global__ void rope_multi(const T *            x,
    const int ix   = i0 / 2 + i1 * s01 + i2 * s02 + i3 * s03;

    ggml_cuda_pdl_sync();
-    if (i0 >= n_dims) {
+    if (i0 < n_offs || i0 >= n_offs + n_dims) {
+        if (inplace) {
+            return;
+        }
        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] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 2 && sector < 3 * sections.v[2]) {  // w
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector % 3 == 0 && sector < 3 * sections.v[0]) {  // t
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, iw / 2.0f);
        }
    } else {
        if (sector < sections.v[0]) {
-            theta_base = pos[i2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sections.v[0] && sector < sec_w) {
-            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 1] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 2] * powf(theta_scale, iw / 2.0f);
        } else if (sector >= sec_w + sections.v[2]) {
-            theta_base = pos[i2 + ne02 * 3] * powf(theta_scale, i0 / 2.0f);
+            theta_base = pos[i2 + ne02 * 3] * powf(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, ext_factor, attn_factor, cos_theta, sin_theta);
+    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>
@@ -344,6 +367,7 @@ static void rope_norm_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -354,6 +378,7 @@ static void rope_norm_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -364,12 +389,12 @@ static void rope_norm_cuda(const T *            x,

    if (freq_factors == nullptr) {
        rope_norm<forward, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        rope_norm<forward, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -386,6 +411,7 @@ static void rope_neox_cuda(const T *            x,
                           const int            s2,
                           const int            s3,
                           const int            n_dims,
+                           const int            n_offs,
                           const int            nr,
                           const int32_t *      pos,
                           const float          freq_scale,
@@ -396,6 +422,7 @@ static void rope_neox_cuda(const T *            x,
                           const float *        freq_factors,
                           const int64_t *      row_indices,
                           const int            set_rows_stride,
+                           const bool           inplace,
                           cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -407,12 +434,12 @@ static void rope_neox_cuda(const T *            x,

    if (freq_factors == nullptr) {
        ggml_cuda_kernel_launch(rope_neox<forward, false, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    } else {
        ggml_cuda_kernel_launch(rope_neox<forward, true, T, D>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, row_indices, set_rows_stride, inplace);
    }
 }

@@ -429,6 +456,7 @@ static void rope_multi_cuda(const T *            x,
                            const int            s2,
                            const int            s3,
                            const int            n_dims,
+                            const int            n_offs,
                            const int            nr,
                            const int32_t *      pos,
                            const float          freq_scale,
@@ -439,6 +467,7 @@ static void rope_multi_cuda(const T *            x,
                            const float *        freq_factors,
                            const mrope_sections sections,
                            const bool           is_imrope,
+                            const bool           inplace,
                            cudaStream_t         stream) {
    GGML_ASSERT(ne00 % 2 == 0);
    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -450,13 +479,13 @@ static void rope_multi_cuda(const T *            x,
    if (freq_factors == nullptr) {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, false, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    } else {
        const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
        ggml_cuda_kernel_launch(rope_multi<forward, true, T>, launch_params,
-            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, pos, freq_scale, ext_factor,
-            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope);
+            x, dst, ne00, ne01, ne02, s01, s02, s03, s1, s2, s3, n_dims, n_offs, pos, freq_scale, ext_factor,
+            attn_factor, corr_dims, theta_scale, freq_factors, sections, is_imrope, inplace);
    }
 }

@@ -552,8 +581,12 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    const int mode       = ((int32_t *) dst->op_params)[2];
    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    const int n_offs     = ((int32_t *) dst->op_params)[15];
    mrope_sections sections;

+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = dst_d == src0->data;
+
    // RoPE alteration for extended context
    float freq_base;
    float freq_scale;
@@ -581,6 +614,7 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,

    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;
@@ -597,31 +631,31 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    if (is_neox) {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_neox_cuda<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,
+                                                  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);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 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);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_neox_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                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);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
    } else if (is_mrope && !is_vision) {
        if (src0->type == GGML_TYPE_F32) {
            rope_multi_cuda<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, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16) {
            rope_multi_cuda<forward>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02, s03, s1,
-                                     s2, s3, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
-                                     corr_dims, freq_factors, sections, is_imrope, stream);
+                                     s2, s3, n_dims, n_offs, nr, pos, freq_scale, freq_base, ext_factor, attn_factor,
+                                     corr_dims, freq_factors, sections, is_imrope, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -640,19 +674,19 @@ void ggml_cuda_op_rope_impl(ggml_backend_cuda_context & ctx,
    } else {
        if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F32) {
            rope_norm_cuda<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,
+                                                  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);
+                                                  set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F32 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, float, half>((const float *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                 s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                 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);
+                                                 set_rows_stride, inplace, stream);
        } else if (src0->type == GGML_TYPE_F16 && dst_type == GGML_TYPE_F16) {
            rope_norm_cuda<forward, half, half>((const half *) src0_d, (half *) dst_d, ne00, ne01, ne02, s01, s02,
-                                                s03, s1, s2, s3, n_dims, nr, pos, freq_scale, freq_base,
+                                                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);
+                                                set_rows_stride, inplace, stream);
        } else {
            GGML_ABORT("fatal error");
        }
@@ -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);
 }
@@ -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;
@@ -1061,9 +1061,11 @@ static bool ggml_backend_et_device_supports_op(ggml_backend_dev_t dev, const ggm
                const bool zero_view_offset = op->src[0]->view_src == nullptr || op->src[0]->view_offs == 0;
                const bool has_sections = ggml_get_op_params_i32(op, 11) > 0 || ggml_get_op_params_i32(op, 12) > 0 ||
                                          ggml_get_op_params_i32(op, 13) > 0;
+                // FIXME: support ggml_rope_set_offset
+                const bool zero_rot_offset  = ggml_get_op_params_i32(op, 15) == 0;

                supported =
-                    zero_view_offset && ndims <= 512 &&
+                    zero_view_offset && zero_rot_offset && ndims <= 512 &&
                    (is_normal || (is_neox && ndims % 16 == 0) || (is_imrope && ndims % 16 == 0 && has_sections));
            } else {
                supported = false;
@@ -3180,6 +3180,10 @@ 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
+    }
+
    int mode = op_params[2];

    // n_dims == ne0/2, so the rotation spans the full row
@@ -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;
@@ -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;
@@ -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);

@@ -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,
@@ -329,6 +329,7 @@ typedef struct {
    uint64_t nb3;
    int32_t  n_past;
    int32_t  n_dims;
+    int32_t  n_offs;
    int32_t  n_ctx_orig;
    float    freq_base;
    float    freq_scale;
@@ -341,8 +342,21 @@ typedef struct {
    int32_t  sect_2;
    int32_t  sect_3;
    bool     src2;
+    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
@@ -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);

@@ -3884,6 +4075,11 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
    const int sect_2 = ((const int32_t *) op->op_params)[13];
    const int sect_3 = ((const int32_t *) op->op_params)[14];

+    const int n_offs = ((const int32_t *) op->op_params)[15];
+
+    // when dst aliases src0, the channels outside the rotated window already hold the correct data
+    const bool inplace = op->data == op->src[0]->data;
+
    ggml_metal_kargs_rope args = {
        /*.ne00        =*/ ne00,
        /*.ne01        =*/ ne01,
@@ -3903,6 +4099,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /*.nb3         =*/ nb3,
        /*.n_past      =*/ n_past,
        /*.n_dims      =*/ n_dims,
+        /*.n_offs      =*/ n_offs,
        /*.n_ctx_orig  =*/ n_ctx_orig,
        /*.freq_base   =*/ freq_base,
        /*.freq_scale  =*/ freq_scale,
@@ -3915,6 +4112,7 @@ int ggml_metal_op_rope(ggml_metal_op_t ctx, int idx) {
        /* sect_2      =*/ sect_2,
        /* sect_3      =*/ sect_3,
        /* src2        =*/ op->src[2] != nullptr,
+        /* inplace     =*/ inplace,
    };

    auto pipeline = ggml_metal_library_get_pipeline_rope(lib, op);
@@ -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);
@@ -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:
@@ -656,13 +656,13 @@ void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & r

 template <typename type4x4>
 void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
    const float d = xb->d;

    float4x4 reg_f;

-    for (int i = 0; i < 16; i++) {
-        reg_f[i/4][i%4] = (qs[i + 16*il] * d);
+    for (int i = 0; i < 4; ++i) {
+        reg_f[i] = float4(qs[4*il + i]) * d;
    }

    reg = (type4x4) reg_f;
@@ -670,12 +670,10 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg

 template <typename type4>
 void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
-    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    device const packed_char4 * qs = (device const packed_char4 *) xb->qs;
    const float d = xb->d;

-    for (int i = 0; i < 4; i++) {
-        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
-    }
+    reg = (type4) (float4(qs[il]) * d);
 }

 template <typename type4x4>
@@ -4688,14 +4686,15 @@ kernel void kernel_rope_norm(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
@@ -4706,6 +4705,10 @@ kernel void kernel_rope_norm(
            dst_data[0] = x0*cos_theta - x1*sin_theta;
            dst_data[1] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -4741,17 +4744,18 @@ kernel void kernel_rope_neox(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

-            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
-            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);

            const float x0 = src[0];
            const float x1 = src[args.n_dims/2];
@@ -4759,6 +4763,10 @@ kernel void kernel_rope_neox(
            dst_data[0]             = x0*cos_theta - x1*sin_theta;
            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -4793,8 +4801,9 @@ kernel void kernel_rope_multi(
    float sin_theta;

    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
-        if (i0 < args.n_dims) {
-            const int ic = i0/2;
+        if (i0 >= args.n_offs && i0 < args.n_offs + args.n_dims) {
+            const int iw = i0 - args.n_offs; // relative idx
+            const int ic = iw/2;

            // mrope theta calculations
            // note: the rest is the same as kernel_rope_neox
@@ -4827,14 +4836,14 @@ kernel void kernel_rope_multi(
            }
            // end of mrope

-            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*iw);

            const float freq_factor = args.src2 ? ((device const float *) src2)[ic] : 1.0f;

-            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, iw, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);

-            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
-            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + (args.n_offs + ic)*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + (args.n_offs + ic)*args.nb0);

            const float x0 = src[0];
            const float x1 = src[args.n_dims/2];
@@ -4842,6 +4851,10 @@ kernel void kernel_rope_multi(
            dst_data[0]             = x0*cos_theta - x1*sin_theta;
            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
        } else {
+            if (args.inplace) {
+                continue;
+            }
+
            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);

@@ -6305,6 +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)]];
@@ -202,6 +202,7 @@ set(GGML_OPENCL_KERNELS
    sqr
    sqrt
    ssm_conv
+    ssm_scan
    gated_delta_net
    sub
    sum_rows
@@ -866,6 +866,9 @@ 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;
@@ -892,6 +895,7 @@ 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_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;
@@ -3154,6 +3158,24 @@ 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
@@ -4442,6 +4464,7 @@ 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(".");
    }
@@ -7301,6 +7324,23 @@ 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}.
@@ -7376,6 +7416,9 @@ 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;
@@ -12257,6 +12300,103 @@ 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);
@@ -12690,7 +12830,10 @@ 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;

@@ -20725,18 +20868,42 @@ 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
@@ -24743,6 +24910,14 @@ 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;
@@ -118,6 +118,17 @@ __kernel void flash_attn_f16(
    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+#if WG_SIZE > FA_SG
+        // WAR on l_k/l_v: a thread that finishes the compute below early — either
+        // it skipped it (my_query_row >= n_q, the continue) or its subgroup simply
+        // ran ahead — wraps around and reloads the tiles while another subgroup is
+        // still reading them. Any WG that is exactly one lockstep subgroup
+        // (WG_SIZE == FA_SG) cannot diverge and hides this; a WG spanning multiple
+        // subgroups (Intel sg=32, or BLOCK_M > 64 on Adreno) corrupts the result.
+        // All threads reach this each iteration (no-op on the first), so it does
+        // not diverge with the continue. Compiled out when WG == one subgroup.
+        barrier(CLK_LOCAL_MEM_FENCE);
+#endif
        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
            const int row = i / DK_VEC;
            const int col = i % DK_VEC;
@@ -119,13 +119,15 @@ __kernel void flash_attn_f32(
    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];

    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
-#if FA_SG < 64
-        // WAR on l_k/l_v: threads with my_query_row >= n_q skip the compute below
-        // (continue) and would race ahead to reload the tiles while active threads
-        // still read them. A single 64-wide Adreno subgroup (WG == sg) runs lockstep
-        // and hides this; a WG that spans multiple narrower subgroups (Intel sg=32)
-        // corrupts the result. All threads reach this each iteration (no-op on the
-        // first), so it does not diverge with the continue. Compiled out at sg=64.
+#if WG_SIZE > FA_SG
+        // WAR on l_k/l_v: a thread that finishes the compute below early — either
+        // it skipped it (my_query_row >= n_q, the continue) or its subgroup simply
+        // ran ahead — wraps around and reloads the tiles while another subgroup is
+        // still reading them. Any WG that is exactly one lockstep subgroup
+        // (WG_SIZE == FA_SG) cannot diverge and hides this; a WG spanning multiple
+        // subgroups (Intel sg=32, or BLOCK_M > 64 on Adreno) corrupts the result.
+        // All threads reach this each iteration (no-op on the first), so it does
+        // not diverge with the continue. Compiled out when WG == one subgroup.
        barrier(CLK_LOCAL_MEM_FENCE);
 #endif
        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
@@ -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,
@@ -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;
+}
@@ -1227,6 +1227,10 @@ static bool is_op_unsupported_case(const ggml_tensor * op) {
        const int32_t * op_params = op->op_params;
        const int n_dims = op_params[1];
        const int mode = op_params[2];
+        if (op_params[15] != 0) {
+            // FIXME: support ggml_rope_set_offset
+            return true;
+        }
        if (mode != GGML_ROPE_TYPE_NORMAL && mode != GGML_ROPE_TYPE_NEOX && mode != GGML_ROPE_TYPE_IMROPE) {
            // GGML_LOG_WARN("OpenVINO backend does not support ROPE with mode %d\n", mode);
            return true;
@@ -47,7 +47,7 @@ struct rpc_tensor {
    uint64_t data;
    char name[GGML_MAX_NAME];

-    char padding[4];
+    int32_t use_count;
 };

 static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
@@ -447,7 +447,7 @@ static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {

    // Avoid sending uninitialized data over the wire
    memset(result.name, 0, sizeof(result.name));
-    memset(result.padding, 0, sizeof(result.padding));
+    result.use_count = 0;

    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
    return result;
@@ -675,7 +675,7 @@ static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
    // this is no-op because we don't have any async operations
 }

-static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
+static void add_tensor(ggml_tensor * tensor, const ggml_cgraph * cgraph, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
    if (tensor == nullptr) {
        return;
    }
@@ -684,10 +684,15 @@ static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors,
    }
    visited.insert(tensor);
    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        add_tensor(tensor->src[i], tensors, visited);
+        add_tensor(tensor->src[i], cgraph, tensors, visited);
    }
-    add_tensor(tensor->view_src, tensors, visited);
-    tensors.push_back(serialize_tensor(tensor));
+    add_tensor(tensor->view_src, cgraph, tensors, visited);
+    rpc_tensor result = serialize_tensor(tensor);
+    const size_t hash_pos = ggml_hash_find(&cgraph->visited_hash_set, tensor);
+    if (hash_pos != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, hash_pos)) {
+        result.use_count = cgraph->use_counts[hash_pos];
+    }
+    tensors.push_back(result);
 }

 static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
@@ -695,7 +700,7 @@ static void serialize_graph(uint32_t device, const ggml_cgraph * cgraph, std::ve
    std::vector<rpc_tensor> tensors;
    std::unordered_set<ggml_tensor*> visited;
    for (uint32_t i = 0; i < n_nodes; i++) {
-        add_tensor(cgraph->nodes[i], tensors, visited);
+        add_tensor(cgraph->nodes[i], cgraph, tensors, visited);
    }
    // serialization format:
    // | device (4 bytes) | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
@@ -1451,6 +1456,10 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {
            GGML_LOG_ERROR("[%s] failed to create graph node %d (id=%" PRId64 ")\n", __func__, i, id);
            return false;
        }
+        if (graph->nodes[i] != nullptr) {
+            const size_t hash_pos = ggml_hash_insert(&graph->visited_hash_set, graph->nodes[i]);
+            graph->use_counts[hash_pos] = tensor_ptrs.at(id)->use_count;
+        }
    }
    ggml_status status = ggml_backend_graph_compute(backends[device], graph);
    GGML_ASSERT(status == GGML_STATUS_SUCCESS && "Unsuccessful graph computations are not supported with RPC");
@@ -0,0 +1,119 @@
+#include "fwht.hpp"
+
+#include <cmath>
+
+template <int N>
+static void fwht_kernel(const float * __restrict__ src, float * __restrict__ dst, const int64_t n_rows,
+                        const float scale, const sycl::nd_item<2> & item) {
+    const sycl::sub_group sg = item.get_sub_group();
+
+    const int64_t r = item.get_global_id(0);
+    if (r >= n_rows) {
+        return;
+    }
+
+    src += r * N;
+    dst += r * N;
+
+    constexpr int el_w = N / WARP_SIZE;
+    static_assert(el_w >= 1 && N % WARP_SIZE == 0, "row must be a whole number of sub-group widths");
+
+    float     reg[el_w];
+    const int lane = sg.get_local_linear_id();
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        reg[i] = src[i * WARP_SIZE + lane] * scale;
+    }
+
+    // Butterflies inside the sub-group. The partner of a lane with bit h clear is the
+    // lower index of the pair, so it takes the sum and the upper takes lower - upper.
+#pragma unroll
+    for (int h = 1; h < WARP_SIZE; h *= 2) {
+#pragma unroll
+        for (int j = 0; j < el_w; ++j) {
+            const float val  = reg[j];
+            const float val2 = dpct::permute_sub_group_by_xor(sg, val, h, WARP_SIZE);
+
+            reg[j] = (lane & h) == 0 ? val + val2 : val2 - val;
+        }
+    }
+
+    // Butterflies across registers: h is a multiple of WARP_SIZE, so the partner of
+    // element i*WARP_SIZE + lane lives in reg[i + h/WARP_SIZE] on the same lane.
+#pragma unroll
+    for (int h = WARP_SIZE; h < N; h *= 2) {
+        const int step = h / WARP_SIZE;
+#pragma unroll
+        for (int j = 0; j < el_w; j += 2 * step) {
+#pragma unroll
+            for (int k = 0; k < step; ++k) {
+                const float x = reg[j + k];
+                const float y = reg[j + k + step];
+
+                reg[j + k]        = x + y;
+                reg[j + k + step] = x - y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int i = 0; i < el_w; ++i) {
+        dst[i * WARP_SIZE + lane] = reg[i];
+    }
+}
+
+template <int N>
+static void launch_fwht(const float * src, float * dst, const int64_t n_rows, const float scale,
+                        dpct::queue_ptr stream) {
+    constexpr int rows_per_block = 4;
+
+    const int64_t num_blocks = (n_rows + rows_per_block - 1) / rows_per_block;
+
+    // dim 1 is the fastest-varying, so a sub-group is exactly one row's WARP_SIZE lanes.
+    const sycl::range<2> global(num_blocks * rows_per_block, WARP_SIZE);
+    const sycl::range<2> local(rows_per_block, WARP_SIZE);
+
+    stream->parallel_for(sycl::nd_range<2>(global, local),
+                         [=](sycl::nd_item<2> item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             fwht_kernel<N>(src, dst, n_rows, scale, item);
+                         });
+}
+
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst) {
+    if (src->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+    if (!ggml_are_same_shape(src, dst)) {
+        return false;
+    }
+    if (!ggml_is_contiguous(src) || !ggml_is_contiguous(dst)) {
+        return false;
+    }
+
+    const int     n    = (int) src->ne[0];
+    const int64_t rows = ggml_nrows(src);
+
+    const float *   src_d  = (const float *) src->data;
+    float *         dst_d  = (float *) dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    const float scale = 1.0f / std::sqrt((float) n);
+
+    switch (n) {
+        case 64:
+            launch_fwht<64>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 128:
+            launch_fwht<128>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 256:
+            launch_fwht<256>(src_d, dst_d, rows, scale, stream);
+            return true;
+        case 512:
+            launch_fwht<512>(src_d, dst_d, rows, scale, stream);
+            return true;
+        default:
+            return false;
+    }
+}
@@ -0,0 +1,12 @@
+#ifndef GGML_SYCL_FWHT_HPP
+#define GGML_SYCL_FWHT_HPP
+
+#include "common.hpp"
+
+// Fast Walsh-Hadamard transform, the fast path for a MUL_MAT whose src0 ggml has
+// tagged GGML_HINT_SRC0_IS_HADAMARD. src0 is not read at all. Returns false if the
+// shape is not one this can serve, in which case the caller must fall through to the
+// ordinary mat-mul dispatch.
+bool ggml_sycl_op_fwht(ggml_backend_sycl_context & ctx, const ggml_tensor * src, ggml_tensor * dst);
+
+#endif  // GGML_SYCL_FWHT_HPP
@@ -58,6 +58,7 @@
 #include "ggml-sycl/backend.hpp"
 #include "ggml-sycl/common.hpp"
 #include "ggml-sycl/element_wise.hpp"
+#include "ggml-sycl/fwht.hpp"
 #include "ggml-sycl/gemm.hpp"
 #include "ggml-sycl/getrows.hpp"
 #include "ggml-sycl/norm.hpp"
@@ -108,7 +109,14 @@ int g_ggml_sycl_enable_host_pinned_mem = 1;
 static ggml_sycl_device_info ggml_sycl_init() {
    ggml_sycl_device_info info = {};

-    info.device_count = dpct::dev_mgr::instance().device_count();
+    // Do not hard crash when there exists no SYCL devices.
+    // We want to allow the user to use non-SYCL tools when SYCL is compiled (such as llama-quantize)
+    try {
+        info.device_count = dpct::dev_mgr::instance().device_count();
+    } catch (sycl::exception const &exc) {
+        GGML_LOG_INFO("%s: no SYCL device available: %s\n", __func__, exc.what());
+        info.device_count = 0;
+    }
    if (info.device_count == 0) {
        GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
        return info;
@@ -4473,6 +4481,18 @@ static bool can_use_mul_mat_vec_q(const ggml_tensor * src0, const ggml_tensor *

 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/2);
+
+    // Handle HADAMARAD hint given from further up the pipeline and pass it to the correct
+    // kernel.
+    //
+    // The op check is not redundant: this backend also routes MUL_MAT_ID through here with a
+    // stack copy of dst, which carries MUL_MAT_ID's own op_params. ggml_mul_mat_set_hint()
+    // asserts GGML_OP_MUL_MAT for the same reason.
+    if (dst->op == GGML_OP_MUL_MAT && ggml_get_op_params_i32(dst, 1) == GGML_HINT_SRC0_IS_HADAMARD &&
+        ggml_sycl_op_fwht(ctx, src1, dst)) {
+        return;
+    }
+
    const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
    int64_t min_compute_capability = INT_MAX;

@@ -6222,6 +6242,8 @@ 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:
@@ -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()
@@ -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];
@@ -962,6 +963,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_cpy_f32_quant[GGML_TYPE_COUNT];
    vk_pipeline pipeline_cpy_quant_f32[GGML_TYPE_COUNT];
    vk_pipeline pipeline_cpy_transpose_16, pipeline_cpy_transpose_32;
+    vk_pipeline pipeline_cpy_transpose_02_16, pipeline_cpy_transpose_02_32;
    // [src0 0=fp32,1=fp16][dst]
    vk_pipeline pipeline_set_rows_i32[2][GGML_TYPE_COUNT];
    vk_pipeline pipeline_set_rows_i64[2][GGML_TYPE_COUNT];
@@ -1644,6 +1646,7 @@ struct vk_op_rope_push_constants {
    uint32_t rope_mode;
    uint32_t nrows;
    uint32_t n_dims;
+    uint32_t n_offs;
    float freq_scale;
    float freq_base;
    float ext_factor;
@@ -3382,10 +3385,10 @@ 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);
    }
 }
@@ -5389,6 +5392,7 @@ 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);
@@ -5525,6 +5529,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_32, "cpy_transpose_32", cpy_transpose_32_len, cpy_transpose_32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_16, "cpy_transpose_16", cpy_transpose_16_len, cpy_transpose_16_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_02_32, "cpy_transpose_02_32", cpy_transpose_02_32_len, cpy_transpose_02_32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_transpose_02_16, "cpy_transpose_02_16", cpy_transpose_02_16_len, cpy_transpose_02_16_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q1_0], "cpy_f32_q1_0", cpy_f32_q1_0_len, cpy_f32_q1_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q2_0], "cpy_f32_q2_0", cpy_f32_q2_0_len, cpy_f32_q2_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {32, 1, 1}, {}, 1);
@@ -8931,6 +8937,18 @@ static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, const
        }
    }

+    // Same, for a 0<->2 swap: src dim2 is the innermost dimension.
+    bool transpose02 = dst && !contig && src->nb[2] == ggml_type_size(to) &&
+                       ggml_is_contiguous(dst) && ggml_are_same_shape(dst, src);
+
+    if (transpose02 && src->type == to) {
+        if (ggml_type_size(to) == 4) {
+            return ctx->device->pipeline_cpy_transpose_02_32;
+        } else if (ggml_type_size(to) == 2) {
+            return ctx->device->pipeline_cpy_transpose_02_16;
+        }
+    }
+
    if (src->type == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
        if (contig) {
            return ctx->device->pipeline_contig_cpy_f32_f32;
@@ -10807,9 +10825,32 @@ 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 &&
@@ -10822,7 +10863,7 @@ 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));
@@ -10836,6 +10877,17 @@ 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
@@ -10862,7 +10914,7 @@ 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;

@@ -10966,6 +11018,34 @@ 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)
@@ -10995,8 +11075,8 @@ 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 };
@@ -11038,6 +11118,10 @@ 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) {
@@ -12192,7 +12276,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
                elements = { ne, 1, 1 };
            }

-            if (pipeline == ctx->device->pipeline_cpy_transpose_32 ||
+            if (pipeline == ctx->device->pipeline_cpy_transpose_02_32 ||
+                pipeline == ctx->device->pipeline_cpy_transpose_02_16) {
+                // 32x32 tiles over dims 0 and 2; dim1 and dim3 are the batch
+                elements[0] = (uint32_t)CEIL_DIV(dst->ne[0], 32);
+                elements[1] = (uint32_t)CEIL_DIV(dst->ne[2], 32);
+                elements[2] = (uint32_t)(dst->ne[1]*dst->ne[3]);
+                elements[0] = std::min(elements[0], ctx->device->properties.limits.maxComputeWorkGroupCount[0]);
+                elements[1] = std::min(elements[1], ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
+                elements[2] = std::min(elements[2], ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
+            } else if (pipeline == ctx->device->pipeline_cpy_transpose_32 ||
                pipeline == ctx->device->pipeline_cpy_transpose_16) {
                // 32x32 tiles
                elements[0] = (uint32_t)CEIL_DIV(dst->ne[0], 32);
@@ -13120,6 +13213,7 @@ static uint32_t ggml_vk_rms_partials_size(ggml_backend_vk_context * ctx, const g
 static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *dst, const ggml_tensor *src0, const bool has_ff, bool backprop, const uint32_t set_rows_stride) {
    const int n_dims        = ((const int32_t *) dst->op_params)[1];
    const int mode          = ((const int32_t *) dst->op_params)[2];
+    const int n_offs        = ((const int32_t *) dst->op_params)[15];
    // const int n_ctx         = ((const int32_t *) dst->op_params)[3];
    const int n_ctx_orig    = ((const int32_t *) dst->op_params)[4];
    const float freq_base   = ((const float *)   dst->op_params)[5];
@@ -13149,7 +13243,7 @@ static vk_op_rope_push_constants ggml_vk_make_rope_constants(const ggml_tensor *
    uint32_t nb13 = dst->nb[3] / ggml_type_size(dst->type);

    vk_op_rope_push_constants rope {
-        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, freq_scale,
+        (uint32_t)mode, (uint32_t)ggml_nrows(src0), (uint32_t)n_dims, (uint32_t)n_offs, freq_scale,
        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale, has_ff,
        { sections[0], sections[1], sections[2], sections[3] }, is_imrope, backprop, set_rows_stride,

@@ -19195,6 +19289,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
                    tensor_clone = ggml_rope_ext_back(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
                }
            }
+            const int n_offs = ((int32_t *) tensor->op_params)[15];
+            if (n_offs != 0) {
+                tensor_clone = ggml_rope_set_offset(tensor_clone, n_offs);
+            }
        } else if (tensor->op == GGML_OP_UNARY) {
            switch (ggml_get_unary_op(tensor)) {
            case GGML_UNARY_OP_EXP:
@@ -0,0 +1,61 @@
+#version 450
+
+#include "types.glsl"
+#include "generic_unary_head.glsl"
+
+// workgroup does 32x32 tile, but uses 32x8 threads
+#define TILE_DIM 32
+layout(local_size_x = 32, local_size_y = 8, local_size_z = 1) in;
+
+// +1 padding avoids shared-memory bank conflicts on the transposed read
+shared uint sh[TILE_DIM][TILE_DIM + 1];
+
+void iter(uvec3 wg_id) {
+    const uint tile_i0 = wg_id.x;   // tiles dst ne10 (== src ne00)
+    const uint tile_i2 = wg_id.y;   // tiles dst ne12 (== src ne02)
+
+    const uint tid_col = gl_LocalInvocationID.x;
+    const uint tid_row = gl_LocalInvocationID.y;
+
+    const uint i1 = wg_id.z % p.ne11;
+    const uint i3 = wg_id.z / p.ne11;
+    const uint i01 = i1;
+    const uint i03 = i3;
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i00 = tile_i0 * TILE_DIM + tid_row + 8 * y;
+        const uint i02 = tile_i2 * TILE_DIM + tid_col;
+        if (i00 < p.ne00 && i01 < p.ne01 && i02 < p.ne02 && i03 < p.ne03) {
+            const uint src_idx = i00 * p.nb00 + i01 * p.nb01 + i02 * p.nb02 + i03 * p.nb03;
+            sh[tid_row + 8 * y][tid_col] = uint(data_a[get_aoffset() + src_idx]);
+        }
+    }
+
+    barrier();
+
+    [[unroll]] for (uint y = 0; y < 4; ++y) {
+        const uint i0 = tile_i0 * TILE_DIM + tid_col;
+        const uint i2 = tile_i2 * TILE_DIM + tid_row + 8 * y;
+        if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
+            const uint dst_idx = i0 * p.nb10 + i1 * p.nb11 + i2 * p.nb12 + i3 * p.nb13;
+            data_d[get_doffset() + dst_idx] = D_TYPE(sh[tid_col][tid_row + 8 * y]);
+        }
+    }
+}
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+void main() {
+    bool need_barrier = false;
+    for (uint z = gl_WorkGroupID.z; z < p.ne11 * p.ne13; z += gl_NumWorkGroups.z) {
+        for (uint y = gl_WorkGroupID.y; y < CEIL_DIV(p.ne12, TILE_DIM); y += gl_NumWorkGroups.y) {
+            for (uint x = gl_WorkGroupID.x; x < CEIL_DIV(p.ne10, TILE_DIM); x += gl_NumWorkGroups.x) {
+                if (need_barrier) {
+                    barrier();
+                }
+                need_barrier = true;
+                iter(uvec3(x, y, z));
+            }
+        }
+    }
+}
@@ -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);

@@ -121,13 +121,13 @@ void main() {
        const uint buf_ib = r * qf_stride + d / 8;
        const uint buf_iqs = d % 8;

-        FLOAT_TYPEV4 vals = is_in_bounds ? FLOAT_TYPEV4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale) : FLOAT_TYPEV4(0.0f);
-        const FLOAT_TYPEV4 abs_vals = abs(vals);
+        vec4 vals = is_in_bounds ? data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale : vec4(0.0f);
+        const vec4 abs_vals = abs(vals);

-        const FLOAT_TYPE thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
-        const FLOAT_TYPE amax = subgroupClusteredMax(thread_max, 8);
-        const FLOAT_TYPE qd = amax / FLOAT_TYPE(127.0);
-        const FLOAT_TYPE qd_inv = qd != FLOAT_TYPE(0.0) ? FLOAT_TYPE(1.0) / qd : FLOAT_TYPE(0.0);
+        const float thread_max = max(max(abs_vals.x, abs_vals.y), max(abs_vals.z, abs_vals.w));
+        const float amax = subgroupClusteredMax(thread_max, 8);
+        const float qd = amax / 127.0f;
+        const float qd_inv = qd != 0.0f ? 1.0f / qd : 0.0f;
        vals = round(vals * qd_inv);

        Qf[buf_ib].qs[buf_iqs] = pack32(i8vec4(vals));
@@ -136,11 +136,11 @@ void main() {
        // the row-sum scaled by qd, used in k_dot_correction.
        if (FaTypeK == FA_TYPE_Q8_0) {
            if (buf_iqs == 0) {
-                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, 0.0);
+                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, 0.0f);
            }
        } else {
-            const FLOAT_TYPE thread_sum = vals.x + vals.y + vals.z + vals.w;
-            const FLOAT_TYPE sum = subgroupClusteredAdd(thread_sum, 8);
+            const float thread_sum = vals.x + vals.y + vals.z + vals.w;
+            const float sum = subgroupClusteredAdd(thread_sum, 8);

            if (buf_iqs == 0) {
                Qf[buf_ib].ds = FLOAT_TYPEV2(qd, sum * qd);
@@ -50,19 +50,21 @@ void rope_norm(const uint i0, const uint i1, const uint i2, const uint i3, rope_
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + 0] = ROPE_D_TYPE(rope_data_a[ix + 0]);
        rope_data_d[idst + 1] = ROPE_D_TYPE(rope_data_a[ix + 1]);

        return;
    }

-    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
+    const uint iw = i0 - p.n_offs; // relative idx

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

    const float x0 = float(rope_data_a[ix + 0]);
    const float x1 = float(rope_data_a[ix + 1]);
@@ -87,25 +89,28 @@ void rope_neox(const uint i0, const uint i1, const uint i2, const uint i3, rope_
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);

        return;
    }

-    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, i0/2.0f);
+    const uint iw = i0 - p.n_offs; // relative idx

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float theta_base = rope_data_pos[i2] * pow(p.theta_scale, iw/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

-    const float x0 = float(rope_data_a[ix + 0]);
-    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
+    const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);

-    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
-    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+    rope_data_d[idst + p.n_offs/2 + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }


@@ -125,53 +130,56 @@ void rope_multi(const uint i0, const uint i1, const uint i2, const uint i3, rope
    }
    idst += p.d_offset;

-    if (i0 >= p.n_dims) {
+    if (i0 < p.n_offs || i0 >= p.n_offs + p.n_dims) {
        rope_data_d[idst + i0/2 + 0] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 0]);
        rope_data_d[idst + i0/2 + 1] = ROPE_D_TYPE(rope_data_a[ix + i0/2 + 1]);

        return;
    }

+    const uint iw = i0 - p.n_offs; // relative idx
+
    const int sect_dims = p.sections[0] + p.sections[1] + p.sections[2] + p.sections[3];
    const int sec_w = p.sections[1] + p.sections[0];
-    const uint sector = (i0 / 2) % sect_dims;
+    const uint sector = (iw / 2) % sect_dims;

    float theta_base = 0.0;
    if (p.is_imrope != 0) {
        if (sector % 3 == 1 && sector < 3 * p.sections[1]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
        } else if (sector % 3 == 2 && sector < 3 * p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
        } else if (sector % 3 == 0 && sector < 3 * p.sections[0]) {
-            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
        } else {
-            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
        }
    } else {
        if (sector < p.sections[0]) {
-            theta_base = rope_data_pos[i2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= p.sections[0] && sector < sec_w) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 1]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= sec_w && sector < sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 2]*pow(p.theta_scale, iw/2.0f);
        }
        else if (sector >= sec_w + p.sections[2]) {
-            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, i0/2.0f);
+            theta_base = rope_data_pos[i2 + p.ne02 * 3]*pow(p.theta_scale, iw/2.0f);
        }
    }

-    const float freq_factor = p.has_ff != 0 ? rope_data_ff[i0/2] : 1.0f;
+    const float freq_factor = p.has_ff != 0 ? rope_data_ff[iw/2] : 1.0f;

    float cos_theta, sin_theta;
-    rope_yarn(theta_base / freq_factor, i0, cos_theta, sin_theta, p);
+    rope_yarn(theta_base / freq_factor, iw, cos_theta, sin_theta, p);

-    const float x0 = float(rope_data_a[ix + 0]);
-    const float x1 = float(rope_data_a[ix + p.n_dims/2]);
+    // idst/ix point at channel i0/2; the first channel of the rotated pair is p.n_offs + iw/2 = i0/2 + p.n_offs/2
+    const float x0 = float(rope_data_a[ix + p.n_offs/2 + 0]);
+    const float x1 = float(rope_data_a[ix + p.n_offs/2 + p.n_dims/2]);

-    rope_data_d[idst + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
-    rope_data_d[idst + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
+    rope_data_d[idst + p.n_offs/2 + 0]          = ROPE_D_TYPE(x0*cos_theta - x1*sin_theta);
+    rope_data_d[idst + p.n_offs/2 + p.n_dims/2] = ROPE_D_TYPE(x0*sin_theta + x1*cos_theta);
 }

 void rope_vision(const uint i0, const uint i1, const uint i2, const uint i3, rope_params p) {
@@ -5,6 +5,7 @@ struct rope_params {
    uint rope_mode;
    uint nrows;
    uint n_dims;
+    uint n_offs;
    float freq_scale;
    float freq_base;
    float ext_factor;
@@ -780,6 +780,10 @@ void process_shaders() {
        if (tname != "f16" && tname != "bf16") {
            string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
        }
+        // Fused dequant+transpose variant for FA quant-KV (per-head-contiguous f16 scratch).
+        if (tname == "q8_0") {
+            string_to_spv("dequant_" + tname + "_transpose", "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}, {"DEQUANT_TRANSPOSE", "1"}}));
+        }

        shader = (tname == "f32" || tname == "f16" || tname == "bf16") ? "get_rows.comp" : "get_rows_quant.comp";

@@ -826,6 +830,8 @@ void process_shaders() {

    string_to_spv("cpy_transpose_16", "copy_transpose.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
    string_to_spv("cpy_transpose_32", "copy_transpose.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});
+    string_to_spv("cpy_transpose_02_16", "copy_transpose_02.comp", {{"A_TYPE", "uint16_t"}, {"D_TYPE", "uint16_t"}});
+    string_to_spv("cpy_transpose_02_32", "copy_transpose_02.comp", {{"A_TYPE", "uint"}, {"D_TYPE", "uint"}});

    for (std::string t : {"q1_0", "q2_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
        string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"S_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
@@ -954,10 +954,11 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key {
    int       vectorized;
    uint32_t  num_cols;
    bool      use_mmvq;
+    bool      src_overlap;

    bool operator==(const ggml_webgpu_mul_mat_vec_pipeline_key & other) const {
        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
-               num_cols == other.num_cols && use_mmvq == other.use_mmvq;
+               num_cols == other.num_cols && use_mmvq == other.use_mmvq && src_overlap == other.src_overlap;
    }
 };

@@ -969,6 +970,7 @@ struct ggml_webgpu_mul_mat_vec_pipeline_key_hash {
        ggml_webgpu_hash_combine(seed, key.vectorized);
        ggml_webgpu_hash_combine(seed, key.num_cols);
        ggml_webgpu_hash_combine(seed, key.use_mmvq);
+        ggml_webgpu_hash_combine(seed, key.src_overlap);
        return seed;
    }
 };
@@ -977,6 +979,7 @@ struct ggml_webgpu_mul_mat_vec_shader_decisions {
    uint32_t wg_size;
    uint32_t outputs_per_wg;
    uint32_t vec_size;
+    bool     src_overlap = false;
 };

 struct ggml_webgpu_quantize_q8_pipeline_key {
@@ -998,10 +1001,11 @@ struct ggml_webgpu_mul_mat_pipeline_key {
    ggml_type src1_type;
    int       vectorized;
    int       use_subgroup_matrix;
+    bool      src_overlap;

    bool operator==(const ggml_webgpu_mul_mat_pipeline_key & other) const {
        return src0_type == other.src0_type && src1_type == other.src1_type && vectorized == other.vectorized &&
-               use_subgroup_matrix == other.use_subgroup_matrix;
+               use_subgroup_matrix == other.use_subgroup_matrix && src_overlap == other.src_overlap;
    }
 };

@@ -1012,6 +1016,7 @@ struct ggml_webgpu_mul_mat_pipeline_key_hash {
        ggml_webgpu_hash_combine(seed, key.src1_type);
        ggml_webgpu_hash_combine(seed, key.vectorized);
        ggml_webgpu_hash_combine(seed, key.use_subgroup_matrix);
+        ggml_webgpu_hash_combine(seed, key.src_overlap);
        return seed;
    }
 };
@@ -1034,6 +1039,7 @@ struct ggml_webgpu_mul_mat_shader_decisions {
    uint32_t subgroup_matrix_n;

    uint32_t mul_mat_wg_size;
+    bool     src_overlap = false;
 };

 /** MUL_MAT_ID **/
@@ -1950,7 +1956,7 @@ class ggml_webgpu_shader_lib {
        return quantize_q8_pipelines[key];
    }

-    webgpu_pipeline get_mul_mat_vec_pipeline(const ggml_webgpu_shader_lib_context & context) {
+    webgpu_pipeline get_mul_mat_vec_pipeline(const ggml_webgpu_shader_lib_context & context, bool src_overlap) {
        ggml_webgpu_mul_mat_vec_pipeline_key key = {};
        key.src0_type                            = context.src0->type;
        key.src1_type                            = context.src1->type;
@@ -1961,6 +1967,7 @@ class ggml_webgpu_shader_lib {
        key.num_cols   = context.dst->ne[1];
        key.use_mmvq =
            ggml_webgpu_can_use_mmvq(context.src0, context.src1, context.supports_dot_product, context.vendor);
+        key.src_overlap = src_overlap;

        auto it = mul_mat_vec_pipelines.find(key);
        if (it != mul_mat_vec_pipelines.end()) {
@@ -2068,6 +2075,11 @@ class ggml_webgpu_shader_lib {
            defines.push_back("Q8_1_T");
        }

+        if (key.src_overlap) {
+            defines.push_back("SRC_OVERLAP");
+            variant += "_src_overlap";
+        }
+
        defines.push_back(std::string("WG_SIZE=") + std::to_string(wg_size));
        defines.push_back(std::string("OUTPUTS_PER_WG=") + std::to_string(outputs_per_wg));
        defines.push_back(context.supports_subgroups ? "USE_SUBGROUP_REDUCTION" : "USE_WORKGROUP_REDUCTION");
@@ -2089,7 +2101,7 @@ class ggml_webgpu_shader_lib {
        return mul_mat_vec_pipelines[key];
    }

-    webgpu_pipeline get_mul_mat_fast_pipeline(const ggml_webgpu_shader_lib_context & context) {
+    webgpu_pipeline get_mul_mat_fast_pipeline(const ggml_webgpu_shader_lib_context & context, bool src_overlap) {
        ggml_webgpu_mul_mat_pipeline_key key = {};
        key.src0_type                        = context.src0->type;
        key.src1_type                        = context.src1->type;
@@ -2098,6 +2110,7 @@ class ggml_webgpu_shader_lib {
                                      1 :
                                      0;
        key.use_subgroup_matrix = context.supports_subgroup_matrix;
+        key.src_overlap         = src_overlap;

        auto it = mul_mat_fast_pipelines.find(key);
        if (it != mul_mat_fast_pipelines.end()) {
@@ -2216,6 +2229,11 @@ class ggml_webgpu_shader_lib {
            variant += "_vectorized";
        }

+        if (key.src_overlap) {
+            defines.push_back("SRC_OVERLAP");
+            variant += "_src_overlap";
+        }
+
        if (!key.use_subgroup_matrix) {
            defines.push_back("WORKGROUP_SIZE_M=" + std::to_string(WEBGPU_MUL_MAT_WG_SIZE_M) + "u");
            defines.push_back("WORKGROUP_SIZE_N=" + std::to_string(WEBGPU_MUL_MAT_WG_SIZE_N) + "u");
@@ -1628,48 +1628,65 @@ static webgpu_encoded_op ggml_webgpu_mul_mat(webgpu_context & ctx,
    // Get or create pipeline
    webgpu_pipeline                   pipeline;
    std::vector<webgpu_dispatch_desc> dispatches;
+    const bool src_overlap = ggml_webgpu_tensor_binding_overlap(ctx->global_ctx, src0, src1) && !use_mmvq;

    if (use_mat_vec) {
        if (use_mmvq) {
            ggml_webgpu_quantize_q8_dispatch(ctx, src0, src1, dst, dispatches);
        }
-        pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx);
+        pipeline = ctx->shader_lib->get_mul_mat_vec_pipeline(shader_lib_ctx, src_overlap);
    } else {
-        pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx);
+        pipeline = ctx->shader_lib->get_mul_mat_fast_pipeline(shader_lib_ctx, src_overlap);
+    }
+
+    uint32_t offset_src0   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type));
+    uint32_t offset_src1   = (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type));
+    size_t   merged_offset = 0;
+    size_t   merged_size   = 0;
+    if (src_overlap) {
+        const ggml_webgpu_merged_binding_range merged_range =
+            ggml_webgpu_tensor_merged_binding_range(ctx, { src0, src1 });
+        merged_offset = merged_range.offset;
+        merged_size   = merged_range.size;
+        offset_src0   = ggml_webgpu_tensor_merged_element_offset(src0, merged_range);
+        offset_src1   = ggml_webgpu_tensor_merged_element_offset(src1, merged_range);
    }

    // Build params
-    std::vector<uint32_t> params = {
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) / ggml_type_size(src0->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src1) / ggml_type_size(src1->type)),
-        (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
-        (uint32_t) dst->ne[0],
-        (uint32_t) dst->ne[1],
-        (uint32_t) src0->ne[0],
-        (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
-        (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
-        (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
-        (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
-        (uint32_t) src0->ne[2],
-        (uint32_t) src0->ne[3],
-        (uint32_t) (src1->ne[2] / src0->ne[2]),
-        (uint32_t) (src1->ne[3] / src0->ne[3])
-    };
+    std::vector<uint32_t> params = { offset_src0,
+                                     offset_src1,
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+                                     (uint32_t) dst->ne[0],
+                                     (uint32_t) dst->ne[1],
+                                     (uint32_t) src0->ne[0],
+                                     (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[1] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[2] / ggml_type_size(src1->type)),
+                                     (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src1->nb[3] / ggml_type_size(src1->type)),
+                                     (uint32_t) src0->ne[2],
+                                     (uint32_t) src0->ne[3],
+                                     (uint32_t) (src1->ne[2] / src0->ne[2]),
+                                     (uint32_t) (src1->ne[3] / src0->ne[3]) };

    // Build bind group entries
    std::vector<wgpu::BindGroupEntry> entries = {};
-
-    entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
    if (use_mmvq) {
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
        auto & mmvq_qq8_entry = dispatches[0].bind_group_entries[1];
        entries.push_back(ggml_webgpu_make_bind_group_entry(1, ggml_webgpu_tensor_buf(dst), mmvq_qq8_entry.offset,
                                                            mmvq_qq8_entry.size));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
+    } else if (src_overlap) {
+        entries.push_back(
+            ggml_webgpu_make_bind_group_entry(0, ggml_webgpu_tensor_buf(src0), merged_offset, merged_size));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, dst));
    } else {
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 0, src0));
        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 1, src1));
+        entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));
    }
-    entries.push_back(ggml_webgpu_make_tensor_bind_group_entry(ctx, 2, dst));

    // Calculate workgroup dimensions
    uint32_t       wg_x           = 1;
@@ -4455,7 +4472,9 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
            supports_op = (op->type == GGML_TYPE_F32 && src0->type == GGML_TYPE_F32) && ggml_is_contiguous_rows(src0);
            break;
        case GGML_OP_ROPE:
-            supports_op = op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16;
+            // FIXME: support ggml_rope_set_offset
+            supports_op =
+                (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) && ((const int32_t *) op->op_params)[15] == 0;
            break;
        case GGML_OP_GLU:
            switch (ggml_get_glu_op(op)) {
@@ -1,3 +1,7 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+
 #ifdef BYTE_HELPERS
 fn get_byte(value: u32, index: u32) -> u32 {
    return (value >> (index * 8)) & 0xFF;
@@ -46,7 +50,7 @@ fn load_f16_as_f32_at_src(byte_offset: u32) -> f32 {

 #ifdef DECLARE_BYTE_LOADERS_SRC0
 fn load_u16_at_src0(byte_offset: u32) -> u32 {
-    let word = src0[byte_offset / 4u];
+    let word = SRC0[byte_offset / 4u];
    let shift = (byte_offset & 0x2u) * 8u;
    return (word >> shift) & 0xFFFFu;
 }
@@ -55,14 +59,14 @@ fn load_u16_at_src0(byte_offset: u32) -> u32 {
 // Caller extracts the 16-bit half it needs via & 0xFFFFu or >> 16u.
 // this is used in k-quants for better performance
 fn load_u32_at_src0_aligned(byte_offset: u32) -> u32 {
-    return src0[(byte_offset & ~3u) / 4u];
+    return SRC0[(byte_offset & ~3u) / 4u];
 }

 fn load_u32_at_src0(byte_offset: u32) -> u32 {
    let word_idx = byte_offset / 4u;
    let shift = (byte_offset & 0x3u) * 8u;
-    let lo = src0[word_idx];
-    let hi = src0[word_idx + 1u];
+    let lo = SRC0[word_idx];
+    let hi = SRC0[word_idx + 1u];
    let shifted = (lo >> shift) | (hi << (32u - shift));
    return select(shifted, lo, shift == 0u);
 }
@@ -73,7 +77,7 @@ fn load_f16_at_src0(byte_offset: u32) -> f16 {
 }

 fn load_f16_as_f32_at_src0(byte_offset: u32) -> f32 {
-    let word = src0[byte_offset / 4u];
+    let word = SRC0[byte_offset / 4u];
    let shift = (byte_offset & 0x2u) * 8u;
    let d_bits = (word >> shift) & 0xFFFFu;
    return unpack2x16float(d_bits)[0];
@@ -1,3 +1,10 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+#ifndef SRC1
+#define SRC1 src1
+#endif
+
 #ifdef VEC
 #define VEC_SIZE 4
 #define SHMEM_TYPE vec4<f16>
@@ -39,7 +46,7 @@ fn init_shmem_src0(thread_id: u32, batch_offset: u32, offset_m: u32, k_outer: u3
        let src0_idx = batch_offset + global_m * params.stride_01 + global_k;
        let src0_val = select( // taking a slight performance hit to avoid oob
            SRC0_TYPE(0.0),
-            src0[src0_idx/VEC_SIZE],
+            SRC0[src0_idx/VEC_SIZE],
            global_m < params.m && global_k < params.k);
        store_shmem(SHMEM_TYPE(src0_val), elem_idx);
    }
@@ -57,7 +64,7 @@ fn init_shmem_src1(thread_id: u32, batch_offset: u32, offset_n: u32, k_outer: u3
        let src1_idx = batch_offset + global_n * params.stride_11 + global_k;
        let src1_val = select(
            SRC1_TYPE(0.0),
-            src1[src1_idx/VEC_SIZE],
+            SRC1[src1_idx/VEC_SIZE],
            global_n < params.n && global_k < params.k);
        store_shmem(SHMEM_TYPE(src1_val), TILE_SRC0_SHMEM + elem_idx);
    }
@@ -1,8 +1,12 @@
 enable f16;

 #define DECLARE_BYTE_LOADERS_SRC0
-#include "common_decls.tmpl"

+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
+#include "common_decls.tmpl"
 #include "mul_mat_decls.tmpl"

 #ifdef VEC
@@ -36,11 +40,17 @@ struct MulMatParams {
    broadcast3: u32
 };

+#ifdef SRC_OVERLAP
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
+#else
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+#define DST_BINDING 2
+#endif

-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 fn get_local_n(thread_id: u32) -> u32 {
    return thread_id / WORKGROUP_SIZE_M;
@@ -4,6 +4,10 @@ enable subgroups;
 enable chromium_experimental_subgroup_matrix;

 #define DECLARE_BYTE_LOADERS_SRC0
+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
 #include "common_decls.tmpl"

 #include "mul_mat_decls.tmpl"
@@ -48,11 +52,17 @@ struct MulMatParams {
 };

 // SRC0_TYPE and SRC1_TYPE are defined in mul_mat_decls, which is included
+#ifdef SRC_OVERLAP
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
+#else
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>; // M rows, K columns
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>; // K rows, N columns (transposed)
-@group(0) @binding(2) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+#define DST_BINDING 2
+#endif

-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<DST_TYPE>; // M rows, N columns (transposed)
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 const WG_M_SG_TILE_SIZE = SUBGROUP_M * SUBGROUP_MATRIX_M * SUBGROUP_MATRIX_M_SIZE;
 const WG_N_SG_TILE_SIZE = SUBGROUP_N * SUBGROUP_MATRIX_N * SUBGROUP_MATRIX_N_SIZE;
@@ -7,6 +7,11 @@ enable f16;
 requires packed_4x8_integer_dot_product;
 #endif

+#ifdef SRC_OVERLAP
+#define SRC0 merged_src
+#define SRC1 merged_src
+#endif
+
 #define DECLARE_BYTE_LOADERS_SRC0
 #include "common_decls.tmpl"

@@ -35,17 +40,22 @@ struct MulMatParams {
    broadcast3: u32
 };

+#if defined(MMVQ)
@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>;
-
-#ifdef MMVQ
@group(0) @binding(1) var<storage, read_write> src1q: array<q8_1>;
+#define DST_BINDING 2
+#elif defined(SRC_OVERLAP)
+@group(0) @binding(0) var<storage, read_write> merged_src: array<SRC0_TYPE>;
+#define DST_BINDING 1
 #else
+@group(0) @binding(0) var<storage, read_write> src0: array<SRC0_TYPE>;
@group(0) @binding(1) var<storage, read_write> src1: array<SRC1_TYPE>;
+#define DST_BINDING 2
 #endif

-@group(0) @binding(2) var<storage, read_write> dst: array<f32>;
+@group(0) @binding(DST_BINDING) var<storage, read_write> dst: array<f32>;
 // "mul_mat_vec_acc.tmpl" requires params.k, params.m, params.stride_01
-@group(0) @binding(3) var<uniform> params: MulMatParams;
+@group(0) @binding(DST_BINDING + 1) var<uniform> params: MulMatParams;

 // Flattened as [row][thread] to keep each row's reduction contiguous in memory.
 var<workgroup> partial_sums: array<f32, OUTPUTS_PER_WG * WG_SIZE>;
@@ -1,3 +1,10 @@
+#ifndef SRC0
+#define SRC0 src0
+#endif
+#ifndef SRC1
+#define SRC1 src1
+#endif
+
 #ifdef U32_DEQUANT_HELPERS
 #define SRC0_TYPE u32

@@ -43,13 +50,13 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
    for (var k = thread_id; k < k_vec; k += WG_SIZE) {
        var x_vals: array<SRC1_TYPE, NUM_COLS>;
        for (var col = 0u;col < NUM_COLS;col += 1) {
-            x_vals[col] = src1[src1_idx_base_vec + col * (params.stride_11 / VEC_SIZE) + k];
+            x_vals[col] = SRC1[src1_idx_base_vec + col * (params.stride_11 / VEC_SIZE) + k];
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
            let output_row = row_base + row;
            if (output_row < params.m) {
                let src0_idx = (src0_batch_offset + output_row * params.stride_01) / VEC_SIZE + k;
-                let w = src0[src0_idx];
+                let w = SRC0[src0_idx];
                for (var col = 0u;col < NUM_COLS;col += 1) {
                    acc[col][row] += inner_dot(w, x_vals[col]);
                }
@@ -76,7 +83,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -116,8 +123,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -160,8 +167,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -205,8 +212,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -253,8 +260,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -302,7 +309,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -347,7 +354,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -409,10 +416,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4u]  = f32(src1[x_base + col * params.stride_11 + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[x_base + col * params.stride_11 + 64u + i]);
-                x_block[col][i + 12u] = f32(src1[x_base + col * params.stride_11 + 96u + i]);
+                x_block[col][i]       = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4u]  = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[x_base + col * params.stride_11 + 64u + i]);
+                x_block[col][i + 12u] = f32(SRC1[x_base + col * params.stride_11 + 96u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -518,8 +525,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 8u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 8u] = f32(src1[x_base + col * params.stride_11 + 32u + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 8u] = f32(SRC1[x_base + col * params.stride_11 + 32u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -610,10 +617,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[col_base + i]);
-                x_block[col][i + 4u]  = f32(src1[col_base + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[col_base + 128u + i]);
-                x_block[col][i + 12u] = f32(src1[col_base + 160u + i]);
+                x_block[col][i]       = f32(SRC1[col_base + i]);
+                x_block[col][i + 4u]  = f32(SRC1[col_base + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[col_base + 128u + i]);
+                x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
            }
        }

@@ -713,10 +720,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var i = 0u; i < 4u; i++) {
-                x_block[col][i]       = f32(src1[col_base + i]);
-                x_block[col][i + 4u]  = f32(src1[col_base + 32u + i]);
-                x_block[col][i + 8u]  = f32(src1[col_base + 128u + i]);
-                x_block[col][i + 12u] = f32(src1[col_base + 160u + i]);
+                x_block[col][i]       = f32(SRC1[col_base + i]);
+                x_block[col][i + 4u]  = f32(SRC1[col_base + 32u + i]);
+                x_block[col][i + 8u]  = f32(SRC1[col_base + 128u + i]);
+                x_block[col][i + 12u] = f32(SRC1[col_base + 160u + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -823,10 +830,10 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        for (var col = 0u; col < NUM_COLS;col += 1) {
            let col_base = x_base + col * params.stride_11;
            for (var l = 0u; l < 4u; l++) {
-                x_block[col][l]       = f32(src1[col_base + l]);
-                x_block[col][l + 4u]  = f32(src1[col_base + 32u + l]);
-                x_block[col][l + 8u]  = f32(src1[col_base + 64u + l]);
-                x_block[col][l + 12u] = f32(src1[col_base + 96u + l]);
+                x_block[col][l]       = f32(SRC1[col_base + l]);
+                x_block[col][l + 4u]  = f32(SRC1[col_base + 32u + l]);
+                x_block[col][l + 8u]  = f32(SRC1[col_base + 64u + l]);
+                x_block[col][l + 12u] = f32(SRC1[col_base + 96u + l]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -899,7 +906,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -960,7 +967,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1039,7 +1046,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1101,7 +1108,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1168,7 +1175,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1234,7 +1241,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1302,7 +1309,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1367,8 +1374,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2u; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4u] = f32(src1[x_base + col * params.stride_11 + i + 16u]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4u] = f32(SRC1[x_base + col * params.stride_11 + i + 16u]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1418,7 +1425,7 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, 16>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < 16u; i++) {
-                x_block[col][i] = f32(src1[x_base + col * params.stride_11 + i]);
+                x_block[col][i] = f32(SRC1[x_base + col * params.stride_11 + i]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1476,8 +1483,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 4] = f32(src1[x_base + col * params.stride_11 + i + 16]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 4] = f32(SRC1[x_base + col * params.stride_11 + i + 16]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -1521,8 +1528,8 @@ fn accumulate_vec_dot(thread_id: u32, row_base: u32, src0_batch_offset: u32, src
        var x_block: array<array<f32, ELEMS_PER_THREAD>, NUM_COLS>;
        for (var col = 0u; col < NUM_COLS;col += 1) {
            for (var i = 0u; i < ELEMS_PER_THREAD / 2; i++) {
-                x_block[col][i]     = f32(src1[x_base + col * params.stride_11 + i]);
-                x_block[col][i + 8] = f32(src1[x_base + col * params.stride_11 + i + 8]);
+                x_block[col][i]     = f32(SRC1[x_base + col * params.stride_11 + i]);
+                x_block[col][i + 8] = f32(SRC1[x_base + col * params.stride_11 + i + 8]);
            }
        }
        for (var row = 0u; row < OUTPUTS_PER_WG; row++) {
@@ -86,6 +86,6 @@ endif()

 target_link_libraries(ggml-zendnn PRIVATE m pthread)

-if (GGML_OPENMP)
-    target_link_libraries(ggml-zendnn PRIVATE OpenMP::OpenMP_CXX)
+if (GGML_OPENMP_ENABLED)
+    target_link_libraries(ggml-zendnn PRIVATE ${GGML_OPENMP_TARGET_CXX})
 endif()
@@ -4200,7 +4200,7 @@ static struct ggml_tensor * ggml_rope_impl(

    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);

-    int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
+    int32_t params[16] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
    memcpy(params +  5, &freq_base,    sizeof(float));
    memcpy(params +  6, &freq_scale,   sizeof(float));
    memcpy(params +  7, &ext_factor,   sizeof(float));
@@ -4212,6 +4212,8 @@ static struct ggml_tensor * ggml_rope_impl(
    } else {
        memset(params + 11, 0,         sizeof(int32_t) * GGML_MROPE_SECTIONS);
    }
+    params[15] = 0; // n_offs, set via ggml_rope_set_offset()
+
    ggml_set_op_params(result, params, sizeof(params));

    result->op     = GGML_OP_ROPE;
@@ -4422,6 +4424,20 @@ struct ggml_tensor * ggml_rope_multi_back(
    result->op = GGML_OP_ROPE_BACK;
    return result;
 }
+
+struct ggml_tensor * ggml_rope_set_offset(
+        struct ggml_tensor  * a,
+        int                   n_offs) {
+    GGML_ASSERT(a->op == GGML_OP_ROPE || a->op == GGML_OP_ROPE_BACK);
+    GGML_ASSERT(n_offs >= 0);
+
+    const int32_t mode = ggml_get_op_params_i32(a, 2);
+    GGML_ASSERT(mode != GGML_ROPE_TYPE_VISION);
+
+    ggml_set_op_params_i32(a, 15, n_offs);
+    return a;
+}
+
 // ggml_clamp

 struct ggml_tensor * ggml_clamp(
@@ -208,6 +208,7 @@ class Keys:
        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
        TEMPERATURE_SCALE            = "{arch}.attention.temperature_scale"
+        ROPE_PATTERN                 = "{arch}.attention.rope_pattern"

        class Indexer:
            HEAD_COUNT = "{arch}.attention.indexer.head_count"
@@ -549,6 +550,7 @@ class MODEL_ARCH(IntEnum):
    GRANITE_MOE      = auto()
    GRANITE_HYBRID   = auto()
    GRANITE_SWITCH   = auto()
+    GRANITE_SWA      = auto()
    CHAMELEON        = auto()
    WAVTOKENIZER_DEC = auto()
    PLM              = auto()
@@ -1265,6 +1267,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.GRANITE_MOE:      "granitemoe",
    MODEL_ARCH.GRANITE_HYBRID:   "granitehybrid",
    MODEL_ARCH.GRANITE_SWITCH:   "graniteswitch",
+    MODEL_ARCH.GRANITE_SWA:      "granite_swa",
    MODEL_ARCH.CHAMELEON:        "chameleon",
    MODEL_ARCH.WAVTOKENIZER_DEC: "wavtokenizer-dec",
    MODEL_ARCH.PLM:              "plm",
@@ -4152,6 +4155,31 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.GRANITE_SWA: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.ATTN_SINKS,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        # MoE (GraniteMoeSWA)
+        MODEL_TENSOR.FFN_GATE_INP,
+        MODEL_TENSOR.FFN_GATE_EXP,
+        MODEL_TENSOR.FFN_GATE_UP_EXP,
+        MODEL_TENSOR.FFN_DOWN_EXP,
+        MODEL_TENSOR.FFN_UP_EXP,
+        # Shared expert - gate+up kept fused in FFN_UP_SHEXP (LLM_FFN_SWIGLU)
+        MODEL_TENSOR.FFN_UP_SHEXP,
+        MODEL_TENSOR.FFN_DOWN_SHEXP,
+    ],
    MODEL_ARCH.CHAMELEON: [
        MODEL_TENSOR.TOKEN_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
@@ -32,6 +32,10 @@ from gguf.constants import (
    GGUFEndian,
 )

+# limits mirroring ggml/src/gguf.cpp (not part of gguf.h)
+GGUF_MAX_STRING_LENGTH  = 1024 * 1024 * 1024
+GGUF_MAX_ARRAY_ELEMENTS = 1024 * 1024 * 1024
+
 logger = logging.getLogger(__name__)

 READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION]
@@ -167,6 +171,10 @@ class GGUFReader:
        offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64]))
        offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64]))
        tensor_count, kv_count = temp_counts
+        if tensor_count > GGUF_MAX_ARRAY_ELEMENTS:
+            raise ValueError(f'Tensor count {tensor_count} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
+        if kv_count > GGUF_MAX_ARRAY_ELEMENTS:
+            raise ValueError(f'KV count {kv_count} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
        offs = self._build_fields(offs, kv_count)

        # Build Tensor Info Fields
@@ -217,6 +225,10 @@ class GGUFReader:

    def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]:
        slen = self._get(offset, np.uint64)
+        if int(slen[0]) > GGUF_MAX_STRING_LENGTH:
+            raise ValueError(f'String length {int(slen[0])} exceeds maximum {GGUF_MAX_STRING_LENGTH}')
+        if offset + 8 + int(slen[0]) > self.data.nbytes:
+            raise ValueError(f'String length {int(slen[0])} exceeds remaining file size {self.data.nbytes - offset - 8}')
        return slen, self._get(offset + 8, np.uint8, slen[0])

    def _get_field_parts(
@@ -241,6 +253,8 @@ class GGUFReader:
            raw_itype = self._get(offs, np.uint32)
            offs += int(raw_itype.nbytes)
            alen = self._get(offs, np.uint64)
+            if int(alen[0]) > GGUF_MAX_ARRAY_ELEMENTS:
+                raise ValueError(f'Array length {int(alen[0])} exceeds maximum {GGUF_MAX_ARRAY_ELEMENTS}')
            offs += int(alen.nbytes)
            aparts: list[npt.NDArray[Any]] = [raw_itype, alen]
            data_idxs: list[int] = []
@@ -824,6 +824,9 @@ class GGUFWriter:
        else:
            self.add_array(key, value)

+    def add_rope_pattern(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.ROPE_PATTERN.format(arch=self.arch), value)
+
    def add_dense_features_dims(self, dense:str, in_f:int, out_f:int) -> None:
        self.add_uint32(Keys.LLM.DENSE_FEAT_IN_SIZE.format(arch=self.arch, dense=dense), in_f)
        self.add_uint32(Keys.LLM.DENSE_FEAT_OUT_SIZE.format(arch=self.arch, dense=dense), out_f)
@@ -458,6 +458,7 @@ class TensorNameMap:
            "transformer.decoder_layer.{bid}.router",           # Grok
            "transformer.blocks.{bid}.ffn.router.layer",        # dbrx
            "model.layers.{bid}.block_sparse_moe.router.layer", # granitemoe
+            "model.layers.{bid}.block_sparse_moe.router",       # granite_swa
            "model.layers.{bid}.feed_forward.router",           # llama4 jamba
            "encoder.layers.{bid}.mlp.router.layer",            # nomic-bert-moe
            "model.layers.{bid}.mlp.router",                    # openai-moe
@@ -1 +1 @@
-3834fd814e74e8af277939dabd69ecc780affd21
+8c63e70982c95ceb862e3a1073a2c1beef75d60a
@@ -102,6 +102,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_GRANITE_MOE,      "granitemoe"       },
    { LLM_ARCH_GRANITE_HYBRID,   "granitehybrid"    },
    { LLM_ARCH_GRANITE_SWITCH,   "graniteswitch"    },
+    { LLM_ARCH_GRANITE_SWA,      "granite_swa"      },
    { LLM_ARCH_CHAMELEON,        "chameleon"        },
    { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
    { LLM_ARCH_PLM,              "plm"              },
@@ -261,6 +262,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
    { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,       "%s.attention.relative_buckets_count"       },
    { LLM_KV_ATTENTION_SLIDING_WINDOW,               "%s.attention.sliding_window"               },
    { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,       "%s.attention.sliding_window_pattern"       },
+    { LLM_KV_ATTENTION_ROPE_PATTERN,                 "%s.attention.rope_pattern"                 },
+
    { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
    { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
    { LLM_KV_ATTENTION_VALUE_SCALE,                  "%s.attention.value_scale"                  },
@@ -1029,6 +1032,8 @@ bool llm_arch_supports_rs_rollback(const llm_arch & arch) {
        case LLM_ARCH_DEEPSEEK4:
        case LLM_ARCH_NEMOTRON_H:
        case LLM_ARCH_NEMOTRON_H_MOE:
+        case LLM_ARCH_LFM2:
+        case LLM_ARCH_LFM2MOE:
            return true;
        default:
            return false;
@@ -107,6 +107,7 @@ enum llm_arch {
    LLM_ARCH_GRANITE_MOE,
    LLM_ARCH_GRANITE_HYBRID,
    LLM_ARCH_GRANITE_SWITCH,
+    LLM_ARCH_GRANITE_SWA,
    LLM_ARCH_CHAMELEON,
    LLM_ARCH_WAVTOKENIZER_DEC,
    LLM_ARCH_PLM,
@@ -267,6 +268,8 @@ enum llm_kv {
    LLM_KV_ATTENTION_SLIDING_WINDOW,
    LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
    LLM_KV_ATTENTION_SCALE,
+    LLM_KV_ATTENTION_ROPE_PATTERN,
+
    LLM_KV_ATTENTION_OUTPUT_SCALE,
    LLM_KV_ATTENTION_VALUE_SCALE,
    LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
@@ -3099,8 +3099,6 @@ ggml_tensor * llm_graph_context::build_attn(
            int       il) const {
    const bool is_swa = hparams.is_swa(il);

-    GGML_UNUSED(v_cur);
-
    auto * k_rot = is_swa ? inp->self_k_rot_swa : inp->self_k_rot;

    if (k_rot) {
@@ -3133,7 +3131,7 @@ ggml_tensor * llm_graph_context::build_attn(
    // MLA-style attention: the cached K is used as V
    ggml_tensor * q = q_cur;
    ggml_tensor * k = mctx_cur->get_k(ctx0, il);
-    ggml_tensor * v = k;
+    ggml_tensor * v = ggml_view_4d(ctx0, k, v_cur->ne[0], k->ne[1], k->ne[2], k->ne[3], k->nb[1], k->nb[2], k->nb[3], 0);

    ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il);
    cb(cur, "kqv_out", il);
@@ -291,7 +291,11 @@ bool llama_hparams::has_rope(uint32_t il) const {
        return false;
    }

-    return true;
+    if (il < n_layer_all) {
+        return rope_pattern[il] != 0;
+    }
+
+    GGML_ABORT("%s: il (%u) out of bounds (n_layer_all: %u)\n", __func__, il, n_layer_all);
 }

 uint32_t llama_hparams::n_layer() const {
@@ -144,6 +144,10 @@ struct llama_hparams {

    std::array<int, 4> rope_sections;

+    // Per-layer RoPE enable flags (1 = use RoPE, 0 = NoPE)
+    // by default, all layers use RoPE (controlled by rope_finetuned)
+    std::array<uint32_t, LLAMA_MAX_LAYERS> rope_pattern;
+
    // Sliding Window Attention (SWA)
    llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
    // the size of the sliding window (0 - no SWA)
@@ -1395,6 +1395,11 @@ void llama_model_loader::get_mapping_range(size_t * first, size_t * last, void *
    }
 }

+void llama_model_loader::unmap_weight(const llama_tensor_weight & w) const {
+    if (!use_mmap) { return; }
+    mappings.at(w.idx)->unmap_fragment(w.offs, w.offs + ggml_nbytes(w.tensor));
+}
+
 void llama_model_loader::load_data_for(struct ggml_tensor * cur) const {
    const auto & w = require_weight(ggml_get_name(cur));

@@ -194,6 +194,9 @@ struct llama_model_loader {

    void get_mapping_range(size_t * first, size_t * last, void ** addr, int idx, ggml_context * ctx) const;

+    // release a weight's mmap pages
+    void unmap_weight(const llama_tensor_weight & w) const;
+
    // for backwards compatibility, does not support ggml-backend
    void load_data_for(struct ggml_tensor * cur) const;

@@ -30,6 +30,7 @@ bool llama_model_saver_supports_arch(llm_arch arch) {
        case LLM_ARCH_MUSE_GLIMMER:
        case LLM_ARCH_MELLUM:
        case LLM_ARCH_LAGUNA:
+        case LLM_ARCH_GRANITE_SWA:
            return false;
        default:
            return true;
@@ -272,6 +273,7 @@ void llama_model_saver::add_kv_from_model() {
    add_kv(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix);
    add_kv(LLM_KV_ATTENTION_GATE_LORA_RANK,          hparams.n_lora_gate);
    add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,  hparams.n_rel_attn_bkts);
+    add_kv(LLM_KV_ATTENTION_ROPE_PATTERN,            hparams.rope_pattern, true);
    add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
    // add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,  ???);
    add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
@@ -246,6 +246,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_minicpm(params);
        case LLM_ARCH_GRANITE_HYBRID:
            return new llama_model_granite_hybrid(params);
+        case LLM_ARCH_GRANITE_SWA:
+            return new llama_model_granite_swa(params);
        case LLM_ARCH_CHAMELEON:
            return new llama_model_chameleon(params);
        case LLM_ARCH_WAVTOKENIZER_DEC:
@@ -1157,6 +1159,7 @@ void llama_model_base::load_hparams(llama_model_loader & ml) {
    std::fill(hparams.n_ff_arr.begin(),      hparams.n_ff_arr.end(),      0);

    std::fill(hparams.rope_sections.begin(), hparams.rope_sections.end(), 0);
+    std::fill(hparams.rope_pattern.begin(),  hparams.rope_pattern.end(), 1);
    std::fill(hparams.is_swa_impl.begin(),   hparams.is_swa_impl.end(), 0);
    std::fill(hparams.is_recr_impl.begin(),  hparams.is_recr_impl.end(),  llm_arch_is_recurrent(ml.get_arch()) ? 1 : 0);
    std::fill(hparams.is_indexer_full_impl.begin(), hparams.is_indexer_full_impl.end(), 0);
@@ -2639,6 +2642,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_GRANITE_MOE:
        case LLM_ARCH_GRANITE_HYBRID:
        case LLM_ARCH_GRANITE_SWITCH:
+        case LLM_ARCH_GRANITE_SWA:
        case LLM_ARCH_CHAMELEON:
        case LLM_ARCH_BAILINGMOE:
        case LLM_ARCH_BAILINGMOE3:
@@ -1270,7 +1270,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
            total_size_org += tensor_size;
            total_size_new += new_size;

-            // update the gguf meta data as we go
+            // update the gguf metadata as we go
            gguf_set_tensor_type(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_type);
            GGML_ASSERT(gguf_get_tensor_size(ctx_outs[cur_split].get(), gguf_find_tensor(ctx_outs[cur_split].get(), metadata[i].name.c_str())) == new_size);
            gguf_set_tensor_data(ctx_outs[cur_split].get(), metadata[i].name.c_str(), new_data);
@@ -1278,6 +1278,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
            // write tensor data + padding
            fout.write((const char *) new_data, new_size);
            zeros(fout, GGML_PAD(new_size, align) - new_size);
+
+            // unmap the tensor to free memory
+            if (ml.use_mmap) { ml.unmap_weight(weight); }
+
        } // no --dry-run
    } // main loop

@@ -10,8 +10,6 @@ void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);

    // MoE parameters
-    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
-    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
--- a/Show More
+++ b/Show More